Goodera is the world's leading technology platform for CSR, Volunteering, \u0026 Sustainability

Goodera

goodera

Let's learn about via these 63 free stories. They are ordered by most time reading created on HackerNoon. Visit the to find the most read stories about any technology. Company Culture /Learn Repo 1. Ex-Twitter Employee Rejected at New Job For Asking 'Political Questions' Think it's easy finding a job after Twitter? Think again. 2. Why 2022 Will Be the Year of the DAOs DAO developers believe that they can eliminate human manipulation or error from how companies work. This way, decisions are geared towards profitability alone. 3. Developers Created a Culture of Shame and a Toxic Workplace Open up Twitter or any dev forum really, have a little scroll, and it will not take you long at all to find developers being unpleasant to each other. 4. "The 4-Day Work Week" and other Insights on Bootstrapped Business with GooseChase CEO Andrew Cross Inspired by scavenger hunts, GooseChase is an online platform that enables you to create exceptional experiences for your communities. Meet CEO Andrew Cross. 5. Beyond Fame and Parties: What top Engineers Really Want From a Startup 8 things that make engineers today say, “this place is for me!” 6. Ageism in Tech: How It Feels To Be The Oldest in the Room The phone interview started with the founder explaining the format, followed by the standard “tell me about yourself.” Before the interview, I had done research to know he wasn’t yet 30. It seemed to be his first startup and publicly available information about this venture was minimal. I was excited that I got the phone interview because I knew my previous marketing experience was a perfect fit. I kept my story short by including only relevant startup marketing projects. 7. MashRadar Reviewed HackerNoon - Everything You Need to Know Hacker Noon is a tech-focused news website that features content from over 10,000 contributors. 8. Fun at Work Matters: Building a Company Culture Having non-productive fun with colleagues at work is a great way of relaxing during work and getting to know the team. 9. Remote Working and Why I Wouldn't Relocate for a Job Again Here is a list of reasons why I won't relocate just because of a job offer. 10. Productivity Hacks: How to Keep Your Employees Motivated Employee wellbeing is a challenge in this world as i89% of employees have worked while feeling unwell. 11. TekPon is Offering "Free" Marketing Services Of $1.5m To Newly Launched Software Companies A tech start-up is offering free marketing services of $1.5m to newly launched software companies 12. How Digital Transformation Impacts Your Company Culture The successful implementation of digital transformation in the enterprise depends largely on the willingness of the corporate culture to change. 13. Why Working in a Startup Is More Rewarding Than a Career in Big-Tech Five reasons of why working in a startup might be more rewarding than position in a big-name company for specialists, who are focused on career achievements. 14. How to Protect Myself Against Sexism in the Workplace In the era of #metoo and Harvey Weinstein, people are becoming more and more aware of harassment and discrimination in the workplace. Society should thank the brave women who stood up and created a better culture for us, but sexism still happens. It can range from unwanted physical advances to unfair pay to getting passed up for a promotion. The work isn’t done yet. We need to continue to fight to make things equal and safe for women in the workplace. To contribute fighting sexism in society, we need to start at our own workplaces. Here are some tips on how to protect yourself from sexism in the workplace. 15. Pandemic Remote Work != Remote Work Have you switched to remote work in the past couple of weeks/days? Well, chances are high you have, or you will. I want to share my words of caution and tips to make pandemic remote work better. 16. 4 Routes to Remarkable Company Culture Everything I learnt about how to do company culture right I learnt in a tiny, cramped office at an up and coming publishing startup in Farringdon Road, London. 17. How to Prevent Burnout and Keep Your Organization on Track As a result of significant events against me, I burned out in 2016. Here are the keys to overcoming (and potentially protecting staff/colleagues from) burnout. 18. What it Really Means to Hire Slow, Fire Fast A look at what it truly means to practically incorporate hire slow, fire fast within your organization. 19. The 10x value isn't all originating from the engineer, though There's been a lot of discussion on Twitter recently about so-called 10x engineers. 20. How to Preserve Your Remote Company Culture Promoting socializing, resisting surveillance, and hosting idea sessions can help to preserve your remote company culture. 21. 6 Questions Devs Should Ask Interviewers To Discern Culture If you want to learn about growing a digital product without targeted advertising, subscribe to my newsletter Engineering Growth 22. The One Objective to Establish a Winning Startup Culture: Get Better Everyday Last week I had a conversation with a couple of friends about the issues they've been facing at their startup. These friends, like many others, have been affected by CoViD-19 related business impact, as well as the Work From Home situation. As you can guess, the biggest issues have been around communication, specially when it involves some sort of negotiation about what product features to prioritise, and how much effort to put into the same. 23. "Companies With Content Employees Thrive, And Clients Can Feel That", Brais Méndez, Docuten CEO B​rais Méndez, CEO of Docuten, discusses the company's beginnings, what the future holds, and exciting innovations in digital signature and e-invoicing. 24. Why Outsourcing Core Business Functions Can Be Dangerous Outsourcing a core business function, no matter how big and strong a company is, can be a very serious decision. 25. Why It’s a Good Idea for Developers to Work with Startups Instead of for Them What if I told you, it’s possible to get the best of both worlds – to capture the advantages of working with a startup while mitigating the downsides? 26. Twitter Employees Fear Elon Musk's Takeover Will Ruin Company Culture Despite being a longtime lightning rod for controversy, even Twitter employees were caught off-guard by recent events. 27. Does Hiring for Culture Affect Diversity? Can your company hire for both diversity and culture? Learn how to prioritize and balance both through talent intelligence and psychological data. 28. How to Overcome Employee Resistance to Change Company Culture (An Infographic) Focusing on cultural transformation can bring big rewards, including an 85% net profit increase in just five years. But employees sometimes resist this change. 29. Managing Workplace Emotions During Organizational Transformation For a long time, organizations have been analyzed as systems and remained mainly indifferent to how various organization members feel about information. 30. Do Away With Command and Control: How to Change Your Work Culture Despite articles and studies claiming it dead, command and control remain in around nine out of ten organizations. 31. What Whiplash Teaches Us About the "Sink or Swim" Method in Engineering I love movies - Particularly ones that can be referenced to bring a point home in a professional discussion. Such is the case with the movie “Whiplash” and the debate over the “sink or swim” methodology in engineer onboarding. 32. "They're The Ones Signing The Check..." Signing the check
The good and the bad 33. How Cultural Sector Can Help The Business Industry Working with people from other fields may help artists and thinkers contribute to the corporate world in useful ways that the current paradigm does not allow. 34. How Companies Should Act When Employees Work Two Full-Time Jobs The pandemic has changed the rules of the game, influencing how we take care of our health, our leisure, and especially how we work. 35. Demystifying Culture - What Does Culture Really Mean? Annie is a software engineering leader on Square’s platform & infrastructure engineering organization. Prior to Square, she worked at a number of startups across a spectrum of industries from consumer products to enterprise solutions, as well as a wide variety of teams from sales to engineering. 36. The Importance Of Staying Humble And Strong Company Culture: Stephen Goldberg, HarperDB CEO Harper DB is a startups of the year nominee from Denver, Colorado. Stephen Goldberg, CEO, shares the origin story and what makes their company culture relevant. 37. How the Noonies is Like the Dundies [or at least I told my team that] Managing people, meh. Building teams, now that's fun. It is really interesting to think about how individuals can make others better. 38. Why Cybersecurity Must Start with Company Culture Cyberattacks, though frequently thought of as ‘sophisticated’, are often anything but. They are actually surprisingly straightforward and mostly financially motivated. We often see the media reporting on large-scale cyberattacks that have managed to breach complex security protocols, at a substantial financial cost to the business involved. 39. Deglorifying CEOs Bright CEOs belong in the past. Web3 is talent-centric and doesn't need a "charismatic leader" to become successful. Well, are we right, or are we right? 40. Omnipresent Requests That All Employees Return to the Office: The Best April Fools Prank Ever? What would you do if your employer suddenly announced that everyone needed to work from the office? Just kidding, it's just a really good joke. 41. Startup Interview with Dmitry Bubnov, CEO Of ENBISYS ENBISYS CEO shares his tips on running software development company and how his team manages to stay abreast of technology development. 42. 5 Ways to Future-Proof Your Brand Your company’s brand is an essential yet intangible asset that narrates how your audience views your organization. It’s no longer sufficient to merely develop your branding. To be sustainable, you must also future-proof your brand, ensuring that you evolve with industry trends and stay relevant to a digital audience. 43. Improve Team Dynamics is Key to Better Workplace Efficiency Understanding group dynamics and how team-building, when done correctly, can dramatically improve workplace efficiency. 44. HackerNoon and the Time We Talked About Pints? HackerNoon is trying to settle a debate: do the international (non-UK) folks order a “pint” of beer, or just generally connect a “pint” with drinking? 45. Proper Communication Helps Make Digital Transformation Smooth The digital transformation process of the business. The role of communication during the fundamental change. Agility as the locomotive of transformation. 46. 7 Ways to Improve Internal Communication to Foster Better Corporate Culture Here are some of the best practices for developing an internal communication strategy that makes sense for your organization. 47. How to Help Your Employees Maintain Work-life Balance in a Hybrid Landscape Companies need to adapt to the changes necessary in a hybrid workspace to set employees up for success. 48. Subtle Culture Problems That Often Go Unnoticed When Peter Drucker said in 2006 that “Culture eats strategy for breakfast,” he did not mean that strategy is not important. Rather an effective strategy is crucial to organizational success. A good strategy is the key to successfully implementing the vision. But a strategy by itself cannot push any organization to achieve tremendous growth and success. What gives life to that strategy are its people. 49. Building Remote Teams: Lessons From The Gitlab Remote Playbook DevOps giant GitLab are one of the leading fully remote tech companies. They have produced a ‘Remote Playbook’ to guide companies looking to adopt remote workin 50. The Unbundled Work Era: The Time Is Now The internet and rise of micro-entrepreneurship 51. The Disregarded Reality of Burnout Crisis in the Tech Industry A self-help blog on how to survive the industry pressure and revolutionize the toxic work culture into something that might become sustainable eventually. 52. 5 Simple Strategies To Build Trust in Remote Teams Before we attempt to answer the challenge of how to build trust over Zoom, Loom, Slack. Let’s try to analyze what built trust 53. Treating Your Internet Friends with Respect Starts Right Here at HackerNoon HQ TL;DR: All Hacker Noon full time employees and long-term part-time contractors are now eligible to enter our fully-funded Mental Health initiative. Plus a few tips on how to make Remote Work work. 54. How to Set Up a Team's Systems and Culture for Strong Code Reviews All software companies must take an intentional and structured approach to code reviews. Here are some tips on how to do that. 55. Office Inferno: Why Aren't We Happy At Work? I just finished watching “The Office”. Funny on the surface, but very satirical and even dark once you go a layer below. 56. Implement These 21 Employee Wellness Program Ideas For Better Retention Employee wellness programs have become a staple in many companies as a way to attract top talent, keep them happy and productive, and decrease employee turnover. 57. Some Tips On Holding a Productive Virtual Offsite Offsites are a big part of remote teams. They allow everyone to socialize, connect more deeply with coworkers, and help build shared experiences and empathy. Even if video calls are a great tool to share information, they can be tiring. It's too easy to miss non-verbal cues. Chance encounters over coffee never happen, and we don’t always experience the same personal connections that come from small-talk. That’s the downside of remote work, and that's why meeting a few times a year is a great way to fill the gap! 58. 5 Things To Ask About the Engineering Team to Avoid Taking a Job From Hell Anyone with enough job experience knows that some jobs can leave you completely exhausted at the end of the day — while others make you excited, proud, and willing to do more. Sometimes people experience these opposite effects from jobs that are in fact quite similar in terms of responsibilities. So what can account for this difference? 59. Secrets of High-Performing Teams: Part II A robust DevOps team facilitates faster development of new products. 60. Ash Coleman, Head of DEI at Credit Karma, talks about Diversity, Equity, and Inclusion at Work Federico Toledo, wraps up season three of Quality Sense with an illuminating interview with a leader in the DEI space. 61. Team Building Activities In The Workplace Help To Establish Trust Let’s take a look at some benefits of team building activities and games at the workplace. 62. Everyone Is Doing Remote Work but Not Everyone Is Doing It Well with Michael Brooks On this episode of the HackerNoon podcast, Amy Tom and Michael Brooks get into everything freelance. They dive into https://golance.com, Michael’s freelance marketplace company. How it came about and Michael’s insights on running it. They talk about wild experiences like getting kicked out of Belarus and taking a shot of Brandy first thing every morning before work in Serbia. It’s a wild ride and you definitely should tune in! Brandy or coffee in hand, your choice ;) 🍻 63. Improving Communication With Global Cultures Using Contextual Communication If we wish to communicate effectively, we must understand how culture affects communication. Here's a communication design to avoid conflicts and uncertainty. Thank you for checking out the 63 most read stories about Company Culture on HackerNoon. to find the most read stories about any technology. Visit the /Learn Repo

