Robust Feature Selection With Weight Cost Maximin Optimization

Recently, applying data-independent random feature extraction methods has been promising for time-series classifications. However, a large portion of the features generated by these methods is redundant. To address this problem, we propose a robust feature selection method that applies a maximin optimization to a weight cost function that makes a trade-off between dimension reduction and preserving the classification accuracy. It can be shown that the weight cost is discrete midpoint convex with respect to the sparsity parameter, and at the same time is concave with respect to the regularization parameter. Based on these findings, we propose an efficient implementation for feature selection with maximin optimization. Our empirical results show that for large classification tasks with a small sample size per class, our feature selection method removes more redundancies than the state-of-the-art methods, while preserving the classification accuracy.