Classification methods and inductive learning rules: what we may learn from theory

Inductive learning methods allow the system designer to infer a model of the relevant phenomena of an unknown process by extracting information from experimental data. A wide range of inductive learning methods is nowadays available, potentially ensuring different levels of accuracy on different problem domains. In this critical review of theoretic results gained in the last decade, we address the problem of designing an inductive classification system with optimal accuracy when domain knowledge is limited and the number of available experiments is-possibly-small. By analyzing the formal properties of consistent learning methods and of accuracy estimators, we wish to convey to the reader the message that the common practice of aggressively pursuing error minimization with different training algorithms and classification families is unjustified