Headed to Lowe&#x27;s this Easter Sunday? Find out if stores will be open or closed in observance of the holiday along with all the details on 2024 store hours.

Is Lowe's open on Easter 2024? Everything to know on store hours this year

Lowe&#x27;s Companies ( NYSE:LOW ) Full Year 2024 Results Key Financial Results Revenue: US$86.4b (down 11% from FY 2023...

Lowe's Companies Full Year 2024 Earnings: EPS Beats Expectations

Get the most recent info and news about Lowe's on HackerNoon, where 10k+ technologists publish stories for 4M+ monthly readers.

Lowe's Companies, Inc., together with its subsidiaries, operates as a home improvement retailer in the United States, Canada, and Mexico. The company offers a line of products for construction, maintenance, repair, remodeling, and decorating. It provides home improvement products in various categories, such as appliances, dÃ©cor, paint, hardware, millwork, lawn and garden, lighting, lumber and building materials, flooring, kitchens and bath, rough plumbing and electrical, seasonal and outdoor living, and tools. It also offers installation services through independent contractors in various product categories; extended protection plans; and in-warranty and out-of-warranty repair services. The company sells its national brand-name merchandise and private branded products to homeowners, renters, and professional customers. As of January 31, 2020, it operated 1,977 home improvement and hardware stores. The company also sells its products through websites comprising Lowes.com and Lowesforpros.com; and through mobile applications. Lowe's Companies, Inc. was founded in 1946 and is based in Mooresville, North Carolina.

Lowe's

MRS. HURTLE AT LOWESTOFT

Funding woes are still plaguing the . According to a survey titled, Startup Perspectives – H1 CY23  capturing the current investment outlook for Indian startups, the Indian startup ecosystem reported the lowest six-month funding in the last four years in H1CY23 at US$3.8 billion across 298 deals – a decline of nearly 36% as compared to H2 CY22 (US$ 5.9 billion). Indian startup ecosystem PwC India During the last few quarters, despite challenging funding market conditions, investors have shown strong support for their portfolio companies by doubling down on their investments in companies that demonstrated positive growth, the report stated. Amit Nawka, Partner – Deals & India Startups Leader, PwC India, said, “A funding winter is just a season in a startup’s journey. There is a slowdown in startup funding despite significant untapped capital reserves held by venture capitalists (VCs). Active VC firms in India have secured new funds in the past year and we can expect the pace of investments to pick up in the next few months. There is a slowdown in startup funding despite significant untapped capital reserves held by venture capitalists (VCs) Amit Nawka, Partner – Deals & India Startups Leader, PwC India In the interim, there has been an increase in the due diligence being carried out by investors before making investments, both in terms of detailing as well as coverage – from typical finance and legal to areas like technology, HR, and business processes – to ensure that the startups have a robust corporate governance framework.” Other statistics reveal that growth and late-stage funding deals accounted for 84% of the funding activity in H1 CY23 (in value terms). These represented 43% of the total count of deals in this period. The average ticket size in growth-stage deals was US$ 19 million and in late-stage deals were US$ 52 million during H1 CY23. Also, 80 M&A deals involving startups were executed in H1 CY23. While this news is saddening for startups, Vinod Khosla, Venture Capitalist at Investor in , has some words of advice. Khosla Ventures, OpenAI “The funding environment is down but all it means is your competition is also not getting funding; capital efficiency will go up. With all the layoffs, it’s a perfect time to hire entrepreneurs. When things are great and there’s euphoria, it’s a bad time to invest for entrepreneurs, and also for investors. The funding environment is down but all it means is your competition is also not getting funding; capital efficiency will go up. With all the layoffs, it’s a perfect time to hire for entrepreneurs Vinod Khosla, Venture Capitalist at Khosla Ventures Investors have been very cautious in India as they focus on revenue too much and not building assets. There is not a single example where I’ve used an IRR calculation to decide whether to invest or not at Khosla Ventures. Too much founder friendliness or focus on revenue is not right. The founder should be treated like a child. Saying yes to everything will spoil them and you’ll do them a disservice. Completely controlling them means they won’t be innovative and creative. A good investor is an assistance to its entrepreneurs.” Amongst those who have been able to secure funding, Fintech, SAAS, and D2C continued to be the most funded sectors in H1 CY23. SaaS, D2C, FinTech, e-commerce B2B and Logi and AutoTech continue to be the top five invested sectors based on the funding received in the first half of the year. These contribute to approximately 89% of the total funding received in H1 CY23 in value terms. The D2C and online gaming sectors each saw investments increase by almost 3x that of H2 CY22. In the FoodTech sector, investments increased by four times in H1 CY23 compared to H2 CY22 in value terms. Bengaluru, NCR, and Mumbai continue to be the key startup cities in India, representing around 83% of the total startup funding activity in H1 CY23. The Indian startup ecosystem is currently facing funding challenges. Despite this, there is optimism for the future The slowdown in funding presents an opportunity for startups to focus on capital efficiency and make strategic hires. Despite the current funding environment, it is crucial for investors to strike a balance between caution and providing necessary assistance to entrepreneurs. This article was originally published by on Navanwita Bora Sachdev The Tech Panda.

