Selectionist Random Vector Functional Link Networks

Abstract

Neural networks are very good, general classifiers, but they can take a very long time to train correctly. A solution to this problem is the Random Vector Functional Link network (RVFL-net), in which the input is combined randomly into a given number of 'enhancement nodes'. Given that these enhancement nodes are random combinations, one might expect that meaningless combinations will be among them. Therefore, this project is aimed at finding out whether it is useful to apply feature selection to the RVFL-net. To this end, 8 different feature selection methods are implemented and their performance is compared to that of an RVFL-net without any feature selection applied and to that of a 'regular' neural network, using a set of 22 datasets. It is found that for most datasets, applying feature selection leads to worse results than when no feature selection is applied, while for some datasets, there is a clear improvement in performance to be gained by applying feature selection. A general conclusion can thus not be drawn. Furthermore, an investigation into which feature selection method is useful in what situation is started. Here, too, a definitive conclusion cannot be drawn, but some interesting patterns are found.