Comparative study of public domain supervised machine-learning accuracy on the UCI database

This paper surveys public domain supervised learning algorithms and performs accuracy (error rate) analysis of their classification performance on unseen instances for twenty-nine of the University California Irvine machine learning datasets. The learning algorithms represent three types of classifier: decision trees, neural networks and rule-based classifiers. The study performs data analysis and examines the effect of irrelevant attributes to explain the performance characteristics of the learning algorithms. The survey concludes with some general recommendations about the selection of public domain machine learning algorithms relative to the properties of the data examined.