BACKSHIFT: Learning causal cyclic graphs from unknown shift interventions

Authors: Dominik Rothenhäusler, Christina Heinze, Jonas Peters, Nicolai Meinshausen

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present empirical results for both synthetic and real data sets. Numerical results for simulated data and applications in flow cytometry and financial data are shown in Section 4.
Researcher Affiliation Academia Dominik Rothenha usler Seminar f ur Statistik ETH Z urich, Switzerland rothenhaeusler@stat.math.ethz.ch Christina Heinze Seminar f ur Statistik ETH Z urich, Switzerland heinze@stat.math.ethz.ch Jonas Peters Max Planck Institute for Intelligent Systems T ubingen, Germany jonas.peters@tuebingen.mpg.de Nicolai Meinshausen Seminar f ur Statistik ETH Z urich, Switzerland meinshausen@stat.math.ethz.ch
Pseudocode Yes Algorithm 1 BACKSHIFT Input: Xj 8j 2 J 1: Compute d x,j 8j 2 J 2: D = FFDIAG( d x,j) 3: ˆD = Permute And Scale( D) 4: ˆB = I ˆD Output: ˆB
Open Source Code Yes We attach R-code with which all simulations and analyses can be reproduced2. 2An R-package called backShift is available from CRAN.
Open Datasets Yes The data published in [22] is an instance of a data set where the external interventions differ between the environments in J and might act on several compounds simultaneously [18]. [22] K. Sachs, O. Perez, D. Pe er, D. Lauffenburger, and G. Nolan. Causal protein-signaling networks derived from multiparameter single-cell data. Science, 308:523 529, 2005.
Dataset Splits No The paper describes grouping data into '74 overlapping blocks of 61 consecutive days each' for financial time series analysis and different 'experimental conditions' or 'settings' for other datasets, but does not specify traditional train/validation/test splits.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running experiments are mentioned.
Software Dependencies No An R-package called backShift is available from CRAN. (The paper mentions an R package but does not provide specific version numbers for R or other software dependencies.)
Experiment Setup Yes Specifically, we generate ten distinct environments with non-Gaussian noise. In each environment, the random intervention variable is generated as (cj)k = βj k, where βj 1, . . . , βj p are drawn i.i.d. from Exp(m I) and Ij 1, . . . , Ij p are independent standard normal random variables. The intervention shift thus acts on all observed random variables. The parameter m I regulates the strength of the intervention.