63 Stories To Learn About Company Culture

is a relatively new way to build user interfaces and APIs for consumers. It's essentially a querying language backed by a strongly typed schema, which makes it easy to write human-readable requests to get exactly the data you need. Compared to REST, whose responses are dictated by the server, GraphQL queries place the power squarely in the hands of the client. GraphQL However, GraphQL is really just a spec that describes an API. Many features one might expect from an API, like authorization/authentication logic, bi-directional communication, or rate limiting, still need to be implemented. Recently launched, is a fully managed GraphQL backend that takes care of all the hard work of setting up a GraphQL API, letting you concentrate on building your app. Slash GraphQL In this article, we'll be building an e-commerce bookstore that relies on Slash GraphQL. To demonstrate the flexibility of the service, we'll also be fetching data from another source— —and integrate it into our responses. the Goodreads API We'll be building our app using , so you should have at least version 12 installed on your machine. Create a directory called and navigate into it from the command line. Node.js slashql_bookstore Creating the GraphQL backend Setting up Slash GraphQL is easy—you can instantly sign-in with your GitHub account. The platform offers a free trial which will work for this article (then moves to a $9.99/mo flat fee for up to 5GB data). Setting up an account on Slash GraphQL After you've created an account, you'll need to create a new backend: Click on "Launch a New Backend". Name the backend "Bookstore". Click "Launch". That's it! You now have a server spun up for you in an AWS region. Defining the GraphQL schema Every GraphQL server has a strongly typed schema, which defines the objects and fields available to consumers. Click on "Schema", and paste the following lines: type Book { : Int!
  isbn: Int!
  title: ! @search(by: [fulltext])
  author: Author!
}

type Author { : Int!
  firstname: !
  lastname: !
} id String id String String Here we've defined two types—`Book` and `Author`—which will represent the data in our API. Of the many niceties Slash GraphQL provides, its support for custom directives is one of the most convenient. Here, for example, by annotating the `title` field with , we're able to provide a fuzzy text with no additional work. @search Click "Deploy" to send this schema to production. Setting up the data Let's move on to inserting data. We can use the API Explorer to not only add data, but query for it, too. In order to populate our GraphQL backend, we'll define a list of books and their authors through a mutation that Slash GraphQL conveniently provides for us, based on our schema: mutation AddData {
  addBook(input: [
    { : , : , : , : { : , : , : }},
    { : , : , : , : { : , : , : }},
    { : , : , : , : { : , : , : }},
    { : , : , : , : { : , : , : }},
    { : , : , : , : { : , : , : }},
    { : , : , : , : { : , : , : }}
    ]) {
  numUids
  }
} id 1 title "The Sound and The Fury" isbn 9780394747743 author id 1 firstname "William" lastname "Faulkner" id 2 title "Light in August" isbn 9780394711898 author id 1 firstname "William" lastname "Faulkner" id 3 title "Giovanni's Room" isbn 9780141032948 author id 2 firstname "James" lastname "Baldwin" id 4 title "The Fire Next Time" isbn 9780140182750 author id 2 firstname "James" lastname "Baldwin" id 5 title "To The Lighthouse" isbn 9780140274165 author id 3 firstname "Virginia" lastname "Woolf" id 6 title "Orlando" isbn 9780241284643 author id 3 firstname "Virginia" lastname "Woolf" Click the giant Play button to have this data added to the backend. Now, let's run a query to verify that this data is present: query GetBooks {
  queryBook {
    title
    isbn
    author {
      firstname
      lastname
    }
  }
} You should see a response back that shows all of the data you previously inserted. Building the server We'll now use to set up a basic API server that will provide users access to this Slash GraphQL source. Our Express server acts as little more than an interface to hand requests off to Slash GraphQL. Its most important lines are these: Express BACKEND_URL = { result = axios({ : , : BACKEND_URL, : { : ,
    }, : .stringify({ : operationsDoc,
      variables,
    }),
  }) result.data
} query = app.get( , jsonParser, (req, res) => { gqlResponseData = fetchGraphQL(query,  {})
  res.render( , { : gqlResponseData.data.queryBook })
}) const "https://rampant-arm.us-west-2.aws.cloud.dgraph.io/graphql" async ( ) function fetchGraphQL operationsDoc, variables const await method 'POST' url headers 'Content-Type' 'application/json' data JSON query return const `
query {
  queryBook {
    title
    isbn
    author {
      firstname
      lastname
    }
  }
}
` '/' async let await 'index' books Here, we've defined a helper function, fetchGraphQL, which sets up the necessary headers and data to send over to Slash GraphQL. Keep in mind that you'll need to change BACKEND_URL to match the URL which Slash GraphQL gives you. Then, we write a GraphQL query that defines the data that we want. Finally, we set up a route to fetch that data, and then render it in our view: <div = >
  {{#each books}}
  <div = > <h2>{{title}}</h2> <h4>{{author.firstname}} {{author.lastname}}</h4> <p><strong>ISBN:</strong> {{isbn}}</p> < each}}
< class "container" class "row" < = > div class "col-md-4" </ > div /div>
  <hr>
  {{/ /div> Our view is backed by , which is a popular templating library. We can feed our JSON data into an HTML fragment and iterate over the keys and values, which makes it much easier to display the information. Here we're using an #each loop to go over every book returned by the GraphQL API, and then plugging the data into HTML tags. You'll notice that we have placeholders that match the names of the JSON keys; these are replaced by the real data when the page is viewed. handlebars After you npm install all the dependencies, run the server with node server.js. Then, head on over to localhost:3000 in your browser: You should get back the exact same response that Slash GraphQL's API Explorer did. Extend the server As a final step, we want to expose some reviewer data about these books. But we don't want to create dummy data here; instead, we'll use . Make sure you set up an account there and get an API key before continuing. the Goodreads API One of the many things that Slash GraphQL excels at is the ability to stitch together disparate sources of data into one GraphQL interface. Perhaps you'll want to make calls to other API endpoints that are under your control. To support this, Slash GraphQL provides the directive, which allows you to implement custom behavior for GraphQL fields and resolvers. @custom First, let's add our rating information to the existing schema: type Rating @remote { : Int!
  average_rating: Float!
} ratings_count Our Rating type defines field names which we expect from . We don't need to define every single field that their REST API provides, only the ones we're interested in. Slash GraphQL will know how to transform that response into a GraphQL one. the book.show_by_isbn API On the field side of things, we'll need to specify the URL we want to call out to, along with the HTTP method: type Book {
  # ...
  rating: Rating @custom(http: { : , : GET
  })
  # ...
} url "https://goodreads-api-munger.glitch.me?isbn=$isbn" method You'll notice here that we're actually calling out to . The Goodreads API doesn't return JSON data in a nice, neat format. We'll be calling out to an intermediary server which will munge the data into a more legible format. The details of this server aren't all that important (but you can check it out if you want to). What is important is noting that this URL can stand in for anything, which can be super helpful if you're interacting with data that you own. a Glitch project The other amazing thing is that Slash GraphQL will substitute data for you. See that part of the URL that says $isbn? Depending on the object returned by our query, Slash GraphQL will substitute in the real field value for that variable. This keeps your URL requests flexible and easy to manage. That's all that we need to do to bring in some custom data. Our GraphQL query can now simply add this rating field to bring in that external data. Let's replace the query in our server.js with this one: query = const `
query {
  queryBook {
    title
    isbn
    rating {
      ratings_count
      average_rating
    }
    author {
      firstname
      lastname
    }
  }
}
` And, let's update the view to show this information: <div = >
  {{#each books}}
    <div = > <h2>{{title}}</h2> <h4>{{author.firstname}} {{author.lastname}}</h4> <p><strong>ISBN:</strong> {{isbn}}</p> <p>Rating: {{rating.average_rating}} (out of {{rating.ratings_count}} votes)</p> < each}}
< class "container" class "row" < = > div class "col-md-4" </ > div /div>
  <hr>
  {{/ /div> If you restart the server and head back to localhost:3000, you should now see the ratings populated as well: Conclusion Slash GraphQL makes it incredibly easy to set up a production-grade GraphQL API, but we've only scratched the surface. In addition to being able to combine disparate data locations, it also supports , , and even . authorization subscriptions encryption You can grab the complete source code for this tutorial . If you haven’t already, be sure to sign up for - feel free to share your creations with the rest of the . over at GitHub Slash GraphQL Dgraph community Also published at https://dev.to/mbogan/create-a-book-rating-system-with-the-goodreads-api-and-slash-graphql-5b78

Book Rating System From Scratch using Goodreads API and Slash GraphQL

Meet , your new digital bookshelf bookclub “A reader lives a thousand lives before he dies, said Jojen. The man who never reads lives only one.” ― George R.R Martin Over the past few years, I have transitioned much of my content consumption away from 30-second headline pieces towards long-form narratives. Reading (and listening to) books has quickly become a fundamental component of my learning experience. I read everything from detailed biographies to science fiction to novels on self-awareness. Why do I love reading books? Why do nearly all of my smartest friends obsess over the books they read? In a world where modern forms of media are evolving at hyper-speed, why are books still around? Why do we even still care about them? Books never expire. The good ones never go bad. Their words are forever. It is this reason, combined with many personal motivations, that I continue to read books. And my impetus is always evolving. I read to get lost in other people’s stories. I read to learn about concepts I never knew existed. I read to gain wisdom. To lose anxiety. To explore different worlds. To understand perspectives. To build empathy. Of course, there are many reasons to consume books: each unique and none wrong. I find myself coming back to this explanation often: “We read to feel like we are not alone.” — William Nicholson Humans love stories. For centuries, communities have been formed and inspired around engaging stories. Today, we’re launching — so you can log and share the books you read. bookclub It is a beautiful, simple alternative to Goodreads. Check out my profile . here So how does it work? In just a few clicks, you set up your bookclub profile (via Twitter/FB login). From there, you can begin filling out your personal library. Adding a book to your profile is very straightforward and easy to do (we query the Amazon API to find the book you read). You simply add a review, rating, and date. From there, your profile is ready to share all over the internet. We have numerous features on the roadmap (quite literally infinite things to build) and will share a more in-depth view of exactly what those are in our next update (reading statistics, community feed, etc.). We’re starting, though, with a simple “digital bookshelf creator.” When we started piecing together bookclub a few months ago, we were inspired by the opportunity to bring people and knowledge together. While bookclub profiles represent just one component of a much larger problem-space, we are still very excited to see how the community feels about this direction. Would love to hear what you think. We’re live on @producthunt today!

