Summary
Recommendation systems often try to understand your different interests by creating multiple interest profiles, but these profiles can end up too similar, making recommendations less effective. The authors found that this happens because typical training only focuses on one positive item at a time, which doesn't help differentiate interests well. They propose MIMA, which groups related items together and assigns each one exclusively to a unique interest, helping the system learn more distinct interests. MIMA also adjusts scores to fairly compare these interests at recommendation time, leading to better results in tests on several datasets and in live business environments.
What this means in practice
- •For e-commerce platform teams: Improve product recommendations by distinguishing multiple user interests through exclusive positive item assignments for more personalized shopping experiences.
- •For music streaming engineers: Enhance personalized playlists by better separating users’ multiple listening preferences with multi-positive grouped item training to avoid similar interest profiles.
Authors
Xingyuan Mao, Alin Fan, Shichao Nie, Junfeng Zhang, Yan Xiao, Tao Luo, Xiaoyi Zeng
Abstract
Multi-interest recommendation represents each user with multiple interest vectors for fine-grained candidate matching, yet it often suffers from interest collapse, where the learned interests converge to similar representations. We highlight the prevailing single-positive paradigm as one important factor behind this issue. Since each instance provides only one positive item, intents are optimized independently, potentially causing the same best-matching interest to be repeatedly updated toward different positives while leaving the others under-supervised. Moreover, existing methods rarely model how strongly a user activates each interest, leaving scores from different interest channels incomparable at inference. To address these problems, we propose MIMA, a Multi-Interest recommendation framework built on Multi-positive exclusive Assignment. MIMA groups items co-occurring within the same request into a positive set, generates complementary interests with a causal Transformer decoder, and exclusively assigns each positive to supervise a distinct interest via Hungarian matching, so that interest differentiation emerges from the training objective itself rather than auxiliary regularization. A lightweight routing module further estimates user-interest activation probabilities to calibrate scores across interest channels. Experiments on three public datasets and an industrial dataset show that MIMA consistently outperforms state-of-the-art baselines, and an online A/B test yields significant business gains.