Adaptive spectral affinity propagation clustering
Lin Tang, Leilei Sun, Chonghui Guo, Zhen Zhang
Abstract
Affinity propagation (AP) is a classic clustering algorithm. To improve the classical AP algorithms, we propose a clustering algorithm namely, adaptive spectral affinity propagation (AdaSAP). In particular, we discuss why AP is not suitable for non-spherical clusters and present a unifying view of nine famous arbitrary-shaped clustering algorithms. We propose a strategy of extending AP in non-spherical clustering by constructing category similarity of objects. Leveraging the monotonicity that the clusters' number increases with the self-similarity in AP, we propose a model selection procedure that can determine the number of clusters adaptively. For the parameters introduced by extending AP in non-spherical clustering, we provide a grid-evolving strategy to optimize them automatically. The effectiveness of Ada-SAP is evaluated by experiments on both synthetic datasets and real-world clustering tasks. Experimental results validate that the superiority of AdaSAP over benchmark algorithms like the classical AP and spectral clustering algorithms.