Full text
| |
| Author(s): |
Dalton Ieda Fazanaro
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2021-05-14 |
| Examining board members: |
Hélio Pedrini;
Daiane Aparecida Zuanetti;
Ricardo Sandes Ehlers;
Fabio Gagliardi Cozman;
Ronaldo Dias
|
| Advisor: | Hélio Pedrini |
| Abstract | |
Bayesian networks are a powerful tool for knowledge engineering. They provide a more comprehensive interpretation of the problem itself by identifying not only the dependence relationships between the variables of it but also the weight of these connections through conditional probability distributions defined over each variable. Furthermore, newly available information about a variable, or a set of them, is automatically propagated across the network to update those probability distributions according to that new evidence, allowing predictions about past or future system states to be made. To model a problem under a Bayesian network knowledge engineering framework, learning and evaluation are essential preliminary routines. Through the learning process, the network topological and parametric structures are revealed, either by a domain expert or an automated structure learning method. In the latter case, many options are readily available for such purpose, differing themselves in the statistical approach taken. The first part of this research developed a comparative analysis between some of the most well-known structure learning algorithms as a proof of concept for the methodology devised to assist in the definition of which method returns the structure that most faithfully models a given problem. Once the network is defined, the evaluation stage then focuses on finding means to improve the predictive accuracy of it. When learned from real-life problems, the Bayesian network usually comprises several semantically meaningful variables with large state spaces, which may induce a less informative model with low precision. The second part of this work explored the unsolved problem of how to reduce the size of those state spaces while conserving the semantic meaning of the respective variables as well as the predictive accuracy of the model, by formulating a state-space clustering strategy based on semantic similarity, an approach that constitutes the main contribution of the developed doctoral investigation. Practical executions and analytical discussions of both research parts showed promising results that merit further development in potential future work (AU) | |
| FAPESP's process: | 17/02073-6 - Network model for crime forecasting |
| Grantee: | Dalton Ieda Fazanaro |
| Support Opportunities: | Scholarships in Brazil - Doctorate |