Learning Bayesian Networks from Ordinal Data

Authors: Xiang Ge Luo, Giusi Moffa, Jack Kuipers

JMLR 2021 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through simulation studies, we illustrate the superior performance of the OSEM algorithm compared to the alternatives and analyze various factors that may influence the learning accuracy. Finally, we demonstrate the practicality of our method with a real-world application on psychological survey data from 408 patients with co-morbid symptoms of obsessive-compulsive disorder and depression.
Researcher Affiliation Academia Xiang Ge Luo EMAIL D-BSSE, ETH Zurich, Mattenstrasse 26, 4058 Basel, Switzerland Giusi Moffa Department of Mathematics and Computer Science, University of Basel, Basel, Switzerland Division of Psychiatry, University College London, London, UK Jack Kuipers EMAIL D-BSSE, ETH Zurich, Mattenstrasse 26, 4058 Basel, Switzerland
Pseudocode Yes Algorithm 1: Ordinal Structural EM (OSEM)
Open Source Code Yes . R code is available at https://github.com/xgluo/OSEM
Open Datasets Yes For score-based approaches, we also evaluate the predictive performance using five real ordinal data sets from Mc Nally et al. (2017) and the UCI machine learning repository (Dua and Graff, 2017) (Table 1).
Dataset Splits Yes For each data set and method, we train the network with 80% of the data points, and conditioned on the structure, we compute the log loss on the remaining 20% test cases.
Hardware Specification No The runtimes are highly dependent on the algorithmic implementation and the software packages chosen (see Appendix B.7). No specific hardware (e.g., GPU, CPU model) is mentioned for running experiments.
Software Dependencies No Table S1 lists several R packages (e.g., pcalg, bnlearn, Bi DAG, MXM, rpcart) used for implementation, but it does not specify version numbers for these packages or the R environment itself.
Experiment Setup Yes We choose the Monte Carlo sample size K to be 5 and the penalty coefficient λ to be 6, which corresponds to the highest sum of average TPR and (1 FPRp) on the ROC curves for N = 500, n = 20 or 30, E[Li] = 4 or 5, and ν = 1 from our simulations studies.