How We’re Reimagining Goodreads

Maker Angst — Job Flippers and Do Gooders

Belle’s product teardown #1 Belle sliding the library ladder — Disney’s Beauty and the Beast I am a huge fan. Most of that I attribute to the fact that it expands the time we have available for reading and learning something new - while stuck in traffic, at a cardio treadmill or washing dishes. Things that used to be boring can now be pleasurably endured as time for learning one more thing, planting the seeds for one more thought. It’s all wins, you get the boring stuff done, and at the same time enrich yourself with new concepts. Audible As learning for me is of utmost importance, it is hard to conceive many other products that deliver this goal in such an enjoyable manner. It is a no-brainer then, that for my first product teardown, aimed at taking a closer look at some of the products I find interesting, I have chosen it. You can think of MOOCs (massive online open courses), or Kindle or even plain old books, but the convenience that audio books offer is difficult to achieve by any of these. MOOCs often require you to be looking at the instructor, kindle and paper books required that you have your hands tied to flipping or clicking through pages. You may argue that there are other aspects MOOCs are better at, such as integrating exercises and the depth of the knowledge being conveyed, or that books force you to do just one thing at the time, increasing focus and absorption. This may be right. But in terms of convenience it is hard to beat audiobooks. Hooked As I am writing I have around 40 items in my online audible library and all except 3 or 4, I have finished reading, and I have earned almost all badges except “the stack”, “the watchtower”, “repeat listener” and the “dabbler” (this one being attributed to users that have lots of books in the library but a lower percentage of completion — e.g an impressive library of 1.000 but-yet-to complete books ), and have 233 hours of listening time to go to achieve the master level (the top rank in the hierarchy of reader types - I know still a long way to go ). All compounded, in terms of non-stop reading I have read 11 days 05 hours and 24 minutes using Audible. Uff!. 😉 😀 Audible badges Total listening time on Audible In terms of Nir Eyal ’s framework I think my level of enthusiasm means I am hooked on this product 😵 I have an internal trigger: I want to keep learning about new topics (non-fiction), or experience fantastic stories (fiction) The external trigger are boring situations or situations that allow the brain to be free to to other things, such has getting stuck in traffic or riding my bicycle The investment I have to do is to review the book, which at the most basic level means swiping the screen to attribute it a certain number of starts, or choose a new from the store, that I can purchase with 1 click book The reward I get is getting to see the book marked as completed and being able to share with my friends that accomplishment The only other app I consider a must-have in my portfolio, and committed to paying a monthly fee being Spotify. In fact, they all see somewhat related to me, as they both are streaming audio content, and sometimes I use them interchangeably. When riding my bicycle I often alternate between continuing an audiobook, and hearing the weekly playlist recommended for me - the annoying part being to swipe the phone screen to look for one or the other app launcher. I am quite satisfied with the experience with Audible so far. I mainly use the app to listen to the books, and the app is pretty clean, putting your latest book right in front of your eyes as the 1st thing you see when you open it. Think it is also pretty good to have other books listed as they give you a sense of context and give you some sort of motivation to finish one more. From this screen you also have access to play and rewind, which are the most import commands to have for any audio player and mostly when you are playing books, but nothing more. When you select a specific book, that screen also displays these commands, in an even bigger format. When I am driving sometimes the external noise prevents me from listening to something carefully and I have to rewind. Or sometimes the variations in the inflections of the narrator make some sentences not so clear in a particular moment. I just hit the 30' rewind again. To be Hooked + As you can see I am quite satisfied with the status quo, and there are not many things that I see as main disturbances in my current experience with Audible. Only when zooming in I can think of things that can push my enthusiasm to yet a different level. I wish there was an even faster way to initialize the app, as it is very uncomfortable and even dangerous to be swiping screens on the phone (that by the way is in a window supporter), while I am driving. I am sure that in voice activated cars this no longer an issue, as at home by asking Amazon’s Alexa I can start reading just with a voice command, but in my case this is still unpleasant. I also wish I could have an easier way to capture, store and share text snippets, as the current features makes them fade to someplace I will only know about if I devote a little more time in looking for it — although most of the time I am content in just listening to the content with no further actions while I am listening (also because many times I am doing something else). Not to mention the the books I read in paper format that are not integrated with anything. I don’t know what is going on between those entities as they are supposed to be one. From a consumer standpoint this is a pain-point as I don’t want to have to duplicate effort in manually adding products to a website when all the other parts of the process are virtual and digital. Nevertheless, I would say that there is one thing, that if improved the has most potential for positive impact on user satisfaction. I read on kindle, audible and paper, and having one library that tracks all of these has just proven to be cumbersome. Although kindle is integrated with Goodreads , audible isn’t, so if I want to keep my library updated I have to go to Goodreads and manually add the book. Even for the paper books — if they were acquired via Amazon — there should be a way of seamlessly putting them in one unique library not two or three. I am not particularly fond of the Audible website, as I think the library display is poor, i.e. doesn’t leverage the nice graphics on the book covers in the same way that the library display in Goodreads. Amazon “library” is even worse, as it is buried under “manage your content and devices” and it took me at least 5 minutes to find it now (also having to sign-off from the uk account and sign in to the .com account where I make my kindle orders). As things stand, would be happy in having my (unique and integrated) library at Goodreads, which user home page by the way is resembling more and more facebook’s timeline (is it just me that noticed?). Hypothesis , and while they seem to be nailing the first one by getting more and more original content narrated many times by the authors themselves in case of non-fiction or acclaimed authors in case of fiction, the second is remains to be addressed, specially when considering the lack of integration with other platforms such as Goodreads. I must say though that to this respect the experience in the audible app is quite different than the one on the audible website, and that some elements such as the gamification badges, could be transferred to whatever ends up being the prevailing social for sharing information on books. Besides the Goodreads challenge that I have accomplished for this year yuh-hu, there are not many elements of gamification yet on this platform. My hypothesis is that content quality (including the narration) and social are the two most elements of the audible product network A colleague from graduate school even started a challenge (within the goodreads challenge) to keep track of how people where accomplishing their goodreads goals. That would be so much easier to accomplish if all the gamification potentialities within goodreads were unleashed. Future and sliding the library ladder I love seeing the progress that is happening in the book lovers space, it seems that books are regaining their spot in the media arena. I am tending towards consuming much more books than television per day, no doubt there. It is always good when people have more and more access to the inspiring thoughts of people living in other continents and places. I am about to finish reading (listening to) Kim Scott’s Radical Candor narrated by herself, and I feel that I have much better insight into how tech companies in Silicon Valley operate (the expected behaviors, etc.) than before. It is quite powerful when your level of insight is one book (and not a thousand miles) away. It is going to be even powerful when navigating those insights is as easy and pleasurable as sliding a fantastic library ladder. Now it is just a matter of who’s going to be providing the library, making the ladder, and helping Belle spread the word about what she is reading. Bonjour

Hooked on Audible, dabbling in Goodreads, wishing for a unique library and a sliding ladder

Recommendations by Concise User Profiles from Review Text: Ethics Statement and References

