Scholarlink

This study investigates different graph construction heuristics applied to attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR) data, analyzing their topological properties using complex network metrics. In addition to this structural analysis, we also evaluate the data classification performance of each graph representation. We compare traditional approaches—such as co-occurrence (COOC), correlation (CORR), epsilon neighborhood (EPS), and k-nearest neighbors (kNN)—with visibility-based strategies (visibility graph - VG). These methods are applied to three biomedical datasets (COVID-19, Oral Cancer, and Diabetes), with their topological properties analyzed using complex network metrics. The topological analysis revealed that COOC and EPS generated structurally poor networks, due to fragmentation or overdensity. In contrast, kNN, CORR, and the two VG variants consistently produced sparse networks with high modularity, indicating a more informative representation. The classification results confirm that there is no universally superior method, with the best performance being dataset-dependent: kNN achieved the highest accuracy for COVID-19, CORR for Oral Cancer, and VG sim for Diabetes. These results collectively validate kNN, CORR, and visibility graphs approaches as effective and structurally expressive tools for spectral data, offering a robust alternative to traditional vector-based analysis in machine learning applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating Graph-Based Representations of ATR-FTIR Data Using Complex Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Barbara Cristina Gama,
  • Robinson Sabino-Silva,
  • Douglas Carvalho Caixeta,
  • Thulio Marquez Cunha,
  • Murillo Guimarães Carneiro

摘要

This study investigates different graph construction heuristics applied to attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR) data, analyzing their topological properties using complex network metrics. In addition to this structural analysis, we also evaluate the data classification performance of each graph representation. We compare traditional approaches—such as co-occurrence (COOC), correlation (CORR), epsilon neighborhood (EPS), and k-nearest neighbors (kNN)—with visibility-based strategies (visibility graph - VG). These methods are applied to three biomedical datasets (COVID-19, Oral Cancer, and Diabetes), with their topological properties analyzed using complex network metrics. The topological analysis revealed that COOC and EPS generated structurally poor networks, due to fragmentation or overdensity. In contrast, kNN, CORR, and the two VG variants consistently produced sparse networks with high modularity, indicating a more informative representation. The classification results confirm that there is no universally superior method, with the best performance being dataset-dependent: kNN achieved the highest accuracy for COVID-19, CORR for Oral Cancer, and VG sim for Diabetes. These results collectively validate kNN, CORR, and visibility graphs approaches as effective and structurally expressive tools for spectral data, offering a robust alternative to traditional vector-based analysis in machine learning applications.