Scalable Probabilistic Causal Structure Discovery

Authors: Dhanya Sridhar, Jay Pujara, Lise Getoor

IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare our method against multiple well-studied approaches on biological and synthetic datasets, showing improvements of up to 20% in F1-score over the best performing baseline in realistic settings.
Researcher Affiliation Academia Dhanya Sridhar1, Jay Pujara2, Lise Getoor1, 1 University of California Santa Cruz 2 University of Southern California {dsridhar,getoor}@soe.ucsc.edu, jay@cs.umd.edu
Pseudocode Yes Table 1 shows the rules used in CAUSPSL.
Open Source Code Yes Code and data at: bitbucket.org/linqs/causpsl.
Open Datasets Yes We validate our approach using three datasets: (1) synthetic linear acyclic models with Gaussian noise; (2) simulated gene expression from the DREAM4 challenge [Marbach et al., 2010; Prill et al., 2010]; (3) perturbation experiments on protein-signaling pathways [Sachs et al., 2005].
Dataset Splits Yes To calculate F1 in DREAM and synthetic settings, rounding thresholds on the continuous outputs of CAUSPSL and Bootstrapped PC are selected using cross-validation with 10 and 100 folds, respectively.
Hardware Specification No The paper does not specify the hardware used to run the experiments, such as specific CPU or GPU models.
Software Dependencies No The paper mentions 'pcalg and bnlearn R packages' and 'PSL [Bach et al., 2017]' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Without a priori preference for rules, we set all CAUSPSL rule weights to 5.0 except for causal and ancestral orientation rules 2 which are set to 10.0, since they encode strong asymmetry constraints. For both PC variants and CAUSPSL, we condition on sets up to size two for DREAM20 and up to size one for DREAM30.