GKD: Semi-supervised Graph Knowledge Distillation for Graph-Independent Inference

The increased amount of multi-modal medical data has opened the opportunities to simultaneously process various modalities such as imaging and non-imaging gain a comprehensive insight into disease prediction domain. Recent studies using Graph Convolutional Networks (GCNs) provide novel semi-supervised approaches for integrating heterogeneous while investigating patients’ associations prediction. However, when meta-data used graph construction is not available at inference time (e.g., coming from distinct population), conventional methods exhibit poor performance. To address this issue, we propose approach named GKD based on knowledge distillation. We train teacher component that employs label-propagation algorithm besides deep neural network benefit non-graph only in training phase. embeds all information soft pseudo-labels. pseudo-labels are then student unseen test which modality unavailable. perform our experiments two public datasets diagnosing Autism spectrum disorder, Alzheimer’s disease, along with thorough analysis synthetic datasets. According these experiments, outperforms previous graph-based learning terms accuracy, AUC, Macro F1.