Is Distance Matrix Enough for Geometric Deep Learning?
Authors: Zian Li, Xiyuan Wang, Yinan Huang, Muhan Zhang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments verify our theory. Our k-Dis GNNs achieve many new state-of-the-art results on MD17. In this section, we evaluate the experimental performance of k-Dis GNNs. Our main objectives are to answer the following questions: Q1 Does 2/3-E/F-Dis GNN outperform their counterparts in experiments (corresponding to Section 6.2)? Q2 As universal models for scalars (when k 2), do k-Dis GNNs also have good experimental performance (corresponding to Section 6.3)? Q3 Does incorporating well-designed edge representations in the k-E-Dis GNN result in improved performance? |
| Researcher Affiliation | Academia | Zian Li1,2, Xiyuan Wang1,2, Yinan Huang1, Muhan Zhang1, 1Institute for Artificial Intelligence, Peking University 2School of Intelligence Science and Technology, Peking University |
| Pseudocode | No | The paper describes algorithms (e.g., 1-WL-E, k-WL, k-FWL, k-Dis GNN message passing) using mathematical notation and textual descriptions, but no clearly labeled 'Algorithm' or 'Pseudocode' blocks are present. |
| Open Source Code | Yes | Our code is available at https://github.com/Graph PKU/Dis GNN. |
| Open Datasets | Yes | MD17 [Chmiela et al., 2017] is a dataset commonly used to evaluate the performance of machine learning models in the field of molecular dynamics. QM9 [Ramakrishnan et al., 2014, Wu et al., 2018] consists of 134k stable small organic molecules with 19 regression targets. Table 4: Training settings. |
| Dataset Splits | Yes | Table 4: Training settings. Train set size 1000, Val. set size 1000 (MD17). Train set size 110000, Val. set size 10000 (QM9). To prevent overfitting, we use early stopping on validation loss and an exponential moving average (EMA) with decay rate 0.99 for model parameters during validation and test. |
| Hardware Specification | Yes | Experiments are conducted on Nvidia RTX 3090 and Nvidia RTX 4090. Experiments are conducted on Nvidia RTX 3090 and Nvidia RTX 4090. The experiments are conducted on Nvidia RTX 3090, and the results are shown in Table 7. Evaluated on Nvidia A100. |
| Software Dependencies | No | We optimize all models using Adam [Kingma and Ba, 2014]. The paper mentions Adam optimizer but does not specify its version or other software dependencies with version numbers. |
| Experiment Setup | Yes | Training Setting. For QM9, we use the mean squared error (MSE) loss for training. For MD17, we use the weighted loss function L(X, z) = (1 ρ)|fθ(X, z) ˆt(X, z)| + ρα=1 ( fθ(X, z)xiα ˆFiα(X, z))2, where the force ratio ρ is fixed as 0.999 (For revised MD17, we set ρ to 0.99 for several targets since the data quality is better). We follow the same dataset split as Gem Net [Gasteiger et al., 2021]. We optimize all models using Adam [Kingma and Ba, 2014] with exponential decay and plateau decay learning rate schedulers, and also a linear learning rate warm-up. Table 4: Training settings. warm-up epochs 25, initial learning rate 0.001, decay on plateau patience (epochs) 15, decay on plateau cooldown (epochs) 15, decay on plateau threshold 0.001, decay on plateau factor 0.7. For rbf dimension, we use 16 for MD17 and 32 for QM9. We choose the number of message passing blocks from {4, 5, 6}. For hidden dimension, we use 512 for 2-Dis GNNs and 320 for 3-Dis GNNs. |