Icml'22 | progcl: rethinking difficult sample mining in graph contrast learning

Recent research shows that , Hard negative sample （ A negative sample more similar to the anchor sample ） It can improve the performance of comparative learning . However , The author finds that the existing difficult negative sample mining techniques in other fields can not promote the graph comparison learning very well （ See the original Table 5）. The author makes an experimental and theoretical analysis of this phenomenon , Find the message passing mechanism that can be attributed to graph neural network . If we only take the similarity as an indicator to measure the difficulty of negative samples, as in other fields , In graph contrast learning, most of the hard negative samples are potential false negative samples （ Same class as anchor sample , Figure 1）. In order to make up for this defect , The authors estimate the probability that a negative sample is a true negative sample by using the mixed distribution , Two schemes are designed （ namely ProGCL-weight and ProGCL-mix） To improve the GCL Performance of .

Thesis link ：https://arxiv.org/abs/2110.02027

Code link ：https://github.com/junxia97/ProGCL