Supervised Community Detection with Line Graph Neural Networks
Authors: Zhengdao Chen, Lisha Li, Joan Bruna
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present a novel family of Graph Neural Networks (GNNs) for solving community detection problems in a supervised learning setting. We show that, in a data-driven manner and without access to the underlying generative models, they can match or even surpass the performance of the belief propagation algorithm on binary and multiclass stochastic block models... We perform the first analysis of the optimization landscape of using (linear) GNNs to solve community detection problems... We present experiments on community detection in synthetic datasets (Sections 6.1, 6.2 and Appendix C.1) as well as real-world datasets (Section 6.3). |
| Researcher Affiliation | Collaboration | Zhengdao Chen Courant Institute of Mathematical Sciences New York University, New York, NY Lisha Li Amplify Partners San Francisco, CA Joan Bruna Courant Institute of Mathematical Sciences New York University, New York, NY |
| Pseudocode | No | The paper describes methods and equations (e.g., (1), (2), (3), (57)) but does not present a formal pseudocode block or algorithm listing. |
| Open Source Code | Yes | Code is available at https://github.com/zhengdao-chen/GNN4CD |
| Open Datasets | Yes | We present experiments on community detection in synthetic datasets (Sections 6.1, 6.2 and Appendix C.1) as well as real-world datasets (Section 6.3)... The stochastic block model is a random graph model with planted community structure... The Geometric Block Model (Sankararaman & Baccelli, 2018) is a random graph generated as follows... We now compare the models on the SNAP datasets, whose domains range from social networks to hierarchical co-purchasing networks. |
| Dataset Splits | No | We assume that a training set {(Gt, yt)}t T is given, with which we train a model that predicts ˆy = Φ(G, θ) by minimizing... Finally, we divide the dataset into training and testing sets by enforcing that no community belongs to both the training and the testing set. While training and test sets are mentioned, specific validation splits (e.g., 80/10/10) or strategies are not detailed. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory used for experiments. It only mentions 'due to computational limitations' in Section 6.3. |
| Software Dependencies | No | We used Adamax (Kingma & Ba, 2014) with learning rate 0.004 across all experiments... Our implementation of GAT is based on https://github.com/Diego999/py GAT. We modified the code so that the number of layers in the network is flexible, and also added spatial batch normalization at the end of each layer... While specific optimizers and activations are mentioned, no software library versions (e.g., TensorFlow, PyTorch) are provided. |
| Experiment Setup | Yes | We used Adamax (Kingma & Ba, 2014) with learning rate 0.004 across all experiments. All the neural network models have 30 layers and 8 features in the middle layers (i.e., bk = 8) for experiments in Sections 6.1 and 6.2, and 20 layers and 6 features for Section 6.3. GNNs and LGNNs have J = 2 across the experiments except the ablation experiments in Section C.3. |