Indian Startup Ecosystem Faces Funding Slump: Lowest 6-Month Funding in 4 Years

The start to this week hasn’t quite gone the way that many in the crypto space would have wanted. Coin prices have dropped to some of their lowest points in almost a year, and there are now fears that we could be in for even bigger drops in the coming days. The issue stems from the broader traditional economy, where countries are seeing massive corrections and governments are looking to stem the tide of inflation and deal with supply chain concerns. While the most ardent cryptocurrency supporter or investor would easily say that we’ve been here before and there isn’t much of a reason to panic, there is a sense that we might be in for a much deeper correction. However, what is the signs that we could look towards to show whether we’re almost out of the woods or there’s more of this mania to come? Record-Nearing Lows Regardless of how you look at it, there’s a very little speck of hope for the crypto market in the long term. , a popular crypto analysis site, shows that the price of Bitcoin currently stands at $31,249. This is the lowest price peg for the leading asset in the past 8 months. Coincheckup The asset even hit $29,761 on Monday - marking the first time that Bitcoin could cross below the $30,000 mark since July 2021. At the time the crypto market was reeling from a massive downturn that began with and stretched up until a massive liquidation from the Grayscale Bitcoin Investment Trust. Tesla abandoning Bitcoin However, the current drop is coming due to fears from the traditional economy. The Federal Reserve announced massive interest rate hikes last week, sending the entire economy into a tailspin. With individuals and companies now finding it more difficult to borrow money, many lack an incentive to invest in cryptocurrencies. The Fed announcement caused the price of Bitcoin to close at the $34,000 price point last week. This marked the sixth week in a row that Bitcoin’s price will close lower - a feat not seen since August 2014. So, considering that many analysts are pointing out bleak outlooks for the Bitcoin and broader crypto markets in the short term, it’s easy to see why. Many of these analysts understand that we seem to be entering uncharted territory, and only strong investors would be rewarded for their tenacity. Inflationary Trend Continues? As explained earlier, the primary cause of this current bearish run is the weakness in the larger macroeconomic environment. Stocks are even more disturbing at this point. Several of the top tech stocks have seen up to 20 per cent drops, with companies like Facebook and Netflix suffering severely. Now that the crypto market is starting to move in line with the , drops in the latter are sure to affect the former. traditional stock market Things are also not so great on the macroeconomic front. The world is still dealing with supply chain constraints, and the conflict in Ukraine isn’t helping matters. With reports that China is now dealing with what looks like a resurgence in coronavirus cases, it is entirely possible that the country takes measures to seal itself once more. All of these will put massive constraints on the global economy. This week, we expect to see the release of April's consumer price index (CPI) figures in the United States. The smart money will bet on these numbers reflecting the dire status of the global economy. President Joe Biden is expected to address the issue of inflation this week as well, just before the release of the CPI figures on Wednesday. The CPI for March was 8.5 per cent - an alarming number according to most experts. If the figures look even worse in April, then the forecast for inflation will be even more realistic. Definitely, that can’t be so great. An Opportunity To Buy The Dip While all of this is bleak, it is worth noting that the current market situations also present a massive opportunity for investors to make money. If there is one thing that crypto has shown from time to time, it is that it’s a very resilient asset class. Short-term outlooks might be bleak, but this doesn’t necessarily mean that the entire market is about to go down the drain. Even if it takes months, crypto will bounce back. Thus, some opportunistic investors will be attracted to the possibility of buying the dip right now and enjoying the massive gains. Institutional interest in crypto remains high, with companies like MicroStrategy and the committing over $1 billion collectively towards investing in Bitcoin across the past month alone. Luna Foundation Guard If there’s anything that can be good for the crypto space, it is increased attention from institutional investors. At the end of the day, these are the individuals and companies with the deepest pockets whose investments - or withdrawal from the market - can cause seismic shifts in the dynamic of things. If they’re still pouring money into the market, then we can all heave a sigh of relief. All in all, it will be impossible to sugarcoat things and say that the market is looking bright in the short term. The CPI report coming this week could serve as a , and then we could see Bitcoin and other altcoins take further dumps. But, the future remains bright. Those who can see that will eventually get gains. further drawdown on the market

