An End-to-End Neural Network Using Fusion Attention Mechanism for Session-Based News Recommendation

News Recommendations mitigate information overload problem while help users more efficiently find news of their interest. Effective modeling of news and users using the textual information from news is crucial for news recommendation algorithms. In this paper, we propose a novel news recommendation method that leverages LLM to enrich news title, and adopts a specialized attention scheme, namely Fusion Attention, to enhance news representation for both user and news modeling. We first design a news encoder that treats the title, the entities, and the derived neighbor entities as three input channels, and generate separate vector representations for each of these channels. For each news, the outputs from the news encoder are then further enhanced by the proposed fusion attention block. The ultimate news representation is obtained by concatenation of the three enhanced channel representations for each news. A user encoder composed of a bidirectional LSTM layer and an aggregate layer is then applied on the historical clicked news of a given user to generate the user representation. Experiments on the MIND datasets validate the effectiveness of our proposed method.