As obras disponibilizadas nesta Biblioteca Digital foram publicadas sob expressa autorização dos respectivos autores, em conformidade com a Lei 9610/98.
A consulta aos textos, permitida por seus respectivos autores, é livre, bem como a impressão de trechos ou de um exemplar completo exclusivamente para uso próprio. Não são permitidas a impressão e a reprodução de obras completas com qualquer outra finalidade que não o uso próprio de quem imprime.
A reprodução de pequenos trechos, na forma de citações em trabalhos de terceiros que não o próprio autor do texto consultado,é permitida, na medida justificada para a compreeensão da citação e mediante a informação, junto à citação, do nome do autor do texto original, bem como da fonte da pesquisa.
A violação de direitos autorais é passível de sanções civis e penais.
Título:NEURO-FUZZY BSP HIERARCHICAL SYSTEM FOR TIME FORECASTING AND FUZZY RULE EXTRACTION DOR DATA MINING APPLICATONS Instituição:PONTIFÍCIA UNIVERSIDADE CATÓLICA DO RIO DE JANEIRO - PUC-RIO Autor(es):ALBERTO IRIARTE LANAS
This dissertation investigates the use of a Neuro-Fuzzy
Hierarchical system for time series forecasting and fuzzy
rule extraction for Data Mining applications. The
objective of this work was to extend the Neuro-Fuzzy BSP
Hierarchical model for the classification of registers and
time series forecasting. The process of classification of
registers in the Data Mining context consists of
extracting association rules that best characterise,
through its accuracy and coverage measures, a certain
group of registers of database (DB). The time series
forecasting other common task in Data Mining, has a main
objective to foresee the behavior of a time series in the
instant t+k (k>=1).
The work consisted of 5 main stages: to elaborate a survey
of the main systems and the most common models in Data
Mining applications; to evaluate the performance of the
original NFHB system in Data Mining applicatons; to
develop an extension of the NFHB model dedicated to the
classification of registers in a DB; to develop a new
Neuro-Fuzzy Genetic hybrid model for the automatic
adjustment of the parameters of the system for time series
forecasting applicatons; and the case estudies.
The study of the area resulted in a survey of the main
Data Mining models. The most common methods used in Data
Mining application are presented such as: neural nets,
crisp and fuzzy decision trees, genetic algorithms,
statistics and neuro-fuzzy systems.
In the stage of evaluation of the original NFHB model, it
verified that besides the traditional learning of the
parameters, common to the neural nets and the neuro-fuzzy
systems, the model possesses the following
characteristics: learning of the structure; recursive
partitioning; larger number of inputs than usually found
on the neuro-fuzzy systems; rule with hierarchy; which are
characteristics adapted for Data Mining applications.
However the rule extraction process and attributes
selection are not appropriate for this type of
applications, as well as the excessive complexity of the
tuning of the model for time series forecasting
An extension of the original NFHB model was then proposed
for applicatons of classification of registers in the Data
Mining context, where the main objective in the extraction
of information in form of interpratable rules. It was
necessary to modify the attributes selection and the
original rule extraction process. The Takagi-Sugeno fuzzy
system of the original NFHB model supplies inadequate
rules, from the Data Mining point of view. The new NFHB
models, endowed with necessary modifications, showed good
performance in extracting valid fuzzy rules that describe
the information contained in the database. The evaluation
metrics, usually used to analyse crips rules (If x1 is
<14.3 and), as coverage and accuracy, were modified to be
applied to the evaluation of the fuzzy rules (If x1 is Low
and) extracted from the NFHB system after the learning
process. The amount and quality of the extracted rules are
important points of the systems dedicated for Data Mining
applicatons, where the target is to obtain the smallest
number of rules and of the best quality. In that sense,
the input selection strategies were implemented (Static
and Adaptive), using different evaluation measures as
Entropy and the jang algorithm.
A new genetic neuro-fuzzy hybrid model for time series
forecasting was created to solve the problem of the
excessive complexity of the model tuning, which comprises
more than 15 parameters. A new model wes proposed, a
genetic neuro-fuzzy hybrid, model capable to develop and
to obtain an appropriate set of parameters for the
forecasting of time series. The new hybrid, model capable
to develop and to obtain an appropriate set of parameters
for the forecasting of time series. The new hybrid model
presented good results with different types of series.
A tool based on the NFHB model was developed for
classification and forecasting applications. Th