Feature Selection via Adaptive Spectral Clustering based on Joint Mutual Information

Feature selection plays an important role in big data mining and pattern recognition, which includes selecting a subset of the most informative features that produces compatible results as the original entire set of features. A new similarity measure is proposed in terms of information theory, in particular joint mutual information between features with respect to class labels. Based on the similarity measure, an adaptive spectral clustering that could determine cluster number adaptively is proposed. Furthermore, we propose a new feature selection algorithm called Adaptive Spectral Clustering based on Joint Mutual Information (ASC-JMI) which can effectively and efficiently deal with both irrelevant and redundant features, and select a high-quality feature subset. The experimental results on five high-dimensional benchmark datasets demonstrate that the ASC-JMI is effective in selecting the informative features and obtains competitive classification performance.