Predicting the Hardness of Learning Bayesian Networks

Authors: Brandon Malone, Kustaa Kangas, Matti Jarvisalo, Mikko Koivisto, Petri Myllymaki

AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results, based on the largest evaluation of stateof-the-art BNS learning algorithms to date, demonstrate that we can predict the runtimes to a reasonable degree of accuracy, and effectively select algorithms that perform well on a particular instance.
Researcher Affiliation Academia Helsinki Institute for Information Technology & Department of Computer Science, University of Helsinki, Finland
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions links to external solvers used (GOBNILP, URLearning) and an 'online supplement' for 'more details' (http://bnportfolio.cs.helsinki.fi/), but does not explicitly state that the source code for their own methodology is being released or provide a clear link to it.
Open Datasets Yes Datasets originating from the UCI repository (Bache and Lichman 2013). 19 datasets. Datasets sampled from benchmark Bayesian networks, downloaded from http://www.cs.york.ac.uk/aig/ sw/gobnilp/.
Dataset Splits Yes We evaluated the portfolios using the standard 10-fold cross-validation technique. That is, the data is partitioned into 10 non-overlapping subsets. In each fold, 9 of the subsets are used to train the model, and the remaining set is used for testing; each subset is used as the testing set once.
Hardware Specification Yes For running the experiments we used a cluster of Dell Power Edge M610 computing nodes equipped with two 2.53GHz Intel Xeon E5540 CPUs and 32-GB RAM.
Software Dependencies Yes We use the GOBNILP solver, version 1.4.1 (http: //www.cs.york.ac.uk/aig/sw/gobnilp/) as a representative for ILP. GOBNILP uses the SCIP framework (http://scip.zib.de/) and an external linear program solver; we used SCIP 3.0.1 and So Plex 1.7.1 (http:// soplex.zib.de/).
Experiment Setup Yes For each dataset and scoring function, we generated scores with parent limits ranging from 2 to up to 6. For each individual run, we used a timeout of 2 hours and a 28-GB memory limit. We considered 5 different scoring functions: BDeu with the Equivalent Sample Size parameter selected from {0.1, 1, 10, 100} and the BIC scores.