A0985
Title: Image generator for tabular data based on non-Euclidean metrics for CNN-based classification
Authors: Han-Ming Wu - National Chengchi University (Taiwan) [presenting]
Abstract: Tabular data is widely used in statistical analysis and machine learning across fields like finance, biomedicine, and environmental sciences. However, traditional methods often struggle with high-dimensional and nonlinear feature relationships. Deep learning models, particularly convolutional neural networks (CNNs), excel at automatic feature extraction but are typically designed for image inputs. To bridge this gap, the purpose is to extend the image generator for tabular data (IGTD) framework, which transforms tabular data into images suitable for CNN-based classification. While the original IGTD uses Euclidean distance, alternative non-Euclidean metrics are incorporated, including one minus correlation, Geodesic distance, Jensen-Shannon distance, and Wasserstein distance, to better capture complex feature structures. A series of experiments on both simulated and real-world datasets systematically compare these distance metrics based on classification accuracy and the structural fidelity of the generated images. Results show that non-Euclidean metrics significantly enhance CNN classification performance on tabular data by providing a more accurate representation of feature relationships. This approach expands the utility of CNNs for structured, high-dimensional data, offering a flexible and interpretable framework applicable across various scientific and industrial domains.
