In the realm of music recommendation, sequential recommenders have shown promise in capturing the dynamic nature of music consumption. A key characteristic of this domain is repetitive listening, where users frequently replay familiar tracks. To capture these repetition patterns, recent research has introduced Personalised Popularity Scores (PPS), which quantify user-specific preferences based on historical frequency. While PPS enhances relevance in recommendation, it often reinforces already-known content, limiting the system’s ability to surface novel or serendipitous items—key elements for fostering long-term user engagement and satisfaction. To address this limitation, we build upon RecJPQ, a Transformer-based framework initially developed to improve scalability in large-item catalogues through sub-item decomposition. We repurpose RecJPQ’s sub-item architecture to model personalised popularity at a finer granularity. This allows us to capture shared repetition patterns across sub-embeddings—latent structures not accessible through item-level popularity alone. We propose a novel integration of sub-ID-level personalised popularity within the RecJPQ framework, enabling explicit control over the trade-off between accuracy and personalised novelty. Our sub-ID-level PPS method (sPPS) consistently outperforms item-level PPS by achieving significantly higher personalised novelty without compromising recommendation accuracy. Code and experiments are publicly available at https://github.com/sisinflab/Sub-id-Popularity.

Since their introduction, Transformer-based models, such as SASRec and BERT4Rec, have become common baselines for sequential recommendations, surpassing earlier neural and non-neural methods. A number of following publications have shown that the effectiveness of these models can be improved by, for example, slightly updating the architecture of the Transformer layers, using better training objectives, and employing improved loss functions. However, the additivity of these modular improvements has not been systematically benchmarked - this is the gap we aim to close in this paper. Through our experiments, we identify a very strong model that uses SASRec's training objective, LiGR Transformer layers, and Sampled Softmax Loss. We call this combination eSASRec (Enhanced SASRec). While we primarily focus on realistic, production-like evaluation, in our preliminarily study we find that common academic benchmarks show eSASRec to be 23% more effective compared to the most recent state-of-the-art models, such as ActionPiece. In our main production-like benchmark, eSASRec resides on the Pareto frontier in terms of the accuracy-coverage tradeoff (alongside the recent industrial models HSTU and FuXi. As the modifications compared to the original SASRec are relatively straightforward and no extra features are needed (such as timestamps in HSTU), we believe that eSASRec can be easily integrated into existing recommendation pipelines and can can serve as a strong yet very simple baseline for emerging complicated algorithms. To facilitate this, we provide the open-source implementations for our models and benchmarks in repository https://github.com/blondered/transformer_benchmark

A large item catalogue is a major challenge for deploying modern sequential recommender models, since it makes the memory footprint of the model large and increases inference latency. One promising approach to address this is RecJPQ, which replaces item embeddings with sub-item embeddings. However, slow inference remains problematic because finding the top highest-scored items usually requires scoring all items in the catalogue, which may not be feasible for large catalogues. By adapting dynamic pruning concepts from document retrieval, we propose the RecJPQPrune dynamic pruning algorithm to efficiently find the top highest-scored items without computing the scores of all items in the catalogue. Our RecJPQPrune algorithm is safe-up-to-rank K since it theoretically guarantees that no potentially high-scored item is excluded from the final top K recommendation list, thereby ensuring no impact on effectiveness. Our experiments on two large datasets and three recommendation models demonstrate the efficiency achievable using RecJPQPrune: for instance, on the Tmall dataset with 2.2M items, we can reduce the median model scoring time by 64× compared to the Transformer Default baseline, and 5.3× compared to a recent scoring approach called PQTopK. Overall, this paper demonstrates the effective and efficient inference of Transformer-based recommendation models at catalogue scales not previously reported in the literature. Indeed, our RecJPQPrune algorithm can score 2 million items in under 10 milliseconds without GPUs, and without relying on Approximate Nearest Neighbour (ANN) techniques.

