Behavioral patterns captured in embeddings learned from interaction data are pivotal across various stages of production recommender systems. However, in the initial retrieval stage, practitioners face an inherent tradeoff between embedding expressiveness and the scalability and latency of serving components, resulting in the need for representations that are both compact and expressive. To address this challenge, we propose a training strategy for learning high-dimensional sparse embedding layers in place of conventional dense ones, balancing efficiency, representational expressiveness, and interpretability. To demonstrate our approach, we modified the production-grade collaborative filtering autoencoder ELSA, achieving up to 10× reduction in embedding size with no loss of recommendation accuracy, and up to 100× reduction with only a 2.5% loss. Moreover, the active embedding dimensions reveal an interpretable inverted-index structure that segments items in a way directly aligned with the model's latent space, thereby enabling integration of segment-level recommendation functionality (e.g., 2D homepage layouts) within the candidate retrieval model itself. Source codes, additional results, as well as a live demo are available at https://github.com/zombak79/compressed\_elsa.

Model Embedding CTR Size per 100M (Compression) Dimension Lift Embeddings SBERT [18] 512 (baseline) 204.8 GB Nomic [14] 768 +4.86% 307.2 GB Nomic (Matryoshka) 64 +1.89% 25.6 GB Nomic (CompresSAE) 4096* +3.44% 25.6 GB *Sparse embeddings with 32 nonzero entries.Figure 1: Comparison of embedding models used for candidate retrieval.We report online recommendation performance on a downstream task, relative to SBERT [18], with anytime-valid 99% confidence intervals.

Recommender systems often use text-side information to improve their predictions, especially in cold-start or zero-shot recommendation scenarios, where traditional collaborative filtering approaches cannot be used. Many approaches to text-mining side information for recommender systems have been proposed over recent years, with sentence Transformers being the most prominent one. However, these models are trained to predict semantic similarity without utilizing interaction data with hidden patterns specific to recommender systems. In this paper, we propose beeFormer, a framework for training sentence Transformer models with interaction data. We demonstrate that our models trained with beeFormer can transfer knowledge between datasets while outperforming not only semantic similarity sentence Transformers but also traditional collaborative filtering methods. We also show that training on multiple datasets from different domains accumulates knowledge in a single model, unlocking the possibility of training universal, domain-agnostic sentence Transformer models to mine text representations for recommender systems. We release the source code, trained models, and additional details allowing replication of our experiments at https://github.com/recombee/beeformer.

Making recommender systems more transparent and auditable is crucial for the future adoption of these systems. Available tools typically present mostly errors of models aggregated over all test users, which is often insufficient to uncover hidden biases and problems. Moreover, the emphasis is primarily on the accuracy of recommendations but less on other important metrics, such as the diversity of recommended items, the extent of catalog coverage, or the opportunity to discover novel items at bestsellers’ expense. In this work, we propose RepSys, a framework for evaluating recommender systems. Our work offers a set of highly interactive approaches for investigating various scenario recommendations, analyzing a dataset, and evaluating distributions of various metrics that combine visualization techniques with existing offline evaluation methods. RepSys framework is available under an open-source license to other researchers.

Recently, the RS research community has witnessed a surge in popularity for shallow autoencoder-based CF methods. Due to its straightforward implementation and high accuracy on item retrieval metrics, EASE is potentially the most prominent of these models. Despite its accuracy and simplicity, EASE cannot be employed in some real-world recommender system applications due to its inability to scale to huge interaction matrices. In this paper, we proposed ELSA, a scalable shallow autoencoder method for implicit feedback recommenders. ELSA is a scalable autoencoder in which the hidden layer is factorizable into a low-rank plus sparse structure, thereby drastically lowering memory consumption and computation time. We conducted a comprehensive offline experimental section that combined synthetic and several real-world datasets. We also validated our strategy in an online setting by comparing ELSA to baselines in a live recommender system using an A/B test. Experiments demonstrate that ELSA is scalable and has competitive performance. Finally, we demonstrate the explainability of ELSA by illustrating the recovered latent space.

Next basket prediction from historical purchases is quite a complex task, even for e-commerce datasets with a low number of items that are being purchased repeatedly. Neural approaches are not much better in predicting next purchases than simple heuristics. This paper focuses on the challenge of how to encode baskets into efficient neural embedding with low reconstruction error while maintaining the similarity of baskets in the latent space. In our representation, replacing a product with a similar product or increasing quantity will not change the embedding of the basket much. We believe that good basket representation is critical for subsequent prediction. Our analysis shows that state-of-the-art next basket prediction approaches have limitations in their representation of baskets. We would like to focus on this aspect in our future research.