News Recommendation with Gated Linear Attention and Simplified Gated Linear Units

As a novel approach, the News Recommendation Algorithm Based on Gated Linear Attention and Simplified Gated Linear Units (NRAGLS) designed to enhance the precision and personalization of digital news dissemination is introduced. At the core of NRAGLS are two advanced techniques: Gated Linear Attention (GLA) and Simplified Gated Linear Units (SGLU), supplemented by tensor normalization to adeptly process and analyze the multifaceted nature of news content and user interactions. This research systematically outlines the development and implementation of the NRAGLS model, emphasizing its capacity to address prevalent challenges within the news recommendation domain, including data sparsity, the dynamic evolution of news content, and the absence of explicit user feedback. Through rigorous experimental validation on the Microsoft News Dataset (MIND) and its variant, MIND-small, the NRAGLS model's performance was benchmarked against four existing recommendation systems. Metrics such as Area Under the Curve (AUC), Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (nDCG) were employed to evaluate each model's ability to predict user engagement accurately and to rank news articles effectively. The NRAGLS model demonstrated superior performance across all metrics, highlighting its robustness in handling the complexities of real-world news recommendation scenarios. Furthermore, this paper explores the resilience of the NRAGLS model to data sparsity and its adaptability to the temporal dynamics of user preferences, key factors in sustaining the relevance and effectiveness of recommendation systems. The findings reveal that the NRAGLS model exhibits remarkable consistency in performance, even under varying levels of data omission, and maintains or improves its recommendation accuracy over time.