How the Use of Machine Learning is Challenging the Retail Apocalypse

Whether retailers like it or not, the future of retail is here, in the form of smart algorithms. Machine learning will change much of the industry's norms, often for the better. Retail trends point to the store of the future being automated using the latest technology. Brick & Mortar, physical retail... however you like to call it, your favourite real-world store is about to get a whole lot more digital. Whether that's the best idea remains to be seen.

Anyone with a wet finger in the air will by now have heard of the “retail apocalypse” sweeping through the developed world’s malls. “People aren’t spending in stores anymore,” your quarter-informed uncle complains, before moaning that youths are too busy Instagramming their avocado brunches to burn crosses on peoples’ lawns.

Indeed, the old retailing models aren’t working as well as they used to. The fact that they were terrible models to start with probably had something to do with it.

The most forward-thinking retailers are adapting, however, and are not only surviving, but thriving. This is in part due to new technologies, which have enabled a wide range of innovations. Least of these technologies is machine learning, which can use troves of customer data to model, analyze, recognize and predict at levels and speeds we meat-sacks could never even hope to achieve (more info on the basics here).

The next few paragraphs give a cursory (if not mildly vulgarized) glance at the areas where the modern store owner might use this technology to stay afloat throughout the next holiday season and the next, proving that artificial intelligence is not only for the realm of digital brands and does indeed have a place in physical retail.

Forecasting

If fed enough usable data (think massive Excel tables on what was bought, when and where, amongst other variables), machine learning can not only tell us what happened, why it happened, and what will happen, but also what is the best thing that could happen (oh how I wish I’d had this power back in high school…) based on the available customer data, the season, the day of the week, activity at other area stores, eCommerce activity, etc…

As such, the technology has the ability to help brands and retailers know what products will be needed where and under what time-frame based on numerous variables, so that consumers can enjoy the benefits of better product availability, more adapted inventory, and even a heads up on current trends. This leads to a more efficient supply chain, with a better use of available resources.

Working with best-case scenarios also allows for better and more flexible pricing, making profit maximization easier. MR = MC. It’s also a better use of promotions; sales are currently more tactical than strategic, with the aim of getting larger market shares or getting rid of expensive stock. But that may not be true for long.

Indeed, better forecasting also means less useless stock and less promotions. Among the many problems facing today’s retail market, unsold stock might be one of its biggest handicaps: Unused inventory costs the U.S retail industry about $50 billion a year. Every dollar spent on what becomes dead inventory is valuable money that could have been put towards training talent, better R&D, or, most obviously, a more liquid inventory.

Forecasting is not sexy, but it saves millions and will save billions. And nothing is better than billions, regardless of currency.

Personalization

There was a time when most store owners would know most of their clients. But those clients wanted cheaper products. And more variety. So the industry concentrated. And expanded. And cities grew. And shoppers became simple numbers, as Traffic and Conversion metrics ruled over Retail-City with an iron fist.

This doesn’t cut it anymore, especially for younger generations, cocooned as they are by brands looking for life-long customers (looking at you, Juul).

Customers want products that fit their needs, right here and right now.

Ironically, statistics can help provide such personalization and retention by using a customer’s previous purchasing behaviors (assuming it was tracked), crossing them with those of all other customers, and drawing inferences about their future needs using advanced machine learning algorithms.

Mass personalization at scale is mostly done online at this time, and as such is out of this article’s scope… wait, what am I on about? There no longer is an offline and online; they have merged into Ma’s infamous “new retail.” Though we cannot alter a store to fit every customer (note that VR startups are trying), we can use machine learning to cater a specific store to a specific neighborhood, and use computer vision (ML with eyes) to inform the digital based on the physical (more on both below).

We’re not yet in a world where you pick up a box of cookies, put it back down in the store, and an email pops up giving you a discount on cookies once you’re home, but it’s coming. The benefits are obvious, and range from up-selling to complementary purchases.

Granular Merchandising

Most regular customers at major grocery stores have a loyalty card, which is linked to a profile with very basic information (age, gender…).

By 1. using this data, 2. using demographics data freely available and 3. using machine learning to draw inferences from the products bought and sold in various areas, it becomes possible to adapt the inventory carried by each store to its surrounding (ie: ramen and EasyMac for the student quarters).

The key question shifts from “where” and “how many” to “who”, “when”, “how often”, “how long” and “for how many cookies?”.

Most grocery stores in large cities have unique layouts with space restrictions, but those can be taken into account. If the store has 40 meters of shelves, and “basics” (tomatoes, beer, ice-cream…) take 30 meters, then the remaining 10 meters can be filled with the “best-sellers” for the categories of people who most visit the store, as well as a few new products. The weight of each can then be adapted and re-adapted as time goes.

