Global and local attention-based multi-label learning with missing labels

Abstract

In multi-label learning algorithms, the classification performance can be significantly improved using global and local label correlation. However, the incompleteness of the label space leads to difficulties in measuring the label correlation. In the process of label recovery, many multi-label learning algorithms focus on label correlation, but ignore the queried instance information. In this paper, we introduce an attention mechanism to jointly exploit label and instance information in order to improve the quality of the recovered labels. Firstly, the attention mechanism is used to encode the label and the instance information for label space reconstruction. Secondly, attention computations are performed on the reconstructed label space to obtain the label completion matrix. Finally, global and local features of label correlation are used to improve the model robustness, and label prediction is completed. Through the analysis of the experimental results of multiple benchmark multi-label datasets, it is demonstrated that the proposed method has certain advantages over other state-of-the-art algorithms.