This paper is available on arxiv under CC 4.0 license.
Authors:
(1) Ghazaleh H. Torbati, Max Planck Institute for Informatics Saarbrucken, Germany & ghazaleh@mpi-inf.mpg.de;
(2) Andrew Yates, University of Amsterdam Amsterdam, Netherlands & a.c.yates@uva.nl;
(3) Anna Tigunova, Max Planck Institute for Informatics Saarbrucken, Germany & tigunova@mpi-inf.mpg.de;
(4) Gerhard Weikum, Max Planck Institute for Informatics Saarbrucken, Germany & weikum@mpi-inf.mpg.de. Table of Links Abstract and Introduction
Related Work
Methodology
Experimental Design
Experimental Results
Conclusion
Ethics Statement and References III. METHODOLOGY A. System Architecture The CUP framework is based on a two-tower architecture for representation learning (one ”tower” for users, the other for items, following the prevalent architecture in neural information retrieval with query and document/passage encodings). The two towers are jointly trained, coupled by the shared loss function and ground truth. Figure 1 shows a pictorial overview. User reviews and items descriptions are fed into BERT followed by a feed-forward network to learn latent representations. Downstream, the vectors are simply compared by a dot product for scores that indicate whether the user likes an item or not. Importantly, unlike prior works, we allow the top-most layer of BERT itself to be fine-tuned as part of the end-to-end training process. The per-item text usually comprises book titles, tags like categories or genre labels, which can be coarse (e.g., “thriller”) or fine-grained (e.g., “Scandinavian crime noir”), and a short description of the book contents. The per-user text can comprise the titles and tags of her training-set books and the entirety of her review texts, which vary widely in length and informativeness, hence the need for smart text selection. In the following, we present the CUP training procedure in Subsection III-B, the test/deployment-time inference in Subsection III-C, the judicious selection of input text snippets in Subsection III-D, and the selection of negative training samples in Subsection III-E. B. Training Another optional mechanism is to process multiple chunks of bounded-length text, which could be of interest for long user reviews. In this case, the per-user representations are learned for each chunk, and there is an additional max-pooling step between the FFN and the dot product, to combine the per-chunk vectors. We use this option for comparison in experiments. As it incurs m times higher computational and energy costs, with m being the number of chunks, we do not favor it. A user-item pair is classified with score calculated by with <, > denoting dot product and σ for the sigmoid function. We use the Adam optimizer to minimize the binary cross entropy loss between predicted labels and the ground truth with sampled negatives. During training we update the topmost layer of BERT, which allows end-to-end training of all components. This is an important difference to earlier works with similar architectures, which use frozen BERT to encode user and item representations. C. Inference Prediction for Ranking. At test time, a prediction is made for user-item pairs. We encode the item description by running it through the trained network, and we compare it to the already learned user vector, which is based on the user’s training-time reviews (none for the given test item). The scores for different test items, computed by the final dot product, yield the ranking of the candidate items. This is a very efficient computation, following established practice in neural information retrieval [37]. Search-based Recommendation. In a deployed system (as opposed to lab experiments with test samples), a typical usage mode would be search-based re-ranking: a user provides context with a tag-based query or an example of a specific liked item, which can be thought of as query-by-example. The user’s expectation is to see a ranked list of recommended items that are similar to her positive sample (as opposed to recommendations from all kinds of categories). The system achieves this by first computing approximate matches to the query item (i.e., similarity-search neighbors), and then re-ranking a shortlist of say top-100 candidates. The CUP framework supports this mode, by using a light-weight BM25 retrieval model, querying all unlabeled items with the category and textual description of the positive point at hand, and keeping the top-100 highest scoring matches. D. Coping with Long and Noisy Texts For constructing user profiles from text, the simplest idea would be to concatenate all available reviews into a long token sequence. Two problems arise, though. User reviews are a noisy mix of descriptive elements (e.g., “the unusual murder weapon”), sentiment expressions (e.g., “it was fun to read”) and personal but irrelevant statements (e.g., “I read only on weekends”). Only the first aspect is helpful for content-based profiling (as the sentiment is already captured by the fact that the user likes the book). Second, the entirety of user-provided text can be too long to be fully digested by the Transformer. Even when it would fit into the token budget, the computational and energy cost is quadratic in the number of input tokens. Therefore, we tightly limit the tokens for each user’s text profile to 128, and devise a suite of light-weight techniques for judiciously selecting the most informative pieces. Variants of our methodology are to create concise profiles by generative language models upfront (e.g., using T5 or ChatGPT), or to digest multiple chunks of text with downstream aggregation by max-pooling or averaging. We include these techniques in our experimental comparisons, but emphasize that both of these alternatives have much higher computational cost and climate footprint. Our techniques for selecting the most informative parts of user reviews to construct concise profiles, are as follows: • Weighted Phrases: selected unigrams, bigrams or trigrams, ordered by descending tf-idf weights, where tf is the frequency of the phrase in all of the given user’s reviews, and idf is pre-computed on Google books n-grams to capture informativeness. • Weighted Sentences: selected sentences, ordered by descending idf weights, where a sentence’s total weight is the sum of the per-word idf weights normalized by sentence length. • Similar Sentences: selected sentences, ordered by descending similarity scores computed via Sentence-BERT [38] for comparing the user-review sentences against the sentences of the item description. To ensure that the selected set is not dominated by a single item, the sentences are picked from different items in a round-robin manner. • ChatGPT-generated Profiles: feeding all reviews of a user, in large chunks, into ChatGPT and instructing it to characterize the user’s book interests with a few short keyphrases. The output size is limited per chunk with a total budget of 128 tokens, to enforce concise profiles. • T5-generated Keywords: using a T5 model fine-tuned for keyword generation, to cast each user’s review text into a set of keywords, concatenated to create the profile. We give examples of selected profile constructions for a Goodreads and an Amazon user in Table VII. An online demo of the CUP framework is accessible on our website[2]. E. Coping with Unlabeled Data A challenge for training in the data-poor regime is how to handle the extreme skew between positive samples and unlabeled data points for sparse users. The crux in many recommender applications is that there are extremely few, if any, explicitly negative samples, such as books rated with low marks. This holds also for the datasets in this work. Therefore, we introduce and experiment with two different techniques to construct negative training samples from unlabeled data: • Uniform random samples. Under the closed world assumption (CWA), aka Selected Completely At Random, negative training points are sample uniformly from all unlabeled data. This is the baseline technique, widely used in prior works. • Weighted pos-neg samples. Prior works on PU learning [39], with positive and unlabeled data and without explicitly negative samples, is largely based on treating unlabeled points as pairs of samples, one positive and one negative with fractional weights. The weights can be based on (learned estimates of) class priors, but the extreme skew in our data renders these techniques ineffective. Instead we leverage the fact that relatedness measures between item pairs can be derived from tagging features or interaction data. We compute item-item relatedness via basic matrix factorization of the user-item matrix for the entire dataset. The relatedness of two items is set to the scalar product between their latent vectors, re-scaled for normalization between 0 and 1. Then, each originally unlabeled sample is cloned, with one instance positive with weight proportional to its average relatedness to the user’s explicitly positive points. The negative clone’s weight is set to the complement. [2] https://sirup.mpi-inf.mpg.de/ This paper is available on arxiv under CC 4.0 license. Authors: (1) Ghazaleh H. Torbati, Max Planck Institute for Informatics Saarbrucken, Germany & ghazaleh@mpi-inf.mpg.de; (2) Andrew Yates, University of Amsterdam Amsterdam, Netherlands & a.c.yates@uva.nl; (3) Anna Tigunova, Max Planck Institute for Informatics Saarbrucken, Germany & tigunova@mpi-inf.mpg.de; (4) Gerhard Weikum, Max Planck Institute for Informatics Saarbrucken, Germany & weikum@mpi-inf.mpg.de. This paper is available on arxiv under CC 4.0 license. Authors: (1) Ghazaleh H. Torbati, Max Planck Institute for Informatics Saarbrucken, Germany & ghazaleh@mpi-inf.mpg.de; (2) Andrew Yates, University of Amsterdam Amsterdam, Netherlands & a.c.yates@uva.nl; (3) Anna Tigunova, Max Planck Institute for Informatics Saarbrucken, Germany & tigunova@mpi-inf.mpg.de; (4) Gerhard Weikum, Max Planck Institute for Informatics Saarbrucken, Germany & weikum@mpi-inf.mpg.de. Table of Links Abstract and Introduction Related Work Methodology Experimental Design Experimental Results Conclusion Ethics Statement and References Abstract and Introduction Abstract and Introduction Related Work Related Work Methodology Methodology Experimental Design Experimental Design Experimental Results Experimental Results Conclusion Conclusion Ethics Statement and References Ethics Statement and References III. METHODOLOGY A. System Architecture A. System Architecture The CUP framework is based on a two-tower architecture for representation learning (one ”tower” for users, the other for items, following the prevalent architecture in neural information retrieval with query and document/passage encodings). The two towers are jointly trained, coupled by the shared loss function and ground truth. Figure 1 shows a pictorial overview. User reviews and items descriptions are fed into BERT followed by a feed-forward network to learn latent representations. Downstream, the vectors are simply compared by a dot product for scores that indicate whether the user likes an item or not. Importantly, unlike prior works, we allow the top-most layer of BERT itself to be fine-tuned as part of the end-to-end training process. The per-item text usually comprises book titles, tags like categories or genre labels, which can be coarse (e.g., “thriller”) or fine-grained (e.g., “Scandinavian crime noir”), and a short description of the book contents. The per-user text can comprise the titles and tags of her training-set books and the entirety of her review texts, which vary widely in length and informativeness, hence the need for smart text selection. In the following, we present the CUP training procedure in Subsection III-B, the test/deployment-time inference in Subsection III-C, the judicious selection of input text snippets in Subsection III-D, and the selection of negative training samples in Subsection III-E. B. Training B. Training Another optional mechanism is to process multiple chunks of bounded-length text, which could be of interest for long user reviews. In this case, the per-user representations are learned for each chunk, and there is an additional max-pooling step between the FFN and the dot product, to combine the per-chunk vectors. We use this option for comparison in experiments. As it incurs m times higher computational and energy costs, with m being the number of chunks, we do not favor it. A user-item pair is classified with score calculated by with <, > denoting dot product and σ for the sigmoid function. We use the Adam optimizer to minimize the binary cross entropy loss between predicted labels and the ground truth with sampled negatives. During training we update the topmost layer of BERT, which allows end-to-end training of all components. This is an important difference to earlier works with similar architectures, which use frozen BERT to encode user and item representations. C. Inference C. Inference Prediction for Ranking. At test time, a prediction is made for user-item pairs. We encode the item description by running it through the trained network, and we compare it to the already learned user vector, which is based on the user’s training-time reviews (none for the given test item). The scores for different test items, computed by the final dot product, yield the ranking of the candidate items. This is a very efficient computation, following established practice in neural information retrieval [37]. Prediction for Ranking. Search-based Recommendation. In a deployed system (as opposed to lab experiments with test samples), a typical usage mode would be search-based re-ranking: a user provides context with a tag-based query or an example of a specific liked item, which can be thought of as query-by-example. The user’s expectation is to see a ranked list of recommended items that are similar to her positive sample (as opposed to recommendations from all kinds of categories). The system achieves this by first computing approximate matches to the query item (i.e., similarity-search neighbors), and then re-ranking a shortlist of say top-100 candidates. The CUP framework supports this mode, by using a light-weight BM25 retrieval model, querying all unlabeled items with the category and textual description of the positive point at hand, and keeping the top-100 highest scoring matches. D. Coping with Long and Noisy Texts D. Coping with Long and Noisy Texts For constructing user profiles from text, the simplest idea would be to concatenate all available reviews into a long token sequence. Two problems arise, though. User reviews are a noisy mix of descriptive elements (e.g., “the unusual murder weapon”), sentiment expressions (e.g., “it was fun to read”) and personal but irrelevant statements (e.g., “I read only on weekends”). Only the first aspect is helpful for content-based profiling (as the sentiment is already captured by the fact that the user likes the book). Second, the entirety of user-provided text can be too long to be fully digested by the Transformer. Even when it would fit into the token budget, the computational and energy cost is quadratic in the number of input tokens. Therefore, we tightly limit the tokens for each user’s text profile to 128, and devise a suite of light-weight techniques for judiciously selecting the most informative pieces. Variants of our methodology are to create concise profiles by generative language models upfront (e.g., using T5 or ChatGPT), or to digest multiple chunks of text with downstream aggregation by max-pooling or averaging. We include these techniques in our experimental comparisons, but emphasize that both of these alternatives have much higher computational cost and climate footprint. Our techniques for selecting the most informative parts of user reviews to construct concise profiles, are as follows: • Weighted Phrases: selected unigrams, bigrams or trigrams, ordered by descending tf-idf weights, where tf is the frequency of the phrase in all of the given user’s reviews, and idf is pre-computed on Google books n-grams to capture informativeness. • Weighted Sentences: selected sentences, ordered by descending idf weights, where a sentence’s total weight is the sum of the per-word idf weights normalized by sentence length. • Similar Sentences: selected sentences, ordered by descending similarity scores computed via Sentence-BERT [38] for comparing the user-review sentences against the sentences of the item description. To ensure that the selected set is not dominated by a single item, the sentences are picked from different items in a round-robin manner. • ChatGPT-generated Profiles: feeding all reviews of a user, in large chunks, into ChatGPT and instructing it to characterize the user’s book interests with a few short keyphrases. The output size is limited per chunk with a total budget of 128 tokens, to enforce concise profiles. • T5-generated Keywords: using a T5 model fine-tuned for keyword generation, to cast each user’s review text into a set of keywords, concatenated to create the profile. We give examples of selected profile constructions for a Goodreads and an Amazon user in Table VII. An online demo of the CUP framework is accessible on our website[2]. E. Coping with Unlabeled Data E. Coping with Unlabeled Data A challenge for training in the data-poor regime is how to handle the extreme skew between positive samples and unlabeled data points for sparse users. The crux in many recommender applications is that there are extremely few, if any, explicitly negative samples, such as books rated with low marks. This holds also for the datasets in this work. Therefore, we introduce and experiment with two different techniques to construct negative training samples from unlabeled data: • Uniform random samples. Under the closed world assumption (CWA), aka Selected Completely At Random, negative training points are sample uniformly from all unlabeled data. This is the baseline technique, widely used in prior works. Uniform random samples. • Weighted pos-neg samples. Prior works on PU learning [39], with positive and unlabeled data and without explicitly negative samples, is largely based on treating unlabeled points as pairs of samples, one positive and one negative with fractional weights. The weights can be based on (learned estimates of) class priors, but the extreme skew in our data renders these techniques ineffective. Instead we leverage the fact that relatedness measures between item pairs can be derived from tagging features or interaction data. We compute item-item relatedness via basic matrix factorization of the user-item matrix for the entire dataset. The relatedness of two items is set to the scalar product between their latent vectors, re-scaled for normalization between 0 and 1. Then, each originally unlabeled sample is cloned, with one instance positive with weight proportional to its average relatedness to the user’s explicitly positive points. The negative clone’s weight is set to the complement. [2] https://sirup.mpi-inf.mpg.de/