What To Watch As Bitcoin Hits Its Lowest Point Since July'21

Using Scanned Mesh Data for Auto-Digitized 3D Modeling: Methodology

Authors:
(1) Vasiliki Kougia, University of Vienna, Faculty of computer science, Vienna, Austria & UniVie Doctoral School Computer Science, Vienna, Austria;
(2) Simon Fetze, University of Vienna, Faculty of computer science, Vienna, Austria;
(3) Thomas Kirchmair, University of Vienna, Faculty of computer science, Vienna, Austria;
(4) Erion Çano, University of Vienna, Faculty of computer science, Vienna, Austria;
(5) Sina Moayed Baharlou, Boston University, Department of Electrical and Computer Engineering, Boston, MA, USA;
(6) Sahand Sharifzadeh, Ludwig Maximilians University of Munich, Faculty of Computer Science, Munich, Germany;
(7) Benjamin Roth, University of Vienna, Faculty of computer science, Vienna, Austria. Table of Links Abstract and Introduction Related Work MemeGraphs Benchmarking Analysis Conclusion, Acknowledgments, and References 4 Benchmarking 4.1 Models To evaluate the results of our automatic augmentation and how it can affect the performance of hateful memes classification, we employed a text-based Transformer with and without the MemeGraphs information as well as a multimodal model. Previous work has shown that using only the text of the memes for classification gives results highly competitive with multimodal methods [14, 13, 28]. On the other hand, unimodal image-based models have lowest results. We use a pre-trained BERT [9] model that takes as input the text of each meme (TxtBERT) [14] and different variants of this approach that include the MemeGraphs input. In order to feed the graphs as input to the model, we represent them as a text sequence (see Section 3), which is given as an extra text input after the [SEP] token (Fig. 3). In the model that we call MemeGraphs[SceneGr], this sequence contained only the scene graphs, which were represented by triplets of the detected objects and the relations between them (Eq. 2), e.g., “0-man has 11-eye. 0-man wearing 12-shirt.”. In the model called MemeGraphs[Know], the text descriptions from the knowledge base corresponding to the detected entities are added as extra input (Eq. 3). For example, the second input sequence for the meme shown in Figure 3 will be “American politician, businessman, and 29th Governor of New Mexico.”. While in the MemeGraphs[SceneGr+Know] model, information from the whole knowledge graphs is added, comprising the scene graphs and the background knowledge concatenated with a full stop into one sequence (see Subsection 3.3). In all the models, the [CLS] token is fed into a final linear layer with a sigmoid activation function that produced the probability of the meme being hateful. In order to compare our MemeGraphs approach with a method that employs learned visual representations, we use ImgBERT, an early fusion multimodal model (see Section 2). 4.2 Experimental Setup Each meme was labeled in the original dataset as offensive or non-offensive. The classification was based on the offensiveness labels that exist in the MultiOFF dataset. The dataset is slightly imbalanced with 42% of the samples being offensive. For training and testing the models, we used the split provided by the authors of the dataset. The split consisted of training, validation, and test sets, which contained 445, 149, and 149 memes respectively. We employed the pre-trained BERT model provided by Hugging Face and fine-tuned it on our dataset.[10] To obtain image embeddings, we experimented with several pre-trained CNN models and based on the best results we chose DenseNet with 161 layers. The embeddings were extracted from the last pooling layer. The max length was set as the average token length of the training texts, depending on the input of each model. The models were trained using batch size 16, the weighted binary cross entropy loss, and AdamW optimizer with an initial learning rate of 2e-5. We used early stopping based on the validation loss with a patience of 3 epochs. We trained all models 20 times with different seeds and the training lasted between 4 and 6 epochs in each run. During inference, a threshold was used to determine if a meme is hateful or not based on the produced probability. This threshold was set to 0.5 for all the models. 