TBCA: a transformer-based candidate awareness method for news recommendation

The objective of news recommendation pertains to the facilitation of Personalized news delivery amid the vast expanse of available news. However, existing methods inadequate exploitation of textual information inherent in news articles, and secondly, when building user interest models from historical clicked news, the interaction behavior between clicked news words is ignored. In this paper, we propose a method using Transformer combined with candidate awareness. This method is adept at capturing the contextual nuances between words within user-generated text and exhibits superior efficacy in handling lengthy text sequences, to improve the ability of user modeling to capture user interests. We use a Transformer-based news encoder to learn semantic features at different levels of abstraction of the input text sequence. In addition, we use Fastformer, a variant of the Transformer network, in conjunction with candidate awareness modeling to capture the interactive behavior of users clicking on word-level news. Empirical analyses conducted on authentic datasets substantiate the efficacy of our approach in enhancing the efficacy of news recommendation systems.