An anomaly detection method for individual service on web-based system by selection of dummy variables in multiple regression

This paper addresses detecting anomalies of individual services from their total resource usage on web-based system. Because the total resource usage is a linear combination of the number of accesses to each service, multiple regression analysis can be applied to estimate a resource usage per an access to each service as regression coefficient. However, the regression coefficients differ from the resource usage per an access of the services, which is caused by unstable resource usage per an access. We propose a method based on a multiple correlation coefficient to identify anomaly time and anomaly services. The proposed method identifies anomaly time when the correlation coefficient is decreased. And the proposed method identifies the anomaly service by judging whether the correlation coefficient is increased or not after the selection of the dummy variable. The experimental result shows that the proposed method can identify all the anomaly time, and improves precision rate and recall rate of detecting anomaly services by 20% at least, respectively.