Novelty Detection Technical Analysis And Research On One-class Normal Data

Novelty detection aims to detect the unexpected data that is not consistent with normal behaviors based on normal/target datasets. It has been applied to many different domains, especially security-critical systems where people would not expect abnormal conditions to happen such as credit card and jet engine. There are a lot of studies and a limited number of literature reviews on novelty detection, but there is a lack of experimental evaluation. This paper first presents a brief review of the existing literature on novelty detection, and then focuses on experimental analysis and performance (False Negative Rate, False Positive Rate and Area Under the Receiver Operating Characteristics Curve) comparison based on ten different benchmark datasets. Finally, it concludes that Gaussian Mixture method outperforms the other three ones, which all the five methods are selected as representative algorithms from different categories of novelty detection techniques. As the current algorithms can not effectively detect the changes of the dataset, the paper proposes a new method named level set based novelty detection method to apply to anomaly detection. It can change the movement of the decision boundary into the partial differential equations. The experiments have been carried out to form the initial boundary based on a normal datasets.