Alexey Vasilev

Maxim Zhelnin, Dmitry Redko, Daniil Volkov et al.

Sequential recommendations (SR) with transformer-based architectures are widely adopted in real-world applications, where SR models require frequent retraining to adapt to ever-changing user preferences. However, training transformer-based SR models often encounters a high computational cost associated with scoring extensive item catalogs, often exceeding thousands of items. This occurs mainly due to the use of cross-entropy loss, where peak memory scales proportionally to catalog size, batch size, and sequence length. Recognizing this, practitioners in the field of recommendation systems typically address memory consumption by integrating the cross-entropy (CE) loss with negative sampling, thereby reducing the explicit memory demands of the final layer. However, a small number of negative samples would degrade model performance, and as we demonstrate in our work, increasing the number of negative samples and the batch size further improves the model's performance, but rapidly starts to exceed industrial GPUs' size (~40Gb). In this work, we introduce the CCE- method, which offers a GPU-efficient implementation of the CE loss with negative sampling. Our method accelerates training by up to two times while reducing memory consumption by more than 10 times. Leveraging the memory savings afforded by using CCE- for model training, it becomes feasible to enhance its accuracy on datasets with a large item catalog compared to those trained with original PyTorch-implemented loss functions. Finally, we perform an analysis of key memory-related hyperparameters and highlight the necessity of a delicate balance among these factors. We demonstrate that scaling both the number of negative samples and batch size leads to better results rather than maximizing only one of them. To facilitate further adoption of CCE-, we release a Triton kernel that efficiently implements the proposed method.

Alexey Vasilev, Anna Volodkevich, Denis Kulandin et al.

Using a single tool to build and compare recommender systems significantly reduces the time to market for new models. In addition, the comparison results when using such tools look more consistent. This is why many different tools and libraries for researchers in the field of recommendations have recently appeared. Unfortunately, most of these frameworks are aimed primarily at researchers and require modification for use in production due to the inability to work on large datasets or an inappropriate architecture. In this demo, we present our open-source toolkit RePlay - a framework containing an end-to-end pipeline for building recommender systems, which is ready for production use. RePlay also allows you to use a suitable stack for the pipeline on each stage: Pandas, Polars, or Spark. This allows the library to scale computations and deploy to a cluster. Thus, RePlay allows data scientists to easily move from research mode to production mode using the same interfaces.

Konstantin Bauman, Alexey Vasilev, Alexander Tuzhilin

Context has been an important topic in recommender systems over the past two decades. A standard representational approach to context assumes that contextual variables and their structures are known in an application. Most of the prior CARS papers following representational approach manually selected and considered only a few crucial contextual variables in an application, such as time, location, and company of a person. This prior work demonstrated significant recommendation performance improvements when various CARS-based methods have been deployed in numerous applications. However, some recommender systems applications deal with a much bigger and broader types of contexts, and manually identifying and capturing a few contextual variables is not sufficient in such cases. In this paper, we study such ``context-rich'' applications dealing with a large variety of different types of contexts. We demonstrate that supporting only a few most important contextual variables, although useful, is not sufficient. In our study, we focus on the application that recommends various banking products to commercial customers within the context of dialogues initiated by customer service representatives. In this application, we managed to identify over two hundred types of contextual variables. Sorting those variables by their importance forms the Long Tail of Context (LTC). In this paper, we empirically demonstrate that LTC matters and using all these contextual variables from the Long Tail leads to significant improvements in recommendation performance.

Anastasiia Zakharova, Dmitriy Alexandrov, Maria Khodorchenko et al.

Machine learning (ML) models trained on datasets owned by different organizations and physically located in remote databases offer benefits in many real-world use cases. State regulations or business requirements often prevent data transfer to a central location, making it difficult to utilize standard machine learning algorithms. Federated Learning (FL) is a technique that enables models to learn from distributed datasets without revealing the original data. Vertical Federated learning (VFL) is a type of FL where data samples are divided by features across several data owners. For instance, in a recommendation task, a user can interact with various sets of items, and the logs of these interactions are stored by different organizations. In this demo paper, we present \emph{Stalactite} - an open-source framework for VFL that provides the necessary functionality for building prototypes of VFL systems. It has several advantages over the existing frameworks. In particular, it allows researchers to focus on the algorithmic side rather than engineering and to easily deploy learning in a distributed environment. It implements several VFL algorithms and has a built-in homomorphic encryption layer. We demonstrate its use on a real-world recommendation datasets.

Anna Volodkevich, Dmitry Anikin, Danil Gusak et al.

