Data mining of inputs: analysing magnitude and functional measures

The problem of data encoding and feature selection for training back-propagation neural networks is well known. The basic principles are to avoid encrypting the underlying structure of the data, and to avoid using irrelevant inputs. This is not easy in the real world, where we often receive data which has been processed by at least one previous user. The data may contain too many instances of some class, and too few instances of other classes. Real data sets often include many irrelevant or redundant input fields. This paper examines the use of weight matrix analysis techniques and functional measures using two real (and hence noisy) data sets. The first part of this paper examines the use of the weight matrix of the trained neural network itself to determine which inputs are significant. A new technique is introduced and compared with two other techniques from the literature. We present our experience and results on some satellite data augmented by a terrain model. The task was to predict the forest supra-type based on the available information. A brute force technique eliminating randomly selected inputs was used to validate our approach. The second part of this paper examines the use of measures to determine the functional contribution of inputs to outputs. Inputs which include minor but unique information to the network are more significant than inputs with higher magnitude contribution but providing redundant information, which is also provided by another input. A comparison is made to sensitivity analysis, where the sensitivity of outputs to input perturbation is used as a measure of the significance of inputs. This paper presents a novel functional analysis of the weight matrix based on a technique developed for determining the behavioral significance of hidden neurons. This is compared with the application of the same technique to the training and test data. Finally, a novel aggregation technique is introduced.