Designing Intelligent Expert Systems to Cope with Liars

To cope with the problem of input distortion by users of Web-based expert systems, we develop methods to distinguish liars from truth-tellers based on verifiable attributes, and redesign the expert systems to control the impact of input distortion. The four methods we propose are termed split tree, consolidated tree, value based split tree, and value based consolidated tree. They improve the performance of expert systems by improving accuracy or reduce misclassification cost. Numerical examples confirm that the most possible accurate recommendation is not always the most economical one. The recommendations based on minimizing misclassification costs are more moderate compared to that based on accuracy. In addition, the consolidated tree methods are more efficient than the split tree methods, since they do not always require the verification of attribute values.