Joint Feature Selection with Dynamic Spectral Clustering

摘要

Current clustering algorithms solved a few of the issues around clustering such as similarity measure learning, or the cluster number estimation. For instance, some clustering algorithms can learn the data similarity matrix, but to do so they need to know the cluster number beforehand. On the other hand, some clustering algorithms estimate the cluster number, but to do so they need the similarity matrix as an input. Real-world data often contains redundant features and outliers, which many algorithms are susceptive to. None of the current clustering algorithms are able to learn the data similarity measure and the cluster number simultaneously, and at the same time reduce the influence of outliers and redundant features. Here we propose a joint feature selection with dynamic spectral clustering (FSDS) algorithm that not only learns the cluster number k and data similarity measure simultaneously, but also employs the \( {\text{L}}_{2,1} \)-norm to reduce the influence of outliers and redundant features. The optimal performance could be reached when all the separated stages are combined in a unified way. Experimental results on eight real-world benchmark datasets show that our FSDS clustering algorithm outperformed the comparison clustering algorithms in terms of two evaluation metrics for clustering algorithms including ACC and Purity.