Recommendations by Concise User Profiles from Review Text: Methodology

Recommendations by Concise User Profiles from Review Text: Experimental Results

This paper is available on arxiv under CC 4.0 license.
Authors:
(1) Ghazaleh H. Torbati, Max Planck Institute for Informatics Saarbrucken, Germany & ghazaleh@mpi-inf.mpg.de;
(2) Andrew Yates, University of Amsterdam Amsterdam, Netherlands & a.c.yates@uva.nl;
(3) Anna Tigunova, Max Planck Institute for Informatics Saarbrucken, Germany & tigunova@mpi-inf.mpg.de;
(4) Gerhard Weikum, Max Planck Institute for Informatics Saarbrucken, Germany & weikum@mpi-inf.mpg.de. Table of Links Abstract and Introduction
Related Work
Methodology
Experimental Design
Experimental Results
Conclusion
Ethics Statement and References Abstract Recommender systems are most successful for popular items and users with ample interactions (likes, ratings etc.). This work addresses the difficult and underexplored case of supporting users who have very sparse interactions but post informative review texts. Our experimental studies address two book communities with these characteristics. We design a framework with Transformer-based representation learning, covering user-item interactions, item content, and user-provided reviews. To overcome interaction sparseness, we devise techniques for selecting the most informative cues to construct concise user profiles. Comprehensive experiments, with datasets from Amazon and Goodreads, show that judicious selection of text snippets achieves the best performance, even in comparison to LLMgenerated rankings and to using LLMs to generate user profiles. Index Terms—recommender, sparse data, user text, language model, transformer I. INTRODUCTION A. Motivation Recommender systems have become a mature technology and are ubiquitous in applications [1]. Methodologies fall into two major families or hybrid combinations: i) interactionbased recommenders that leverage binary signals (e.g., membership in personal playlists or libraries) or numeric ratings for user-item pairs, and ii) content-based recommenders that exploit item features and user-provided content, ranging from meta-data attributes (e.g., item categories) all the way to textually rich reviews. In many settings, where user-item signals are easily collectable and abundant, interaction-based recommenders achieve excellent performance and are hard to beat. Some applications combine interaction data with informative item features, again relying on massive data (e.g., [2]). In contrast, the data-poor regime of long-tail users and long-tail items has received less attention. Some works advocate countering data sparseness by clustering long-tail items (e.g., [3]). However, this is not sufficient to gear up for users with few interactions but diverse tastes. Approaches for cold-start and cross-domain recommendations and zero-shot learning (e.g., [4]–[7]), are often focused on the item side only: handling new items by transferring (latent) knowledge from existing ones with dense data. Note that sparseness is not just a lack of user-item interactions, but also refers to the users and items themselves: users with merely a handful of items do not exhibit a clear picture of their tastes, and items that are completely unseen at training time are a challenging case. Moreover, even users with a good number of items can be a major problem when the preferred items are highly diverse, such as less known books from all kinds of genres. To address these difficulties, the most promising approach is to leverage review texts by users. In settings where users spend substantial time per item, such as books or travel destinations (as opposed to items with short attention spans, like music or restaurants), even data-poor users with few interactions may leave rich texts that express their interests and tastes. The approach has been studied before (e.g., [8]–[11]), but the performance is much lower than for settings with ample interaction data. This paper aims to advance the state of the art on this theme, by tackling both issues of data sparseness and diversity of tastes via textbased concise user profiles. B. Research Questions Our approach constructs and encodes concise user profiles from reviews of data-poor text-rich users, focusing on the book domain for concreteness (using data from Goodreads and Amazon). Unlike movies or restaurants, books have a much longer tail of popularity and exhibit huge diversity of user tastes, making books a particularly challenging domain for data-poor recommendations. Also, negative reviews are very rare (most have ratings like 4 or 5 stars), so that we deal with binary data and have no negative samples. This approach raises several research questions (RQs): RQ1: Learning with Language Models. How can we best incorporate language models, like BERT, T5 or ChatGPT, for item ranking or for encoding the gist of a user’s reviews, enhanced with the LM’s world knowledge? RQ2: Long Text. Since neural architectures can consume only limited input size (e.g., 512 tokens), we have to select a subset of snippets. Even with a large capacity, the computational and energy cost is typically quadratic in the number of input tokens, so that being selective has benefits. What are the most informative cues that should be fed into the model? RQ3: Review Aspects. User reviews express a mix of aspects, including personal background (e.g., “I’m a retired teacher”), sentiment expressions (e.g., “what a great story”), general emotions (e.g., “brings back lovely memories”), and comments about the book contents. As sentiments do not add value over the already known fact that the book was liked, only the content comments yield cues towards new recommendations. How should this issue be considered in constructing a concise user profile with informative cues? C. Approach We devise a general framework, called CUP, for constructing Concise User Profiles, and encoding them in a recommender system, which allows us to configure a variety of methods. We propose a Tranformer-based architecture that supports end-to-end learning of item and user encodings, making use of a language model. Our choice for the language model is BERT; alternatives such as T5 can be easily plugged in but incur higher computational cost. The end-to-end learning operates on sufficiently short, judiciously constructed profiles, and leverages the available user item interactions as well. On top of the Tranformer, we place feed-forward layers for a supervised classifier. The classifier and the representation learning are jointly trained end-to-end. The prediction scores for user-item pairs yield the per-user ranking of items. This architecture can take as input a spectrum of user and item information including user reviews and item descriptions and metadata (e.g., categories). The rationale for this design is twofold: i) ensuring that inference-time computations are efficient, leveraging trainingtime-computed vectors for users and items and solely computing a scalar product at run-time, and ii) being able to express a wide spectrum of user profiling techniques within the same architecture. Alternative architectures, such as CNNbased, or LLM-based with smart prompts, are baselines in our experiments. We specifically devise new techniques for coping with the case where users have few reviews with long and noisy text (see RQ2 and RQ3), such that their total length exceeds the number of tokens that the encoders can consume. Our techniques are based on information-retrieval-style statistics (like idf weights) for n-grams or sentences, or on neural similarity measures (using SentenceBERT), or on running review texts through a generative language model, like T5 or ChatGPT, for generating user profiles. Table VII illustrates the role of concise user profiles, by giving examples for two users of Amazon books and the Goodreads community, with profiles computed by different methods. In experiments, we limit the number of tokens per user profile, as a stress-test. The main argument for this choice is to keep the system as light-weight as possible. The computational cost of the Transformer, and thus also the environmental footprint, are quadratic in the number of input tokens. Section V-D experimentally shows that a small token budget is competitive to larger ones, while having advantages in efficiency. D. Contributions The salient contributions of this work are the following: • a new framework, called CUP, for Transformer-based recommenders that leverage language models and concise user profiles from reviews; • judicious techniques for selecting and encoding informative cues from the long and noisy text of user reviews, outperforming prior baselines as well as methods based on large language models (LLMs); • comprehensive experiments with focus on data-poor but text-rich users with highly diverse preferences, and difficult predictions (e.g., recommending new authors, not just new books by previously seen authors). We make code, data and experimental details available on our project web page[1]. [1] https://personalization.mpi-inf.mpg.de/CUP This paper is available on arxiv under CC 4.0 license. Authors: (1) Ghazaleh H. Torbati, Max Planck Institute for Informatics Saarbrucken, Germany & ghazaleh@mpi-inf.mpg.de; (2) Andrew Yates, University of Amsterdam Amsterdam, Netherlands & a.c.yates@uva.nl; (3) Anna Tigunova, Max Planck Institute for Informatics Saarbrucken, Germany & tigunova@mpi-inf.mpg.de; (4) Gerhard Weikum, Max Planck Institute for Informatics Saarbrucken, Germany & weikum@mpi-inf.mpg.de. This paper is available on arxiv under CC 4.0 license. Authors: (1) Ghazaleh H. Torbati, Max Planck Institute for Informatics Saarbrucken, Germany & ghazaleh@mpi-inf.mpg.de; (2) Andrew Yates, University of Amsterdam Amsterdam, Netherlands & a.c.yates@uva.nl; (3) Anna Tigunova, Max Planck Institute for Informatics Saarbrucken, Germany & tigunova@mpi-inf.mpg.de; (4) Gerhard Weikum, Max Planck Institute for Informatics Saarbrucken, Germany & weikum@mpi-inf.mpg.de. Table of Links Abstract and Introduction Related Work Methodology Experimental Design Experimental Results Conclusion Ethics Statement and References Abstract and Introduction Abstract and Introduction Related Work Related Work Methodology Methodology Experimental Design Experimental Design Experimental Results Experimental Results Conclusion Conclusion Ethics Statement and References Ethics Statement and References Abstract Recommender systems are most successful for popular items and users with ample interactions (likes, ratings etc.). This work addresses the difficult and underexplored case of supporting users who have very sparse interactions but post informative review texts. Our experimental studies address two book communities with these characteristics. We design a framework with Transformer-based representation learning, covering user-item interactions, item content, and user-provided reviews. To overcome interaction sparseness, we devise techniques for selecting the most informative cues to construct concise user profiles. Comprehensive experiments, with datasets from Amazon and Goodreads, show that judicious selection of text snippets achieves the best performance, even in comparison to LLMgenerated rankings and to using LLMs to generate user profiles. Index Terms —recommender, sparse data, user text, language model, transformer Index Terms I. INTRODUCTION A. Motivation A. Motivation Recommender systems have become a mature technology and are ubiquitous in applications [1]. Methodologies fall into two major families or hybrid combinations: i) interactionbased recommenders that leverage binary signals (e.g., membership in personal playlists or libraries) or numeric ratings for user-item pairs, and ii) content-based recommenders that exploit item features and user-provided content, ranging from meta-data attributes (e.g., item categories) all the way to textually rich reviews. In many settings, where user-item signals are easily collectable and abundant, interaction-based recommenders achieve excellent performance and are hard to beat. Some applications combine interaction data with informative item features, again relying on massive data (e.g., [2]). In contrast, the data-poor regime of long-tail users and long-tail items has received less attention. Some works advocate countering data sparseness by clustering long-tail items (e.g., [3]). However, this is not sufficient to gear up for users with few interactions but diverse tastes. Approaches for cold-start and cross-domain recommendations and zero-shot learning (e.g., [4]–[7]), are often focused on the item side only: handling new items by transferring (latent) knowledge from existing ones with dense data. Note that sparseness is not just a lack of user-item interactions, but also refers to the users and items themselves: users with merely a handful of items do not exhibit a clear picture of their tastes, and items that are completely unseen at training time are a challenging case. Moreover, even users with a good number of items can be a major problem when the preferred items are highly diverse, such as less known books from all kinds of genres. To address these difficulties, the most promising approach is to leverage review texts by users. In settings where users spend substantial time per item, such as books or travel destinations (as opposed to items with short attention spans, like music or restaurants), even data-poor users with few interactions may leave rich texts that express their interests and tastes. The approach has been studied before (e.g., [8]–[11]), but the performance is much lower than for settings with ample interaction data. This paper aims to advance the state of the art on this theme, by tackling both issues of data sparseness and diversity of tastes via textbased concise user profiles. B. Research Questions B. Research Questions Our approach constructs and encodes concise user profiles from reviews of data-poor text-rich users, focusing on the book domain for concreteness (using data from Goodreads and Amazon). Unlike movies or restaurants, books have a much longer tail of popularity and exhibit huge diversity of user tastes, making books a particularly challenging domain for data-poor recommendations. Also, negative reviews are very rare (most have ratings like 4 or 5 stars), so that we deal with binary data and have no negative samples. This approach raises several research questions (RQs): RQ1: Learning with Language Models. How can we best incorporate language models, like BERT, T5 or ChatGPT, for item ranking or for encoding the gist of a user’s reviews, enhanced with the LM’s world knowledge? Learning with Language Models. RQ2: Long Text. Since neural architectures can consume only limited input size (e.g., 512 tokens), we have to select a subset of snippets. Even with a large capacity, the computational and energy cost is typically quadratic in the number of input tokens, so that being selective has benefits. What are the most informative cues that should be fed into the model? Long Text. RQ3: Review Aspects. User reviews express a mix of aspects, including personal background (e.g., “I’m a retired teacher”), sentiment expressions (e.g., “what a great story”), general emotions (e.g., “brings back lovely memories”), and comments about the book contents. As sentiments do not add value over the already known fact that the book was liked, only the content comments yield cues towards new recommendations. How should this issue be considered in constructing a concise user profile with informative cues? Review Aspects. C. Approach C. Approach We devise a general framework, called CUP, for constructing Concise User Profiles, and encoding them in a recommender system, which allows us to configure a variety of methods. We propose a Tranformer-based architecture that supports end-to-end learning of item and user encodings, making use of a language model. Our choice for the language model is BERT; alternatives such as T5 can be easily plugged in but incur higher computational cost. The end-to-end learning operates on sufficiently short, judiciously constructed profiles, and leverages the available user item interactions as well. On top of the Tranformer, we place feed-forward layers for a supervised classifier. The classifier and the representation learning are jointly trained end-to-end. The prediction scores for user-item pairs yield the per-user ranking of items. This architecture can take as input a spectrum of user and item information including user reviews and item descriptions and metadata (e.g., categories). The rationale for this design is twofold: i) ensuring that inference-time computations are efficient, leveraging trainingtime-computed vectors for users and items and solely computing a scalar product at run-time, and ii) being able to express a wide spectrum of user profiling techniques within the same architecture. Alternative architectures, such as CNNbased, or LLM-based with smart prompts, are baselines in our experiments. We specifically devise new techniques for coping with the case where users have few reviews with long and noisy text (see RQ2 and RQ3), such that their total length exceeds the number of tokens that the encoders can consume. Our techniques are based on information-retrieval-style statistics (like idf weights) for n-grams or sentences, or on neural similarity measures (using SentenceBERT), or on running review texts through a generative language model, like T5 or ChatGPT, for generating user profiles. Table VII illustrates the role of concise user profiles, by giving examples for two users of Amazon books and the Goodreads community, with profiles computed by different methods. In experiments, we limit the number of tokens per user profile, as a stress-test. The main argument for this choice is to keep the system as light-weight as possible. The computational cost of the Transformer, and thus also the environmental footprint, are quadratic in the number of input tokens. Section V-D experimentally shows that a small token budget is competitive to larger ones, while having advantages in efficiency. D. Contributions D. Contributions The salient contributions of this work are the following: • a new framework, called CUP, for Transformer-based recommenders that leverage language models and concise user profiles from reviews; • judicious techniques for selecting and encoding informative cues from the long and noisy text of user reviews, outperforming prior baselines as well as methods based on large language models (LLMs); • comprehensive experiments with focus on data-poor but text-rich users with highly diverse preferences, and difficult predictions (e.g., recommending new authors, not just new books by previously seen authors). We make code, data and experimental details available on our project web page[1]. [1] https://personalization.mpi-inf.mpg.de/CUP

