Semi-supervised multi-label feature selection with adaptive structure learning and manifold learning

Abstract High-dimensional multi-label data brings challenges and difficulties in multi-label learning. Therefore, feature selection as an effective dimension reduction technique is widely used in multi-label learning. Due to the high cost of collecting sufficient labels, semi-supervised multi-label feature selection has received much attention in recent years. Nevertheless, existing semi-supervised multi-label feature selection methods mainly use manifold assumption to explore the label correlations. However, due to a large amount of unlabeled data, the label correlations are inaccurate and insufficient. Therefore, we need more multi-label data structure information to guide feature selection. In this paper, we propose a unified learning framework combining adaptive global structure learning and manifold learning(SFAM). Adaptive global structure learning promises the selected features to preserve the global and sparse reconstruction structure. Manifold learning is responsible for exploring the local structure and label correlations. The combination of these two items can compensate for the negative influence brought by each other and help to select representative characteristics. The objective function of this method is not smooth and difficult to solve, so an efficient iterative algorithm is designed to make it suitable for practical applications. We evaluate the performance of SFAM on real-world data sets and compare the results with state-of-the-art supervised and semi-supervised feature selection algorithms as well as the baseline using all features. Experimental results show that SFAM has excellent performance.