In this paper we introduce a methodology that utilizes Large Language Models (LLMs) for efficient data annotation which is essential for student models training. We specifically focus on classification tasks, assessing the impact of various annotation methods. We employ LLMs, such as GPT-4 and 7B parameter open-source models, as teacher models to facilitate the annotation process. We proceed to evaluate the efficiency of smaller, more deployment-efficient student models like BERT in production environments. We show two alternative methods. One using a cascade of LLMs distillation, starting from GPT-4 and following by 7B parameter LLM and BERT. And a second method, which shows the feasibility of annotating data without relying on proprietary model like GPT-4, by exclusively using a 7B parameter open-source model, leveraging self-training methods. This strategic use of smaller LLMs in conjunction with smaller student models presents an efficient and cost-effective solution for enhancing classification tasks, offering insights into the potential of utilizing large-scale language models in E-Commerce production settings.

Product reviews play a key role in e-commerce platforms. Studies show that many users read product reviews before a purchase and trust them to the same extent as personal recommendations. However, in many cases, the number of reviews per product is large and extracting useful information becomes a challenging task. Severalwebsites have recently added an option to post tips-short, concise, practical, and self-contained pieces of advice about the products. These tips are complementary to the reviews and usually add a new non-trivial insight about the product, beyond its title, attributes, and description. Yet, most if not all major e-commerce platforms lack the notion of a tip as a first-class citizen and customers typically express their advice through other means, such as reviews. In this work, we propose an extractive method for tip generation from product reviews. We focus on five popular e-commerce domains whose reviews tend to contain useful non-trivial tips that are beneficial for potential customers. We formally define the task of tip extraction in e-commerce by providing the list of tip types, tip timing (before and/or after the purchase), and connection to the surrounding context sentences. To extract the tips, we propose a supervised approach and leverage a publicly available dataset, annotated by human editors, containing 14,000 product reviews. To demonstrate the potential of our approach, we compare different tip generation methods and evaluate them both manually and over the labeled set. Our approach demonstrates particularly high performance for popular products in the Baby, Home Improvement, and Sports & Outdoors domains, with precision of over 95% for the top 3 tips per product. In addition, we evaluate the performance of our methods on previously unseen domains. Finally, we discuss the practical usage of our approach in real-world applications. Concretely, we explain how tips generated from user reviews can be integrated in various use cases within e-commerce platforms and benefit both buyers and sellers.

In e-commerce, the watchlist enables users to track items over time and has emerged as a primary feature, playing an important role in users' shopping journey. Watchlist items typically have multiple attributes whose values may change over time (e.g., price, quantity). Since many users accumulate dozens of items on their watchlist, and since shopping intents change over time, recommending the top watchlist items in a given context can be valuable. In this work, we study the watchlist functionality in e-commerce and introduce a novel watchlist recommendation task. Our goal is to prioritize which watchlist items the user should pay attention to next by predicting the next items the user will click. We cast this task as a specialized sequential recommendation task and discuss its characteristics. Our proposed recommendation model, Trans2D, is built on top of the Transformer architecture, where we further suggest a novel extended attention mechanism (Attention2D) that allows to learn complex item-item, attribute-attribute and item-attribute patterns from sequential-data with multiple item attributes. Using a large-scale watchlist dataset from eBay, we evaluate our proposed model, where we demonstrate its superiority compared to multiple state-of-the-art baselines, many of which are adapted for this task.

We address the product question generation task. For a given product description, our goal is to generate questions that reflect potential user information needs that are either missing or not well covered in the description. Moreover, we wish to cover diverse user information needs that may span a multitude of product types. To this end, we first show how the T5 pre-trained Transformer encoder-decoder model can be fine-tuned for the task. Yet, while the T5 generated questions have a reasonable quality compared to the state-of-the-art method for the task (KPCNet), many of such questions are still too general, resulting in a sub-optimal global question diversity. As an alternative, we propose a novel learning-to-diversify (LTD) fine-tuning approach that allows to enrich the language learned by the underlying Transformer model. Our empirical evaluation shows that, using our approach significantly improves the global diversity of the underlying Transformer model, while preserves, as much as possible, its generation relevance.

Recent advances in the e-commerce fashion industry have led to an exploration of novel ways to enhance buyer experience via improved personalization. Predicting a proper size for an item to recommend is an important personalization challenge, and is being studied in this work. Earlier works in this field either focused on modeling explicit buyer fitment feedback or modeling of only a single aspect of the problem (e.g., specific category, brand, etc.). More recent works proposed richer models, either content-based or sequence-based, better accounting for content-based aspects of the problem or better modeling the buyer's online journey. However, both these approaches fail in certain scenarios: either when encountering unseen items (sequence-based models) or when encountering new users (content-based models). To address the aforementioned gaps, we propose PreSizE - a novel deep learning framework which utilizes Transformers for accurate size prediction. PreSizE models the effect of both content-based attributes, such as brand and category, and the buyer's purchase history on her size preferences. Using an extensive set of experiments on a large-scale e-commerce dataset, we demonstrate that PreSizE is capable of achieving superior prediction performance compared to previous state-of-the-art baselines. By encoding item attributes, PreSizE better handles cold-start cases with unseen items, and cases where buyers have little past purchase data. As a proof of concept, we demonstrate that size predictions made by PreSizE can be effectively integrated into an existing production recommender system yielding very effective features and significantly improving recommendations.

