Data‐Driven Classification Using Boundary Observations

Classification is often a critical task for business managers in their decision‐making processes. It is generally more difficult for a classification scheme to produce accurate results when the input domains of the various output classes coincide, to some degree, with one another. In an attempt to address this issue, this article discusses a data‐driven algorithm that identifies the region of coincidence, or overlap, for two‐group classification problems by empirically determining the convex boundary for each group. The results are extendable to multigroup classification. The class membership of a new observation is then determined by its relative position with respect to each of these boundaries. Due to minimal data storage requirements, this boundary‐point classification technique can adapt to changing conditions far more easily than other approaches. Test results demonstrate that the new classification technique has similar performance to a back‐propagation neural network under static conditions and significantly outperforms a back‐propagation neural network under dynamic conditions.