A HEURISTIC FOR OPTIMIZING THE SPATIAL LAYOUT OF ATTRIBUTES IN THE CONVERSION OF TABULAR DATA INTO IMAGES FOR CLASSIFICATION

This work investigates the optimization of the spatial layout of attributes in the conversion of tabular data into images for classification by convolutional neural networks. The proposal formulates the construction of the layout as a combinatorial optimization problem, in which more similar attributes are allocated in positions close to a two-dimensional grid. An objective function based on the similarity between attributes and the distance between their positions is used, solved by a local search heuristic. The method was evaluated in the public datasets Madelon and BreastCancer, compared to random and sequential arrays. The results indicated that the optimized layout produced lower structural cost and the best mean values of accuracy and f1-score in both sets, with more evident gains in Madelon. Such results suggest that the spatial organization of attributes is a relevant component of the tabular-image approach and a promising topic for integration of Operations Research and machine learning.