Many modern sequential recommender systems use deep neural networks, which can effectively estimate the relevance of items but require a lot of time to train. Slow training increases the costs of training, hinders product development timescales, and prevents the model from being regularly updated to adapt to changing user preferences. The training of such sequential models involves appropriately sampling past user interactions to create a realistic training objective. The existing training objectives have limitations. For instance, next item prediction never uses the beginning of the sequence as a learning target, thereby potentially discarding valuable data. However, the item masking used by the state-of-the-art BERT4Rec recommender model is only weakly related to the goal of the sequential recommendation; therefore, it requires much more time to obtain an effective model. Hence, we propose a novel Recency-based Sampling of Sequences (RSS) training objective (which is parameterized by a choice of recency importance function) that addresses both limitations. We apply our method to various recent and state-of-the-art model architectures—such as GRU4Rec, Caser, and SASRec. We show that the models enhanced with our method can achieve performances exceeding or very close to the effective BERT4Rec, but with much less training time. For example, on the MovieLens-20M dataset, RSS applied to the SASRec model can result in a 60% improvement in NDCG over a vanilla SASRec, and a 16% improvement over a fully trained BERT4Rec model, despite taking 93% less training time than BERT4Rec. We also experiment with two families of recency importance functions and show that they perform similarly. We further empirically demonstrate that RSS-enhanced SASRec successfully learns to distinguish differences between recent and older interactions—a property that the original SASRec model does not exhibit. Overall, we show that RSS is a viable (and frequently better) alternative to the existing training objectives, which is both effective and efficient for training sequential recommender model when the computational resources for training are limited.

A large catalogue size is one of the central challenges in training recommendation models: A large number of items makes them memory- and computationally inefficient to compute scores for all items during training, forcing these models to deploy negative sampling. However, negative sampling increases the proportion of positive interactions in the training data, and therefore, models trained with negative sampling tend to overestimate the probabilities of positive interactions—a phenomenon we call overconfidence . While the absolute values of the predicted scores/probabilities are not important for the ranking of retrieved recommendations, overconfident models may fail to estimate nuanced differences in the top-ranked items, resulting in degraded performance. In this article, we show that overconfidence explains why the popular SASRec model underperforms when compared to BERT4Rec. This is contrary to the BERT4Rec authors’ explanation that the difference in performance is due to the bi-directional attention mechanism. To mitigate overconfidence, we propose a novel Generalised Binary Cross-Entropy Loss function (gBCE) and theoretically prove that it can mitigate overconfidence. We further propose the gSASRec model, an improvement over SASRec that deploys an increased number of negatives and the gBCE loss. Through detailed experiments on three datasets, we show that gSASRec does not exhibit the overconfidence problem. As a result, gSASRec can outperform BERT4Rec (e.g., +9.47% NDCG on the MovieLens-1M dataset) while requiring less training time (e.g., -73% training time on MovieLens-1M). Moreover, in contrast to BERT4Rec, gSASRec is suitable for large datasets that contain more than 1 million items. Finally, we show how addressing overconfidence can improve model calibration—the ability of a model to predict actual interaction probabilities accurately. By applying gBCE to the SASRec model on MovieLens-1M dataset, we reduce the models’ expected calibration error by 98.9% (from 0.966 to 0.01).

In the realm of music recommendation, sequential recommender systems have shown promise in capturing the dynamic nature of music consumption. Nevertheless, traditional Transformer-based models, such as SASRec and BERT4Rec, while effective, encounter challenges due to the unique characteristics of music listening habits. In fact, existing models struggle to create a coherent listening experience due to rapidly evolving preferences. Moreover, music consumption is characterized by a prevalence of repeated listening, i.e. users frequently return to their favourite tracks, an important signal that could be framed as individual or personalized popularity. This paper addresses these challenges by introducing a novel approach that incorporates personalized popularity information into sequential recommendation. By combining user-item popularity scores with model-generated scores, our method effectively balances the exploration of new music with the satisfaction of user preferences. Experimental results demonstrate that a Personalized Most Popular recommender, a method solely based on user-specific popularity, outperforms existing state-of-the-art models. Furthermore, augmenting Transformer-based models with personalized popularity awareness yields superior performance, showing improvements ranging from 25.2% to 69.8%. The code for this paper is available at https://github.com/sisinflab/personalized-popularity-awareness.

