Profiling users with tag-enhanced spherical metric learning for recommendation

With the increasing complexity of user-item interactions on the Internet, it is important to profile users and model their preferences in recommender systems. Traditional methods, including metric learning, rely on historical user-item interactions to model preferences but struggle in sparse data scenarios. While item tags offer valuable auxiliary information to enhance representations, their shared nature across items makes it challenging to effectively profile users with tags, which requires preserving user personalization through high-quality tag representations. Moreover, traditional optimization for user/item representations always takes place in Euclidean space, where the unconstrained nature of embedding norms tends to lean toward trivial solutions. This may bias the system towards common or popular preferences, thus suppressing the variety in tag-aware user profiles. To this end, we propose to profile users with tag-enhanced spherical metric learning for recommendation, named UTRec. Specifically, we propose an adaptive tag selection mechanism to ensure the quality of tag representations and learn tag-enhanced representations of users/items, thereby effectively profiling users. Additionally, we introduce a spherical optimization strategy for tag-enhanced recommendations to alleviate the limitations imposed by lazy learning and traditional optimization, ensuring the accuracy and diversity of user and item representations within the spherical space. Numerous experiments have been conducted on four real-world datasets, where our proposed tag-enhanced UTRec framework can bring consistent performance gains and achieve a 13.67% improvement regarding both Recall and NDCG metrics.