Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Graphon-Explainer: Generating Model-Level Explanations for Graph Neural Networks using Graphons

Authors: Sayan Saha, Sanghamitra Bandyopadhyay

TMLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The effectiveness of our method is validated through thorough theoretical analysis and extensive experimentation on both synthetic and real-world datasets on the task of graph classification. Results demonstrate its capability to effectively learn and generate diverse graph patterns identified by a trained GNN, thus enhancing its interpretability for end-users.
Researcher Affiliation Academia Sayan Saha EMAIL Machine Intelligence Unit Indian Statistical Institute, Kolkata; Sanghamitra Bandyopadhyay EMAIL Machine Intelligence Unit Indian Statistical Institute, Kolkata
Pseudocode Yes Algorithm 1 Generating Explanations using Graphon-Explainer; Algorithm 4 Estimating the Graphon of a Target Class and Generating Graphs from the Graphon; Algorithm 5 Adjacency Score Computation; Algorithm 2 2shapes dataset generation; Algorithm 3 4Shapes dataset generation
Open Source Code Yes Our code repository is provided at https://github.com/amisayan/Graphon-Explainer.
Open Datasets Yes Real Datasets: For real datasets, we conduct experiments using the IMDB-BINARY (IMDB-B), REDDIT-BINARY (REDDIT-B), and MUTAG datasets. ... We also demonstrate the effectiveness of our method on the BA2Motif (Luo et al., 2020) dataset
Dataset Splits No The paper mentions training a GNN classifier on each dataset and reports classifier test accuracy in Table 1, implying a train/test split. However, it does not provide specific details on the methodology (e.g., percentages, sample counts, or explicit splitting strategy) used for these dataset splits in the main text.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processors, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions using the Adam optimizer (Kingma & Ba, 2014) and refers to code repositories for comparison methods (XGNN, GNNInterpreter), but it does not specify version numbers for any key software components or libraries (e.g., Python, PyTorch, TensorFlow, CUDA).
Experiment Setup Yes On all datasets except MUTAG, we implement a deep GCN architecture comprising three GCN layers, each with a dimension of 64. This is followed by a global mean pooling layer, a batch normalization layer, and finally a linear layer to obtain the class logits. We utilize Leaky Re LU as the activation function between the GCN layers and within the MLP. The model parameters are initialized using the Kaiming method. Training is conducted using the Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 0.01. Additionally, we use a learning rate scheduler to halve the learning rate every hundred epochs. ... On the MUTAG dataset, the model architecture consists of three graph convolutional layers... The first and third layers are Graph Convolutional Networks (GCN), while the second layer is a Graph Attention Network (GAT) with two attention heads... A residual connection from the input features to the output of the third layer ensures better gradient flow and feature retention. The model employs a Jumping Knowledge mechanism in concatenation mode, aggregating features from all three layers to capture hierarchical information. The aggregated features are globally pooled, batch normalized, and passed through a dropout layer with a 0.5 probability to prevent overfitting. Finally, these processed features are fed into a linear classifier that outputs the class logits. This architecture is trained the same way as the above architectures.