4.3 Benchmarking results For each model, we obtained predictions for the test set from all twenty differently initialized runs. We evaluated each prediction set by calculating the F1 score defining the minority class (offensive class) as positive. In Table 1, we report the average F1 score over the twenty runs for each method and the corresponding standard error. In Table 2, the test score of the model that achieved the best score on the development set is shown. We observe that MemeGraphs[SceneGr+Know] outperforms the other models when predicting with the best checkpoint (Table 2). On the other hand, when looking at the average, the model with only the background knowledge as input achieves the best F1 score (MemeGraphs[Know]). We generally observe that the methods that incorporate our automatic augmentation (MemeGraphs) outperform the simple text-based fine-tuned BERT. Also, ImgBERT is outperformed showing that when the visual information is employed in the form of scene graphs the performance is improved, compared to when using only the image embeddings. [10] https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertForSequenceClassification This paper is available on arxiv under CC 4.0 license. Authors: (1) Vasiliki Kougia, University of Vienna, Faculty of computer science, Vienna, Austria & UniVie Doctoral School Computer Science, Vienna, Austria; (2) Simon Fetze, University of Vienna, Faculty of computer science, Vienna, Austria; (3) Thomas Kirchmair, University of Vienna, Faculty of computer science, Vienna, Austria; (4) Erion Çano, University of Vienna, Faculty of computer science, Vienna, Austria; (5) Sina Moayed Baharlou, Boston University, Department of Electrical and Computer Engineering, Boston, MA, USA; (6) Sahand Sharifzadeh, Ludwig Maximilians University of Munich, Faculty of Computer Science, Munich, Germany; (7) Benjamin Roth, University of Vienna, Faculty of computer science, Vienna, Austria. Authors: Authors: (1) Vasiliki Kougia, University of Vienna, Faculty of computer science, Vienna, Austria & UniVie Doctoral School Computer Science, Vienna, Austria; (2) Simon Fetze, University of Vienna, Faculty of computer science, Vienna, Austria; (3) Thomas Kirchmair, University of Vienna, Faculty of computer science, Vienna, Austria; (4) Erion Çano, University of Vienna, Faculty of computer science, Vienna, Austria; (5) Sina Moayed Baharlou, Boston University, Department of Electrical and Computer Engineering, Boston, MA, USA; (6) Sahand Sharifzadeh, Ludwig Maximilians University of Munich, Faculty of Computer Science, Munich, Germany; (7) Benjamin Roth, University of Vienna, Faculty of computer science, Vienna, Austria. Table of Links Abstract and Introduction Abstract and Introduction Related Work Related Work MemeGraphs MemeGraphs Benchmarking Benchmarking Analysis Analysis Conclusion, Acknowledgments, and References Conclusion, Acknowledgments, and References 4 Benchmarking 4.1 Models To evaluate the results of our automatic augmentation and how it can affect the performance of hateful memes classification, we employed a text-based Transformer with and without the MemeGraphs information as well as a multimodal model. Previous work has shown that using only the text of the memes for classification gives results highly competitive with multimodal methods [14, 13, 28]. On the other hand, unimodal image-based models have lowest results. We use a pre-trained BERT [9] model that takes as input the text of each meme (TxtBERT) [14] and different variants of this approach that include the MemeGraphs input. In order to feed the graphs as input to the model, we represent them as a text sequence (see Section 3), which is given as an extra text input after the [SEP] token (Fig. 3). In the model that we call MemeGraphs[SceneGr], this sequence contained only the scene graphs, which were represented by triplets of the detected objects and the relations between them (Eq. 2), e.g., “0-man has 11-eye. 0-man wearing 12-shirt.”