Label Reconstruction based Laplacian Score for semi-supervised feature selection

In many real world learning tasks, we often face the situation that there is no shortage of unlabeled samples but only a small number of samples are labeled. How to make full use of the limited label information to improve the learning performance is widely studied. In this paper, we consider one of learning methods: semi-supervised feature selection, and we present a novel semi-supervised method, called Label Reconstruction based Laplacian Score (LRLS). The basic assumption of our method is that the labels share the same similarity with the samples. We utilize the geodesic distance to measure the similarity between two samples. Then, we reconstruct the labels of the unlabeled samples by using label reconstruction technique. The weight matrix can be obtained from these labels and the Laplacian score can be calculated. We select the features according to the score. The experimental results have demonstrated the effectiveness of our proposed method.