Transformer-based recommender systems, such as BERT4Rec or SASRec, achieve state-of-the-art results in sequential recommendation. However, it is challenging to use these models in production environments with catalogues of millions of items: scaling Transformers beyond a few thousand items is problematic for several reasons, including high model memory consumption and slow inference. In this respect, RecJPQ is a state-of-the-art method of reducing the models’ memory consumption; RecJPQ compresses item catalogues by decomposing item IDs into a small number of shared sub-item IDs. Despite reporting the reduction of memory consumption by a factor of up to 50 ×, the original RecJPQ paper did not report inference efficiency improvements over the baseline Transformer-based models. Upon analysing RecJPQ’s scoring algorithm, we find that its efficiency is limited by its use of score accumulators for each item, which prevents parallelisation. In contrast, LightRec (a non-sequential method that uses a similar idea of sub-ids) reported large inference efficiency improvements using an algorithm we call PQTopK. We show that it is also possible to improve RecJPQ-based models’ inference efficiency using the PQTopK algorithm. In particular, we speed up RecJPQ-enhanced SASRec by a factor of 4.5 × compared to the original SASRec’s inference method and by the factor of 1.56 × compared to the method implemented in RecJPQ code on a large-scale Gowalla dataset with more than million items. Further, using simulated data, we show that PQTopK remains efficient with catalogues of up to tens of millions of items, removing one of the last obstacles to using Transformer-based models in production environments with large catalogues.

Sequential Recommender Systems use the order of user-item interactions to predict the next item in the sequence. This task is similar to Language Modelling, where the goal is to predict the next token based on the sequence of past tokens. Therefore, adaptations of language models, and, in particular, Transformer-based models, achieved state-of-the-art results for a sequential recommendation. However, despite similarities, the sequential recommendation problem poses a number of specific challenges not present in Language Modelling. These challenges include the large catalogue size of real-world recommender systems, which increases GPU memory requirements and makes the training and the inference of recommender models slow. Another challenge is that a good recommender system should focus not only on the accuracy of recommendation but also on additional metrics, such as diversity and novelty, which makes the direct adaptation of language model training strategies problematic. Our research focuses on solving these challenges. In this doctoral consortium abstract, we briefly describe the motivation and background for our work and then pose research questions and discuss current progress towards solving the described problems.

Sequential Recommendation is a popular recommendation task that uses the order of user-item interaction to model evolving users' interests and sequential patterns in their behaviour. Current state-of-the-art Transformer-based models for sequential recommendation, such as BERT4Rec, generate sequence embeddings and compute scores for catalogue items, but the increasing catalogue size makes training these models costly. The Joint Product Quantisation (JPQ) method, originally proposed for passage retrieval, markedly reduces the size of the retrieval index with minimal effect on model effectiveness, by replacing passage embeddings with a limited number of shared sub-embeddings. This paper introduces RecJPQ, a novel adaptation of JPQ for sequential recommendations, which replaces each item embedding with a concatenation of a limited number of shared sub-embeddings and, therefore, limits the number of learnable model parameters. The main idea of RecJPQ is to split items into sub-item entities before training the main recommendation model, which is inspired by splitting words into tokens and training tokenizers in language models. We apply RecJPQ to SASRec, BERT4Rec, and GRU models on three large-scale sequential datasets. Our results showed that RecJPQ could notably reduce the model size (e.g., 48× reduction for the Gowalla dataset with no effectiveness degradation). RecJPQ can also improve model performance through a regularization effect (e.g. +0.96% NDCG@10 improvement on the Booking.com dataset). Overall, RecJPQ allows the training of state-of-the-art transformer models in industrial applications, where datasets with millions of items are common.

