Multi-view clustering (MVC), which can exploit complementary information of different views to enhance the clustering performance, has attracted people’s increasing attentions in recent years. However, existing multi-view clustering methods typically solve a non-convex problem, therefore are easily stuck into bad local minima. In addition, noisy data and outliers affect the clustering process negatively. In this paper, we propose self-paced multi-view clustering via a novel soft weighted regularizer (SPMVC) to address these issues. Specifically, SPMVC progressively selects samples to train the MVC model from simplicity to complexity in a self-paced manner. A novel soft weighted regularizer is proposed to further reduce the negative impact of outliers and noisy data. Experimental results on real-world data sets demonstrate the effectiveness of the proposed method.

Introduction

The aim of clustering [1] is to divide a set of objects into different groups such that similar objects will be grouped into the same cluster, while dissimilar ones are placed into different clusters. Clustering has been widely used in different fields, including pattern recognition, social network analysis, astronomical data analysis, information retrieval, and bioinformatics, etc. In the past couple of decades, a large number of clustering models have been proposed, such as k-means [2], fuzzy clustering [3], density-based clustering [4], [5], distribution-based clustering [6], [7], mean shift clustering [8], [9], consensus clustering [10]–[۱۲], clustering based on deep neural networks [13], [14] etc. However, these conventional algorithms can only deal with single view clustering problems. In real-world clustering tasks, data sets are often described by multiple views, each providing a specific aspect of data. To take full advantage of complementary information from different views, multi-view clustering was proposed [15]. Recently, a number of multi-view clustering methods [16]–[۲۳] have been proposed and have been proved to be effective in solving multi-view clustering problems. However, existing multi-view clustering methods typically solve a non-convex optimization problem [24], which results in the consequence that they get trapped in bad local minima easily. To address the non-convexity issue, an effective and efficient way is to use curriculum learning [25] and self-paced learning [26]. The core idea of curriculum learning and self-paced learning is imitating the mechanisms of cognition of humans.