Banca de QUALIFICAÇÃO: GUSTAVO TORRES CUSTODIO

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : GUSTAVO TORRES CUSTODIO
DATE: 11/10/2024
TIME: 14:00
LOCAL: por participação remota
TITLE:

Clustering-based ensemble guidance

PAGES: 80
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUMMARY:

Classification problems are very common in Machine Learning. They include tasks ranging from spam detection to credit analysis. Often, simple classifiers are not sufficient to learn how to separate classes satisfactorily. A common way to solve these problems is to employ ensembles of classifiers, where multiple classifiers simultaneously determine the class of new data. Some studies use clustering combined with ensembles to separate the dataset into clusters, thereby simplifying the attribute space for each classifier. In this work, we propose CBEG (Clustering-Based Ensemble Guidance), a method that clusters data to train ensembles of classifiers. In CBEG, each classifier is trained using data from a different cluster. The clustering is performed in such a way that the distance between classes in each cluster is as high as possible. The idea is that each classifier will find it easier to learn the decision boundary between different classes for its cluster. CBEG also includes a fusion method for classifiers where the degree of membership to a cluster is used to determine the voting weight of each classifier. In this case, the idea is that if a data point has a higher membership to a cluster, the classifier of that cluster is better adapted to it and should have a higher voting weight. Experiments with multiple datasets were conducted, comparing the performance of CBEG with other ensemble methods such as Random Forests and XGBoost. In the experiments feature selection was performed, so that each individual classifier of the CBEG was trained with different features. We observed that when the number of attributes is reduced, the performance of CBEG tends to be less affected than the performance of other ensemble methods. This work also includes a systematic literature review analyzing how ensembles of classifiers can be improved using clustering methods and evolutionary algorithms. In the future, we plan to incorporate evolutionary algorithms into the work to improve the quality of the clustering performed and, consequently, the quality of the classification. We also plan to add base classifier selection methods for CBEG, as only SVMs were used in the current experiments.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1918407 - DEBORA MARIA ROSSI DE MEDEIROS
Membro Titular - Examinador(a) Interno ao Programa - 1934625 - JESUS PASCUAL MENA CHALCO
Membro Titular - Examinador(a) Externo à Instituição - ANA CAROLINA LORENA - ITA
Membro Suplente - Examinador(a) Externo ao Programa - 1762420 - KARLA VITTORI
Membro Suplente - Examinador(a) Externo à Instituição - MARCIO BASGALUPP - UNIFESP