Reproduce “gSASRec: Reducing Overconfidence in Sequential Recommendation Trained with Negative Sampling” 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
Aleksandr Petrov, Craig Macdonald
This work introduces gSASRec, an improved sequential recommendation model designed to reduce overconfidence when trained with negative sampling. It achieves state-of-the-art or comparable performance on various datasets, including large catalogues where models like BERT4Rec are infeasible, through a theoretically grounded Generalised Binary Cross-Entropy (gBCE) loss.
Key insights
- The performance advantage of BERT4Rec over SASRec is primarily due to its use of full Softmax loss without negative sampling, rather than its bi-directional attention mechanism, which contradicts previous assumptions.
- Overconfidence in models trained with negative sampling and BCE loss causes predicted probabilities for highly-ranked items to erroneously converge near 1, obscuring genuine ranking differences.
- The gBCE loss can theoretically calibrate predicted probabilities to align with the true prior item distribution when the calibration parameter is set correctly based on the negative sampling rate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.