Mirage: Model-agnostic Graph Distillation for Graph Classification
Authors: Mridul Gupta, Sahil Manchanda, HARIPRASAD KODAMANA, Sayan Ranu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive benchmarking on real-world datasets underscores MIRAGE s superiority, showcasing enhanced generalization accuracy, data compression, and distillation efficiency when compared to state-of-the-art baselines. |
| Researcher Affiliation | Academia | Mridul Gupta Yardi School of Artificial Intelligence Indian Institute of Technology, Delhi Hauz Khas, New Delhi, Delhi, India mridul.gupta@scai.iitd.ac.in Sahil Manchanda Department of Computer Science Indian Institute of Technology, Delhi Hauz Khas, New Delhi, Delhi, India sahilm1992@gmail.com Hariprasad Kodamana Department of Chemical Engineering Yardi School of Artificial Intelligence Indian Institute of Technology, Delhi Hauz Khas, New Delhi, Delhi, India kodamana@iitd.ac.in Sayan Ranu Department of Computer Science Yardi School of Artificial Intelligence Indian Institute of Technology, Delhi Hauz Khas, New Delhi, Delhi, India sayanranu@cse.iitd.ac.in |
| Pseudocode | Yes | Algorithm. 1 in the appendix outlines the pseudocode of our data distillation and Algorithm. 2 outlines the modeling algorithm. |
| Open Source Code | Yes | The codebase of MIRAGE is shared at https://github.com/idea-iitd/Mirage. |
| Open Datasets | Yes | To evaluate MIRAGE, we use datasets from Open Graph Benchmark (OGB) (Hu et al., 2020) and TU Datasets (DD, IMDB-B and NCI1) (Morris et al., 2020) spanning a variety of domains. |
| Dataset Splits | Yes | The OGB datasets come with the train-validation-test splits, which are also used in DOSCOND and KIDD. For TU Datasets, we randomly split the graphs into 80%/10%/10% for training-validation-test. |
| Hardware Specification | Yes | All experiments are performed on an Intel Xeon Gold 6248 processor with 96 cores and 1 NVIDIA A100 GPU with 40GB memory, and 377 GB RAM with Ubuntu 18.04. |
| Software Dependencies | No | The paper mentions 'Ubuntu 18.04' as the operating system. However, it does not specify versions for any other software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages critical for reproducibility. |
| Experiment Setup | Yes | For the model hyper-parameters, we perform grid search to optimize performance on the whole dataset. The same parameters are used to train and infer on the distilled dataset. The hyperparameters used are shown in Table Db. ... All experiments are performed on an Intel Xeon Gold 6248 processor... In all experiments, we have trained using the Adam optimizer with a learning rate of 0.0001 and choose the model based on the best validation loss. ... We stop the training of a model if it does not improve the validation loss for more than 15 epochs. Table D (b) Model parameters: GCN/GAT/GIN Layers {2, 3}, Hidden Dimension {64, 128}, Dropout [0, 0.6], Reduce Type {sum,mean}. |