EUVIMEDEUVIMED entdecken
Crossref

Faster and Memory-Efficient Training of Sequential Recommendation Models for Large Catalogs

Maxim Zhelnin, Dmitry Redko, Daniil Volkov, Anna Volodkevich, Petr Sokerin, Valeriy Shevchenko, Egor Shvetsov, Alexey Vasilev, Daria Denisova, Ruslan Izmailov, Alexey Zaytsev

ACM Transactions on Information Systems · 19.11.2026

DOI: 10.1145/3862474 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ACM Transactions on Information Systems, online erschienen am 19.11.2026. Verfasst von Maxim Zhelnin und 10 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Transformer-based sequential recommender systems are retrained frequently, yet their training is dominated by cross-entropy (CE) over item catalogs that can reach millions of items. Its logit tensor scales with catalog size × batch size × sequence length, forcing practitioners to trade these quantities against each other or use negative sampling. However, how to allocate memory among negatives, batch size, and sequence length remains an open question. We study this trade-off with nearly 1,000 SASRec training runs on six datasets and find that no single memory-related hyperparameter should be maximized in isolation. Scaling both the number of negatives and batch size outperforms maximizing only one, while the choice between full CE and CE with negative sampling is dataset-dependent. We further introduce a memory-efficient and GPU-optimized replacement for CE with negative sampling that avoids materializing the logit tensor. It speeds up training by up to 3.5× and reduces memory by 4–24× relative to PyTorch CE with negative sampling. Reallocating memory saved by our replacement and Cut Cross-Entropy, which we adapt to sequential recommender systems, into larger batches, longer sequences, and more negatives improves NDCG@10 by up to 54.7% under the same memory budget. We release the Triton kernel and a public dataset of all training runs at https://github.com/On-Point-RND/MemoryEfficientSRS .

Zitieren

APA

Zhelnin, M., Redko, D., Volkov, D., Volodkevich, A., Sokerin, P., Shevchenko, V., Shvetsov, E., Vasilev, A., Denisova, D., Izmailov, R., & Zaytsev, A. (2026). Faster and Memory-Efficient Training of Sequential Recommendation Models for Large Catalogs. ACM Transactions on Information Systems. https://doi.org/10.1145/3862474

Vancouver

Zhelnin M, Redko D, Volkov D, Volodkevich A, Sokerin P, Shevchenko V, et al. Faster and Memory-Efficient Training of Sequential Recommendation Models for Large Catalogs. ACM Transactions on Information Systems. 2026. doi:10.1145/3862474

BibTeX
@article{Zhelnin2026,
  author = {Zhelnin, Maxim and Redko, Dmitry and Volkov, Daniil and Volodkevich, Anna and Sokerin, Petr and Shevchenko, Valeriy and Shvetsov, Egor and Vasilev, Alexey and Denisova, Daria and Izmailov, Ruslan and Zaytsev, Alexey},
  title = {Faster and Memory-Efficient Training of Sequential Recommendation Models for Large Catalogs},
  journal = {ACM Transactions on Information Systems},
  year = {2026},
  doi = {10.1145/3862474}
}
Kostenlos starten →