. In the model called MemeGraphs[Know], the text descriptions from the knowledge base corresponding to the detected entities are added as extra input (Eq. 3). For example, the second input sequence for the meme shown in Figure 3 will be “American politician, businessman, and 29th Governor of New Mexico.”. While in the MemeGraphs[SceneGr+Know] model, information from the whole knowledge graphs is added, comprising the scene graphs and the background knowledge concatenated with a full stop into one sequence (see Subsection 3.3). In all the models, the [CLS] token is fed into a final linear layer with a sigmoid activation function that produced the probability of the meme being hateful. In order to compare our MemeGraphs approach with a method that employs learned visual representations, we use ImgBERT, an early fusion multimodal model (see Section 2). 4.2 Experimental Setup Each meme was labeled in the original dataset as offensive or non-offensive. The classification was based on the offensiveness labels that exist in the MultiOFF dataset. The dataset is slightly imbalanced with 42% of the samples being offensive. For training and testing the models, we used the split provided by the authors of the dataset. The split consisted of training, validation, and test sets, which contained 445, 149, and 149 memes respectively. We employed the pre-trained BERT model provided by Hugging Face and fine-tuned it on our dataset.[10] To obtain image embeddings, we experimented with several pre-trained CNN models and based on the best results we chose DenseNet with 161 layers. The embeddings were extracted from the last pooling layer. The max length was set as the average token length of the training texts, depending on the input of each model. The models were trained using batch size 16, the weighted binary cross entropy loss, and AdamW optimizer with an initial learning rate of 2e-5. We used early stopping based on the validation loss with a patience of 3 epochs. We trained all models 20 times with different seeds and the training lasted between 4 and 6 epochs in each run. During inference, a threshold was used to determine if a meme is hateful or not based on the produced probability. This threshold was set to 0.5 for all the models. 4.3 Benchmarking results For each model, we obtained predictions for the test set from all twenty differently initialized runs. We evaluated each prediction set by calculating the F1 score defining the minority class (offensive class) as positive. In Table 1, we report the average F1 score over the twenty runs for each method and the corresponding standard error. In Table 2, the test score of the model that achieved the best score on the development set is shown. We observe that MemeGraphs[SceneGr+Know] outperforms the other models when predicting with the best checkpoint (Table 2). On the other hand, when looking at the average, the model with only the background knowledge as input achieves the best F1 score (MemeGraphs[Know]). We generally observe that the methods that incorporate our automatic augmentation (MemeGraphs) outperform the simple text-based fine-tuned BERT. Also, ImgBERT is outperformed showing that when the visual information is employed in the form of scene graphs the performance is improved, compared to when using only the image embeddings. [10] https://huggingface.co/docs/transformers/model_doc/bert#transformers.BertForSequenceClassification This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

