Efficient Computation of Higher-Order Subgraph Attribution via Message Passing
Authors: Ping Xiong, Thomas Schnake, Grégoire Montavon, Klaus-Robert Müller, Shinichi Nakajima
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show the significant acceleration of the proposed algorithms and demonstrate the high usefulness and scalability of our novel generalized subgraph attribution method. In this section, we conduct two experiments demonstrating (1) the massive gain in computation time by our efficient s GNN-LRP, and (2) the usefulness of the generalized subgraph attribution in relevant node-ordering tasks. |
| Researcher Affiliation | Academia | 1Technische Universit at Berlin (TU Berlin) 2BIFOLD Berlin Institute for the Foundations of Learning and Data 3Department of Artificial Intelligence, Korea University, Seoul 136-713, Korea 4Max Planck Institut f ur Informatik, 66123 Saarbr ucken, Germany 5RIKEN Center for AIP, Japan. |
| Pseudocode | Yes | Algorithm 1 s GNN-LRP, α = 0, Algorithm 2 s GNN-LRP, α (0, 1], Algorithm 3 Model Activation Task, and Algorithm 4 Model Pruning Task. |
| Open Source Code | Yes | Detailed experimental setting is given in Appendix H, and our implementation is available at our Git Hub repository.2 (footnote 2: https://github.com/xiong-ping/sgnn_lrp_ via_mp.) |
| Open Datasets | Yes | We used the following five popular datasets: BA-2motif (Luo et al., 2020), MUTAG (Debnath et al., 1991), Mutagenicity (Kazius et al., 2005b), REDDIT-BINARY (Yanardag & Vishwanathan, 2015), and Graph-SST2 (Yuan et al., 2020b). We downloaded the dataset from the repository of Schnake et al. (2021)... |
| Dataset Splits | No | The paper does not explicitly provide training/validation/test dataset splits. For each dataset, it describes a train/test split, for example, for BA-2motif: 'we used the 0-400 and 500-900 as training dataset and the rest as testing dataset.' However, no separate validation set is specified. |
| Hardware Specification | Yes | Experiments were performed on a Xeon E5-2620 CPU with 8GB memory. |
| Software Dependencies | No | The paper describes the models used (e.g., 'GIN models', 'GCN') and optimizers ('SGD', 'Adam'), but it does not provide specific software dependency versions (e.g., Python, PyTorch, or TensorFlow versions) that would be needed to replicate the experiment. |
| Experiment Setup | Yes | We trained GIN models with 2,3,4,5,6,7 layers, and all models has the same GIN block, which is a 2-layer MLP. The input feature dimension is 1 and the output feature dimension in the MLP blocks N (l) = 20, l = 1, , L 1. The activation function is Re LU. We employed the SGD optimizer with a decreasing learning rate γ = 0.00001/(1.0+(epoch/epochs) for 10000, 10000, 5000, 1000, 1000, 1000 epochs, respectively. |