Customer return detection with features selection

We address the semiconductor industry problem of detecting microchips that escape production tests but are returned by customers as non-functional. This problem deals with analyzing high dimensional unbalanced databases collecting only a very small number of customer return samples. We show how to construct a model for effectively discriminating, based on wafer probe test data, potential customer returns from other good chips at the cost of a low overkill, where a model is a pair consisting of a selected set of wafer probe tests with minimal redundancy and a 1-class-SVM (Support Vector Machine) with optimal kernel parameters. We report about an experimentation on real data from EWS (Electronic Wafer Sort) test and customer returns showing the capability of predicting customer returns at cost of a relatively low overkill.