Wasserstein Embedding for Graph Learning

Authors: Soheil Kolouri, Navid Naderializadeh, Gustavo K. Rohde, Heiko Hoffmann

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our new graph embedding approach on various benchmark graph-property prediction tasks, showing state-of-the-art classification performance while having superior computational efficiency.
Researcher Affiliation Collaboration HRL Laboratories, LLC., University of Virginia
Pseudocode No The paper describes methods and numerical details in text and equations but does not present them in a structured pseudocode or algorithm block format.
Open Source Code Yes The code is available at https://github.com/navid-naderi/WEGL.
Open Datasets Yes We first evaluate our algorithm on the molecular property prediction task on the ogbg-molhiv dataset. This dataset is part of the Open Graph Benchmark (Hu et al., 2020)... We also consider a set of social network, bioinformatics and molecule graph datasets (Kersting et al., 2020).
Dataset Splits Yes To train and evaluate our proposed method, we use the scaffold split provided by the dataset... we perform 10-fold cross-validation with random splitting on the entire dataset
Hardware Specification Yes We carry out our experiments for WEGL and WWL on a 2.3 GHz Intel R Xeon R E5-2670 v3 CPU, while we use a 16 GB NVIDIA R Tesla R P100 GPU for GIN.
Software Dependencies Yes Auto-ML: Auto-Sklearn 2.0: The next generation. ar Xiv preprint ar Xiv:2007.04074, 2020.
Experiment Setup Yes We perform a grid search over a set of hyperparameters and report the configuration that leads to the best validation performance... HYPERPARAMETERS We use the following set of hyperparameters to perform a grid search over in each of the experiments: Random Forest: min_samples_leaf {1, 2, 5}, min_samples_split {2, 5, 10}, and n_estimators {25, 50, 100, 150, 200}. Gradient Boosted Decision Tree (GBDT): min_samples_leaf {1, 2, 5}, min_samples_split {2, 5, 10}, n_estimators {25, 50, 100, 150, 200}, and max_depth {1, 3, 5}. SVM-Linear and SVM-RBF: C {10 2, . . . , 105}. Multi-Layer Perceptron (MLP): hidden_layer_sizes {(128), (256), (128, 64), (256, 128)}. Auto-ML: Auto-Sklearn 2.0 searches over a space of 42 hyperparameters using Bayesian optimization techniques, as mentioned in Feurer et al. (2020). Number of Diffusion Layers in equation 7 and equation 14: L {3, . . . , 8}. Initial Node Feature Dimensionality (for ogbg-molhiv only): {100, 300, 500}.