This paper reports the results of a project using neural networks and data supplied by the IT department of Columbus State University in Georgia to predict which students will drop out after their first year of college. The author began with two-layer networks, that is, networks without any hidden layers. When data from the first two semesters were provided, these networks were able to predict whether a given student would drop out with an accuracy of 75 percent. However, accuracy fell to 10 percent when the second semester data was not used. To get better results without second semester data, the author experimented with networks with a single hidden layer. Using a feed-forward cascade three-layer network increased accuracy to 69 percent. (The confusion matrix for the most accurate of these networks is not reproduced in the paper, even though the author says it is. Rather, the matrix for a less accurate network is given. However, the results are described in the body of the paper.)
The neural network was then used to extract weights for the input data categories, thus removing data that was not found relevant to the outcome. The items that were assigned the highest weight included number of fall semester courses, fall grade point average, selected minor, aggregate score on the Columbus State entrance test, and highest education level attained by both parents.
Although the author does not draw any conclusions as to how an advisor might make use of the results, readers can draw their own conclusions. It would be interesting to compare the results with those obtained from a decision tree generating system.