Robust path-based clustering for the unsupervised and semi-supervised learning settings
Hong Chang, Dit Yan Yeung · 2004
Spectral clustering and path-based clustering have shown excellent performance on some clustering tasks involving highly nonlinear and elongated clusters in addition to compact clusters. However, they are not robust enough towards noise and outliers in the data. Inthis paper, inspired by robust statistical techniques, we propose a robust path-based spectral clustering method that outperforms both spectral clustering and path-based clustering especially in the presence of noise and outliers. Moreover, we show that a simple extension allows our method to work in the semi-supervised clustering setting by incorporating pair-wise side information. The effectiveness of our approach is demonstrated through a series of experiments involving noisy data, slightly coupled clusters, and even complex clustering partition with the aid of side information.