This study addresses mainstream bias in recommender systems, specifically how niche users receive low utility from these models. We identify two root causes: discrepancy modeling and unsynchronized learning.
To combat the bias against niche users, we propose an end-to-end adaptive local learning framework, introducing a loss-driven Mixture-of-Experts module and an adaptive weight module.
Our experimental results demonstrate that our proposed method outperforms state-of-the-art alternatives, effectively benefiting both niche and mainstream users in recommendation scenarios.
The proposed approach combines powerful elements to enhance learning performance across varying user interests, aligning recommendations closer to user preferences.
[
Collection
]
[
|
...
]