Advanced search
Start date
Betweenand

Inference and learning algorithms for probabilistic logic programming

Grant number: 16/18841-0
Support type:Research Grants - Research Partnership for Technological Innovation - PITE
Duration: August 01, 2017 - July 31, 2019
Field of knowledge:Physical Sciences and Mathematics - Computer Science
Cooperation agreement: IBM Brasil
Principal Investigator:Fabio Gagliardi Cozman
Grantee:Fabio Gagliardi Cozman
Home Institution: Escola Politécnica (EP). Universidade de São Paulo (USP). São Paulo, SP, Brazil
Company: IBM Brasil - Indústria, Máquinas e Serviços Ltda
Assoc. researchers:Denis Deratani Mauá
Associated scholarship(s):17/19007-6 - Inference and learning algorithms for probabilistic logic programming, BP.MS

Abstract

The goal of this project is to develop inference and learning techniques for probabilistic logic programs, with an eye on the scalable automatic induction of probabilistic rules from large knowledge bases. Such techniques have applications in information search and retrieval, automated diagnosis, decision and recommendation systems - applications that benefit from large and accurate knowledge bases. We intend to work in two directions: (1) to study the theoretical properties of probabilistic logic programs, as there are still several open questions concerning their complexity; and, more importantly, (2) to develop better inference and learning algorithms for probabilistic logic programs, as there is largely an open territory when it comes to rule learning from large-scale datasets. Concerning the theoretical study, the PI and Associate Researcher have already investigated the semantics and complexity of such programs, and we intend to do further study of non-stratified and disjunctive programs. Concerning algorithmic development, we intend to implement our algorithms on top of the the ProbLog package, a freely available package that runs inference for probabilistic logic programs and that can learn probabilities from data. There are packages that can also learn the rules themselves (for instance, the ProbFOIL package, however their current computational performance is insufficient for processing large-scale datasets. The strategy for this research will be to enhance existing algorithms by importing some of the insights that have been recently employed in machine learning: namely, the focus on layered architectures with layer-wise learning. The project will be successful if: (1) it characterizes the semantics and the complexity of a large class of probabilistic logic programs (a class large enough to encode existing knowledge in the NELL base); (2) it enhances the ProbLog package so that it can run inference for large-scale knowledge bases (we intend to use facts in the NELL knowledge base as the main testing ground; success will be attained if we can learn new rules with better accuracy than rival methods). (AU)

Distribution map of accesses to this page
Click here to view the access summary to this page.