Multi-label active learning through serial–parallel neural networks

Abstract

Multi-label active learning is an extension of supervised learning with high-dimensional label spaces and interactive scenarios. Its key issues include the exploitation of label correlations, handling of missing labels, and selection of query labels. Various techniques have been proposed for this purpose; however, there is still room for performance improvement. In this study, we propose a multi-label active learning through serial–parallel neural networks (MASP) algorithm with simple and effective mechanisms. For label correlations, the serial part of the network extracts features that are common to all the labels. This mechanism is more effective than the explicit feature extraction or compressed sensing methods. Regarding the missing labels, the network sets the corresponding losses to zero for backpropagation. Thus, it does not require label completions that may introduce additional errors. For label queries, the parallel part of the network provides independent pairwise predictions for each label. Such pairwise predictions present appropriate information for computing label uncertainty. Three sets of experiments were conducted on 22 benchmark datasets using 14 popular algorithms for comparison. The results show that our algorithm achieves state-of-the-art active learning performance.