Parameter-free Clustering with Adaptive Nearest Neighbors Learning

Abstract Clustering techniques that leverage pairwise similarity graphs are foundational across fields like pattern recognition, machine learning, and bioinformatics, offering a means to simplify complex data while preserving crucial information. Nonetheless, conventional approaches often fail to construct similarity graphs with a number of connected components that match the intended cluster count, leading to suboptimal learning outcomes. To overcome these limitations, we introduce a new method named Reliable Clustering with Adaptive Neighbors (rCAN). This approach integrates rank constraints on the Laplacian graph alongside an adaptive parameter tuning mechanism, ensuring the similarity graph aligns its connected components with the target cluster count. Furthermore, we conduct comprehensive theoretical analyses and extensive experimental evaluations, revealing that rCAN surpasses traditional methods in both clustering accuracy and robustness.