On the Relationship between Sum-Product Networks and Bayesian Networks

Han Zhao, Mazen Melibari, Pascal Poupart
Proceedings of the 32nd International Conference on Machine Learning, PMLR 37:116-124, 2015.

Abstract

In this paper, we establish some theoretical connections between Sum-Product Networks (SPNs) and Bayesian Networks (BNs). We prove that every SPN can be converted into a BN in linear time and space in terms of the network size. The key insight is to use Algebraic Decision Diagrams (ADDs) to compactly represent the local conditional probability distributions at each node in the resulting BN by exploiting context-specific independence (CSI). The generated BN has a simple directed bipartite graphical structure. We show that by applying the Variable Elimination algorithm (VE) to the generated BN with ADD representations, we can recover the original SPN where the SPN can be viewed as a history record or caching of the VE inference process. To help state the proof clearly, we introduce the notion of \em normal SPN and present a theoretical analysis of the consistency and decomposability properties. We conclude the paper with some discussion of the implications of the proof and establish a connection between the depth of an SPN and a lower bound of the tree-width of its corresponding BN.

Cite this Paper


BibTeX
@InProceedings{pmlr-v37-zhaoc15, title = {On the Relationship between Sum-Product Networks and Bayesian Networks}, author = {Zhao, Han and Melibari, Mazen and Poupart, Pascal}, booktitle = {Proceedings of the 32nd International Conference on Machine Learning}, pages = {116--124}, year = {2015}, editor = {Bach, Francis and Blei, David}, volume = {37}, series = {Proceedings of Machine Learning Research}, address = {Lille, France}, month = {07--09 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v37/zhaoc15.pdf}, url = {https://proceedings.mlr.press/v37/zhaoc15.html}, abstract = {In this paper, we establish some theoretical connections between Sum-Product Networks (SPNs) and Bayesian Networks (BNs). We prove that every SPN can be converted into a BN in linear time and space in terms of the network size. The key insight is to use Algebraic Decision Diagrams (ADDs) to compactly represent the local conditional probability distributions at each node in the resulting BN by exploiting context-specific independence (CSI). The generated BN has a simple directed bipartite graphical structure. We show that by applying the Variable Elimination algorithm (VE) to the generated BN with ADD representations, we can recover the original SPN where the SPN can be viewed as a history record or caching of the VE inference process. To help state the proof clearly, we introduce the notion of \em normal SPN and present a theoretical analysis of the consistency and decomposability properties. We conclude the paper with some discussion of the implications of the proof and establish a connection between the depth of an SPN and a lower bound of the tree-width of its corresponding BN.} }
Endnote
%0 Conference Paper %T On the Relationship between Sum-Product Networks and Bayesian Networks %A Han Zhao %A Mazen Melibari %A Pascal Poupart %B Proceedings of the 32nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2015 %E Francis Bach %E David Blei %F pmlr-v37-zhaoc15 %I PMLR %P 116--124 %U https://proceedings.mlr.press/v37/zhaoc15.html %V 37 %X In this paper, we establish some theoretical connections between Sum-Product Networks (SPNs) and Bayesian Networks (BNs). We prove that every SPN can be converted into a BN in linear time and space in terms of the network size. The key insight is to use Algebraic Decision Diagrams (ADDs) to compactly represent the local conditional probability distributions at each node in the resulting BN by exploiting context-specific independence (CSI). The generated BN has a simple directed bipartite graphical structure. We show that by applying the Variable Elimination algorithm (VE) to the generated BN with ADD representations, we can recover the original SPN where the SPN can be viewed as a history record or caching of the VE inference process. To help state the proof clearly, we introduce the notion of \em normal SPN and present a theoretical analysis of the consistency and decomposability properties. We conclude the paper with some discussion of the implications of the proof and establish a connection between the depth of an SPN and a lower bound of the tree-width of its corresponding BN.
RIS
TY - CPAPER TI - On the Relationship between Sum-Product Networks and Bayesian Networks AU - Han Zhao AU - Mazen Melibari AU - Pascal Poupart BT - Proceedings of the 32nd International Conference on Machine Learning DA - 2015/06/01 ED - Francis Bach ED - David Blei ID - pmlr-v37-zhaoc15 PB - PMLR DP - Proceedings of Machine Learning Research VL - 37 SP - 116 EP - 124 L1 - http://proceedings.mlr.press/v37/zhaoc15.pdf UR - https://proceedings.mlr.press/v37/zhaoc15.html AB - In this paper, we establish some theoretical connections between Sum-Product Networks (SPNs) and Bayesian Networks (BNs). We prove that every SPN can be converted into a BN in linear time and space in terms of the network size. The key insight is to use Algebraic Decision Diagrams (ADDs) to compactly represent the local conditional probability distributions at each node in the resulting BN by exploiting context-specific independence (CSI). The generated BN has a simple directed bipartite graphical structure. We show that by applying the Variable Elimination algorithm (VE) to the generated BN with ADD representations, we can recover the original SPN where the SPN can be viewed as a history record or caching of the VE inference process. To help state the proof clearly, we introduce the notion of \em normal SPN and present a theoretical analysis of the consistency and decomposability properties. We conclude the paper with some discussion of the implications of the proof and establish a connection between the depth of an SPN and a lower bound of the tree-width of its corresponding BN. ER -
APA
Zhao, H., Melibari, M. & Poupart, P.. (2015). On the Relationship between Sum-Product Networks and Bayesian Networks. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:116-124 Available from https://proceedings.mlr.press/v37/zhaoc15.html.

Related Material