Product reviews play a key role in e-commerce platforms. Studies show that many users read product reviews before purchase and trust them as much as personal recommendations. However, in many cases, the number of reviews per product is large and finding useful information becomes a challenging task. A few websites have recently added an option to post tips - short, concise, practical, and self-contained pieces of advice about products. These tips are complementary to the reviews and usually add a new non-trivial insight about the product, beyond its title, attributes, and description. Yet, most if not all major e-commerce platforms lack the notion of a tip as a first class citizen and customers typically express their advice through other means, such as reviews. In this work, we propose an extractive method for tip generation from product reviews. We focus on five popular e-commerce domains whose reviews tend to contain useful non-trivial tips that are beneficial for potential customers. We formally define the task of tip extraction in e-commerce by providing the list of tip types, tip timing (before and/or after the purchase), and connection to the surrounding context sentences. To extract the tips, we propose a supervised approach and provide a labeled dataset, annotated by human editors, over 14,000 product reviews using a dedicated tool. To demonstrate the potential of our approach, we compare different tip generation methods and evaluate them both manually and over the labeled set. Our approach demonstrates especially high performance for popular products in the Baby, Home Improvement and Sports & Outdoors domains, with precision of over 95% for the top 3 tips per product.

The importance of e-commerce platforms has driven forward a growing body of research work on e-commerce search. We present the first large-scale and in-depth study of query reformulations performed by users of e-commerce search; the study is based on the query logs of eBay's search engine. We analyze various factors including the distribution of different types of reformulations, changes of search result pages retrieved for the reformulations, and clicks and purchases performed upon the retrieved results. We then turn to address a novel challenge in the e-commerce search realm: predicting whether a user will reformulate her query before presenting her the search results. Using a suite of prediction features, most of which are novel to this study, we attain high prediction quality. Some of the features operate prior to retrieval time, whereas others rely on the retrieved results. While the latter are substantially more effective than the former, we show that the integration of these two types of features is of merit. We also show that high prediction quality can be obtained without considering information from the past about the user or the query she posted. Nevertheless, using these types of information can further improve prediction quality.

E-Commerce marketplaces support millions of daily transactions, and some disagreements between buyers and sellers are unavoidable. Resolving disputes in an accurate, fast, and fair manner is of great importance for maintaining a trustworthy platform. Simple cases can be automated, but intricate cases are not sufficiently addressed by hard-coded rules, and therefore most disputes are currently resolved by people. In this work we take a first step towards automatically assisting human agents in dispute resolution at scale. We construct a large dataset of disputes from the eBay online marketplace, and identify several interesting behavioral and linguistic patterns. We then train classifiers to predict dispute outcomes with high accuracy. We explore the model and the dataset, reporting interesting correlations, important features, and insights.

With mobile devices, users are taking ever-growing numbers of photos every day. These photos are uploaded to social sites such as Facebook and Flickr, often automatically. Yet, the portion of these uploaded photos being publicly shared is low, and on a constant decline. Deciding which photo to share takes considerable time and attention, and many users would rather forfeit the social interaction and engagement than sift through their piles of uploaded photos. In this paper, we introduce a novel task of recommending socially-engaging photos to their creators for public sharing. This will turn a tedious manual chore into a quick, software-assisted process. We provide extensive analysis over a large-scale dataset from the Flickr photo sharing website, which reveals some of the traits of photo sharing in such sites. Additionally, we present a ranking algorithm for the task that comprises three steps:(a) grouping of near-duplicate photos;(b) ranking the photos in each group by their "shareability"; and(c) ranking the groups by their likelihood to contain a shareable photo. A large-scale experiment allows us to evaluate our algorithm and show its benefits compared to competitive baselines and algorithmic alternatives.

User-generated reviews are a key driving force behind some of the leading websites, such as Amazon, TripAdvisor, and Yelp. Yet, the proliferation of user reviews in such sites also poses an information overload challenge: many items, especially popular ones, have a large number of reviews, which cannot all be read by the user. In this work, we propose to extract short practical tips from user reviews. We focus on tips for travel attractions extracted from user reviews on TripAdvisor. Our method infers a list of templates from a small gold set of tips and applies them to user reviews to extract tip candidates. For each attraction, the associated candidates are then ranked according to their predicted usefulness. Evaluation based on labeling by professional annotators shows that our method produces high-quality tips, with good coverage of cities and attractions.

In this paper we concentrate on finding the best migration plan, that is, a partial ordering of live migrations that realizes a move from the current to the desired placement, takes the minimal possible time, and maintains the placement constrains throughout the process. This is not an easy task since additional resources and intermediate migrations may be needed in order to maintain feasibility; in fact, we show that even for a simple model, capturing only the critical aspects of the problem, computing the optimal migration plan is NP hard. We develop algorithms that find feasible migration plans and prove that their overall migration time is within an adaptive constant factor from the optimal possible time. Then, using data from real cloud placements, we evaluate the expected performance of these algorithms in realistic scenarios. Our results indicate that the algorithms perform very well under these realistic conditions.

Virtual Machine (VM) placement has to carefully consider the aggregated resource consumption of co-located VMs in order to obey service level agreements at lower possible cost. In this paper, we focus on satisfying the traffic demands of the VMs in addition to CPU and memory requirements. This is a much more complex problem both due to its quadratic nature (being the communication between a pair of VMs) and since it involves many factors beyond the physical host, like the network topologies and the routing scheme. Moreover, traffic patterns may vary over time and predicting the resulting effect on the actual available bandwidth between hosts within the data center is extremely difficult. We address this problem by trying to allocate a placement that not only satisfies the predicted communication demand but is also resilient to demand time-variations. This gives rise to a new optimization problem that we call the Min Cut Ratio-aware VM Placement (MCRVMP). The general MCRVMP problem is NP-Hard, hence, we introduce several heuristics to solve it in reasonable time. We present extensive experimental results, associated with both placement computation and run-time performance under time-varying traffic demands, to show that our heuristics provide good results (compared to the optimal solution) for medium size data centers.