Generic Predictive Systems: An Empirical Evaluation Using the Learning Base System (LBS)

The predictive accuracy of expert systems is often evaluated in specific situations, i.e. with noise-free and domain-specific data, or with speciJic predictive protocols such as data fitting or resampling. Consequently the predictive tests depend on the situation and comparisons between models are dificult. This paper examines the general problem of prediction in the context of one particular system, the Bayesian classifier: The effects of varying components of the predictive task are determined on an ecological data set. The components of the predictive task varied were: Object of prediction, type of data, predictive protocol, decision system and measures of predictive accuracy. The results show the ways in which success or failure of prediction depends on the specification of the predictive paradigm.