Multi-label Feature Extraction With Distance-Based Graph Attention Network

Feature extraction deals with information redundancy in data with a large number of features. Existing feature extraction approaches to multi-label data usually consider label correlations, while rarely consider sample correlations. In this paper, we propose a multi-label feature extraction with the distance-based graph attention network (DBGAT) algorithm. First, to easily extract the neighbors of the sample later, we construct an adjacency matrix according to the distance between samples and the number of neighbors specified by the user. Second, to obtain the importance of neighbor features to instances, we get the weight coefficients of each instance and its neighbors through the attention network. Third, a new representation for each instance is obtained by weighted summation of neighboring instances. The difference in the weight coefficient reflects the degree of influence of different neighbors on the new feature. We tested the new algorithm and eight other popular algorithms on twelve datasets. Experiments show that this method improves the accuracy of multi-label classification.