Within this paper, we analyse the nature of knowledge dis-covery in database. We conclude that it is similar to that of knowledge acquisition, yet unique in that it employs pre-existing data collected for reasons other than analysis. The post-hoc nature of KDD means that the database is often unfit for analysis using traditional machine-learning techniques. We present a methodology for KDD that attempts to over-come this problem. Knowledge elicitation techniques are employed to define the structure of an appropriate learning dataset and to relate this structure to the raw database. The raw database is then redescribed in terms of the new structure before machine learning tools are applied. We also present CASTLE, a software workbench designed to support this methodology, and illustrate it's usage upon a worked example drawn from the Sisyphus-I room allocation problem.
展开▼