RecSys Study: Simple Models Outperform Transformers on Multiple Benchmarks

_reachsumit · x · 2026-08-21

A paper accepted at RecSys 2026 questions whether widely used sequential recommendation benchmarks actually require higher-order sequence modeling. The study employs two simple, recency-weighted pairwise probe models: Sequential Rules (SeqRules) and Probabilistic Collaborative Transition Model (PCTM). Results show that on several Amazon datasets and MovieLens-1M, these simple probes outperform Transformer-based recommenders like eSASRec and SASRec by up to 38%, only trailing on MovieLens-20M. This suggests that many popular benchmarks may be poorly suited for measuring gains from higher-order sequence modeling.

Original post →

More from Research

Research channel →