Interpretable News Clustering and Topic Attribution for the U.S. Presidential Election: An AIME and LLM Approach

This paper presents an automated news-analysis method for the U.S. presidential election that combines approximate inverse model explanations (AIME), an explainable AI (XAI) technique, with a large language model (LLM). News articles related to the election were collected using NewsAPI and classified into thematic clusters using k-means clustering. Thereafter, AIME was applied to compute the global feature importance for each cluster, extracting key descriptive terms. Subsequently, an LLM was used to automatically generate appropriate labels for each cluster, significantly enhancing topic interpretability. Visualizing the daily distribution of these clusters enabled the identification of surges in specific topics on specific days and facilitated a detailed analysis of associated news articles. This approach addresses the challenge of interpretability in conventional news analysis clustering, enabling both automated political news classification and time-series trend analysis. This method provides a novel framework for tracking evolving news coverage during extended political events, such as presidential elections, and identifying key events. This methodology is not restricted to election analysis and can be applied to various other domains, including finance, sports, and social news.