The aim of this research is to establish a coherent framework for data mining in the relational model. Observing that data mining depends on two partitions, the classifier and the estimator, this paper defines the classifier=estimator (CE) framework. The classifier indicates the target of the data mining investigation. The classifier may be diffcult to express from the relational instance or may involve an oracle beyond the extant data. The estimator is typically simply expressible using the relational instance. The degree to which the estimator refines the classifier partition can be used to measure how well the data instance matches the concept being investigated. The CE framework is shown to generalize a variety of data mining and database concepts, including rough sets, functional dependency, multivalued dependency, and association rules. Furthermore, the CE framework suggests a wider range of data mining questions. The CE framework is shown to naturally express qualitative and quantitative measures of the quality of approximation. Additionally, the CE framework allows a question to be posed at a number of different conceptual scopes from local to global interests.