PRISM fixes similarity bias in sequential recommendation with dual-perspective attention
Transformer recommenders overlook key item relations—PRISM's perspective lenses catch them
Sequential recommendation models rely on capturing user preferences from interaction sequences, and self-attention has become the dominant mechanism for learning inter-item relations. But the researchers behind PRISM noticed a critical flaw: dot-product attention suffers from similarity bias, systematically favoring items that look alike while suppressing heterogeneous relations that often carry strong preference signals. This bias limits accuracy across multiple transformer-based recommenders.
To fix this, the team introduces PRISM (Perspective-based Relational Insight Synthesis Module), which re-examines item relations through multiple lenses. It uses K Perspective Lenses to calibrate attention from distinct viewpoints—an Affinity View refines homogeneous relations while a Contrast View exposes heterogeneous ones hidden by similarity bias. In experiments across seven real-world benchmarks, PRISM consistently outperforms state-of-the-art baselines, showing that recovering these overlooked relations leads to better predictions. The paper was accepted at CIKM 2026, and the code is publicly available.
- Transformers in sequential recommenders show similarity bias: dot-product attention overweights similar items and underweights heterogeneous relations
- PRISM uses K Perspective Lenses combining Affinity View and Contrast View to capture both homogeneous and heterogeneous item relations
- Outperforms SOTA baselines across 7 real-world benchmarks; accepted at CIKM 2026 and open-sourced
Why It Matters
More accurate sequential recommendations mean better personalization in e-commerce, media, and search, where understanding nuanced item transitions drives user engagement.