Learning of Tractable Probabilistic Models with Application to Multilabel Classification

• Grant number: 16/01055-1
• Support Opportunities: Regular Research Grants
• Duration: May 01, 2016 - April 30, 2018
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Denis Deratani Mauá
• Grantee: Denis Deratani Mauá
• Host Institution: Instituto de Matemática e Estatística (IME). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated researchers: Fabio Gagliardi Cozman

Abstract

Bayesian networks allow for the compactrepresentation of uncertain knowledge. There is strong evidence from computational complexity that performing inference in them takes timeexponential in the network treewidth, which measures the complexity of the model.Bounding the treewidth of the networks decreases their representational power. We can counter that decrease without increasing the complexity ofinference by using latent variables in the model, that is, variables which are not observed. However, current methods for learning low-treewidth Bayesian networks require fully observable variables. They also have problems scaling to large domains (hundreds or even thousands of variables) that appear in real world problems.This document describes a research proposal on methods for learning low-treewidth Bayesian networks from data in the presence of latentvariables, and scalable to large domains. These methods will be then be used to construct multilabel classifiers, that is, to learn functions that assigns relevant classes to objects. Unlike traditional (singlelabel) classification, class relevances (i.e., whether an object belongs to a particular class) are often highly correlated, which requires moresophisticates models.Another approach to building efficient multilabel classifiers is to adopt another class of probabilistic models with tractable inference. Sum-product networks are arithmetic circuits targeted at the representation of probabilistic models. Inference in a sum-product network takes linear time in its size.In this research we will also investigate the construction of multilabel classifiers based on sum-product networks, and compare it with Bayesiannetwork classifiers. (AU)

Scientific publications

(5)

(References retrieved automatically from Web of Science and SciELO through information on FAPESP grants and their corresponding numbers as mentioned in the publications by the authors)

MAUA, DENIS DERATANI; CONATY, DIARMAID; COZMAN, FABIO GAGLIARDI; POPPENHAEGER, KATJA; DE CAMPOS, CASSIO POLPO. Robustifying sum-product networks. INTERNATIONAL JOURNAL OF APPROXIMATE REASONING, v. 101, n. SI, p. 163-180, OCT 2018. (16/01055-1, 16/18841-0)

DE FARIA, FRANCISCO H. O. VIEIRA; GUSMAO, ARTHUR COLOMBINI; DE BONA, GLAUBER; MAUA, DENIS DERATANI; COZMAN, FABIO GAGLIARDI. Speeding up parameter and rule learning for acyclic probabilistic logic programs. INTERNATIONAL JOURNAL OF APPROXIMATE REASONING, v. 106, p. 32-50, MAR 2019. (17/19007-6, 16/18841-0, 16/25928-4, 16/01055-1, 15/21880-4)

COZMAN, FABIO GAGLIARDI; MAUA, DENIS DERATANI. On the Semantics and Complexity of Probabilistic Logic Programs. JOURNAL OF ARTIFICIAL INTELLIGENCE RESEARCH, v. 60, p. 221-262, 2017. (16/18841-0, 16/01055-1)

COZMAN, FABIO GAGLIARDI; MAUA, DENIS DERATANI. On the complexity of propositional and relational credal networks. INTERNATIONAL JOURNAL OF APPROXIMATE REASONING, v. 83, p. 298-319, APR 2017. (16/01055-1)

MAUA, DENIS DERATANI; COZMAN, FABIO GAGLIARDI. The effect of combination functions on the complexity of relational Bayesian networks. INTERNATIONAL JOURNAL OF APPROXIMATE REASONING, v. 85, p. 178-195, JUN 2017. (16/01055-1)