The rapid growth of online news platforms has intensified the need for intelligent, personalized recommendation systems capable of efficiently filtering and delivering relevant content. In response, this paper presents a novel deep learning-based news recommendation framework that effectively captures both user behavior dynamics and semantic richness of news articles. Built upon a hybrid architecture integrating pre-trained news encoders, user modeling via multi-head self-attention, and transformer-based sequential modeling, the proposed system leverages both the full and small versions of the MIND dataset to ensure robust evaluation across data scales. Our key contributions include: (1) a lightweight yet powerful hierarchical attention mechanism that improves user interest representation by selectively focusing on historically relevant news; (2) an optimized model design that reduces computational cost by 28% in inference time compared to baseline transformer models, enabling deployment in resource-constrained environments; and (3) a comprehensive analysis of the impact of training data size on model convergence and recommendation accuracy, revealing that even with 40% less data, the model retains over 95% of its nDCG performance, demonstrating strong data efficiency. Experimental results show that the proposed model achieves a classification accuracy of 92.22%, AUC of 77.62%, MRR of 52.44, and nDCG@5 of 0.618 and nDCG@10 of 0.583, outperforming state-of-the-art methods by +5.8% in nDCG@5 and +6.3% in nDCG@10. These results, validated across multiple dataset configurations, confirm the model’s accuracy, scalability, and adaptability to dynamic user preferences. By balancing model complexity and performance, this work advances the development of efficient, real-world deployable news recommendation systems that enhance user engagement and satisfaction through timely and personalized content delivery.