TabSOM: A tabular-to-image encoding method based on self-organizing maps
TabSOM turns tabular rows into image-like inputs while preserving both fixed feature positions and pairwise feature links.
The method uses self-organizing maps to place each feature on a canvas and to derive a graph of relationships between features. Its images stack channels for feature values and spatial connections, aiming to give CNNs and vision transformers more than marginal feature values. In the authors’ benchmarks against twelve tabular-to-image methods, TabSOM ranked first or second on every tested binary-classification dataset and had the lowest variance. The paper also introduces SOM-based interpretability tools, with feature-importance results compared against Random Forest, XGBoost, and SHAP. ArXiv · AI/CL/LG's note
The method uses self-organizing maps to place each feature on a canvas and to derive a graph of relationships between features. Its images stack channels for feature values and spatial connections, aiming to give CNNs and vision transformers more than marginal feature values. In the authors’ benchmarks against twelve tabular-to-image methods, TabSOM ranked first or second on every tested binary-classification dataset and had the lowest variance. The paper also introduces SOM-based interpretability tools, with feature-importance results compared against Random Forest, XGBoost, and SHAP. ArXiv · AI/CL/LG's note
score 4