Reproduce “eSASRec: Enhancing Transformer-based Recommendations in a Modular Fashion” on OpenResearch

About this paper

Daria Tikhonovich, Nikita Zelinskiy, Aleksandr V. Petrov, Mayya Spirina, Andrei Semenov, Andrey V. Savchenko, Sergei Kuliev

Researchers systematically evaluated modular enhancements for Transformer-based sequential recommender systems, identifying "eSASRec" as a powerful baseline built with LiGR Transformer layers and Sampled Softmax loss. This enhanced model achieves Pareto-optimal performance comparable to industrial state-of-the-art models in realistic settings, while significantly outperforming many academic state-of-the-art models on common benchmarks.

Key insights

• Optimizing a simple SASRec architecture with robust components like LiGR Transformer layers and Sampled Softmax loss (eSASRec) creates a highly competitive baseline that can surpass many published academic SOTA models.
• Realistic evaluation methodologies, including time-based data splits and Beyond-Accuracy metrics, are critical for reliable performance assessment, as academic benchmarks can inflate perceived model effectiveness.
• The original SASRec Shifted Sequence training objective remains highly effective when combined with strong modular enhancements, and Mixed Negative sampling provides an effective control for the accuracy-coverage trade-off.

Summary by alphaXiv.