Recommendations by Concise User Profiles from Review Text: Abstract and Introduction

189 Stories To Learn About Reading Books

188 Stories To Learn About Analytics

209 Stories To Learn About Mental Health

The Story Girl by L. M. Montgomery, is part of the HackerNoon Books Series. You can jump to any chapter in this book here . A TRAGEDY OF CHILDHOOD CHAPTER VIII. A TRAGEDY OF CHILDHOOD The Story Girl went to Charlottetown for a week in June to visit Aunt Louisa. Life seemed very colourless without her, and even Felicity admitted that it was lonesome. But three days after her departure Felix told us something on the way home from school which lent some spice to existence immediately. “What do you think?” he said in a very solemn, yet excited, tone. “Jerry Cowan told me at recess this afternoon that he HAD SEEN A PICTURE OF GOD—that he has it at home in an old, red-covered history of the world, and has looked at it OFTEN.” To think that Jerry Cowan should have seen such a picture often! We were as deeply impressed as Felix had meant us to be. “Did he say what it was like?” asked Peter. “No—only that it was a picture of God, walking in the garden of Eden.” “Oh,” whispered Felicity—we all spoke in low tones on the subject, for, by instinct and training, we thought and uttered the Great Name with reverence, in spite of our devouring curiosity—“oh, WOULD Jerry Cowan bring it to school and let us see it?” “I asked him that, soon as ever he told me,” said Felix. “He said he might, but he couldn’t promise, for he’d have to ask his mother if he could bring the book to school. If she’ll let him he’ll bring it to-morrow.” “Oh, I’ll be almost afraid to look at it,” said Sara Ray tremulously. I think we all shared her fear to some extent. Nevertheless, we went to school the next day burning with curiosity. And we were disappointed. Possibly night had brought counsel to Jerry Cowan; or perhaps his mother had put him up to it. At all events, he announced to us that he couldn’t bring the red-covered history to school, but if we wanted to buy the picture outright he would tear it out of the book and sell it to us for fifty cents. We talked the matter over in serious conclave in the orchard that evening. We were all rather short of hard cash, having devoted most of our spare means to the school library fund. But the general consensus of opinion was that we must have the picture, no matter what pecuniary sacrifices were involved. If we could each give about seven cents we would have the amount. Peter could only give four, but Dan gave eleven, which squared matters. “Fifty cents would be pretty dear for any other picture, but of course this is different,” said Dan. “And there’s a picture of Eden thrown in, too, you know,” added Felicity. “Fancy selling God’s picture,” said Cecily in a shocked, awed tone. “Nobody but a Cowan would do it, and that’s a fact,” said Dan. “When we get it we’ll keep it in the family Bible,” said Felicity. “That’s the only proper place.” “Oh, I wonder what it will be like,” breathed Cecily. We all wondered. Next day in school we agreed to Jerry Cowan’s terms, and Jerry promised to bring the picture up to Uncle Alec’s the following afternoon. We were all intensely excited Saturday morning. To our dismay, it began to rain just before dinner. “What if Jerry doesn’t bring the picture to-day because of the rain?” I suggested. “Never you fear,” answered Felicity decidedly. “A Cowan would come through ANYTHING for fifty cents.” After dinner we all, without any verbal decision about it, washed our faces and combed our hair. The girls put on their second best dresses, and we boys donned white collars. We all had the unuttered feeling that we must do such honour to that Picture as we could. Felicity and Dan began a small spat over something, but stopped at once when Cecily said severely, “How DARE you quarrel when you are going to look at a picture of God to-day?” Owing to the rain we could not foregather in the orchard, where we had meant to transact the business with Jerry. We did not wish our grown-ups around at our great moment, so we betook ourselves to the loft of the granary in the spruce wood, from whose window we could see the main road and hail Jerry. Sara Ray had joined us, very pale and nervous, having had, so it appeared, a difference of opinion with her mother about coming up the hill in the rain. “I’m afraid I did very wrong to come against ma’s will,” she said miserably, “but I COULDN’T wait. I wanted to see the picture as soon as you did.” We waited and watched at the window. The valley was full of mist, and the rain was coming down in slanting lines over the tops of the spruces. But as we waited the clouds broke away and the sun came out flashingly; the drops on the spruce boughs glittered like diamonds. “I don’t believe Jerry can be coming,” said Cecily in despair. “I suppose his mother must have thought it was dreadful, after all, to sell such a picture.” “There he is now!” cried Dan, waving excitedly from the window. “He’s carrying a fish-basket,” said Felicity. “You surely don’t suppose he would bring THAT picture in a fish-basket!” Jerry HAD brought it in a fish-basket, as appeared when he mounted the granary stairs shortly afterwards. It was folded up in a newspaper packet on top of the dried herring with which the basket was filled. We paid him his money, but we would not open the packet until he had gone. “Cecily,” said Felicity in a hushed tone. “You are the best of us all. YOU open the parcel.” “Oh, I’m no gooder than the rest of you,” breathed Cecily, “but I’ll open it if you like.” With trembling fingers Cecily opened the parcel. We stood around, hardly breathing. She unfolded it and held it up. We saw it. Suddenly Sara began to cry. “Oh, oh, oh, does God look like THAT?” she wailed. Felix and I spoke not. Disappointment, and something worse, sealed our speech. DID God look like that—like that stern, angrily frowning old man with the tossing hair and beard of the wood-cut Cecily held. “I suppose He must, since that is His picture,” said Dan miserably. “He looks awful cross,” said Peter simply. “Oh, I wish we’d never, never seen it,” cried Cecily. We all wished that—too late. Our curiosity had led us into some Holy of Holies, not to be profaned by human eyes, and this was our punishment. “I’ve always had a feeling right along,” wept Sara, “that it wasn’t RIGHT to buy—or LOOK AT—God’s picture.” As we stood there wretchedly we heard flying feet below and a blithe voice calling, “Where are you, children?” The Story Girl had returned! At any other moment we would have rushed to meet her in wild joy. But now we were too crushed and miserable to move. “Whatever is the matter with you all?” demanded the Story Girl, appearing at the top of the stairs. “What is Sara crying about? What have you got there?” “A picture of God,” said Cecily with a sob in her voice, “and oh, it is so dreadful and ugly. Look!” The Story Girl looked. An expression of scorn came over her face. “Surely you don’t believe God looks like that,” she said impatiently, while her fine eyes flashed. “He doesn’t—He couldn’t. He is wonderful and beautiful. I’m surprised at you. THAT is nothing but the picture of a cross old man.” Hope sprang up in our hearts, although we were not wholly convinced. “I don’t know,” said Dan dubiously. “It says under the picture ‘God in the Garden of Eden.’ It’s PRINTED.” “Well, I suppose that’s what the man who drew it thought God was like,” answered the Story Girl carelessly. “But HE couldn’t have known any more than you do. HE had never seen Him.” “It’s all very well for you to say so,” said Felicity, “but YOU don’t know either. I wish I could believe that isn’t like God—but I don’t know what to believe.” “Well, if you won’t believe me, I suppose you’ll believe the minister,” said the Story Girl. “Go and ask him. He’s in the house this very minute. He came up with us in the buggy.” At any other time we would never have dared catechize the minister about anything. But desperate cases call for desperate measures. We drew straws to see who should go and do the asking, and the lot fell to Felix. “Better wait until Mr. Marwood leaves, and catch him in the lane,” advised the Story Girl. “You’ll have a lot of grown-ups around you in the house.” Felix took her advice. Mr. Marwood, presently walking benignantly along the lane, was confronted by a fat, small boy with a pale face but resolute eyes. The rest of us remained in the background but within hearing. “Well, Felix, what is it?” asked Mr. Marwood kindly. “Please, sir, does God really look like this?” asked Felix, holding out the picture. “We hope He doesn’t—but we want to know the truth, and that is why I’m bothering you. Please excuse us and tell me.” The minister looked at the picture. A stern expression came into his gentle blue eyes and he got as near to frowning as it was possible for him to get. “Where did you