Multi-task recommendation based on dynamic knowledge graph

Introducing knowledge graphs into recommender systems effectively solves sparsity and cold start problems. However, existing KG recommendation methods such as MKR mostly rely on static knowledge graphs, ignoring that nodes and edges in the graph dynamically change over time, leading to problems such as insufficient timeliness, inability to describe context dependencies, data redundancy, and noise. We propose a multi-task learning method for recommendation enhancement based on dynamic knowledge graphs, MTRDKG, which models the dynamic knowledge graph as a series of continuous time events. Specifically, after node events (node-level or interactions between nodes) occur, the memory state of the nodes is updated through a temporal graph network (TGN), and node temporal embeddings are generated to capture the nodes’ attributes, contextual relationships, and dynamic change information. We use the node embeddings generated by TGN in conjunction with the recommended items as a shared part, aiming to integrate dynamic knowledge graph information into the recommendation task, thereby improving the recommendation effect. Extensive experiments were conducted with four real-world datasets and state-of-the-art baseline methods. The results show that MTRDKG outperforms existing methods in terms of recommendation accuracy and knowledge graph embedding quality, especially in dealing with datasets of different sparsity levels.