Prometheus : Directly Learning Acyclic Directed Graph Structures for Sum-Product Networks

Priyank Jaini, Amur Ghose, Pascal Poupart
; Proceedings of the Ninth International Conference on Probabilistic Graphical Models, PMLR 72:181-192, 2018.

Abstract

In this paper, we present Prometheus, a graph partitioning based algorithm that creates multiple variable decompositions efficiently for learning Sum-Product Network structures across both continuous and discrete domains. Prometheus proceeds by creating multiple candidate decompositions that are represented compactly with an acyclic directed graph in which common parts of different decompositions are shared. It eliminates the correlation threshold hyperparameter often used in other structure learning techniques, allowing Prometheus to learn structures that are robust in low data regimes. Prometheus outperforms other structure learning techniques in 30 discrete and continuous domains. We also describe a sampling based approximation of Prometheus that scales to high-dimensional domains such as images.

Cite this Paper


BibTeX
@InProceedings{pmlr-v72-jaini18a, title = {Prometheus : Directly Learning Acyclic Directed Graph Structures for Sum-Product Networks}, author = {Jaini, Priyank and Ghose, Amur and Poupart, Pascal}, booktitle = {Proceedings of the Ninth International Conference on Probabilistic Graphical Models}, pages = {181--192}, year = {2018}, editor = {Václav Kratochvíl and Milan Studený}, volume = {72}, series = {Proceedings of Machine Learning Research}, address = {Prague, Czech Republic}, month = {11--14 Sep}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v72/jaini18a/jaini18a.pdf}, url = {http://proceedings.mlr.press/v72/jaini18a.html}, abstract = {In this paper, we present Prometheus, a graph partitioning based algorithm that creates multiple variable decompositions efficiently for learning Sum-Product Network structures across both continuous and discrete domains. Prometheus proceeds by creating multiple candidate decompositions that are represented compactly with an acyclic directed graph in which common parts of different decompositions are shared. It eliminates the correlation threshold hyperparameter often used in other structure learning techniques, allowing Prometheus to learn structures that are robust in low data regimes. Prometheus outperforms other structure learning techniques in 30 discrete and continuous domains. We also describe a sampling based approximation of Prometheus that scales to high-dimensional domains such as images.} }
Endnote
%0 Conference Paper %T Prometheus : Directly Learning Acyclic Directed Graph Structures for Sum-Product Networks %A Priyank Jaini %A Amur Ghose %A Pascal Poupart %B Proceedings of the Ninth International Conference on Probabilistic Graphical Models %C Proceedings of Machine Learning Research %D 2018 %E Václav Kratochvíl %E Milan Studený %F pmlr-v72-jaini18a %I PMLR %J Proceedings of Machine Learning Research %P 181--192 %U http://proceedings.mlr.press %V 72 %W PMLR %X In this paper, we present Prometheus, a graph partitioning based algorithm that creates multiple variable decompositions efficiently for learning Sum-Product Network structures across both continuous and discrete domains. Prometheus proceeds by creating multiple candidate decompositions that are represented compactly with an acyclic directed graph in which common parts of different decompositions are shared. It eliminates the correlation threshold hyperparameter often used in other structure learning techniques, allowing Prometheus to learn structures that are robust in low data regimes. Prometheus outperforms other structure learning techniques in 30 discrete and continuous domains. We also describe a sampling based approximation of Prometheus that scales to high-dimensional domains such as images.
APA
Jaini, P., Ghose, A. & Poupart, P.. (2018). Prometheus : Directly Learning Acyclic Directed Graph Structures for Sum-Product Networks. Proceedings of the Ninth International Conference on Probabilistic Graphical Models, in PMLR 72:181-192

Related Material