Relation pruning and discriminative sampling over knowledge graph for long-tail recommendation

Long-tail recommendations have gained significant attention owing to their potential economic market. However, the scarcity of interaction data for long-tail users/items and the popularity bias present challenges in capturing high-quality embeddings for long-tail users/items. This predicament further exacerbates the long-tail recommendation problem, as current approaches tend to exhibit a bias towards making recommendations for short-head users/items, resulting in a detrimental cycle for long-tail recommendation. To this end, we propose a novel knowledge graph-based approach called LTailKG to improve long-tail recommendations. LTailKG leverages the semantic information in knowledge graph to produce high-quality embeddings and augmented samples for generating satisfactory long-tail recommendations. First, LTailKG parameterizes each node and relation as vector representations. Next, LTailKG presents a relation pruning-based graph contrastive learning operation to generate additional self-supervised signals for long-tail users/items, thereby producing high-quality embeddings for them. Furthermore, LTailKG introduces a knowledge graph-driven discriminative sampling operation to select augmented positive and negative samples from the uninteracted item set, which enables LTailKG to excel at not only identifying long-tail items that better align with the user's interests but also extracting the genuine preferences of long-tail users. Extensive experiments on real-world datasets demonstrate the superiority of LTailKG over state-of-the-art approaches in terms of long-tail recommendation.