Multilabel Classification Based on Graph Neural Networks

Abstract

Typical Laplacian embedding focuses on building Laplacian matrices prior to minimizing weights of connected graph components. However, for multilabel problems, it is difficult to determine such Laplacian graphs owing to multiple relations between vertices. Unlike typical approaches that require precomputed Laplacian matrices, this chapter presents a new method for automatically constructing Laplacian graphs during Laplacian embedding. By using trace minimization techniques, the topology of the Laplacian graph can be learned from input data, subsequently creating robust Laplacian embedding and influencing graph convolutional networks. Experiments on different open datasets with clean data and Gaussian noise were carried out. The noise level ranged from 6% to 12% of the maximum value of each dataset. Eleven different multilabel classification algorithms were used as the baselines for comparison. To verify the performance, three evaluation metrics specific to multilabel learning are proposed because multilabel learning is much more complicated than traditional single-label settings; each sample can be associated with multiple labels. The experimental results show that the proposed method performed better than the baselines, even when the data were contaminated by noise. The findings indicate that the proposed method is reliably robust against noise.

Graph embedding and geometric deep learning relevance to network biology and structural chemistry

“…The graph convolution network (GCN) is a widely-used method to embed multi-label graphs (Ye and Wang, 2022 ). However, Gao et al ( 2019 ), pointed out that for multi-label learning problems, the supervision component of GCN just minimises the cross-entropy loss between the last layer outputs and the ground-truth label distribution, which often misses important information such as label correlations and prevents obtaining high performance.…”