Hierarchical Aggregation Based Knowledge Graph Embedding for Multi-task Recommendation

Recently, knowledge graph has been used for alleviating the problems such as sparsity faced by the recommendation. Multi-task learning, which is an important emerged frontier research direction, helps complement the available information of different tasks and improves recommendation performance effectively. However, the existing multi-task methods ignore high-order information between entities. At the same time, the existing multi-hop neighbour aggregation methods suffer from the problem of over-smoothing. Also, the existing knowledge graph embedding methods in multi-task recommendation ignore the attribute triples in knowledge graph and recommendation tends to neglect the learning of user attributes. To mitigate these problems, we propose a multi-task recommendation model, called AHMKR. We use hierarchical aggregation and high-order propagation to alleviate the over-smoothing problem and obtain a better entity representation that integrates high-order information for multi-task recommendation. We leverage the text information of attribute triples, to improve the performance of knowledge graph in expanding the features of recommendation items. For users, we conduct fine-grained user learning based on the user attributes to capture user preferences in a more accurate matter. The experiments on the real-world datasets demonstrate the good performance of AHMKR.