Examining Statistical Process Control as a Method of Temporal Data Mining

A methodology based on statistical process control was examined for the data mining problem of anomaly detection. This methodology does not suffer from many of the limitations of other data mining techniques often proposed for anomaly detection. This research demonstrated statistical process control has sound theoretical backing, has a linear time complexity, is accurate in classifying anomalies, and is able to identify novel information. Furthermore, it was shown that the contemporaneous use of numerous univariate statistical process control charts can address the prevalent problem of class imbalance. This research found that statistical process control based techniques are an effective method of temporal anomaly detection. Statistical process control based algorithms were developed and tested on a food industry complaint database of both frequent and infrequent Poisson distributed events. In applications of statistical process control, Shewhart charts are often used in conjunction with either exponential weighted moving average or cumulative sum charts for detecting both large and small shifts in a process quickly. This research compared exponentially weighted moving average charts and cumulative sum charts, each used with Shewhart charts, for the purpose of data mining. For the trial database considered, the cumulative sum based method was preferred finding significantly more events of interest. Considerations for the design, setup, and maintenance of statistical process control based anomaly detection algorithms were also examined. The relationships between a confusion matrix, often used for binary classification, and typical measures of statistical process control performance, e.g. average run length and average time to signal, were derived. In addition, a general process of adapting statistical process control charts for a data mining task was developed