Multi-source multi-label feature selection with missing features

Abstract

Feature dimensionality reduction on Multi-Source Multi-Label (MSML) data is a critical and challenging task. Because practical situations always produce massive MSML data, but they usually contain more missing feature values in the high-dimensional feature space and present severely skewed label distributions in the multi-label space, which aggravate the difficulties in the tackling of high-dimensional feature selection on MSML data. However, much attention in feature selection has been directed either toward multi-label data or multi-source data, while little attention is focused on MSML data, not to mention those containing missing features. Motivated by this, we present a new feature selection method for MSML data with missing features, called MMFSMF. Specifically, to overcome the issue of feature missing, we first supplement the feature matrix by constructing a feature correlation matrix during the modeling process of multi-label learning. At the meanwhile, we utilize a multi-label oversampling mechanism to address the persistent problem of label skewness in multi-label data. Secondly, in terms of the above processing, we introduce a refined infinite feature selection algorithm to perform feature dimensionality reduction in each multi-label data source, considering both label correlations and label-specific features. Thirdly, to address feature redundancy among multiple data sources, we apply a new inter-source feature fusion method. Finally, experiments conducted on nine synthetic MSML datasets with missing features demonstrate that MMFSMF achieves superior performances compared to all competing ones.