Graph Neural Networks for Gaussian Graphs Classification

🎯 Objective

The project's goal is to implement Graph Neural Networks (GNN) for graph classification in unbalanced settings that consider edge weights, utilizing attention mechanisms and other techniques to improve the interpretability of the predictions.

🏁 Usage

Install the necessary dependencies

pip install -r requirements.txt

🏁 Dataset

The data used for simulation are two synthetic datasets simulated using Gaussian graph model for balanced and unbalanced cases (credits to Alessia Mapelli). Each case consists of 500 undirected graphs with 20 nodes. Each of the .RData files is structured the following way:

• Adj_matrices: Contains the weighted adjacency matrix for each sample.
• Sim_data1: Contains the node observations and graph class for each sample.

🚀 Pipeline

• Exploratory Data Analysis (EDA): the step involves investigation of the dataset focusing on the node statistics and features, edge statistics and weights distribution and general topology analysis.
• Data Preprocessing: this step we implement scaling of the node features and weights of the adjacency matrices. In addition, we implement additional features extracted from the encoding of the node identifier and embeddings provided by Node2Vec.
• Model Training: we implement training of various GNN models, including GCN, GAT, GraphConv and GIN, including the implementation of some of them from scratch. We also tran MLP to compare the graph models with a topology-agnostic model.
• Explainability (XAI): using Graph Neural Network Explainer (GNNE) and Integrated Gradients (IG), we visualize the most influential nodes and edges to the final performance of the model.

📝 Results

The main reults are provided in the Presentation.pdf file.

Extra References

• Morris, C. (2022). Graph Neural Networks: Graph classification. In L. Wu, P. Cui, J. Pei, & L. Zhao (Eds.), Graph Neural Networks: Foundations, Frontiers, and Applications (pp. 179-193). Singapore: Springer Nature. https://doi.org/10.1007/978-981-16-6054-2_9
• Sui, Y., Wang, X., Wu, J., He, X., & Chua, T.-S. (2021). Deconfounded training for graph neural networks. CoRR, abs/2112.15089. Retrieved from https://arxiv.org/abs/2112.15089
• Gong, L., & Cheng, Q. (2018). Adaptive edge features guided graph attention networks. CoRR, abs/1809.02709. Retrieved from http://arxiv.org/abs/1809.02709
• Wang, Y., Zhao, Y., Shah, N., & Derr, T. (2022). Imbalanced graph classification via Graph-of-Graph Neural Networks. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM '22) (pp. 2067-2076). New York, NY, USA: Association for Computing Machinery. https://doi.org/10.1145/3511808.3557356