Adversarial Graph Augmentation to Improve Graph Contrastive Learning
Authors: Susheel Suresh, Pan Li, Cong Hao, Jennifer Neville
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally validate AD-GCL2 by comparing with the state-of-the-art GCL methods and achieve performance gains of up-to 14% in unsupervised, 6% in transfer, and 3% in semi-supervised learning settings overall with 18 different benchmark datasets for the tasks of molecule property regression and classification, and social network classification. |
| Researcher Affiliation | Collaboration | Susheel Suresh Purdue University suresh43@purdue.edu Pan Li Purdue University panli@purdue.edu Cong Hao Georgia Tech callie.hao@gatech.edu Jennifer Neville Purdue University and Microsoft Research jenneville@microsoft.com |
| Pseudocode | Yes | (See Appendix D for a summary of AD-GCL in its algorithmic form.) |
| Open Source Code | Yes | 2https://github.com/susheels/adgcl |
| Open Datasets | Yes | We use datasets from Open Graph Benchmark (OGB) [52], TU Dataset [73] and ZINC [74] for graph-level property classification and regression. |
| Dataset Splits | Yes | AD-GCL-OPT assumes the augmentation search space has some weak information from the downstream task. A full range of analysis on how λreg impacts AD-GCL will be investigated in Sec. 5.2. We compare AD-GCL with three unsupervised/selfsupervised learning baselines for graph-level tasks, which include randomly initialized untrained GIN (RU-GIN) [72], Info Graph [18] and Graph CL [24]. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. |
| Software Dependencies | No | The paper mentions using GIN as the encoder and implies standard machine learning frameworks but does not list specific software dependencies with version numbers. |
| Experiment Setup | Yes | We consider two types of AD-GCL, where one is with a fixed regularization weight λreg = 5 (Eq.8), termed AD-GCL-FIX, and another is with λreg tuned over the validation set among {0.1, 0.3, 0.5, 1.0, 2.0, 5.0, 10.0}, termed AD-GCL-OPT. |