A SOM-based method for feature selection

This paper presents a method, called Feature Competitive Algorithm (FCA), for feature selection, which is based on an unsupervised neural network, the self-organising map (SOM). The FCA, is capable of selecting the most important features describing target concepts from a given whole set of features via the unsupervised learning. The FCA is simple to implement and fast in feature selection as the learning can be done automatically and no need for training data. A quantitative measure, called average distance distortion ratio, is figured out to assess the quality of the selected feature set. An asymptotic optimal feature set can then be determined on the basis of the assessment. This addresses an open research issue in feature selection. This method has been applied to a real case, a software document collection consisting of a set of UNIX command manual pages. The results obtained from a retrieval experiment based on this collection demonstrated some very promising potential.