Differentiable and Transportable Structure Learning

Authors: Jeroen Berrevoets, Nabeel Seedat, Fergus Imrie, Mihaela Van Der Schaar

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we empirically validate D-Struct with respect to edge accuracy and structural Hamming distance in a variety of settings.
Researcher Affiliation Academia 1DAMTP, University of Cambridge, UK 2UCLA, CA, USA 3The Alan Turing Institute, UK.
Pseudocode Yes Algorithm 1: Outer-loop of dual ascent procedure for D-Struct(NOTEARS-MLP) and Algorithm 2: training step for DStruct(NOTEARS-MLP) cfr. Alg. 1
Open Source Code Yes Please find our online code repository at: https://github.com/jeroenbe/d-struct or https://github.com/vanderschaarlab Our code is based on code provided by Zheng et al. [36], and we annotated our code where we used their implementation.
Open Datasets No The paper describes generating synthetic data based on 'Erdos-Reny ı (ER) graphs' and 'Scale-Free graphs' with structural equations as in Zheng et al. [36]. While the generation method is cited, the specific datasets used in the experiments are not provided via a direct link, DOI, or named repository, nor are they described as a pre-existing public dataset.
Dataset Splits No The paper mentions drawing 'samples' and splitting them for D-Struct's internal mechanism (e.g., 'split the samples into two equal-sized subsets' for transportability evaluation), but does not specify standard training, validation, and test splits for model evaluation or reproduction.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory, or cloud resources) used to run the experiments.
Software Dependencies No The paper mentions 'Python', 'PyTorch', and the 'L-BFGS-B optimizer' being used, and that their code is based on Zheng et al. [36]'s implementation, but it does not specify version numbers for any of these software components or libraries.
Experiment Setup Yes Specifically, we let K distinct DSFs learn a DAG from one of the K datasets, agnostic from each other. If the chosen DSF requires hyperparameters (such as λ1,2 and ρ in eq. (4)), we have to also include these in D-Struct s set of required hyperparameters. While it is possible to set different hyperparameter values for each of the DSFs separately (which is potentially helpful when there is a lot of variety in the K distinct datasets), we fix these across DSFs in light of simplicity. A discussion on D-Struct s hyperparameters can be found in Appendix A.1. We set our hyperparameters to those which yielded best performance (deduced from Fig. 7 for α, and K = 3 when not varied over as this yielded the most stable results overall).