MLFBERT: Advancing News Recommendation with Multi-Layer Fusion over DistilBERT

Traditional news recommendation methods usually encode news text and user interests directly into vectors and then calculate the matching degree of the two. Such approaches may lose low-level matching signals and different levels of textual information. We propose a BERT-based news recommendation model, MLFBERT. This model consists of a BERT-based news encoder and a user encoder. The news encoder uses a PLM Fusion Module and additive attention to generate embeddings of available news items. The user encoder builds a user profile by aggregating the representations of their news browsing history. Recommendations are made by computing matching scores between the user profile and candidate news. Experiments on the MIND news dataset demonstrate that our model outperforms existing methods on multiple evaluation metrics.