Reproduce “Decoupled Side Information Fusion for Sequential Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yueqi Xie, Peilin Zhou, Sunghun Kim
Decoupled Side Information Fusion for Sequential Recommendation (DIF-SR) presents a framework for integrating side information into self-attention based sequential recommenders by performing fusion within the attention layer rather than at the input. This approach demonstrates superior performance over existing methods on four real-world datasets and improves models like SASRec and BERT4Rec when integrated.
Key insights
- Fusing side information at the attention layer, rather than at the input embedding level, resolves the "rank bottleneck" and allows for higher expressiveness in attention matrices.
- Decoupled attention mechanisms enable flexible gradient flows for different types of side information, allowing the model to adaptively learn their individual importance.
- Auxiliary attribute prediction tasks, when synergistically combined with decoupled attention, further enhance the interaction and leverage of side information for core item representation learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.