DUPLEX: Dual GAT for Complex Embedding of Directed Graphs

Authors: Zhaoru Ke, Hang Yu, Jianguo Li, Haipeng Zhang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments on diverse tasks and datasets, showcasing the superior performance of our approach. The results demonstrate its efficacy in modeling low-degree nodes, generalizing to multiple tasks, and handling unseen nodes.
Researcher Affiliation Collaboration 1School of Information Science and Technology, Shanghai Tech University, China 2Ant Group, China.
Pseudocode No No section or figure explicitly labeled 'Pseudocode' or 'Algorithm', nor are structured steps formatted like code or an algorithm present. Figure 2 shows an architecture diagram, not pseudocode.
Open Source Code Yes The code is available at https://github.com/alipay/DUPLEX.
Open Datasets Yes The experiments are conducted on five public datasets of digraphs, namely Cora-ml, Citeseer, Cora, Epinions, and Twitter (overview in Table 2 and more details in Appendix E). [...] We use the version of Citeseer and Cora-ml dataset provided by (Zhang et al., 2021), Cora dataset provided by (ˇSubelj & Bajec, 2013) and other two datasets from the Stanford Large Network Dataset Collection2. (Footnote 2 points to https://snap.stanford.edu/data/)
Dataset Splits Yes Embarking on our exploration with link prediction, we adopt the approach delineated by Zhang et al.(Zhang et al., 2021) to randomly split the datasets into training, validation, and testing subsets, maintaining a ratio of 16:1:3. [...] Nodes across all datasets are partitioned into training, validation, and testing subsets with a 3:1:1 split.
Hardware Specification No No specific hardware (e.g., GPU models, CPU types, memory amounts) used for running experiments is mentioned in the paper.
Software Dependencies No We implemented DUPLEX using DGL and Py Torch, employing Adam optimizer with a learning rate of 1e-3. No specific version numbers for DGL or PyTorch are specified.
Experiment Setup Yes We implemented DUPLEX using DGL and Py Torch, employing Adam optimizer with a learning rate of 1e-3. DUPLEX consists of two 3-layer GATs with attention head 1, one for the amplitude embedding and the other for the phase embedding. We sampled four types of node pairs (forward edge, reverse edge, bidirectional edge, and no edge) at a ratio 1:1:1:x for self-supervised loss computation, where x tends to be small due to fewer bidirectional edges, and in our experiments we just control x 1. We set the hidden dim to 128 and the dropout rate to 0.5. We tuned the initial loss weight λ in {0.1,0.3} and the decay rate q in {0,1e-4,1e-2}. We ran our model with maximum 3000 epochs with early stopping for all experiments. For the baseline methods, we utilize the publicly available code repositories, and tune the hyperparameters as follows.