This paper focuses on machine learning-based fact-checking systems, specifically ensemble methods that combine diverse classifiers, to address psychological biases (such as confirmation bias) that make them vulnerable to the spread of fake news on social media. The performance of existing ensemble methods relies heavily on the diversity of the constituent classifiers, but their tendency to learn overlapping patterns makes it difficult to select models with true diversity. To address this, we propose HierarchySelect, a novel automatic classifier selection method that prioritizes diversity among classifiers and considers performance. HierarchySelect calculates pairwise diversity between classifiers and applies hierarchical clustering to group them into different levels of granularity. It selects a pool of classifiers exhibiting different diversity at each level and selects the most diverse pool to form an ensemble. By incorporating evaluation metrics that reflect the performance of each classifier, it also ensures the generalization performance of the ensemble. We validate the performance of our method by comparing it to existing methods through experiments using six diverse datasets and 40 heterogeneous classifiers.
