Robust Affinity Propagation using Preference Estimation

Abstract

Affinity propagation is a novel unsupervised learning algorithm for exemplar-based clustering without the priori knowledge of the number of clusters (NC). In this article, the influence of the “preference” on the accuracy of AP output is addressed. We present a robust AP clustering method, which estimates what preference value could possibly yield an optimal clustering result. To demonstrate the performance promotion, we apply the robust AP on picture clustering, using local SIFT, global MPEG-7 CLD, and the proposed preference as the input of AP. The experimental results show that over 40% enhancement of ARI accuracy for several image datasets.