Complex Network-Based Data Classification Using Minimum Spanning Tree Metric and Optimization

Autor(es)

Saire, Josimar Chire ; Zhao, Liang ; IEEE

• Tipo de documento: Artigo Científico
• Fonte: 2023 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS, IJCNN; v. N/A, p. 7-pg., 2023-01-01.

Resumo

Generally speaking, data classification techniques divide the data or feature space into sub-spaces, where each one corresponds to a class. For this purpose, physical features, such as distance and distributions, are considered. Such an approach has difficulty to characterize complex patterns embedded in the training and testing data sets. On the other hand, complex networks are a powerful tool to capture internal relationships and class structures to realize High-Level Classification. Several complex network-based classification techniques have been proposed. The main advantage of high-level classification is its ability to classify data according to the pattern formation. In this work, we present a new network-based classification technique considering a unique measure extracted from the Minimum Spanning Tree (MST) to characterize the data pattern represented by the network constructed for each class. Promising numerical results have been obtained in comparison to classic high-level classification techniques. Still in this paper, we apply the proposed technique to chest X-ray image classification for COVID-19 diagnosis and snoring signal classification. Moreover, the proposed model presents the following distinguished features in comparison to the previous high-level classification techniques: 1) Since only one network measure is used to characterize data pattern, no wight parameters among network measures is need to be determined. Consequently, the model is largely simplified and, at the same time, can obtain better classification result; 2) an optimization approach is introduced to construct more suitable network for each class of the training data; 3) the network measure based on MST is more sensitive to the insertion of new nodes. This is an important feature because each time only one testing data sample is inserted in each class network for classification. (AU)

• Processo FAPESP: 19/07665-4 - Centro de Inteligência Artificial
• Beneficiário: Fabio Gagliardi Cozman
• Modalidade de apoio: Auxílio à Pesquisa - Programa eScience e Data Science - Centros de Pesquisa em Engenharia