FNRS-MSC: Federated News Recommender System Integrating Multiple Self-Attention and Convolution

News recommendation is an important feature to help users discover news that matches their bias. In response to users' growing privacy concerns, various recommendation models integrating joint learning have been developed to preserve user privacy. However, these approaches often prioritize protecting user privacy over recommendation performance. This study introduces a federated news recommendation method, FNRS-MSC (Federated News Recommender System Integrating Multiple Self-Attention and Convolution), to address this challenge. The approach is based on a federated learning recommendation model that integrates an attention mechanism in a pre-trained language model (BERT), which is capable of extracting essential features and reducing dimensionality to enhance news representations. In addition, a one-dimensional convolutional layer is integrated for the user encoder in a multiple self-attention model (NRMS) to address the problem of ignoring news features in the user model. Experimental results on two real datasets show that the approach outlined in this study is effective in both protecting user privacy and providing high-quality recommendations.