get that thing?” he asked. THING! We began to breathe easier. “We bought it from Jerry Cowan. He found it in a red-covered history of the world. It SAYS it’s God’s picture,” said Felix. “It is nothing of the sort,” said Mr. Marwood indignantly. “There is no such thing as a picture of God, Felix. No human being knows what he looks like—no human being CAN know. We should not even try to think what He looks like. But, Felix, you may be sure that God is infinitely more beautiful and loving and tender and kind than anything we can imagine of Him. Never believe anything else, my boy. As for this—this SACRILEGE—take it and burn it.” We did not know what a sacrilege meant, but we knew that Mr. Marwood had declared that the picture was not like God. That was enough for us. We felt as if a terrible weight had been lifted from our minds. “I could hardly believe the Story Girl, but of course the minister KNOWS,” said Dan happily. “We’ve lost fifty cents because of it,” said Felicity gloomily. We had lost something of infinitely more value than fifty cents, although we did not realize it just then. The minister’s words had removed from our minds the bitter belief that God was like that picture; but on something deeper and more enduring than mind an impression had been made that was never to be removed. The mischief was done. From that day to this the thought or the mention of God brings up before us involuntarily the vision of a stern, angry, old man. Such was the price we were to pay for the indulgence of a curiosity which each of us, deep in our hearts, had, like Sara Ray, felt ought not to be gratified. “Mr. Marwood told me to burn it,” said Felix. “It doesn’t seem reverent to do that,” said Cecily. “Even if it isn’t God’s picture, it has His name on it.” “Bury it,” said the Story Girl. We did bury it after tea, in the depths of the spruce grove; and then we went into the orchard. It was so nice to have the Story Girl back again. She had wreathed her hair with Canterbury Bells, and looked like the incarnation of rhyme and story and dream. “Canterbury Bells is a lovely name for a flower, isn’t it?” she said. “It makes you think of cathedrals and chimes, doesn’t it? Let’s go over to Uncle Stephen’s Walk, and sit on the branches of the big tree. It’s too wet on the grass, and I know a story—a TRUE story, about an old lady I saw in town at Aunt Louisa’s. Such a dear old lady, with lovely silvery curls.” After the rain the air seemed dripping with odours in the warm west wind—the tang of fir balsam, the spice of mint, the wild woodsiness of ferns, the aroma of grasses steeping in the sunshine,—and with it all a breath of wild sweetness from far hill pastures. Scattered through the grass in Uncle Stephen’s Walk, were blossoming pale, aerial flowers which had no name that we could ever discover. Nobody seemed to know anything about them. They had been there when Great-grandfather King bought the place. I have never seen them elsewhere, or found them described in any floral catalogue. We called them the White Ladies. The Story Girl gave them the name. She said they looked like the souls of good women who had had to suffer much and had been very patient. They were wonderfully dainty, with a strange, faint, aromatic perfume which was only to be detected at a little distance and vanished if you bent over them. They faded soon after they were plucked; and, although strangers, greatly admiring them, often carried away roots and seeds, they could never be coaxed to grow elsewhere. “My story is about Mrs. Dunbar and the Captain of the FANNY,” said the Story Girl, settling herself comfortably on a bough, with her brown head against a gnarled trunk. “It’s sad and beautiful—and true. I do love to tell stories that I know really happened. Mrs. Dunbar lives next door to Aunt Louisa in town. She is so sweet. You wouldn’t think to look at her that she had a tragedy in her life, but she has. Aunt Louisa told me the tale. It all happened long, long ago. Interesting things like this all did happen long ago, it seems to me. They never seem to happen now. This was in ‘49, when people were rushing to the gold fields in California. It was just like a fever, Aunt Louisa says. People took it, right here on the Island; and a number of young men determined they would go to California. “It is easy to go to California now; but it was a very different matter then. There were no railroads across the land, as there are now, and if you wanted to go to California you had to go in a sailing vessel, all the way around Cape Horn. It was a long and dangerous journey; and sometimes it took over six months. When you got there you had no way of sending word home again except by the same plan. It might be over a year before your people at home heard a word about you—and fancy what their feelings would be! “But these young men didn’t think of these things; they were led on by a golden vision. They made all their arrangements, and they chartered the brig Fanny to take them to California. “The captain of the Fanny is the hero of my story. His name was Alan Dunbar, and he was young and handsome. Heroes always are, you know, but Aunt Louisa says he really was. And he was in love—wildly in love,—with Margaret Grant. Margaret was as beautiful as a dream, with soft blue eyes and clouds of golden hair; and she loved Alan Dunbar just as much as he loved her. But her parents were bitterly opposed to him, and they had forbidden Margaret to see him or speak to him. They hadn’t anything against him as a MAN, but they didn’t want her to throw herself away on a sailor. “Well, when Alan Dunbar knew that he must go to California in the Fanny he was in despair. He felt that he could NEVER go so far away for so long and leave his Margaret behind. And Margaret felt that she could never let him go. I know EXACTLY how she felt.” “How can you know?” interrupted Peter suddenly. “You ain’t old enough to have a beau. How can you know?” The Story Girl looked at Peter with a frown. She did not like to be interrupted when telling a story. “Those are not things one KNOWS about,” she said with dignity. “One FEELS about them.” Peter, crushed but not convinced, subsided, and the Story Girl went on. “Finally, Margaret ran away with Alan, and they were married in Charlottetown. Alan intended to take his wife with him to California in the Fanny. If it was a hard journey for a man it was harder still for a woman, but Margaret would have dared anything for Alan’s sake. They had three days—ONLY three days—of happiness, and then the blow fell. The crew and the passengers of the Fanny refused to let Captain Dunbar take his wife with him. They told him he must leave her behind. And all his prayers were of no avail. They say he stood on the deck of the Fanny and pleaded with the men while the tears ran down his face; but they would not yield, and he had to leave Margaret behind. Oh, what a parting it was!” There was heartbreak in the Story Girl’s voice and tears came into our eyes. There, in the green bower of Uncle Stephen’s Walk, we cried over the pathos of a parting whose anguish had been stilled for many years. “When it was all over, Margaret’s father and mother forgave her, and she went back home to wait—to WAIT. Oh, it is so dreadful just to WAIT, and do nothing else. Margaret waited for nearly a year. How long it must have seemed to her! And at last there came a letter—but not from Alan. Alan was DEAD. He had died in California and had been buried there. While Margaret had been thinking of him and longing for him and praying for him he had been lying in his lonely, faraway grave.” Cecily sprang up, shaking with sobs. “Oh, don’t—don’t go on,” she implored. “I CAN’T bear any more.” “There is no more,” said the Story Girl. “That was the end of it—the end of everything for Margaret. It didn’t kill HER, but her heart died.” “I just wish I’d hold of those fellows who wouldn’t let the Captain take his wife,” said Peter savagely. “Well, it was awful said,” said Felicity, wiping her eyes. “But it was long ago and we can’t do any good by crying over it now. Let us go and get something to eat. I made some nice little rhubarb tarts this morning.” We went. In spite of new disappointments and old heartbreaks we had appetites. And Felicity did make scrumptious rhubarb tarts! About HackerNoon Book Series: We bring you the most important technical, scientific, and insightful public domain books. This book is part of the public domain. L. M. Montgomery (2004). The Story Girl. Urbana, Illinois: Project Gutenberg. Retrieved https://www.gutenberg.org/cache/epub/5342/pg5342-images.html This eBook is for the use of anyone anywhere at no cost and with almost no restrictions whatsoever. You may copy it, give it away or re-use it under the terms of the Project Gutenberg License included with this eBook or online at www.gutenberg.org , located at https://www.gutenberg.org/policy/license.html .

