Recommendation Models Based on Heterogeneous Information Networks and Multi-task Learning

To address the problem that it is difficult to adequately capture the content information of nodes in recommendation systems based on heterogeneous information networks, and to tackle the problem of link loss in information mining of heterogeneous networks based on meta-path, a recommendation method was proposed based on heterogeneous information networks and multi-task learning. The influence degree of different neighbor instances on nodes on each meta-path view was first calculated to mine the internal information of meta-path, then attention mechanism was used to learn the semantic information of heterogeneous graphs to get the embedding of nodes in heterogeneous networks, and finally multi-task learning method was adopted to optimize both recommendation and link prediction tasks to alleviate the link loss calculated problem. Experimental results on three public heterogeneous data sets show that the model can fully leverage the heterogeneous information network semantic and structure information, and its performance in the recommendation and link prediction task are both better than the baselines.