Computer systems that learn: an empirical study of the effect of noise on the performance of three classification methods

Classification learning systems are useful in many domain areas. One problem with the development of these systems is feature noise. Learning from examples classification methods from statistical pattern recognition, machine learning, and connectionist theory are applied to synthetic data sets possessing a known percentage of feature noise. Linear discriminant analysis, the C5.0 tree classification algorithm, and a backpropagation neural network tool are used as representative techniques from these three categories. K-fold cross validation is used to estimate the sensitivity of the true classification accuracy to level of feature noise present in the data sets. Results indicate that the backpropagation neural network outperforms both linear discriminant analysis and C5.0 tree classification when appreciable (10% or more of the cases) feature noise is present. These results are confirmed when the same type of empirical analysis is applied to a realworld data set previously analyzed and reported in the statistical and machine learning literature.