Intent Disentanglement and Feature Self-supervision for Novel Recommendation

One key property in recommender systems is the long-tail distribution user-item interactions where most items only have few user feedback. Improving recommendation of tail can promote novelty and bring positive effects to both users providers, thus a desirable systems. Current novel studies over-emphasize importance without differentiating degree users' intent on popularity often incur sharp decline accuracy. Moreover, none existing methods has ever taken extreme case items, i.e., cold-start any interaction, into consideration. In this work, we first disclose mechanism that drives user's interaction towards popular or niche by disentangling her conformity influence (popularity) personal interests (preference). We then present unified end-to-end framework simultaneously optimize accuracy targets based disentangled preference. further develop new paradigm for which exploits self-supervised learning technique model correlation between collaborative features content features. conduct extensive experimental results three real-world datasets. The demonstrate our proposed yields significant improvements over state-of-the-art baselines terms accuracy, novelty, coverage, trade-off.