A new time adjusting control limits chart for concept drift detection

Abstract

Time varying dynamic systems whose underlying changing distribution should be continuously monitored to track abnormal behaviors are one of the most recent challenges in many practical applications. When the data arrive in a continuous way, the target concept to be monitored may change accordingly causing a problem of concept drift. Thus, distinguishing between true alarms and changes due to the nonstationarity of the loading data is required. Traditional control charts cannot handle such processes since they do not use a change dependent procedure and they are not designed for concept drift problems. This article proposes the first two-stage time adjusting control chart for monitoring data stream processes with concept drift. Stage I updates the control limits each time an adjustment condition is satisfied based on an incremental linear combination of the historical and the new data. Stage II validates the shift detected in Stage I by testing whether the two subsamples around the drift belong to the same distribution. Experiments based on several drift situations and using different performance measures show that the proposed adaptive chart is more robust than the most recent competitive time varying charts existing in the literature. Moreover, it is more efficient in terms of shift recognition by reducing the false negative detections in several cases.