Learning Disentangled Multi-intent Representations for Scalable Recommendation

Intent learning aims to capture user intentions to enhance user understanding and item recommendation, emerging as a key research focus in recent years. However, existing approaches often rely on complex and computationally expensive alternating optimization processes, which limit both performance and scalability. To address these challenges, we propose a novel intent learning method, termed DMI , which unifies behavior representation learning within an end-to-end trainable clustering framework to achieve effective and efficient recommendations. Specifically, user behavior sequences are encoded, and cluster centers—representing latent intents—are initialized as learnable neural parameters. A novel learnable clustering module is designed to separate different cluster centers, thereby disentangling users' complex intentions. Moreover, by explicitly modeling the relationships between user and item intents, DMI enforces behavior embeddings to align with cluster centers in an end-to-end manner, guiding the network to extract user intent from behavioral data. Since DMI eliminates the need for full-graph training, it enables the simultaneous optimization of recommendation and clustering through mini-batch updates. By integrating both global and local learning, DMI effectively captures high-quality node embeddings. Compared to the runner-up model, DMI significantly reduces computational costs while improving NDCG@3 by 13.68% on the DBLP dataset.