| Recently,the main models in recommendation are collaborative filtering algorithm and deep neural network,but the former has the limitation of cold start and sparsity,while the latter has poor explainability.Therefore,researchers use the side information of the knowledge graph to improve recommendation performance.Rich knowledge graph data overcome the problem of cold start and data sparsity and have stronger explanatory power.In this paper,we creatively propose a model by unifying Knowledge graph and user Preferences of Recommendation in multi-task feature learning(KPR).The joint learning unit of preference and feature cross unit are used as two links of multi-task learning,which makes the recommendation module and knowledge graph module learn:im parallel and alternately.The model is divided into two modules:recommendation module and knowledge graph embedding module,which are bridged by two feature cross units.The proposed units can automatically learn high-order feature interactions of items of recommendation and entities in the knowledge graph to obtain potential information in two associated features.In addition,the most important module of this paper is the collection of user preferences and the joint learning unit of knowledge graph relationship.We design two strategies for preference induction:one is hard strategy,which just select one out of preference matrix by Gumbel Softmax sampling.The other approach combines all preferences with attentions,which is called soft strategy.In addition,two ways are proposed in high-order feature combination of recommendation module,one is based on element-wise product,the other is second-order linear feature interactions.The experiment was carried out on three datasets of movie,book and music.By comparing the four methods of KPR,we can come to a conclusion that the soft strategy is better than the hard strategy in most scenarios,and the second-order linear feature interaction is better than the element-wise product.In addition,compared with the latest model in the same field,KPR has a better performance in sparse dataset,and the result of knowledge graph module is significantly better than other models. |
