Minimal I-MAP MCMC for Scalable Structure Discovery in Causal DAG Models
Authors: Raj Agrawal, Caroline Uhler, Tamara Broderick
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 6 we empirically compare our model to order MCMC and partition MCMC (Kuipers & Moffa, 2017), the state-of-the-art version of structure MCMC. In experiments we observe O(p3) time scaling for our method, and we demonstrate better mixing and ROC performance for our method on several datasets. |
| Researcher Affiliation | Academia | 1Computer Science and Artificial Intelligence Laboratory 2Institute for Data, Systems and Society 3Laboratory for Information and Decision Systems, Massachusetts Institute of Technology. Correspondence to: Raj Agrawal <r.agrawal@csail.mit.edu, r.agrawal.raj@gmail.com>. |
| Pseudocode | Yes | Algorithm 1 Minimal I-MAP MCMC; Algorithm 2, denoted as update minimal I-MAP (UMI), is used as a step in Algorithm 1 and describes how to compute a minimal I-MAP ˆG from a minimal I-MAP ˆG when and differ by an adjacent transposition without recomputing all edges; see also Solus et al. (2017). |
| Open Source Code | No | The paper states: "In terms of software, we used the code provided by Kuipers & Moffa (2017) to run partition and order MCMC. We used the method and software of Kangas et al. (2016) for counting linear extensions for bias correction, and we implemented minimal I-MAP MCMC using the R-package bnlearn." This indicates they used third-party code but does not state that their own source code is provided or made publicly available. |
| Open Datasets | Yes | The third dataset is from the Dream4 in-silico network challenge (Schaffter et al., 2011) on gene regulation. |
| Dataset Splits | No | The paper does not explicitly mention the use of a validation set or specific training/validation/test splits. It discusses burn-in and thinning for the MCMC chains but not dataset partitioning for model training vs. evaluation. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU/CPU models, memory, or cloud computing instance types used for running the experiments. |
| Software Dependencies | No | The paper mentions implementing minimal I-MAP MCMC using the "R-package bnlearn" but does not specify version numbers for R, bnlearn, or any other software dependencies. |
| Experiment Setup | Yes | For each dataset, we ran the Markov chains for 105 iterations, including a burn-in of 2 × 104 iterations, and thinned the remaining iterations by a factor of 100. [...] To achieve this end, we choose a prior of the form P(G) = P(G) exp(-γ|A|) where P(G) can include any structural information known about the DAG. [...] Input: Data D, number of iterations T, significance level α, initial permutation π0, sparsity strength γ, thinning rate τ. |