Accurate, Even at Random: Exploring Multiple Distinct Data Clusterings via Random Projection

The clustering process helps users uncover hidden structures in their data without relying on predefined labels, offering insights through subjective interpretation based on users' needs and interests. Unlike classification tasks, clustering is inherently subjective, varying with different user perspectives. In real-world applications, data often contains multiple meaningful clusterings, each potentially valuable to different users. However, most existing clustering techniques generate only a single clustering outcome, which can be restrictive. This paper proposes a method to discover multiple alternative clusterings, allowing users to choose the ones most relevant to their interests. Initially, a diverse set of clusterings is generated based on various data projections to encompass a wide range of possible solutions. Similar clusterings are then identified and merged into representative ones, providing users with a concise set of non-redundant options. The proposed method is evaluated using both artificial and real-world datasets, including visual assessments with images. Extensive comparisons of the method with other techniques are also conducted. The results demonstrate that the proposed method effectively discovers multiple interesting and meaningful clustering solutions with significantly improved performance.