Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders

TL;DR

This paper identifies the causes of popularity bias in generative recommenders and introduces Ghost, a new system that reduces bias through innovative optimization and tokenization techniques, improving fairness at some utility cost.

Contribution

The study provides a theoretical analysis of popularity bias in GRs and proposes Ghost, a novel system with asymmetric unlikelihood optimization and skeleton-based tokenization to mitigate bias.

Findings

Ghost significantly reduces popularity bias in experiments.

Ghost achieves fairer recommendations across multiple datasets.

Slight decrease in overall recommendation utility with Ghost.

Abstract

Recently, Generative Recommenders (GRs), characterized by a unified end-to-end framework, have exhibited astonishing potential in transforming the recommendation paradigm. Despite their effectiveness, we recognize that GRs are still susceptible to the long-standing issue of popularity bias that has pervaded the recommendation community. Although a few studies have attempted to extend traditional debiasing methods to GRs, their effectiveness is marginal, and the fundamental reason why GRs suffer from popularity bias remains under-explored. To bridge this gap, this study focuses on two core aspects in GRs: the optimization of generative framework and the item tokenization based on semantic index. Based on theoretical analyses, we identify that the severe popularity bias emerges from the confluence of a token-level optimization flaw and the undifferentiated property of item tokenization. Accordingly, this study develops a novel generative recommender system, called Ghost, by designing the asymmetric unlikelihood optimization and the skeleton-founded tokenization. Extensive empirical evaluations across three datasets, alongside multiple SOTA baselines, reveal that Ghost substantially alleviates popularity bias and promotes fairer recommendations, while incurring slight degradation to the overall recommendation utility.