A Novel SOM Ensemble Classification Technique for Limited Data

Abstract

Small datasets are common in research areas where measurement of data is expensive or difficult (e.g., healthcare). For small datasets, training of a data model and classification using the same remains a challenge. To address the issue of limited data training and classification, the authors propose to use a novel ensemble of self-organizing maps (SOMs). SOMs are shallow neural networks, and as a result, the number of parameters is not very high. This avoids over-fitting and over-parameterization of the network. For the SOM ensemble, a novel early stopping training technique is proposed that requires a small number of samples for training. In the ensemble, one SOM is used for every class. For benchmarking the proposed technique, its performance is compared with that of other state-of-the-art techniques like deep learning, CATBOOST decision tree approach, etc. For reproducibility of results, three popular and extensively benchmarked digit image datasets have been used (i.e., MNIST, USPS+, and MADB). The proposed technique generally outperforms all other techniques with which it is compared.