Prediction of melting points of organic compounds using extreme learning machines

Nonlinear regression methods such as artificial neural networks have been extensively used in prediction of properties of compounds from their molecular structure. Recently a new fast algorithm for training artificial neural networks known as the extreme learning machine was developed. In this paper we apply a simple ensemble of extreme learning machines to a large data set of melting points of organic molecules. The results obtained by extreme learning machines (cross-validated test set root-mean-square error = 45.4 K) are slightly better than those obtained using k nearest neighbor regression with genetic parameter optimization (cross-validated test set error = 46.2 K) and significantly better than those obtained by artificial neural networks trained using gradient descent (test set error = 49.3 K). The training of the extreme learning machine involves only linear regression resulting in faster training. Ensembling the extreme learning machines removes the dependence of results on initial random weights and improves the prediction. We also discuss the similarity between an ensemble of extreme learning machines and the random forest algorithm.