Multiple-case outlier detection in least-squares regression model using Quantum-inspired Evolutionary Algorithm

摘要

In ordinary statistical methods, multiple outliers in least-squares regression model are detected sequentially one after another, where smearing and masking effects give misleading results. If the potential multiple outliers can be detected simultaneously, smearing and masking effects can be avoided. Such multiple-case outlier detection is of combinatorial nature and 2V - 1 sets of possible outliers need to be tested, where N is the number of data points. This exhaustive search is practically impossible. Like other combinatorial applications, evolutionary algorithms may play a vital role in multiple-case outlier detection problem. In this paper, we have used Quantum-inspired Evolutionary Algorithm (QEA) for multiple-case outlier detection in least-squares regression model. An information-criterion-based fitness function incorporating extra penalty for number of potential outliers has been used for identifying the most appropriate set of potential outliers. Experimental results with four data sets from statistical literature show that the QEA effectively detects the most appropriate set of outliers.