A study of composite or hybrid classifiers for knowledge discovery

There are several well-known techniques for knowledge discovery such as neural networks, discriminant analysis, etc. This article compares the forecasting accuracy of five models (two parametric and three nonparametric) and proposes composite (or hybrid) methods based on combining the individual methods. Basically, a composite classifier makes use of ‘trained’ models whose ‘expertise’ is combined to obtain an optimal classifier. All the methods are illustrated and evaluated using marketing data related to the problem of deciding whether a supermarket should carry a new product. All methods except one performed considerably better on predicting the reject decision accurately than the accept decision. However, when hybrid methods were devised and tested, the performance on accept decisions improved dramatically, while the overall performance was almost the same as before. Moreover, no pure method dominated any of the four hybrid methods. We also performed a cost analysis to show that, depending upon the ratio of the costs of Type 1 and Type 2 errors, the hybrid methods could outperform all the pure methods in terms of total profit for the supermarket.