Explainable News Recommendations

In the digital age, online news platforms have transformed the way users access news, offering vast amounts of content tailored to individual preferences. Personalized news recommender systems play a crucial role in helping users find relevant articles. However, as the reliance on these systems grows, users often encounter a lack of transparency regarding why specific articles are suggested. This has driven a growing interest in developing explainable news recommendation systems that not only provide accurate recommendations but also offer clear, user-friendly explanations. While recent advancements in deep learning and natural language processing have greatly improved recommendation performance, these systems often operate as "black boxes", obscuring the reasoning behind their decisions. The primary objective of this thesis is to bridge the gap between recommendation accuracy and explainability in personalized news recommendation systems by developing models that deliver high-quality, personalized recommendations while providing interpretable insights into the reasoning behind these suggestions. Striking this balance is essential because although deep learning models excel at predicting user preferences, they often lack transparency, leaving users uncertain about the rationale behind the recommended content. News articles typically cover a wide range of topics, from politics to entertainment, providing an ideal foundation for topic-centric explanations that enhance the transparency of recommendation models. These explanations allow users to better understand how their interests align with the recommended content. Furthermore, news consumption is shaped not only by individual preferences but also by broader patterns of user behavior, both locally from individual users and globally from collective user trends. Thus, a key contribution is made to integrating local and global behaviors into the recommendation process, enabling the system to deliver more contextually relevant and accurate recommendations. Additionally, leveraging advancements in large language models (LLMs), this research enhances the generation of accurate recommendations and provides explanations by summarizing relevant topics for users. Finally, this research offers a unified framework that systematically compares different recommendation models, assesses their performance, and generates topic-centric explanations.