Logical second order models: Achieving synergy between computer power and human reason

• Authors: Gwangyong Gim;Thomas Whalen
• Venue: Information Sciences: an International Journal
• Year: 1999

This research investigates computer generated hybrid second order models of two numerically based approaches to risk classification: logistic regression and neural networks. If-Then rules to assess bankruptcy risk were generated by a rule induction algorithm in three ways: (1) using actual bankruptcy as the criterion, (2) using a criterion based on a Logit model of bankruptcy probability, and (3) using a criterion based on a neural model of membership in the fuzzy set of firms at risk for bankruptcy. Human experts then modified these If-Then rules. The outcome variable categories in the training sample are the outputs from the respective first order model. Best results on the test sample came from neural-based hybrid rules modified by the human expert; human modification improved neural-hybrid rules but degraded the performance of Logit-hybrid rules. Rules generated by induction from raw data were the worst, and their performance was not significantly affected by the human modification. Results indicate a powerful synergy between the robustness of neural nets, the perspicuity of If-Then rules, and the knowledge of the human expert. This new approach is expected to overcome two major drawbacks of the rule induction algorithm. Since the dependent variable is computed by the numerical system as a function of the inputs, the same set of input variables always is associated with the same output variable in the database given to the rule induction algorithm, eliminating the problem of contradictory inputs, and the smoothing effect of using the output of the numerical system should also improve generalizability, since this induction process will be less likely to be misled by anomalous cases in the raw data.