Table2Image: Interpretable Tabular Data Classification with Realistic Image Transformations

TL;DR

Table2Image presents a novel pipeline that converts tabular data into realistic image representations to enable CNN-based classification, addressing multicollinearity through VIF-informed initialization and expanding interpretability with DualSHAP, which fuses tabular and image explanations. The method achieves competitive results against GBDTs and contemporary neural models across OpenML-CC18 and TabZilla benchmarks while remaining lightweight. Key contributions include a realistic tabular-to-image mapping, VIF-based weight initialization, and a dual-modal interpretability framework that leverages SHAP and distribution alignment to enhance transparency. Practically, this work broadens the applicability of deep learning to tabular domains and lays groundwork for robust, multimodal AI with reliable explanations.

Abstract

Recent advancements in deep learning for tabular data have shown promise, but challenges remain in achieving interpretable and lightweight models. This paper introduces Table2Image, a novel framework that transforms tabular data into realistic and diverse image representations, enabling deep learning methods to achieve competitive classification performance. To address multicollinearity in tabular data, we propose a variance inflation factor (VIF) initialization, which enhances model stability and robustness by incorporating statistical feature relationships. Additionally, we present an interpretability framework that integrates insights from both the original tabular data and its transformed image representations, by leveraging Shapley additive explanations (SHAP) and methods to minimize distributional discrepancies. Experiments on benchmark datasets demonstrate the efficacy of our approach, achieving superior accuracy, area under the curve, and interpretability compared to recent leading deep learning models. Our lightweight method provides a scalable and reliable solution for tabular data classification.

Figures (10)

• Figure 1: Table2Image framework.
• Figure 2: VIF initialization.
• Figure 3: An overview of the DualSHAP framework.
• Figure 4: The result of our interpretability framework using the balance-scale dataset. The feature importance scores for one sample of the balance-scale dataset are plotted, comparing the scores obtained from $P$, $Q$ in Appendix \ref{['appendix:a']}, and SHAP.
• Figure 5: Pixel unshuffle operation.

Theorems & Definitions (1)

• Theorem 1.1