Learning Fast-Inference Bayesian Networks

Authors: Vaidyanathan Peruvemba Ramaswamy, Stefan Szeider

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that our new learning methods produce BNs that support significantly faster exact probabilistic inference than BNs learned with treewidth bounds. We consider several variants of bss learning algorithms, tested them on 16 real-world benchmark data sets with up to 1041 variables, and compare them with bounded treewidth BN learning algorithms.
Researcher Affiliation Academia Vaidyanathan Peruvemba Ramaswamy Algorithms and Complexity Group TU Wien, Vienna, Austria Stefan Szeider Algorithms and Complexity Group TU Wien, Vienna, Austria
Pseudocode No The paper describes algorithms in prose (e.g., in Section 4.1 'Modified Heuristics') but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes We provide the source code as a public github repository [Peruvemba Ramaswamy and Szeider, 2021b].
Open Datasets Yes We tested the algorithms on a subset of the bnlearn repository. These networks are commonly used as benchmarks in the literature. 3https://www.bnlearn.com/bnrepository/
Dataset Splits No The paper does not specify train/validation/test dataset splits or cross-validation methodology for their experiments. It focuses on evaluating BN learning algorithms on full benchmark datasets.
Hardware Specification Yes We tested the various proposed methods on 4-core Intel Xeon E5540 2.53 GHz CPU (internal cluster), with each process having access to 8GB RAM.
Software Dependencies No The paper mentions software like BLIP, Java, Network X, Python, UWr Max Sat, and Merlin, but does not provide specific version numbers for any of these dependencies, which is required for reproducibility.
Experiment Setup Yes We run the algorithms for a total time of 90 minutes and record the reasoning time and score at the end. We denote by BN-SLIMbss(X) the algorithm composed of running the heuristic X for 30 minutes and then running the bounded state space local improvement algorithm on top of the heuristic solution for another 60 minutes.