Is Feature Selection Essential for ANN Modeling?

Abstract

In modeling approaches, artificial neural networks (ANNs) have a special place to address the nonlinear phenomena or curved manifold. Often one or other feature selection approach is used prior to ANN to feed the input variables for its models. The function of ‘selected’ versus ‘arbitrary’ features on the outcome of ANN models is investigated with a variety of objectively selected and arbitrarily chosen variables from chemical databases namely thiazolidinones, anilinoquinolines and piperazinoquinolines. For each database, its biological activity is considered as the dependent variable and the molecular descriptors from DRAGON software are used as explanatory variables. The selection sets are obtained from feature selection approaches namely, combinatorial protocol in multiple linear regression, stepwise regression and genetic algorithm. Apart from these, a large number of arbitrary sets have been created by randomly picking the descriptors from corresponding databases. The features of all sets have shown a variety of inter- and intra- set diversities. A three-layer back propagation ANN with Levenberg-Marquardt optimization algorithm has been used for modeling the phenomena. Regardless of the origin of the feature sets, the ANN models from a very large number of sets have well explained the activity and qualified themselves to be predictive models. Also, no specific pattern is apparent between the quality of ANN model and the origin of its input feature set. Since these results are unusual, the study is extended to a few more databases. All the results emphasized the innate ability of ANN in developing complex network of relations among features to estimate the target variable. This has prompted us to suggest that prior feature selection is not essential for ANN and it is a desirable option for meaningful outputs in terms of the rationale behind the inputs.