In this paper we introduce a new approach to case-based classification, i.e., classification based soley on previously classified cases. Our approach is characterized by on-the-fly generation of concepts, done at the time of new instance presentation. Best concepts are formed that would classify the new instance as being from each of a set of competing classes. Concepts are represented as constraints that cases must satisfy. Concepts are rated according to an evaluation function that captures aspects of necessity and sufficiency. We define a case representation and case-space structure that make the process reasonably efficient.

Our approach has the advantage that it is not vulnerable to the problem of concept drift, which can arise when concepts formed in the past do not reflect recent experience. Since concepts are formed at classification time, we can justify classification decisions in terms of contstraints that a new instance satisfies, something most, purely case-based approaches can not do.