The combination of deep learning and knowledge graphs has attracted the attention of researchers in the field of recommender system. However, some models use sparse feature vectors as input which will not only increase the training difficulty of the model, but also cause the model to fall into a local optimum. In addition, most models can′t fully describe the relationship between the items in the recommender system and the entities in the knowledge graphs. The vector representation is not accurate and it will affect the performance of the recommender system. Based on these issues, a multi-task recommendation model based on two-stage deep learning (TMR) is proposed. First, using KCNN to extract the features of the items′ name, convert it into a dense vector, and then combine the property of the item as the initial representation of the item feature vector. Second, alternate training is used to obtain extra information in the knowledge graph, and then DeepFM is used as a feature extraction layer to mine the feature interaction between user and item. The experiments demonstrate that our proposed TMR model achieves better results in several evaluation indicators and improves the performance of the recommender system, comparing with other state-of-the-art methods.

deep learning; recommender system; knowledge graph; multi-task