Generative Recommenders Still Suffer from Popularity Bias
A recent study published on arXiv indicates that Generative Recommenders (GRs), even with their sophisticated unified end-to-end framework, still exhibit susceptibility to popularity bias. The researchers pinpoint two primary reasons for this: a flaw in token-level optimization and the uniformity in item tokenization related to semantic indexing. Existing debiasing techniques have not been successful when applied to GRs. To tackle these challenges, the study introduces an innovative generative method designed to eliminate the bias that has persistently affected recommendation systems.
Key facts
- Generative Recommenders (GRs) use a unified end-to-end framework.
- GRs are susceptible to popularity bias.
- Traditional debiasing methods have marginal effectiveness on GRs.
- Two core causes identified: token-level optimization flaw and undifferentiated item tokenization.
- Item tokenization is based on semantic index.
- The study develops a novel generative method to address popularity bias.
- The research is published on arXiv with ID 2605.16825.
- The announcement type is cross.
Entities
Institutions
- arXiv