A TRAGEDY OF CHILDHOOD

An Afternoon at the Stone House

Just a Happy Day

A Danger Averted

116 Stories To Learn About Web Scraping

Hello everyone, my name is Dmitrii Ivashchenko and I'm a software engineer at MY.GAMES. In this article, we'll discuss the differences between garbage collection in Unity and .NET. One of the main advantages of the C# programming language is automatic memory management. It eliminates the need for developers to manually free memory from unused objects and significantly reduces development time. However, this can also lead to a large number of problems related to memory leaks and unexpected application crashes. Aside from general recommendations for avoiding garbage collection issues in C#, it's important to highlight the specifics of memory management implementation in Unity, which result in additional constraints. First, let's examine how memory management works in .NET. Memory management in .NET The (CLR) is the runtime environment for managed code in .NET. Any high-level .NET code written in languages like C#, F#, Visual Basic, and so on, is compiled into Intermediate Language (IL-code), which is then executed in the CLR. In addition to executing IL code, the CLR also provides several other necessary services, such as type safety, security boundaries, and automatic memory management. Common Language Runtime Managed heap When a new CLR process is initialized, it reserves a contiguous region of address space for that process — this space is referred to as the . The managed heap stores a pointer to the address where the next object will be allocated in the heap. Initially, this pointer is set to the base address of the managed heap. All reference types (often value types, too) are placed in the managed heap. managed heap When an application creates the first reference type, memory is allocated for the type at the base address of the managed heap. When creating the next object, the garbage collector allocates memory for it in the address space immediately following the first object. As long as address space is available, the garbage collector continues to allocate space for new objects in this way. Allocating memory from managed heap is faster than allocating unmanaged memory. For example, if you want to allocate memory in C++, you would need to make a system call to the operating system to do this. When using CLR, the memory is already reserved from the OS when the application is launched. Since the runtime allocates memory for an object by adding a value to the pointer, it happens almost as quickly as allocating memory from the stack. Additionally, because new objects that are allocated sequentially are also stored sequentially in the managed heap, an application can access objects very quickly. CLR GC algorithm The garbage collection (GC) algorithm in CLR is based on several considerations: New objects have a shorter life span, while older objects have a longer one. It's faster to compact the memory for a small portion of the managed heap than for the entire managed heap at once. New objects are usually related to each other and are available to the application at approximately the same time. Based on these time-tested assumptions, the CLR garbage collector algorithm is built as follows. There are three generations of objects: Generation 0: all new objects go into this generation. Generation 1: objects from generation 0 that survive one garbage collection are moved into this generation. Generation 2: objects from generation 1 that survive a second garbage collection are moved into this generation. Each generation has its own address space in memory and is processed independently of the others. When an application is launched, the garbage collector places all created objects in the generation 0 space. Once there is no longer enough space for another object, a garbage collection is triggered for generation 0. So, how can we determine the objects no longer in use? To do this, the garbage collector uses the of the application. This list usually includes: list of roots Static fields of classes. Local variables and method parameters. References to objects stored in the thread stack. References to objects stored in processor registers. Objects awaiting finalization. Objects associated with event handlers may also be included in the list of roots. The garbage collector analyzes the list of roots and builds a graph of accessible objects starting from the root objects. Objects that cannot be reached from the root objects are considered dead, and the garbage collector frees up the memory occupied by these objects. After dealing with any dead objects from generation 0, the garbage collector moves the remaining live objects to the address space of generation 1. Once becoming a part of generation 1, objects are less frequently considered by the garbage collector as candidates for removal. Over time, if cleaning up generation 0 does not provide enough space for creating new objects, garbage collection for generation 1 is performed. The graph is constructed again, dead objects are deleted again, and the surviving objects from generation 1 are moved to generation 2. If constant memory leaks are occurring, the garbage collector will consume the available address space of all three generations and decide to request additional space from the operating system. If memory leaks persist, the OS will continue to allocate additional memory to the process up to a certain limit, but when its available memory runs out, it will be forced to stop the process. Some details have intentionally been omitted in this text for ease of understanding, but nonetheless, this allows us to see the complexity and thoughtfulness of memory management organization in .NET. Now, equipped with this knowledge, let's take a look at how this is organized in Unity. Memory management in Unity Since the Unity game engine is written in C++, it obviously uses some amount of memory for its runtime which is not accessible to users (this is called ). It's also important to highlight a special type of memory called , which is used when you use Unity Collections structures like or . All other used memory space is and uses a garbage collector to allocate and free memory. native memory C# unmanaged memory NativeArray<T> NativeList<T> managed memory Due to the absence of CLR, memory management in Unity applications is handled by the scripting runtime (Mono or IL2CPP). However, it should be noted that these environments are not as efficient at memory management as .NET. One of the most frustrating consequences of this is fragmentation. M emory fragmentation Memory fragmentation in Unity is the process where the available memory space is divided into scattered blocks. Fragmentation occurs when an application constantly allocates and deallocates memory, leading to the division of free space in memory into many continuous blocks of different sizes. Memory fragmentation can be of two types: : this occurs when the free space in memory is divided into multiple non-contiguous blocks of different sizes distributed throughout the memory. As a result, there can be enough overall free memory available to accommodate a new object, but there might not be a suitable contiguous block of free space to accommodate it. External fragmentation : this occurs when allocated memory blocks have more space than required to store an object. This leaves unused memory space inside the allocated blocks, which can also lead to inefficient resource utilization. Internal fragmentation Memory fragmentation in Unity is particularly relevant due to the use of a conservative garbage collector that does not support memory compaction. Without compaction, freed memory blocks remain scattered, which can cause performance issues, especially in the long term. The Boehm-Demers-Weiser garbage collector Unity uses the conservative Boehm-Demers-Weiser (BDW) garbage collector, which stops the execution of your program and resumes normal execution only after completing its work. BDW's work algorithm can be described as follows: : the garbage collector pauses program execution for garbage collection. Stop the world : it scans the root pointers, determining all live objects directly accessible from the program code. Root scanning : it follows references from root objects to determine all available objects, creating a graph of accessible objects. Object tracing : it counts the number of references to each object. Reference counting : the garbage collector frees memory occupied by objects that have no references (dead objects). Memory reclamation : after garbage collection is complete, program execution continues. World resumption This is a conservative garbage collector, which means it does not require precise information about the location of all object pointers in memory. Instead, it assumes that any value in memory that can be a pointer to an object is a valid pointer. This allows the BDW garbage collector to work with programming languages that do not provide precise pointer information. Incremental garbage collection Starting with Unity 2019.1, the BDW is used in incremental mode by default. This means that the garbage collector distributes its workload across multiple frames instead of stopping the main CPU thread (stop-the-world garbage collection) to process all objects in the managed heap. Thus, Unity takes shorter breaks while executing your application, instead of having one long interruption, allowing the garbage collector to process objects in the managed heap. Incremental mode does not speed up garbage collection as a whole, but since it distributes the workload across multiple frames, performance spikes related to GC are reduced. Incremental garbage collection can be problematic for your application. In this mode, the garbage collector divides its work, including the marking phase. The marking phase is the scanning of all managed objects by the garbage collector to determine which objects are still in use and which can be cleared. Dividing the marking phase works well when most references between objects do not change between work fragments. However, too many changes can overload the incremental garbage collector and create a situation where the marking phase will never complete. In this case, the garbage collector will switch to performing a full, non-incremental collection. To inform the garbage collector of each reference change, Unity uses write barriers; this adds some overhead when changing references, which affects the performance of managed code. Let's compare: Unity (Boehm-Demers-Weiser) .NET GC Algorithm Conservative garbage collector Generational garbage collector Runtime Environment Mono or IL2CPP .NET Core, .NET Framework, .NET 5+ Root Scanning Less precise root scanning Precise root scanning Object Tracing Yes Yes Reference Counting Yes No Memory Compaction No Yes (except for large objects) Generations No Yes (0, 1, and 2) Speed Slower due to overhead and lack of compactification Faster due to generational approach and precise root scanning Stop-the-world Yes Yes (but with less impact on performance due to incremental garbage collection) Fragmentation Handling Prone to fragmentation (due to lack of object compaction) Reduced fragmentation (due to object compaction) Thus, with no CLR GC in Unity, we get a mechanism prone to heap fragmentation and slow, cumbersome garbage collection in the space of all objects not divided into generations. Let's consider the consequences from this that should be taken into account when developing games in Unity. Avoiding Unity GC limitation pitfalls Due to Unity Incremental GC limitations, the following pain points can be identified: GC calls are more expensive than in .NET (an entire object graph has to be processed, not just its subset). Frequent object creation and deletion leads to memory fragmentation; spaces between objects can only be filled with new objects of equal or smaller size. Frequent changes in object relationships make incremental GC mode difficult to work with (GC cycles take more frames and reduce FPS). Too frequent changes in object relationships (in each frame) result in GC switching to non-incremental mode; instead of spreading GC launches over multiple frames, we get one big world stop until garbage collection is complete. To avoid these pain points, be particularly careful with unnecessary memory allocations in the heap, which can cause garbage collection spikes: avoid passing value type variables instead of reference type variables. This creates a temporary object and potential garbage that is implicitly associated with it, converting value types to object types. A simple example: Boxing: int day = 15;
int month = 10;
int year = 2023;
Debug.Log(string.Format("Date: {0}/{1}/{2}", day, month, year)); It may seem that since , , and are value types and defined inside the function, there will be no allocations or garbage. But if we look into the implementation of , we will see that the arguments are of the type, which means that boxing of value types will occur and they will be placed in the heap: day month year string.Format() object /// Replaces the format items in a specified string with the string
/// representation of three specified objects.
public static string Format(string format, object arg0, object arg1, object arg2); the the C# language, strings are reference types, not value types. Try to minimize the creation and manipulation of strings. Use the class for working with strings during runtime. Strings: StringBuilder although the operator itself does not create garbage, creating a new object does: Coroutines: yield WaitForSeconds private IEnumerator BadExample()
{
  while (true)
  {
    // Creating a new WaitForSeconds object here causes garbage generation.
    yield return new WaitForSeconds(1f);
  }
}

private IEnumerator GoodExample()
{
  // Cache the WaitForSeconds object to avoid garbage generation.
  WaitForSeconds waitForOneSecond = new WaitForSeconds(1f);

  while (true)
  {
    // Reuse the cached WaitForSeconds object.
    yield return waitForOneSecond;
  }
} in general, try to avoid using closures in C# if possible. Minimize the use of anonymous methods and method references in performance-related code, especially if the code runs on every frame. Method references in C# are reference types and they are located on the heap. This means that when you pass a method reference as an argument, temporary memory allocations may occur. This happens regardless of whether an anonymous method is passed or a method is already defined. Moreover, when you turn an anonymous method into a closure, the amount of memory that needs to be passed to the closure method greatly increases. Closures and anonymous methods. both of these tools generate garbage due to hidden boxing. If performance is critical, avoid using LINQ and regular expressions. LINQ and regular expressions: please note that some functions allocate memory on the heap. Keep references to arrays in cache instead of allocating them inside a loop. Also, use functions that don't generate garbage. For example, prefer using instead of comparing strings with (as returning a new string creates garbage). Use non-allocating alternatives to Unity API methods, such as . Unity functions: GameObject.CompareTag GameObject.tag Physics.RaycastNonAlloc If you know that at a certain point in the game the garbage collection process will not affect the gaming experience (for example, on a loading screen), you can initiate garbage collection by calling . However, the best practice to strive for is zero-allocation. This means that you reserve all the memory you need at the start of the game or when loading a specific game scene, and then reuse that memory throughout the entire game cycle. In combination with object pools for reuse and the use of structures instead of classes, these techniques solve most memory management issues in the game. System.GC.Collect

Stories, News & Mentions about GOODERA

Goodera’s Stories on HackerNoon