Comparing Meme Analysis Methods: MemeGraphs vs. ImgBERT

What Does It Take to Get a Medical Device to Market?

Understanding the Privacy Risks of Popular Search Engine Advertising Systems: 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 V. EXPERIMENTAL RESULTS A. Comparison of CUP against Baselines • LLM-Rec also performs poorly. Solely relying on the LLM’s latent knowledge about books is not sufficient when coping with long-tail items that appear sparsely in the LLM’s training data. Popularity and position bias [31] further aggravate this adverse effect. Thus, to address RQ1 (see Section I-B), merely and fully resorting to LLMs alone is not a viable solution. We also ran extended CUP configurations where precomputed CF vectors were incorporated (see Section III-B). These achieved only very minor gains in some of the settings, and there was even a detrimental effect on some cases. We omit these hybrid methods here. We will revisit them in Subsection V-C for sensitivity studies on data density. Another observation is that the weighted training almost always improves performance. Therefore, the following subsections focus on results with weighted training samples. B. Comparison of CUP Configurations • On the GR-1K-rich data, the overall performance is higher than for AM across all configurations. This can be attributed to the fact that GR has longer reviews, and these texts tend to be more informative than the ones in the AM data (which sometimes refer to packaging, shipment and other non-content aspects). C. Influence of Interaction Density The slices AM-1K-rich and GR-1K-rich are constructed with a text-centric stress-test in mind. Both are extremely sparse in terms of user-item interactions. To study the influence of sparseness in isolation, we created two larger samples of both datasets, one still sparse by design and the other denser in terms of users sharing items. We refer to these as AM-10Ksparse and AM-10K-dense, and analogously for GR. All these samples cover 10K users: 10x more users than our previous text-rich slices, to allow more potential for learning from interactions. Table I in Section IV-B shows statistics. The datasets are constructed as follows. To ensure connectivity in the interaction graphs, we first select 500 users and sample 2000 books connected to these users, both uniformly at random. This is the seed for constructing two variants of data, based on the cumulative item degrees of users, that is, the sum of the #users per book for all books that the user has in her collection: • 10K-dense: We randomly select 10K (minus the initial 500) users from all users in proportion to the users’ cumulative item degrees. This favors users with many or popular books. • 10K-sparse: We select users inversely proportional to their cumulative item degrees, thus favoring sparse users. In this sensitivity study, we focus on comparing CF (based on matrix factorization) against our default CUPtext configuration (with idf-selected sentences) and the hybrid combination CUPtext +CF. Tables V and VI show the results for the AM10K data and GR-10K data, comparing the sparse vs. the dense variants with search-based evaluation (corresponding results with standard evaluation are on the project web page). Key insights are: • As the 10K-sparse data is already much denser than the stress-test 1K-rich slices, CF can achieve decent results on the sparse data. Especially in search-based mode, CF is almost on par with CUP. Note that these gains come from the seen items alone, as CF is bound to fail on unseen items. • CUPtext is still the clear winner on the sparse variant. The hybrid configuration, with text and CF vectors combined, is inferior to learning from text alone. • For the 10K-dense data, CF alone performs well, as this is the traditional regime for which CF has been invented. An interesting point arises for the search-based mode and with dense data. Here, the negative test points are closer to the positive sample (after the BM25 search), and pose higher difficulty for text cues alone to discriminate them. Thus, CF becomes more competitive. On GR, hybrid CUP+CF performs best. D. Efficiency of CUP Two architectural choices make CUP efficient: • input length restricted to 128 tokens, and • fine-tuning only the last layer of BERT and the FFN layers. For more analysis, we measured the training time and resulting NDCG on the GR-1K-rich data, comparing different input size budgets and choices of which parameters are tuned and which ones are kept frozen. Figure 2 shows the results. We observe that the 128-token configuration has the lowest training cost: significantly less time per epoch than the other variants and fast convergence (reaching its best NDCG already after 15 epochs in ca. 3000 seconds). The 256- and 512-token models eventually reach higher NDCG, but only by a small margin and after much longer training time. This confirms that concise user profiles, with a small token budget, are the best approach in terms of benefit/cost ratio. As for configurations with more or less tunable parameters, we observe that the variant with frozen BERT takes much longer to converge and is inferior to the preferred CUP method even after more than 50 epochs. The other extreme, allowing all of BERT parameters to be altered together with the FFN parameters, performs best after enough training epochs, better than the CUP configuration that tunes only the last BERT layer. However, it takes almost twice as much time per epoch. So again, from the benefit/cost perspective, our design choice hits a sweet spot in the spectrum of model configurations. 