Offline evaluation of recommender systems is often affected by hidden, under-documented choices in data preparation. Seemingly minor decisions in filtering, handling repeats, cold-start treatment, and splitting strategy design can substantially reorder model rankings and undermine reproducibility and cross-paper comparability. In this paper, we introduce SplitLight, an open-source exploratory toolkit that enables researchers and practitioners designing preprocessing and splitting pipelines or reviewing external artifacts to make these decisions measurable, comparable, and reportable. Given an interaction log and derived split subsets, SplitLight analyzes core and temporal dataset statistics, characterizes repeat consumption patterns and timestamp anomalies, and diagnoses split validity, including temporal leakage, cold-user/item exposure, and distribution shifts. SplitLight further allows side-by-side comparison of alternative splitting strategies through comprehensive aggregated summaries and interactive visualizations. Delivered as both a Python toolkit and an interactive no-code interface, SplitLight produces audit summaries that justify evaluation protocols and support transparent, reliable, and comparable experimentation in recommender systems research and industry.

Anton Klenitskiy, Alexey Vasilev

Recently sequential recommendations and next-item prediction task has become increasingly popular in the field of recommender systems. Currently, two state-of-the-art baselines are Transformer-based models SASRec and BERT4Rec. Over the past few years, there have been quite a few publications comparing these two algorithms and proposing new state-of-the-art models. In most of the publications, BERT4Rec achieves better performance than SASRec. But BERT4Rec uses cross-entropy over softmax for all items, while SASRec uses negative sampling and calculates binary cross-entropy loss for one positive and one negative item. In our work, we show that if both models are trained with the same loss, which is used by BERT4Rec, then SASRec will significantly outperform BERT4Rec both in terms of quality and training speed. In addition, we show that SASRec could be effectively trained with negative sampling and still outperform BERT4Rec, but the number of negative examples should be much larger than one.

Anton Klenitskiy, Anna Volodkevich, Anton Pembek et al.

Sequential recommender systems are an important and demanded area of research. Such systems aim to use the order of interactions in a user's history to predict future interactions. The premise is that the order of interactions and sequential patterns play an essential role. Therefore, it is crucial to use datasets that exhibit a sequential structure to evaluate sequential recommenders properly. We apply several methods based on the random shuffling of the user's sequence of interactions to assess the strength of sequential structure across 15 datasets, frequently used for sequential recommender systems evaluation in recent research papers presented at top-tier conferences. As shuffling explicitly breaks sequential dependencies inherent in datasets, we estimate the strength of sequential patterns by comparing metrics for shuffled and original versions of the dataset. Our findings show that several popular datasets have a rather weak sequential structure.

Mikhail Shirokikh, Ilya Shenbin, Anton Alekseev et al.

We develop and evaluate neural architectures to model the user behavior in recommender systems (RS) inspired by click models for Web search but going beyond standard click models. Proposed architectures include recurrent networks, Transformer-based models that alleviate the quadratic complexity of self-attention, adversarial and hierarchical architectures. Our models outperform baselines on the ContentWise and RL4RS datasets and can be used in RS simulators to model user response for RS evaluation and pretraining.

Abdulaziz Samra, Evgeney Frolov, Alexey Vasilev et al.

Data sparsity has been one of the long-standing problems for recommender systems. One of the solutions to mitigate this issue is to exploit knowledge available in other source domains. However, many cross-domain recommender systems introduce a complex architecture that makes them less scalable in practice. On the other hand, matrix factorization methods are still considered to be strong baselines for single-domain recommendations. In this paper, we introduce the CDIMF, a model that extends the standard implicit matrix factorization with ALS to cross-domain scenarios. We apply the Alternating Direction Method of Multipliers to learn shared latent factors for overlapped users while factorizing the interaction matrix. In a dual-domain setting, experiments on industrial datasets demonstrate a competing performance of CDIMF for both cold-start and warm-start. The proposed model can outperform most other recent cross-domain and single-domain models. We also provide the code to reproduce experiments on GitHub.

Anna Volodkevich, Danil Gusak, Anton Klenitskiy et al.

The goal of modern sequential recommender systems is often formulated in terms of next-item prediction. In this paper, we explore the applicability of generative transformer-based models for the Top-K sequential recommendation task, where the goal is to predict items a user is likely to interact with in the "near future". We explore commonly used autoregressive generation strategies, including greedy decoding, beam search, and temperature sampling, to evaluate their performance for the Top-K sequential recommendation task. In addition, we propose novel Reciprocal Rank Aggregation (RRA) and Relevance Aggregation (RA) generation strategies based on multi-sequence generation with temperature sampling and subsequent aggregation. Experiments on diverse datasets give valuable insights regarding commonly used strategies' applicability and show that suggested approaches improve performance on longer time horizons compared to widely-used Top-K prediction approach and single-sequence autoregressive generation strategies.

Nikita Severin, Danil Kartushov, Vladislav Urzhumov et al.

