Can One Embedding Fit All? A Multi-Interest Learning Paradigm Towards Improving User Interest Diversity Fairness

TL;DR

This paper addresses fairness gaps in recommender systems caused by varying user interest diversity, demonstrating that users with broader interests often receive poorer recommendations. It introduces a multi-interest framework that represents each user/item with a center embedding and multiple virtual embeddings derived from shared global interest parameters, enabling better alignment for diverse preferences. The approach, validated across four datasets and two backbones, improves the fairness-utility trade-off, enhances embedding alignment, and yields more diverse recommendations, with only a modest increase in parameters. The work provides a practical, model-agnostic avenue to reduce group disparities in recommendations and suggests future refinements in interest generation and the alignment-uniformity trade-off.

Abstract

Recommender systems (RSs) have gained widespread applications across various domains owing to the superior ability to capture users' interests. However, the complexity and nuanced nature of users' interests, which span a wide range of diversity, pose a significant challenge in delivering fair recommendations. In practice, user preferences vary significantly; some users show a clear preference toward certain item categories, while others have a broad interest in diverse ones. Even though it is expected that all users should receive high-quality recommendations, the effectiveness of RSs in catering to this disparate interest diversity remains under-explored. In this work, we investigate whether users with varied levels of interest diversity are treated fairly. Our empirical experiments reveal an inherent disparity: users with broader interests often receive lower-quality recommendations. To mitigate this, we propose a multi-interest framework that uses multiple (virtual) interest embeddings rather than single ones to represent users. Specifically, the framework consists of stacked multi-interest representation layers, which include an interest embedding generator that derives virtual interests from shared parameters, and a center embedding aggregator that facilitates multi-hop aggregation. Experiments demonstrate the effectiveness of the framework in achieving better trade-off between fairness and utility across various datasets and backbones.

Figures (9)

• Figure 1: Why diverse interests matter? Real-world RS examples (A) Food recommendation (B) Dating recommendation.
• Figure 2: Group recommendation performance (Recall $\uparrow$): the pattern that users with more diverse interests generally receive lower recommendation quality is consistent across various datasets, models, diversity metrics, and group partitions. A larger group ID indicates a higher level of user interest diversity.
• Figure 3: Group-level embedding alignment ($\downarrow$) of ml-1m dataset based on LightGCN and CAGCN$^*$.
• Figure 4: Multi-interest motivation: single embedding is insufficient to capture users' diverse interests.
• Figure 5: Multi-interest framework (interest number equals two): rather than a single embedding, each user/item is represented by multiple embeddings (i.e., center and virtual). Center embeddings and global interest embeddings are learnable parameters while the interest (virtual) embeddings are calulated without assigning extra parameters.

Theorems & Definitions (2)

• definition 1
• definition 2