Indicators of input contributions: analysing the weight matrix

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 fields. This paper examines the use of the weight matrix of the trained neural network itself to determine which inputs are significant. A novel 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.