TabSOM: A tabular-to-image encoding method based on self-organizing maps

In the era of data-driven decision-making, structured tabular data remains the backbone of many critical applications in healthcare, finance, and industry. Despite the success of deep learning models in unstructured data like images and text, their application to tabular data has faced significant hurdles. Traditional models such as gradient boosting machines and random forests excel at handling tabular data but lack transparency. Conversely, deep neural networks often struggle to capture complex feature interactions and rely on feature engineering or embedding techniques that obscure interpretability.

Recent advances have explored converting tabular data into image representations to leverage powerful convolutional neural networks (CNNs) and vision transformers (ViTs). These methods typically employ dimensionality reduction techniques like t-SNE, UMAP, or PCA to embed features into a 2D space, then generate images where spatial proximity reflects feature similarity. While effective in some cases, these approaches mainly encode marginal feature distributions, neglecting the relationships among features that are crucial for understanding the data structure and improving model robustness.

This paper introduces TabSOM, a novel encoding framework that harnesses the topology-preserving properties of self-organizing maps (SOMs). Unlike conventional embeddings, SOMs maintain the topological relationships of the original feature space, enabling a more meaningful spatial layout. The authors develop a collision-free feature placement strategy using the Hungarian algorithm, ensuring each feature occupies a distinct position on the image canvas based on its component plane. They further construct a feature relationship graph derived from the SOM's prototype geometry, capturing pairwise interactions that are embedded as spatial connections.

To enrich the representation, TabSOM employs multi-scale Gaussian kernels to render node channels at different bandwidths, balancing local detail and regional coherence. An additional relational edge channel encodes feature interactions, providing a comprehensive multi-channel image input for CNN classifiers. Extensive experiments across twelve public binary classification datasets demonstrate that TabSOM consistently outperforms existing methods in predictive accuracy, achieving the highest or second-highest AUC-ROC scores with minimal variance, indicating high stability.

Beyond performance, TabSOM emphasizes interpretability. The framework introduces two SOM-based tools: class-separation importance scores, which quantify how features differentiate classes, and prototype-inspired partial dependence plots, which visualize the influence of features within the learned topological structure. These explanations align well with established importance measures like SHAP and feature importance from tree models, validating their reliability.

Overall, TabSOM bridges the gap between high predictive performance and interpretability in tabular data modeling. By integrating topological preservation, feature relationship encoding, and multi-scale spatial rendering, it offers a robust, explainable, and scalable solution. Its potential impact spans healthcare diagnostics, financial risk assessment, and industrial process monitoring, where understanding model decisions is as vital as their accuracy. Future work aims to extend the framework to multi-class tasks, incorporate dynamic graph learning, and optimize computational efficiency, further broadening its applicability in real-world scenarios.