Feature selection for brain-computer interface using nearest neighbor information

We consider the feature selection problem for a braincomputer interface (BCI). A BCI collects data from sensors, and the data are discriminated using information in a high-dimensional space. We show how relevant features in a high dimensional space can be selected using a simple nearest neighbor method for estimating an information-theoretic measure, Jensen-Shannon divergence. Conventional nonparametric estimation using nearest neighbors already works very well for the feature selection problem and outperforms many other methods. In this paper, we show how this nearest neighbor method can be further exploited by properly trimming the non-informative direction for a distance calculation, and estimate the Jensen-Shannon divergence more accurately. Through experiments with synthetic data, we show how the proposed method outperforms a conventional nearest neighbor method as well as other feature selection methods with a large margin.