Compatible intent-based interest modeling for personalized recommendation

The conventional uniform embeddings lack diversity to infer users’ interests and make suboptimal recommendations for users. Fortunately, users’ interactions imply a complex and hybrid composition of users’ interests with multiple compatible intents. Therefore, this work strives to investigate fine-grained interest modeling from the diversified composition of interest with the intent hypothesis. We propose a cross-intent transformer embedding (CITE) for personalized recommendation, which extracts collaborative filtering (CF) signals by propagating interests within intent subgraphs and between compatible intents. In the scenario of interaction sparsity, intent-aware interest propagation employs graph convolution to ensure interest consistency in each intent subgraph. It builds intent-aware embeddings with interaction confidences learned iteratively on each intent subgraph. In addition, the transformer evaluates inter-intent compatibility to perform cross-intent interest propagation. It updates intent embeddings with CF signals between intents. The resulting multiple fine-grained intent embeddings model the hybrid composition of users’ interests for personalized recommendation. Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed CITE and verify the active role of the compatible intents for interest modeling.