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 V. EXPERIMENTAL RESULTS A. Comparison of CUP against Baselines A. Comparison of CUP against Baselines • LLM-Rec also performs poorly. Solely relying on the LLM’s latent knowledge about books is not sufficient when coping with long-tail items that appear sparsely in the LLM’s training data. Popularity and position bias [31] further aggravate this adverse effect. Thus, to address RQ1 (see Section I-B), merely and fully resorting to LLMs alone is not a viable solution. We also ran extended CUP configurations where precomputed CF vectors were incorporated (see Section III-B). These achieved only very minor gains in some of the settings, and there was even a detrimental effect on some cases. We omit these hybrid methods here. We will revisit them in Subsection V-C for sensitivity studies on data density. Another observation is that the weighted training almost always improves performance. Therefore, the following subsections focus on results with weighted training samples. B. Comparison of CUP Configurations • On the GR-1K-rich data, the overall performance is higher than for AM across all configurations. This can be attributed to the fact that GR has longer reviews, and these texts tend to be more informative than the ones in the AM data (which sometimes refer to packaging, shipment and other non-content aspects). C. Influence of Interaction Density C. Influence of Interaction Density The slices AM-1K-rich and GR-1K-rich are constructed with a text-centric stress-test in mind. Both are extremely sparse in terms of user-item interactions. To study the influence of sparseness in isolation, we created two larger samples of both datasets, one still sparse by design and the other denser in terms of users sharing items. We refer to these as AM-10Ksparse and AM-10K-dense, and analogously for GR. All these samples cover 10K users: 10x more users than our previous text-rich slices, to allow more potential for learning from interactions. Table I in Section IV-B shows statistics. The datasets are constructed as follows. To ensure connectivity in the interaction graphs, we first select 500 users and sample 2000 books connected to these users, both uniformly at random. This is the seed for constructing two variants of data, based on the cumulative item degrees of users, that is, the sum of the #users per book for all books that the user has in her collection: • 10K-dense: We randomly select 10K (minus the initial 500) users from all users in proportion to the users’ cumulative item degrees. This favors users with many or popular books. • 10K-sparse: We select users inversely proportional to their cumulative item degrees, thus favoring sparse users. In this sensitivity study, we focus on comparing CF (based on matrix factorization) against our default CUPtext configuration (with idf-selected sentences) and the hybrid combination CUPtext +CF. Tables V and VI show the results for the AM10K data and GR-10K data, comparing the sparse vs. the dense variants with search-based evaluation (corresponding results with standard evaluation are on the project web page). Key insights are: • As the 10K-sparse data is already much denser than the stress-test 1K-rich slices, CF can achieve decent results on the sparse data. Especially in search-based mode, CF is almost on par with CUP. Note that these gains come from the seen items alone, as CF is bound to fail on unseen items. • CUPtext is still the clear winner on the sparse variant. The hybrid configuration, with text and CF vectors combined, is inferior to learning from text alone. • For the 10K-dense data, CF alone performs well, as this is the traditional regime for which CF has been invented. An interesting point arises for the search-based mode and with dense data. Here, the negative test points are closer to the positive sample (after the BM25 search), and pose higher difficulty for text cues alone to discriminate them. Thus, CF becomes more competitive. On GR, hybrid CUP+CF performs best. D. Efficiency of CUP D. Efficiency of CUP Two architectural choices make CUP efficient: • input length restricted to 128 tokens, and • fine-tuning only the last layer of BERT and the FFN layers. For more analysis, we measured the training time and resulting NDCG on the GR-1K-rich data, comparing different input size budgets and choices of which parameters are tuned and which ones are kept frozen. Figure 2 shows the results. We observe that the 128-token configuration has the lowest training cost: significantly less time per epoch than the other variants and fast convergence (reaching its best NDCG already after 15 epochs in ca. 3000 seconds). The 256- and 512-token models eventually reach higher NDCG, but only by a small margin and after much longer training time. This confirms that concise user profiles, with a small token budget, are the best approach in terms of benefit/cost ratio. As for configurations with more or less tunable parameters, we observe that the variant with frozen BERT takes much longer to converge and is inferior to the preferred CUP method even after more than 50 epochs. The other extreme, allowing all of BERT parameters to be altered together with the FFN parameters, performs best after enough training epochs, better than the CUP configuration that tunes only the last BERT layer. However, it takes almost twice as much time per epoch. So again, from the benefit/cost perspective, our design choice hits a sweet spot in the spectrum of model configurations.

LOWE'S

lowes.com

10K+ employees

Since 1946

COMPANY RANKING

LOWE'S

EVERGREEN INDEX #6455

Representing this company? Click here to claim and customize this page!

Lowe's WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

LOWE'S

lowes.com

10K+ employees

Since 1946

COMPANY RANKING

LOWE'S

EVERGREEN INDEX #6455

RELATED COMPANIESRANK

STORIES ABOUT LOWE'S

Representing this company? Click here to claim and customize this page!

Lowe's WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!