Directed Graph Contrastive Learning
Authors: Zekun Tong, Yuxuan Liang, Henghui Ding, Yongxing Dai, Xinke Li, Changhu Wang
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on various benchmarks reveal our dominance over the state-of-the-art approaches. We conduct extensive experiments to evaluate the effectiveness of our model. Overall accuracy. The performance comparisons between our model and baselines on five datasets are reported in Table 1. |
| Researcher Affiliation | Collaboration | 1National University of Singapore 2Byte Dance 3ETH Zürich 4Peking University |
| Pseudocode | Yes | the pseudo-code is given in the Supplementary Material. |
| Open Source Code | Yes | Our implement can be obtained at https://github.com/flyingtango/Di GCL. |
| Open Datasets | Yes | We use several widely-used datasets including directed graph datasets: CORA-ML [3], CITESEER [37] and AM-PHOTO [38]; undirected graph datasets: PUBMED [27] and DBLP [30]. |
| Dataset Splits | Yes | For train/validation/test split, following the rules in GCN [17], we choose 20 labels per class for training set, 500 labels for validation set, and rest for the test set. |
| Hardware Specification | No | The paper mentions 'OOM means out of memory on a 12GB GPU' in a table footnote, but does not specify the exact GPU model or other detailed hardware specifications. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies used in the experiments. |
| Experiment Setup | Yes | For curriculum learning scheme, we select the log pacing function and the start and ending difficulties are set in Section 3.2. We train all models according to their default settings, then calculate mean test accuracy with STD in percent (%) averaged over 20 random dataset splits with random weight initialization. Note that we follow the [43] and set α = 0.1 in this paper. We take k = 100 and the tolerance as 1e-6 throughout. We empirically make one of the views unperturbed in this paper, i.e., α1 = 0. We set the start point αa = 0.9α (0.9 is used because the boundary value cannot be obtained) and ending point αb = 0 in this paper, i.e., da = 1/9,db = 1. S() denotes cosine similarity function and τ denotes the temperature parameter. |