Look-up in Google Scholar
Title: Network-based high-level classification using betweenness centrality and attribute-attribute interaction
Other Titles: Classificação de alto nível baseada em redes usando betweenness centrality e interação atributo-atributo; Clasificación de alto nivel basada en redes utilizando centralidad de intermediación e interacción atributo-atributo
Advisor(s): Liang, Zhao
OCDE field: https://purl.org/pe-repo/ocde/ford#1.02.00
Issue Date: 2021
Institution: Universidade de São Paulo
Abstract: The democratization of technology caused by the internet, cloud technology, social media increase dramatically the quantity of data collected. Machine learning is an artificial intelligence field that produces valuable information from data. Supervised learning is a type of machine learning that focuses on using labeled data to learn to predict the label of future data. In this area, the High-level classification algorithms use the structure of the relation between the data for classification instead of physical attributes like distance. Complex networks are a data structure that provides metrics to evaluate the data as a system. They provide measures to the connectivity, communicability, or sparseness of the data interaction. In this study, we explore complex network properties and attributeattribute interaction to develop new high-level classification techniques. Firstly, we exploit a mixed metric betweenness centrality that captures global and local characteristics for classification. This method reduces the number of metrics evaluated and performs an improvement compared to other high-level algorithms. Then, we explore a new complex network building methodology to capture structural measures using attribute-attribute interaction. This interaction builds and evaluates each attribute independently and combining them using an optimized weighted equation. Finally, we analyze the results obtained by these metrics in synthetic and real datasets and compare them to other classical low-level and high-level algorithms. The proposed techniques present some promising characteristics as the reduction of metrics used for classification, resilience in front of non-normalized data, and a new network evaluation metric derived from the building methodology.

La accesibilidad tecnológica gracias a internet, la computación en la nube y las redes sociales ha generado un aumento significativo en la cantidad de datos disponibles. El aprendizaje automático, una rama de la inteligencia artificial, extrae conocimientos valiosos de estos datos. En particular, el aprendizaje supervisado utiliza datos etiquetados para predecir etiquetas futuras. En este contexto, los algoritmos de clasificación de alto nivel se basan en la estructura y relaciones entre los datos en lugar de características físicas como la distancia. Las redes complejas, estructuras de datos que evalúan la conectividad y dispersión de la información, son fundamentales. Este estudio propone nuevas técnicas de clasificación de alto nivel: primero, usando la centralidad de intermediación para capturar características de red, reduciendo el número de métricas evaluadas y mejorando la precisión. Luego, empleando una metodología innovadora de construcción de redes complejas basada en interacciones atributo-atributo, combinando atributos de manera ponderada. Los resultados, analizados en conjuntos de datos sintéticos y reales, muestran promesas como la reducción de métricas, la adaptabilidad a datos no normalizados y una nueva métrica de evaluación de red.
DOI identifier: https://doi.org/10.11606/D.59.2021.tde-16112023-144830
Discipline: Computación Aplicada
Grade or title grantor: Universidade de São Paulo. Faculdade de Filosofia, Ciências e Letras de Ribeirão Preto
Grade or title: Magíster en Ciencias
Juror: Liang, Zhao; Guimarães Carneiro, Murillo; Bertini Júnior, João Roberto
Register date: 24-Apr-2024

Files in This Item:
File Description SizeFormat 
VilcaZunigaEW.pdf
  Restricted Access
Disertación1.92 MBAdobe PDFView/Open Request a copy
Autorizacion.pdf
  Restricted Access
Autorización del registro186.76 kBAdobe PDFView/Open Request a copy


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.