Confidence correction for trained graph convolutional networks

Junqing Yuan, Huanlei Guo, Chenyi Zhou, Jiajun Ding, Zhenzhong Kuang, Zhou Yu, Yuan Liu

Abstract: Highlights•A simple and effective mechanism for the under-confidence problem of existing GCNs.•CCM is formulated as an optimization problem and is solved by greedy strategy.•CCM is model-agnostic and can be applied on most of existing GCNs.•CCM can help to calibrate the confidence and improve the classification accuracy.