A Personalized and Explainable News Recommendation Framework Leveraging User Clickstreams and Content Semantics

MindReader introduces an interpretable and personalized news recommender system that integrates clickstream data to summarize transformer-based content semantics and provide token-level biased attribution. Unlike classical recommender systems, which often operate as black boxes, MindReader offers actionable interpretability by utilizing Shapley Additive explanations (SHAP) to reveal the word-level contributions behind each recommendation decision. This model combines both user reading history and article content extraction through a unified framework, incorporating temporal patterns and semantic embeddings. MindReader demonstrates state-of-the-art AUC and coherence scores on real-world news datasets, outperforming several strong baselines, including TF-IDF and neural content models. Human evaluation confirms its superiority. SHAP-based overlays closely align with user attention patterns, while error case analysis highlights resilience against linguistic noise and clickbait content. A key differentiator of MindReader lies in its commitment to not only achieving high performance but also ensuring transparency and trust qualities vital for the deployment of AI in sensitive areas such as journalism, education, and civic communication. This transparency allows users to not only see what content is recommended but also understand the reasoning behind it. In alignment with the UN Sustainable Development Goals (SDG 4 – Quality Education, SDG 9 – Innovation and Infrastructure, and SDG 16 – Strong Institutions), MindReader advocates for an interpretable AI framework for public information dissemination. The architecture proposed in this work offers a scalable, user-centric, and SDG-compliant approach to the implementation of explainable recommended systems.