Dual Attention Network Based on Knowledge Graph for News Recommendation

With the continuous development of Internet technology, there are more and more online news readers, and the amount of news is huge and growing explosively. In addition, the news language is highly concentrated and contains a large number of entities. Currently, the commonly used recommendation methods are difficult to make full use of knowledge information and discover the potential interests of users. To solve the above problems, we propose a Dual attention network based on knowledge graph for news recommendation (DAKRec), which takes news titles, the entities contained in them and the contexts as input, and uses knowledge graph to extract news features. In order to better characterize the diversity of users' interests, a dual attention network is constructed to obtain the weight of users' historical news through word-level attention mechanism and item-level attention mechanism (integrating news words, entities, and contexts). Finally, the multi-head attention module is used to connect historical news and candidate news, and the click-through rate is calculated through a fully connected multilayer perceptron after feature fusion. Through a large number of experiments, we prove that our model DAKRec is better than the advanced DKN model and the other comparison models (FM, DMF) in AUC and MSE, further improves the recommendation performance.