A Hierarchical Contextual Graph-Transformer Network for Multi-Objective Personalized News Recommendation

Personalized news recommendation has become the key to tackle the plethora of digital content, providing timely and relevant information to users. Conventional systems like collaborative filtering and simple content-based filtering tend to be limited by issues like cold-start problems, absence of rich semantic knowledge, and inability to track dynamic user preferences. Current deep learning-based models, which are good at extracting rich representations, often specialize in either sequential behavior or relational structure but do not fully combine the two to support better personalization. Addressing such limitations, this study introduces a Hierarchical Contextual Graph-Transformer Network (HCGTN) that interweaves domain-adapted BERT for semantic news content representation, Graph Neural Networks for capturing intricate user-item-entity relations, and Transformer layers for capturing user interaction sequences over time. A new multiobjective ranking loss concurrently optimizes accuracy, diversity, and long-term engagement, while a semantic similarity-based cold-start approach enhances recommendations for new users and pieces. Experimental benchmarking on standard news recommendation datasets proves that HCGTN surpasses current baselines, such as deep neural and graph-based methods, in CTR prediction, diversity, and coverage. The suggested method provides an effective and scalable system for future-generation personalization in news delivery, tuning relevance to content variety to enhance user satisfaction and retention.