14 May 2021
14:00 Doctoral defense Fully distance
Theme
A Knowledge Engineering Method in Bayesian Networks to Reduce Sample Spaces
Student
Dalton Ieda Fazanaro
Advisor / Teacher
Helio Pedrini
Brief summary
Bayesian networks are statistical models capable of detecting the probabilistic causal relationships between variables with semantic significance. Given the evidence about the states of a subset of these variables in a problem modeled by Bayesian networks, the distribution of the probabilities to the others is immediately updated, a property that allows to assist the user in the acquisition of information under uncertainty, both in terms of diagnosis and of prediction. There is even continuous research in Artificial Intelligence on how to obtain intelligible explanations of these updates from the Bayesian networks. However, real problems are often characterized by high dimensional sample spaces, a fact that reduces the model's precision and / or explicability. Previous work on reducing these spaces focused exclusively on the preservation of precision metrics, an approach to which semantically dissimilar states were grouped, generating less significant Bayesian networks. The research addresses this unresolved problem of how to preserve both the semantic relevance of the sample spaces of the variables in a Bayesian network and the precision of this, while such spaces are reduced. As a case study, a Bayesian network model built on the database of the Police Department of the city of Chicago, USA, was used. Three methods for reducing sample spaces are proposed, through clustering based on frequency, semantic similarity or a hybrid version of both. The analysis of the results concluded that the hybrid version presents better performance, being worthy of studies for its improvement and consequent use for Knowledge Engineering in Bayesian Networks.
Examination Board
Headlines:
Hélio Pedrini IC / UNICAMP
Ricardo Sandes Ehlers ICMC / USP
Fabio Gagliardi Cozman POLI / USP
Ronaldo Dias IMECC / UNICAMP
Daiane Aparecida Zuanetti DES / UFSCar
Substitutes:
André Santanché IC / UNICAMP
Caio Lucidius Naberezny Azevedo IMECC / UNICAMP
Luís Gustavo Nonato ICMC / USP