Predicting the Hardness of Learning Bayesian Networks

Online Supplement

This webpage is an online supplement of the paper

Predicting the Hardness of Learning Bayesian Networks.
Brandon Malone, Kustaa Kangas, Matti Järvisalo, Mikko Koivisto, and Petri Myllymäki.
In Carla E. Brodley and Peter Stone, editors, Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI 2014), pages 2460-2466. AAAI Press, 2014.
[pdf] [abstract/bibtex]

Benchmark data files

• Derived datasets, as used for cross-validation in [Malone, Järvisalo, and Myllymäki, UAI 2015], are available here.

Solver running times

Features values and solving running times (in .csv)

Comparison of A* Variants

Figure 1: Comparison of A*-ec with A* (left) and A*-ed3 (right).

Comparison of parameterizations of A* and ILP and the VBS

Figure 2: Comparison of parameterizations of A* (left) and ILP (right) and the VBS over all parameterizations of all solvers.

Predictions Errors using M5' Trees

Figure 3: Prediction errors using M5' trees and different sets of features: A* (left) and ILP (right).

Predicted vs Actual Runtimes

Figure 4: Predicted and actual runtimes for ILP using Basic features (left) and all non-probing features and ILP probing (right).

Prediction Errors using REP Trees

Figure 5: Prediction errors using REP trees and different sets of features: A* (left) and ILP (right).

Figure 6: Prediction errors using REP trees and different sets of features, including multiplicative pairs: A* (left) and ILP (right).