Attribute and Structure Preserving Graph Contrastive Learning
Authors: Jialu Chen, Gang Kou
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the effectiveness of the proposed framework on various real-world networks with different levels of homophily. The results demonstrate the superior performance of our model over the representative baselines. In this section, we conduct extensive experiments to validate the proposed ASP. |
| Researcher Affiliation | Academia | Jialu Chen, Gang Kou* School of Business Administration, Faculty of Business Administration, Southwestern University of Finance and Economics lukac@smail.swufe.edu.cn, kougang@swufe.edu.cn |
| Pseudocode | Yes | The detailed description of our framework is provided in Algorithm 1. Algorithm 1: ASP algorithm |
| Open Source Code | Yes | 1Our implementation is available at https://github.com/ Jialu Chen China/ASP |
| Open Datasets | Yes | To evaluate the performance of different methods, we adopt seven public real-world datasets with different levels of homophily. These datasets can be categorized into two types: homophilous datasets and non-homophilous datasets. For homophilous datasets, we choose three popular public datasets: Cora, Citeseer and Pubmed (Yang, Cohen, and Salakhutdinov 2016), where nodes represent documents and edges represent citation links. Node attributes of these datasets are bag-of-words representation of documents. For non-homophilous datasets, we choose four public datasets: Actor, Cornell, Texas, and Wisconsin (Pei et al. 2020). |
| Dataset Splits | Yes | For Cora, Citeseer and Pubmed, we use the public fixed split introduced by Yang, Cohen, and Salakhutdinov (2016). For Actor, Cornell, Texas, and Wisconsin, we randomly split nodes of each class into 60%, 20%, and 20% for training, validation and testing as introduced by Pei et al. (2020). |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for the experiments (e.g., CPU, GPU models, memory, or cloud instances). It only mentions the software frameworks used for implementation. |
| Software Dependencies | Yes | We implement our proposed framework and some baselines using Pytorch (Paszke et al. 2017) and Pytorch Geometric (Fey and Lenssen 2019). |
| Experiment Setup | Yes | The hyperparameters we tune include: (1) the initial learning rate {1e 1, 1e 2, 1e 3, 1e 4}, (2) k {20, 30, 40, 50, 60, 70} for k nearest neighbors, (3) the aggregation hops l {10, 20, 30}. For homophilous datasets, the output dimension of our encoder network is fixed to 256, while tuned within {24, 32, 64, 128} for non-homophilous datasets. We set a patience of 20 and a maximum of 500 epochs for early stopping. |