AI-Based Perspective Diversity and Blind Spot Detection in Personalized News Recommendation Systems

Personalized news recommendation systems suggest articles based on a user’s reading historyand interests. However, they may repeatedly recommend similar types of content, leavingsome important categories underrepresented. This study proposes an AI-based approachto identify these gaps and recommend articles from categories that users have read lessoften. The system uses a Large Language Model to classify news articles into five categories:Government and Policy, Business and Industry, Technology and Innovation, Society andPublic, and Education and Research. It then identifies missing or underrepresented categoriesin a user’s reading history and selects relevant articles from those categories. The proposedmethod was tested using a curated dataset of 300 Indian news articles and 10 simulated userprofiles. Its performance was compared with a conventional content-based recommendationmethod using normalized Shannon entropy to measure category diversity and TF-IDF cosinesimilarity to measure content relevance. The results showed that the proposed methodincreased average normalized entropy from 0.5050 before recommendations to 0.7841 afterrecommendations. The baseline method increased it to 0.5471. The proposed methodalso maintained measurable similarity between recommended articles and users’ readinginterests. These findings suggest that identifying gaps in reading histories can improvecategory diversity in an offline recommendation setting. However, the experiment usedsimulated profiles and does not establish whether the approach improves user satisfaction orreal-world news consumption.