Prediction with Multiple, Multi-class Models and Dempster-Shafer Theory

Mathematical models are frequently developed to try to predict whether an individual has one of several conditions given a number of measured factors. Frequently, multiple models are developed for the same sets of data and, in general, some work better than others and usually one is selected as the primary model. But is there some way to use multiple models to aid prediction? In earlier work, we explored this question by considering the use of multiple predictive models as “evidence” and formulated a multi-model approach to prediction based on DempsterShafer’s Theory of Evidence. In that work, we looked strictly at models which predicted yes-or-no. In this paper, we take a similar approach, but consider the more general case of models which are used to predict one of several categories. Again, we build the Dempster-Shafer Models from the accuracy measures of the models. We then illustrate how the results might be used in conjunction with those models for prediction and present an evaluation of this approach and compare it to predictions by the individual