Banca de DEFESA: GRETA AUGAT ABIB

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : GRETA AUGAT ABIB
DATE: 17/06/2025
TIME: 14:00
LOCAL: meet.google.com/iwh-ygou-tqy
TITLE:

Feature Selection in Graph Attributed Data

PAGES: 75
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUMMARY:

Graphs serve as fundamental data structures to represent relationships and interactions between entities. Using graphs to represent data helps uncover complex relationships and patterns that could be missed in models that concentrate on individual data points. Attributed graphs enhance this capability by associating data features with vertices or edges of the graph. However, when addressing complex real-world problems, datasets often involve numerous features. Feature selection emerges as a critical technique that aims to identify a pertinent subset of features for specific tasks such as classification, prediction, or anomaly detection. However, the computational demands of feature selection are heightened by the size and complexity of these datasets. Furthermore, the domain of attributed graphs faces a deficiency in adequate feature selection methods, leading to suboptimal results in various data analysis tasks. This study addresses this challenge by framing the selection of features in attributed graph data as a graph similarity problem. We evaluated the applicability of our graph-based feature selection approach through two case studies based on Brazilian census data: one focused on identifying key census features associated with \coo emissions in Brazil; the other aimed to uncover socioeconomic determinants of Brazilian homicide rates. The findings underscore the effectiveness of the proposed approach, demonstrating its consistent ability to outperform traditional feature selection methods in multiple scenarios. For \coo emissions, the method achieved superior predictive accuracy by consistently achieving lower RMSE values compared to the baseline algorithms, while maintaining competitive performance in SAE, particularly beyond certain neighborhood size thresholds in KNN. In the case of homicide rates, the approach excelled by identifying relevant socio-economic features, achieving superior RMSE values and demonstrating strong competitiveness in SAE, especially for smaller neighborhood sizes. These outcomes, supported by statistical and qualitative analyses, highlight the robustness and adaptability of the proposed method. By delivering improved predictive accuracy and reliability across diverse and complex datasets, this work significantly advances feature selection methodologies for attributed graphs and their real-world applications.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1673092 - RONALDO CRISTIANO PRATI
Membro Titular - Examinador(a) Interno ao Programa - 1918407 - DEBORA MARIA ROSSI DE MEDEIROS
Membro Titular - Examinador(a) Externo ao Programa - 1762420 - KARLA VITTORI
Membro Titular - Examinador(a) Externo ao Programa - 1761107 - RICARDO SUYAMA
Membro Titular - Examinador(a) Externo à Instituição - FLÁVIA CRISTINA BERNARDINI - UFF
Membro Suplente - Examinador(a) Externo à Instituição - MARCOS GONCALVES QUILES - UNIFESP