Efficient Intervention Design for Causal Discovery with Latents

Authors: Raghavendra Addanki, Shiva Kasiviswanathan, Andrew Mcgregor, Cameron Musco

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we compare the total number of interventions required to recover causal graph G parameterized by p-colliders (See Section 4) vs. maximum degree utilized by [Kocaoglu et al., 2017b]. ... In our plots (Figure 2), we compare the maximum undirected degree (d) with the maximum number of p-colliders between any pair of nodes (which defines ). We ran each experiment 10 times and plot the mean value along with one standard deviation error bars.
Researcher Affiliation Collaboration Raghavendra Addanki 1 Shiva Prasad Kasiviswanathan 2 Andrew Mc Gregor 1 Cameron Musco 1 1College of Information and Computer Sciences, University of Massachusetts, Amherst, USA. 2Amazon.
Pseudocode Yes Algorithm 1 SSMATRIX (V, m)
Open Source Code No The paper does not contain any explicit statement about releasing source code for the described methodology or provide a link to a code repository.
Open Datasets No We demonstrate our results by considering sparse random graphs generated from the families of: (i) Erdös Rényi random graphs G(n, c/n) for constant c, (ii) Random bipartite graphs generated using G(n1, n2, c/n) model, with partitions L, R and edges directed from L to R, (iii) Random directed trees with degrees of nodes generated from power law distribution.
Dataset Splits No The paper describes generating synthetic graphs for experiments but does not provide specific train/validation/test dataset splits or any other detailed data partitioning methodology.
Hardware Specification No The paper does not provide specific hardware details (such as GPU/CPU models or memory specifications) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers, needed to replicate the experiment.
Experiment Setup Yes Setup. We demonstrate our results by considering sparse random graphs generated from the families of: (i) Erdös Rényi random graphs G(n, c/n) for constant c, (ii) Random bipartite graphs generated using G(n1, n2, c/n) model, with partitions L, R and edges directed from L to R, (iii) Random directed trees with degrees of nodes generated from power law distribution. In each of the graphs that we consider, we include latent variables by sampling 5% of pairs and adding a latent between them. ... For random bipartite graphs, ... we use equal partition sizes n1 = n2 = n/2 and plot the results for G(n/2, n/2, c/n) for constant c = 5.