Tabular Data Classification Method Based on Image Transformation and Capsule Networks

Due to the fundamental differences in structure between tabular data and image data, CNNs (Convolutional Neural Networks) are challenging to apply directly to the analysis and classification of tabular data. To address this issue, we propose an innovative method for classifying tabular data, combining tabular-to-image transformation techniques with Capsule Networks. First, we employ a consensus clustering method based on multiple subspace weighting to group features, ensuring the stability and consistency of the groupings. Next, we introduce a method to convert each data instance into a two-dimensional pixel image, using the ant colony algorithm to optimize pixel layout. Finally, the generated images are fed into an improved Capsule Network for classification prediction. Experimental results from 11 multidimensional tabular datasets demonstrate that the proposed method significantly improves classification accuracy compared to existing tabular data transformation methods. Furthermore, due to the smaller size of the generated images, the complexity of the model is reduced, which improves computational efficiency.