Explainable Multi-agent Network for Multi-classification in Small Tabular Data

Deep learning has gained tremendous success in recent years in most data types, but still outperformed by classical machine learning on tabular datasets, the most common data type, especially smaller ones. An explainable deep learning model with accurate prediction power on tabular datasets can make it easier to use the advantages of such methods to allow more flexibility in architecture design to be applied on this type of data, the use of transfer learning on similar tasks, and build trust in model predictions. We present an explainable multi-agent network for multi-classification of small tabular datasets, that uses a discrimination network, an attention block, and a classification network. Our model has an accuracy close to the best results of OpenML benchmark on all datasets used with an average difference of 4.2%, and outperformed a similar deep learning approach.