It’s also possible to go even further than loyalty cards; computer vision could potentially identify different demographics if given enough training footage, helping inform strategic decisions about inventory.

It could also be used to identify what draws peoples’ eyes and what doesn’t, as well as create heat-maps, helping change and adapt store layouts. The cameras are already in most stores and are pretty useless as it is—why not make good use of them for once?

Think of it as the equivalent of 1,000 interns per store writing down what people look like and what they’re looking at each day.

Those interns are arguably cheaper in the U.S. because there are no labor laws there, but for civilized countries, computer vision might very well be the way to go. Check out RetailNext for more info.

Cashierless Stores

The use of mobile is, of course, a fantastic idea. If we throw privacy out the window, the store could track where we went before, where we go after, and link cookies to store behavior.

There are many other implications to these types of stores becoming prevalent:

• theft prevention,
• customer loyalty,
• automation…

The idea of identifying clients thanks to facial recognition, for example, would be particularly attractive to high fashion stores where high spenders could be identified, and their habits, past purchases and shipping address could appear on a screen for all sales associates to see.

This is already the case in luxury stores like Louis Vuitton, but in a much more analogue way.

Check out AiFi, Standard Cognition, Zippin and Trigo Vision for more info on cashierless stores.

Stock Visibility

Yes, this once again will refer to computer vision. Turns out, the real world is not 1s and 0s and needs to be seen to be analyzed. Who knew?

There are a few options to optimize stock visibility, but all of them go through a phase of training an algorithm to recognize ranges of products based on their packaging, and knowing how to count how many there are, how many there aren’t and, more importantly, how many there should be.

The latter is important because the till may know how many products there are in stock and how many were sold, but it doesn’t know if theft or counting errors (ie: suppliers skimming off the top) led to fewer products in store, in reality.

Indeed, cameras could be used to drastically limit theft by identifying when and where they occur. If we can see that an item has been picked up, we can also see that it’s been hidden. We could also potentially check for suspicious behavior based on past data (as shown here by startup StopLift). The ROI is also very easy to calculate (stores know very precisely how much they lose to theft and errors), so this tech is likely to be one of the first implemented.

Stock visibility is also key to optimizing something many retailers are keen to automate: restocking. It’s indeed repetitive and involves products which tend to be of the same shape, which is ideal for a robot going through aisles (the robotics of it all is another story entirely).

Employees could then be better used for higher-value tasks. Or so they say: them being fired is probably just as likely.

One of the most original such ideas comes from startup Focal Systems ($2.8M), which aims to equip existing shopping carts with computer tablets that use computer vision to monitor the shelves as the cart moves through the store, and also display digital ads to the shopper (1 stone, 2 birds).

Why use robotics when the customer is the cheapest robot out there?

Natural Language Processing

Speaking of robots, much in the same way that an algorithm can be taught to “see”, it can also be taught to “hear” and “speak.” This goes even further than computer vision, as language is full of subtleties and ultra-local differences that are hard to predict.

This hasn’t stopped a few stores from replacing greeters with in-store conversational interfaces, which would suggest products and answer questions for customers.

I believe this is just a gimmick, but it will be added here for the sake of fairness. Whether the excitement wears off once retail robots become “the norm” remains to be determined.

This hasn’t stopped companies from investing in such ideas however. Beyond the example above, a startup called TwentyBN has mixed computer vision to analyze what shoppers are doing with some very basic conversational notions to create a bot that might push customers to spend more.

Conclusion

Much of the above may sound like science fiction. And to most retailers, it is. Though there’s a lot of talk about the blurring between online and offline, most companies first need to concentrate on optimizing physical retail as it is (workforce, supply chain, pricing…). God knows there’s a lot of work.

Beyond building healthy foundations, none of the above is possible without vast amounts of data. This data exists for retailers, but knowing how to gather it, clean it, and use it is far from easy!

Privacy issues are also likely to abound: tracking every movement in a store and compiling data, which could be mixed with online behavior, is not the most ethical thing one can do.

Nevertheless, and despite all hurdles, there are hope that the technology will be implemented far and wide, in one form or another, within the next 10 years. Not because it will increase profit (well, that too), but because, perhaps counter-intuitively, it will make retailing a more human experience.

This article was originally written for The Pourquoi Pas, an online magazine providing in-depth analyses of today’s technological challenges.

Previously published at https://www.thepourquoipas.com/post/guide-machine-learning-retail