Ad-gcl:advantageous graph augmentation to improve graph contractual learning

??? Be augmented by confrontation view： In the final analysis, there are two view Comparison of . And another one. view It's through drop Side by side . therefore , The selling point is ：T(.).   （ What I am curious about is how to use Bernoulli distribution .）

target:

Various tasks face the problem of lack of tag data ：

therefore , need SSL.

however ,SSL Tasks may focus too much on redundant information ：

  This may result in poor results for downstream tasks ：

  And the design of this paper AD-GCL You can avoid GCN Capture too much redundant information ：

And to achieve this goal is to enhance the strategy through the designed graph ： 

 

  Besides , This article also has theoretical analysis ：

 

• This is the author's original text , Sure get To the author's starting point ：

Purdue University Li pan ： A good picture shows what it is ？| vector | Robustness | gaussian _ Netease subscription

• These two articles are OK ： Figure neural network self supervised learning toolbox - AD-GCL（ Two ） - Simple books

【 Paper reading 】AD-GCL：Adversarial Graph Augmentation to Improve Graph Contrastive Learning_Cziun The blog of -CSDN Blog

Because the graphics in the real world / Label scarcity is a common problem in network data , There is an urgent need for graphical neural networks （GNN） Conduct self supervised learning . Figure control learning （Graphcontrastive learning,GCL）, Through training GNN Maximize the correspondence between different extended representations of the same graph . Even without using labels , It is also possible to produce robust and transferable GNN. However , By tradition GCL Trained GNN There is usually a risk of capturing redundant graphic features , And these characteristics can be fragile , Performance below standard in downstream tasks . ad locum , We put forward a new principle , It's called antagonism gcl（AD-gcl）, This principle is optimized gcl Countermeasure map enhancement strategy used in , send SGNNS It can avoid capturing redundant information during training . We will AD-GCL Combined with theoretical explanation , An example based on trainable edge descent graph is designed . We have verified through experiments AD！ By working with the most advanced GCL Methods for comparison , In unsupervised learning 、 Transfer learning 、 Semi-supervised learning 、 Attribute regression and classification and social network classification , Achieved as high as 14%、6% and 3% The performance gain of , share 18 Different benchmark datasets , For molecular attribute regression and classification and social network classification tasks