Input significance analysis: feature ranking through synaptic weights manipulation for ANNS-based classifiers

Abstract

Due to the ANNs architecture, the ISA methods that can manipulate synaptic weights selected
are Connection Weights (CW) and Garson’s Algorithm (GA). The ANNs-based classifiers that
can provide such manipulation are Multi-Layer Perceptron (MLP) and Evolving Fuzzy Neural
Networks (EFuNNs). The goals for this work are firstly to identify which of the two
classifiers works best with the filtered/ranked data, secondly is to test the FR method by using
a selected dataset taken from the UCI Machine Learning Repository and in an online
environment and lastly to attest the FR results by using another selected dataset taken from
the same source and in the same environment. There are three groups of experiments
conducted to accomplish these goals. The results are promising when FR is applied, some
efficiency and accuracy are noticeable compared to the original data.

Keywords: artificial neural networks, input significance analysis; feature selection; feature
ranking; connection weights; Garson’s algorithm; multi-layer perceptron; evolving fuzzy
neural networks.