Applying manufacturing batch techniques to customer fraud detection
by Zhihong Zhou
M.Phil. Industrial Engineering and Engineering Management
xi, 42 leaves : ill. ; 30 cm
In this research, we focused on the connection between the manufacturing industry and service industry. We mainly considered two critical issues, proposing robust discriminnant rules and applying Multi-way Principal Component Analysis (MPCA) to differentiate between normal and abnormal behaviors in fraud detection problems....[ Read more ]
In this research, we focused on the connection between the manufacturing industry and service industry. We mainly considered two critical issues, proposing robust discriminnant rules and applying Multi-way Principal Component Analysis (MPCA) to differentiate between normal and abnormal behaviors in fraud detection problems.
In producing low-volume, high-value products, batch processes always play an important role. Monitoring these batch-type processes is crucial to ensure safe operations and consistent high-quality products. Known techniques, such as MPCA, linear discriminant analysis (LDA), and the Mahalanobis-Taguchi system (MTS) can be applied in monitoring and controlling a batch process, where the model is based on training sample batches in a historical batch database. However, as the historical batch data are not always correctly classified between normal and abnormal batches, we have to modify the existing techniques that are more robust to the contaminated training samples to maintain the monitoring performance. Both a detailed comparison of the proposed robust approaches and a practical guideline on their implementation are provided.
For the second issue, we applied MPCA techniques in batch process to solve the fraud detection problem. In order to detect fraud behaviors as soon as possible before the customer information is complete, we prefer to apply Batch Library Method(BLM), used in manufacturing batch processes, to predict the incomplete part of customer information. Furthermore, we define the overall type I error as a normal customer calling information is finally classified as an abnormal one. The overall type I error is too large to be ignored in our application since large overall type I errors will make our detection unreliable. Consequently, we propose a procedure to revise the BLM to control the overall type I error. Finally, a real case study in a well-known telecom company demonstrates the performance of the Modified Batch Library Method (MBLM).