A large catalogue size is one of the central challenges in training recommendation models: a large number of items makes them memory and computationally inefficient to compute scores for all items during training, forcing these models to deploy negative sampling. However, negative sampling increases the proportion of positive interactions in the training data, and therefore models trained with negative sampling tend to overestimate the probabilities of positive interactions – a phenomenon we call overconfidence. While the absolute values of the predicted scores/probabilities are not important for the ranking of retrieved recommendations, overconfident models may fail to estimate nuanced differences in the top-ranked items, resulting in degraded performance. In this paper, we show that overconfidence explains why the popular SASRec model underperforms when compared to BERT4Rec. This is contrary to the BERT4Rec authors’ explanation that the difference in performance is due to the bi-directional attention mechanism. To mitigate overconfidence, we propose a novel Generalised Binary Cross-Entropy Loss function (gBCE) and theoretically prove that it can mitigate overconfidence. We further propose the gSASRec model, an improvement over SASRec that deploys an increased number of negatives and the gBCE loss. We show through detailed experiments on three datasets that gSASRec does not exhibit the overconfidence problem. As a result, gSASRec can outperform BERT4Rec (e.g. +9.47% NDCG on the MovieLens-1M dataset), while requiring less training time (e.g. -73% training time on MovieLens-1M). Moreover, in contrast to BERT4Rec, gSASRec is suitable for large datasets that contain more than 1 million items.

Many modern sequential recommender systems use deep neural networks, which can effectively estimate the relevance of items but require a lot of time to train. Slow training increases expenses, hinders product development timescales and prevents the model from being regularly updated to adapt to changing user preferences. Training such sequential models involves appropriately sampling past user interactions to create a realistic training objective. The existing training objectives have limitations. For instance, next item prediction never uses the beginning of the sequence as a learning target, thereby potentially discarding valuable data. On the other hand, the item masking used by BERT4Rec is only weakly related to the goal of the sequential recommendation; therefore, it requires much more time to obtain an effective model. Hence, we propose a novel Recency-based Sampling of Sequences training objective that addresses both limitations. We apply our method to various recent and state-of-the-art model architectures – such as GRU4Rec, Caser, and SASRec. We show that the models enhanced with our method can achieve performances exceeding or very close to state-of-the-art BERT4Rec, but with much less training time.

BERT4Rec is an effective model for sequential recommendation based on the Transformer architecture. In the original publication, BERT4Rec claimed superiority over other available sequential recommendation approaches (e.g. SASRec), and it is now frequently being used as a state-of-the art baseline for sequential recommendation. However, not all subsequent publications confirmed its superiority and have proposed other models that were shown to outperform BERT4Rec in effectiveness. In this paper we systematically review all publications that compare BERT4Rec with another popular Transformer-based model, namely SASRec, and show that BERT4Rec results are not consistent within these publications. To understand the reasons behind this inconsistency, we analyse the available implementations of BERT4Rec and show that we fail to reproduce results of the original BERT4Rec publication when using their default configuration parameters. However, we are able to replicate the reported results with the original code if training for a much longer amount of time (up to 30x) compared to the default configuration. We also propose our own implementation of BERT4Rec based on the HuggingFace Transformers library, which we demonstrate replicates the originally reported results on 3 out 4 datasets, while requiring up to 95% less training time to converge. Overall, from our systematic review and detailed experiments, we conclude that BERT4Rec does indeed exhibit state-of-the-art effectiveness for sequential recommendation, but only when trained for a sufficient amount of time. Additionally, we show that our implementation can further benefit from adapting other Transformer architectures that are available in the HuggingFace Transformers library (e.g. using disentangled attention, as provided by DeBERTa, or larger hidden layer size cf. ALBERT). For example, on the MovieLens-1M dataset, we demonstrate that both these models can improve BERT4Rec performance by up to 9%. Moreover, we show that an ALBERT-based BERT4Rec model achieves better performance on that dataset than state-of-the-art results reported in the most recent publications.