Scalable Intervention Target Estimation in Linear Models

Authors: Burak Varici, Karthikeyan Shanmugam, Prasanna Sattigeri, Ali Tajer

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, simulation results on both real and synthetic data demonstrate the gains of the proposed approach for scalable causal structure recovery.
Researcher Affiliation Collaboration Burak Varıcı Rensselaer Polytechnic Institute varicb@rpi.edu Karthikeyan Shanmugam IBM Research AI karthikeyan.shanmugam2@ibm.com
Pseudocode Yes Algorithm 1 Causal Intervention Target Estimator (CITE) Algorithm 2 Functions for the main algorithm
Open Source Code Yes Implementation of the algorithm and the code to reproduce the simulation results are available at https://github.com/bvarici/intervention-estimation.
Open Datasets Yes Protein signaling data. We first consider the dataset in [22] for discovering the protein signaling network of 11 nodes. It consists of measurements of proteins and phospolipids under different interventional environments. Perturb-seq gene expression data. We analyze the performance of our algorithm on the perturb-seq dataset by in [24].
Dataset Splits No The paper mentions sample sizes (e.g., "5000 samples", "1755 observational and 4091 interventional samples") but does not provide specific training/validation/test dataset splits needed to reproduce the experiment.
Hardware Specification Yes All the simulations are run on a Mac Book Pro with 2.7 GHz Dual-Core i5 core and 8 GB RAM.
Software Dependencies No The paper mentions using an "ADMM-based method" from [12] but does not provide specific software names with version numbers (e.g., Python 3.8, PyTorch 1.9).
Experiment Setup Yes We generate 100 realizations of Erd os-Rényi [21] DAGs with expected neighborhood size c = 1.5, and |I| = 5. We sample the entries of B, i.e., the edge weights, independently at random according to the uniform distribution on [ 1, 0.25] [0.25, 1]. The additive Gaussian noise terms have distribution N(0, Ip). We select the intervention set I by randomly selecting 5 nodes from [p]. We consider three different models to intervene on the nodes in I: (i) shift intervention model in which mean of the noise ϵi is shifted from 0 to 1, (ii) variance increase model in which the variance of the noise ϵi is increased from 1 to 2, and (iii) randomized intervention model, in which (B(2))pa(i),i = 0 and the noise variance varies from 1 to 1.5.