DEXTER: A System that Experiments with Choices of Training Data Using Expert Knowledge in the Domain of DNA Hydration

摘要

In this paper, we describe a system, DEXTER, that uses knowledge to suggest inductive learning experiments in the domain of DNA hydration pattern prediction. These experiments vary the training data presented to a classifier learner. Such experiments are necessary in this domain, since, as in many other scientific domains, data are noisy, the relevance of particular attributes is not well established, and the number of training cases is limited. In each experiment, DEXTER chooses a set of training cases, attributes and classes to learn. To generate an experiment, it examines the results of previous experiments, and uses domain knowledge and domain independent heuristics to select and modify a previous experiment. For the domain expert interested in using the induced rules to understand data, DEXTER's explicit use of knowledge provides several advantages that other data selection techniques do not. In particular, the variation of classifiers induced in different experiments yields insights into the roles and interactions of particular attributes in determining hydration. In addition, many of the classifiers induced from DEXTER's choices of data are of accuracy greater than or equal to those induced using the entire set of available data or data chosen by several other techniques. This work is of theoretical and pragmatic importance to molecular biophysicists. The learned hydration predictors provide insights about factors influencing DNA hydration. Also, the hydration predictors could lead to a tool for automatically predicting water positions around DNA molecules for which crystallographic data are not available.