Research On Anomaly Detection Based On Ensemble Learning Algorithms

Anomaly detection is an important problem not only in many reasearch fields but in diverse application domains.Anomaly detection techniques have been extensively used in areas such as network intrusion detection,credit card fraud detection and so on.In most application scenarios,unevenly distributed importance of anomalies have been long existent.How to capture anomaly importance and derive importance sensitive anomaly detection model has become an interesting research problem.Another prob-lem we tried to solve is how to improve isolation forest to better handle complicated anomalies.There exists many state-of-art anoamly detection methods.However,a majority of model-based anomaly detection methods take the form of singly elaborated model which may suffer from poor generalization ability.As a hot subject in machine learning,ensemble learning is shown to have better generalization ability.What's more,in many applications,ensemble is more accurate than single-model.Under the aforementioned background,we discuss how to apply ensemble method to anomaly detection effctively.Our major contributions are as follows:? We proposed a gradient boosting based anomaly detection method with customiz-able importance indicator.By deriving customizable weighted loss function,we tailored the gradient boosting method for anomaly detection.? We proposed importance sensitive weighted and balanced random forests for anomaly detection.By mapping the problem of unevenly distributed importance to the im-balanced classification problem,we extend weighted random forest and balanced random forest to better handle anomaly detecton problems.? We developed anomaly sensitive splitting criteria for isolation forest model.With this new criteria,our model beats original isolation forest on benchmark anomaly detection tasks.