Avaliação de medidas de rede complexas para classificação de dados via caracterização de importância

Abstract

The use of complex networks has gained attention in various domains including data classification. Because they essentially consider physical characteristics of data, traditional classifiers have difficulty classifying new data based on its structural or topological pattern, as do classification techniques based on complex networks like importance characterization or pattern conformity. This work consists of investigating the classification model via importance characterization in order to empirically analyze other measures of complex networks in comparison to the PageRank measure. The environment developed to carry out these tests and subsequently the analyses is made from a graph constructed using the k-nearest neighbors network, and the use of measures of complex networks capable of characterizing some type of importance between the vertices of the network. For this purpose, 4 measures of complex networks were selected, namely: degree, closeness, PageRank and average shortest path. For the experiments we used 2 artificial datasets with varying levels of noise and another 4 real-world datasets. The obatined results demonstrate that other network measures can have a competitive result compared to PageRank, and that some network measures can be very sensitive in relation to the network construction phase. Concisely, this work has potential to assist other works related to data classification via importance characterization or even those regarding the appropriate selection of complex network measures for classification tasks.