Sparse Mutual Granularity-Based Feature Selection and its Application of Schizophrenia Patients
Hengrong Ju, Tao Yin, Jiashuang Huang, Weiping Ding, Xibei Yang
Abstract
<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">K</i> -nearest neighborhood information granularity-based feature selection is derived from the well-known <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> -nearest neighbor ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> NN) classification technique, which is widely employed in data mining. However, the current <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> -nearest neighborhood-based information granules cannot effectively handle data with different density distributions. To address this problem, a sparse mutual granularity-based feature selection approach is developed. First, a personalized information granule is constructed based on the optimal <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> values for each sample. The optimal <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> value is obtained through the number of correlated samples, where the correlations between the samples are learned by the sparse constraint function. The achieved optimal <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> can define the size of the granularity-based model and increase the classification accuracy. Second, a mutual-information strategy is introduced in the granularity process. Irrelevant samples in the granularity-based model are removed, which improves the classification performance. Third, an improved heuristic feature selection algorithm is developed to address the nonmonotonic problem. Compared with the classical heuristic method, the proposed feature selection method can improve the performance of the obtained subset and avoid degradation caused by non-monotonicity. The experimental results on the UCI datasets show that the sparse mutual granularity-based feature selection approach is effective for managing data with different density distributions. Finally, the proposed feature selection approach is applied to select significant brain regions in several schizophrenia datasets. It contributes to the prediction of schizophrenia and also provides a new direction for the improvement of medical-image analysis.