Sequential recommender systems have achieved significant success in modeling temporal user behavior but remain limited in capturing rich user semantics beyond interaction patterns. Large Language Models (LLMs) present opportunities to enhance user understanding with their reasoning capabilities, yet existing integration approaches create prohibitive inference costs in real time. To address these limitations, we present a novel knowledge distillation method that utilizes textual user profile generated by pre-trained LLMs into sequential recommenders without requiring LLM inference at serving time. The resulting approach maintains the inference efficiency of traditional sequential models while requiring neither architectural modifications nor LLM fine-tuning.

Danil Gusak, Anna Volodkevich, Anton Klenitskiy et al.

Modern sequential recommender systems, ranging from lightweight transformer-based variants to large language models, have become increasingly prominent in academia and industry due to their strong performance in the next-item prediction task. Yet common evaluation protocols for sequential recommendations remain insufficiently developed: they often fail to reflect the corresponding recommendation task accurately, or are not aligned with real-world scenarios. Although the widely used leave-one-out split matches next-item prediction, it permits the overlap between training and test periods, which leads to temporal leakage and unrealistically long test horizon, limiting real-world relevance. Global temporal splitting addresses these issues by evaluating on distinct future periods. However, its applications to sequential recommendations remain loosely defined, particularly in terms of selecting target interactions and constructing a validation subset that provides necessary consistency between validation and test metrics. In this paper, we demonstrate that evaluation outcomes can vary significantly across splitting strategies, influencing model rankings and practical deployment decisions. To improve reproducibility in both academic and industrial settings, we systematically compare different splitting strategies for sequential recommendations across multiple datasets and established baselines. Our findings show that prevalent splits, such as leave-one-out, may be insufficiently aligned with more realistic evaluation strategies. Code: https://github.com/monkey0head/time-to-split

Valeriy Shevchenko, Nikita Belousov, Alexey Vasilev et al.

In the rapidly evolving domain of Recommender Systems (RecSys), new algorithms frequently claim state-of-the-art performance based on evaluations over a limited set of arbitrarily selected datasets. However, this approach may fail to holistically reflect their effectiveness due to the significant impact of dataset characteristics on algorithm performance. Addressing this deficiency, this paper introduces a novel benchmarking methodology to facilitate a fair and robust comparison of RecSys algorithms, thereby advancing evaluation practices. By utilizing a diverse set of $30$ open datasets, including two introduced in this work, and evaluating $11$ collaborative filtering algorithms across $9$ metrics, we critically examine the influence of dataset characteristics on algorithm performance. We further investigate the feasibility of aggregating outcomes from multiple datasets into a unified ranking. Through rigorous experimental analysis, we validate the reliability of our methodology under the variability of datasets, offering a benchmarking strategy that balances quality and computational demands. This methodology enables a fair yet effective means of evaluating RecSys algorithms, providing valuable guidance for future research endeavors.

Anton Klenitskiy, Konstantin Polev, Daria Denisova et al.

Many current state-of-the-art models for sequential recommendations are based on transformer architectures. Interpretation and explanation of such black box models is an important research question, as a better understanding of their internals can help understand, influence, and control their behavior, which is very important in a variety of real-world applications. Recently sparse autoencoders (SAE) have been shown to be a promising unsupervised approach for extracting interpretable features from language models. These autoencoders learn to reconstruct hidden states of the transformer's internal layers from sparse linear combinations of directions in their activation space. This paper is focused on the application of SAE to the sequential recommendation domain. We show that this approach can be successfully applied to the transformer trained on a sequential recommendation task: learned directions turn out to be more interpretable and monosemantic than the original hidden state dimensions. Moreover, we demonstrate that the features learned by SAE can be used to effectively and flexibly control the model's behavior, providing end-users with a straightforward method to adjust their recommendations to different custom scenarios and contexts.

Anton Pembek, Artem Fatkulin, Anton Klenitskiy et al.

Many sequential recommender systems suffer from the cold start problem, where items with few or no interactions cannot be effectively used by the model due to the absence of a trained embedding. Content-based approaches, which leverage item metadata, are commonly used in such scenarios. One possible way is to use embeddings derived from content features such as textual descriptions as initialization for the model embeddings. However, directly using frozen content embeddings often results in suboptimal performance, as they may not fully adapt to the recommendation task. On the other hand, fine-tuning these embeddings can degrade performance for cold-start items, as item representations may drift far from their original structure after training. We propose a novel approach to address this limitation. Instead of entirely freezing the content embeddings or fine-tuning them extensively, we introduce a small trainable delta to frozen embeddings that enables the model to adapt item representations without letting them go too far from their original semantic structure. This approach demonstrates consistent improvements across multiple datasets and modalities, including e-commerce datasets with textual descriptions and a music dataset with audio-based representation.