Interpretable Prototype-based Graph Information Bottleneck
Authors: Sangwoo Seo, Sungwon Kim, Chanyoung Park
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments, including qualitative analysis, demonstrate that PGIB outperforms state-of-the-art methods in terms of both prediction performance and explainability. ... We conducted extensive experiments to evaluate the effectiveness and interpretability of the reasoning process of PGIB in graph classification tasks. Our results show that PGIB outperforms recent stateof-the-art methods, including existing prototype learning-based and IB-based methods. Moreover, we evaluated the ability of PGIB in capturing the label information by evaluating the classification performance using only the detected subgraph Gsub. We also conducted a qualitative analysis that visualizes the subgraph Gsub and prototype graph Gp, suggesting the ability of PGIB in detecting the key subgraph. |
| Researcher Affiliation | Academia | Sangwoo Seo1 Sungwon Kim1 Chanyoung Park1 1KAIST {sangwooseo@kaist.ac.kr, swkim@kaist.ac.kr, cy.park@kaist.ac.kr} |
| Pseudocode | No | The paper describes its method in text and with diagrams, but does not provide pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | The source code of PGIB is available at https://github.com/sang-woo-seo/PGIB. |
| Open Datasets | Yes | We use the MUTAG [16], PROTEINS [1], NCI1 [24], and DD[5] datasets. These are datasets related to molecules or bioinformatics, and are widely used for evaluations on graph classification. ... We use four molecular properties from the ZINC [8] dataset, which consists of 250,000 molecules, for graph interpretation. |
| Dataset Splits | Yes | Each dataset is split into training, validation, and test sets with a ratio of 80%, 10%, and 10%, respectively. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory specifications) used for running the experiments. |
| Software Dependencies | No | GIN [28] is used as the encoder for all models used in the experiment. ... All models are trained for 300 epochs using the Adam optimizer with a learning rate of 0.005. The paper mentions software components like GIN and Adam optimizer, but does not provide specific version numbers for these or other libraries/frameworks. |
| Experiment Setup | Yes | All models are trained for 300 epochs using the Adam optimizer with a learning rate of 0.005. ... For simplicity, the hyperparameters α1, α2, and α3 in Equation 16 are set to 0.0001, 0.01 to 0.1 and 5, respectively. The prototype merge operation starts at epoch 100 and is performed every 50 epochs thereafter. We set the number of prototypes per class to 7 and combine 30% of the most similar prototype pairs. |