Active Structure Learning of Causal DAGs via Directed Clique Trees

Authors: Chandler Squires, Sara Magliacane, Kristjan Greenewald, Dmitriy Katz, Murat Kocaoglu, Karthikeyan Shanmugam

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show via synthetic experiments that our algorithm can scale to much larger graphs than most of the related work and achieves better worst-case performance than other scalable approaches.
Researcher Affiliation Collaboration Chandler Squires LIDS, MIT MIT-IBM Watson AI Lab csquires@mit.edu Sara Magliacane MIT-IBM Watson AI Lab IBM Research sara.magliacane@gmail.com Kristjan Greenewald MIT-IBM Watson AI Lab IBM Research kristjan.h.greenewald@ibm.com Dmitriy Katz MIT-IBM Watson AI Lab IBM Research dkatzrog@us.ibm.com Murat Kocaoglu MIT-IBM Watson AI Lab IBM Research murat@ibm.com Karthikeyan Shanmugam MIT-IBM Watson AI Lab IBM Research karthikeyan.shanmugam2@ibm.com
Pseudocode Yes Algorithm 1 DCT POLICY
Open Source Code Yes A code base to recreate these results can be found at https://github.com/csquires/dct-policy.
Open Datasets No The paper uses synthetic graphs generated by a described procedure but does not refer to a publicly available dataset with a specific link or citation for access.
Dataset Splits No No explicit training, validation, or test dataset splits (e.g., percentages, sample counts, or cross-validation setup) were mentioned for the synthetic graph generation and evaluation.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for experiments were explicitly mentioned.
Software Dependencies No No specific software dependencies with version numbers were mentioned (e.g., library names with versions).
Experiment Setup Yes For our evaluation on smaller graphs, we generate random connected moral DAGs using the following procedure, which is a modification of Erdös-Rényi sampling that guarantees that the graph is connected. We first generate a random ordering σ over vertices. Then, for the n-th node in the order, we set its indegree to be Xn = max(1, Bin(n 1, ρ)), and sample Xn parents uniformly from the nodes earlier in the ordering. Finally, we chordalize the graph by running the elimination algorithm (Koller & Friedman, 2009) with elimination ordering equal to the reverse of σ.