Modeling of Scholastic Aptitude Tests

摘要

We categorize nonlinearities and present some simple non-connectionist and connectionist methods of dealing with the different kinds of nonlinearities. In a case study we apply these methods to predict the result of a medical exam from a pre-university aptitude test. This seems to be a feasible task for a neural network model. A linear model fit to about one third of the 26,000 data points reaches a linear correlation coefficient of about 0.6 uniformly over the whole data set. A network trained on the residual error of the linear model does not manage to outperform the linear model. To simplify the problem, we reduce it to a pass/fail prediction of the exam. A three-step architecture, known as boosting, is especially suited for hard classification problems with many data points. Applying this technique to the data set, the final percentage of correct decision is not higher than the one of the linear baseline. We analyze the learning process and give reasons for the 'failure' of the connectionist model for this data.