Learning to Predict Graphs with Fused Gromov-Wasserstein Barycenters

Authors: Luc Brogat-Motte, Rémi Flamary, Celine Brouard, Juho Rousu, Florence D’Alché-Buc

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments show the strength of the method and its ability to interpolate in the labeled graph space on simulated data and on a difficult metabolic identification problem where it can reach very good performance with very little engineering.
Researcher Affiliation Academia 1LTCI, Télécom Paris, Institut Polytechnique de Paris, France 2Ecole Polytechnique, Institut Polytechnique de Paris, CMAP, UMR 7641, Palaiseau, France 3Université de Toulouse, INRAE, UR MIAT, France 4Department of Computer Science, Aalto University, Finland.
Pseudocode Yes Algorithm 1 Neural network-based model training One stochastic gradient descent step
Open Source Code Yes A Python implementation of the method is available on github1. https://github.com/lmotte/graphprediction-with-fused-gromov-wasserstein
Open Datasets Yes Here we consider a set of 4138 labeled data, that have been extracted and processed in Dührkop et al. (2015), from the GNPS public spectral library (Wang et al., 2016).
Dataset Splits Yes The two hyperparameters (ridge regularization parameter λ and the output metric s parameter) are selected using a validation set (1/5 of the training set) and Top-1 accuracy.
Hardware Specification No The paper describes its methods and experimental setup but does not provide specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments.
Software Dependencies No The paper mentions 'Python implementation... using the POT library: Python Optimal Transport (Flamary et al., 2021), and Pytorch library (Paszke et al., 2019)' but does not specify exact version numbers for Python, POT, or PyTorch.
Experiment Setup Yes We use M = 10 templates, with 5 nodes, and initialize them drawing Ci R5 5, Fi R5 1 uniformly in [0, 1]5 5 and [0, 1]5 1. The weights α(x; W) RM are implemented using a three-layer ( 100 neurons in each hidden layer) fully connected neural network with Re LU activation functions, and a final softmax layer. We use β = 1/2 as FGW s balancing parameter and a prediction size of n = 40 during training.