Research On News Recommendation Method Based On Knowledge Graph And Personalized Attention Network

Nowadays,many online news platforms produce a huge amount of news and attract a large number of users to read them.It is very difficult for users to find the news they are interested in from the mass news,so it is especially important to alleviate the news information overload,improve users’ experience and recommend the news that really interests them.At present,the traditional news recommendation scenario faces the following two challenges.(1)Not all news clicked by users can reflect users’ preferences,and the same news should have different representations for different users,and users’ interest preferences need to be considered.(2)Different words in news headlines may have different information weights,and traditional news recommendation only learns through individual news information,ignoring the word weights in news representations,resulting in poor recommendation results.In order to solve the above problems,this paper proposes a news recommendation method based on knowledge graph and personalized attention mechanism,and the main work and contributions are as follows.(1)A news recommendation model based on knowledge graph and personalized attention mechanism,NR-KGPAN(News Recommendation-Knowledge Graph and Personalized Attention Network),is proposed.The model contains two core modules: user encoder and news encoder.Compared with traditional approaches,this model adds knowledge graph to the news encoder to enrich entity information,treats different types of news data as different news perspectives,and selects important words by using word-level and news-level attention networks.In the user encoder,the user’s personalized needs are considered,and the news-level attention mechanism is used to select highly informative news,thus improving the final recommendation accuracy.(2)Considering that users’ long-and short-term interest preferences have a large impact on current recommendations,we incorporate long-and short-term news preferences in the news encoder of the NR-KGPAN model,and use the bi-directional time-series recurrent neural network Bi-LSTM(Bidirectional-Long Short-term Memory)to capture the back-and-forth semantic relationships between texts.In the user encoder,the output of the CNN and Bi-LSTM are jointly used as the personalized attention network layer The output of CNN and Bi-LSTM are jointly used as the input of the personalized attention network layer in the user encoder,and the output is passed through the gated neural network GRU(Gate Recurrent Unit)to capture the user’s long-and short-term interest preferences.(3)The proposed method is compared with several benchmark models on the Microsoft News Dataset(MIND).The mean-reverse ranking(MRR),model evaluation metric(AUC),and normalized discounted cumulative gain(NDGG)are utilized as evaluation metrics.The implementation results show that the model in this paper has a significant improvement in recommendation accuracy compared with that of the benchmark model,which proves the effectiveness of the proposed model in news recommendation.