Research On Anomaly Automatic Detection Method Of Risk Management Log Supervision

Along with the development of online financial business, Risk management system as a financial business immune system, has become increasingly important. The risk management system takes on a large number of financial services, which requires a high stability of the risk management system. Therefore, a supervision platform is needed to monitor the stable operation of the risk management system. This paper mainly analyzes the business log of Baidu risk management system, takes the outreach and business process in the business log as the monitoring items, collects the statistical data of the business monitoring items in real time, uses the machine learning method to implement the discovery and detection and alarm of the business anomaly. This paper mainly carried out the following work:(1) Research and design of log supervision platform. Mainly includes the overall structure and logical design of the log supervision platform,from the functional point of view, designing supervision platform from different modules; from the logical design point of view, hierarchical designing log supervision platform. This supervision platform is used to manage the monitoring items of the risk management system, the person in charge of the monitoring items, the anomaly discovery of the monitoring items, the detection and the processing and feedback of anomaly.(2) Found anomalies based on improved K-Means clustering.Obtaining the monitoring items and the statistical data from the business log. First, the standardizing the data, and the K-Means clustering is used to find the abnormal cluster of the monitoring items. In order to help people analyze the abnormal instances in the abnormal cluster. Combining the feedback online, and further determines whether it is an abnormal instance.(3) Design and evaluation of anomaly detection model based on dynamic classification. First, based on the historical data of each monitoring item, constructing the training set and test set of the monitoring items. The classification model is constructed by the machine learning classification algorithm, and the classification model is compared and evaluated. Using the optimal classification model for exception detection of monitored items. Second, the monitoring items continue to generate new data, whenever the monitoring items to generate new exceptions, the exception will be added to the training set, re-training classification model to reduce the abnormal rate of false positives and false negative rate.