Balancing Computational Cost and Accuracy in Inference of Continuous Bayesian Networks

Maarten C. Vonk, Sebastiaan Brand, Ninoslav Malekovic, Thomas Bäck, Alfons Laarman, Anna V. Kononova
Proceedings of The 12th International Conference on Probabilistic Graphical Models, PMLR 246:361-381, 2024.

Abstract

Bayesian networks allow a parsimonious encoding of joint probability distributions via directed acyclic graphs. While discrete Bayesian network inference is well-established, conducting inference on continuous Bayesian networks often requires discretization. In this paper, continuous Bayesian networks are subjected to various supervised and unsupervised discretization methods. Subsequently, the discretized Bayesian networks are encoded into decision diagrams, facilitating efficient inference. The trade-off between the quality of discretization/inference and the computational cost of inference with decision diagrams is explored by contrasting both metrics on a Pareto front. Through empirical evaluation across a range of causal and non-causal Bayesian networks, we investigate the impact of different discretization methods on this trade-off. We corroborate the significantly improved scalability of using decision diagrams for inference as opposed to traditional inference methods and extend this finding to discretized continuous networks. Coupled with insights on the accuracy-compute cost trade-off, we advocate for discretization as a viable method for Bayesian network inference on continuous networks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v246-vonk24a, title = {Balancing Computational Cost and Accuracy in Inference of Continuous Bayesian Networks}, author = {Vonk, Maarten C. and Brand, Sebastiaan and Malekovic, Ninoslav and B{\"a}ck, Thomas and Laarman, Alfons and Kononova, Anna V.}, booktitle = {Proceedings of The 12th International Conference on Probabilistic Graphical Models}, pages = {361--381}, year = {2024}, editor = {Kwisthout, Johan and Renooij, Silja}, volume = {246}, series = {Proceedings of Machine Learning Research}, month = {11--13 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v246/main/assets/vonk24a/vonk24a.pdf}, url = {https://proceedings.mlr.press/v246/vonk24a.html}, abstract = {Bayesian networks allow a parsimonious encoding of joint probability distributions via directed acyclic graphs. While discrete Bayesian network inference is well-established, conducting inference on continuous Bayesian networks often requires discretization. In this paper, continuous Bayesian networks are subjected to various supervised and unsupervised discretization methods. Subsequently, the discretized Bayesian networks are encoded into decision diagrams, facilitating efficient inference. The trade-off between the quality of discretization/inference and the computational cost of inference with decision diagrams is explored by contrasting both metrics on a Pareto front. Through empirical evaluation across a range of causal and non-causal Bayesian networks, we investigate the impact of different discretization methods on this trade-off. We corroborate the significantly improved scalability of using decision diagrams for inference as opposed to traditional inference methods and extend this finding to discretized continuous networks. Coupled with insights on the accuracy-compute cost trade-off, we advocate for discretization as a viable method for Bayesian network inference on continuous networks.} }
Endnote
%0 Conference Paper %T Balancing Computational Cost and Accuracy in Inference of Continuous Bayesian Networks %A Maarten C. Vonk %A Sebastiaan Brand %A Ninoslav Malekovic %A Thomas Bäck %A Alfons Laarman %A Anna V. Kononova %B Proceedings of The 12th International Conference on Probabilistic Graphical Models %C Proceedings of Machine Learning Research %D 2024 %E Johan Kwisthout %E Silja Renooij %F pmlr-v246-vonk24a %I PMLR %P 361--381 %U https://proceedings.mlr.press/v246/vonk24a.html %V 246 %X Bayesian networks allow a parsimonious encoding of joint probability distributions via directed acyclic graphs. While discrete Bayesian network inference is well-established, conducting inference on continuous Bayesian networks often requires discretization. In this paper, continuous Bayesian networks are subjected to various supervised and unsupervised discretization methods. Subsequently, the discretized Bayesian networks are encoded into decision diagrams, facilitating efficient inference. The trade-off between the quality of discretization/inference and the computational cost of inference with decision diagrams is explored by contrasting both metrics on a Pareto front. Through empirical evaluation across a range of causal and non-causal Bayesian networks, we investigate the impact of different discretization methods on this trade-off. We corroborate the significantly improved scalability of using decision diagrams for inference as opposed to traditional inference methods and extend this finding to discretized continuous networks. Coupled with insights on the accuracy-compute cost trade-off, we advocate for discretization as a viable method for Bayesian network inference on continuous networks.
APA
Vonk, M.C., Brand, S., Malekovic, N., Bäck, T., Laarman, A. & Kononova, A.V.. (2024). Balancing Computational Cost and Accuracy in Inference of Continuous Bayesian Networks. Proceedings of The 12th International Conference on Probabilistic Graphical Models, in Proceedings of Machine Learning Research 246:361-381 Available from https://proceedings.mlr.press/v246/vonk24a.html.

Related Material