SCOD: A novel semi-supervised outlier detection framework

Nowadays, outlier detection has been widely used in various areas of data mining. For example, it could help to detect whether a credit card is fraudulent used. Due to its less time consuming and higher effectiveness, existing works usually focus on semi-supervised anomaly detection methods, which mainly utilize the distance between examples to detect outliers. However, such methods suffer low detection accuracy facing massive data with high density. To approach this problem, we propose a novel semi-supervised cluster-based outlier detection method (SCOD) which combines density-based method and clustering. By doing so, we could not only distinguish various behaviors in huge size of data, but also can detect anomalies with dense distance-based neighbors. In this way, the detection accuracy can be significantly improved. In addition, we use real-world datasets to evaluate our approach and the evaluation confirms the precision of SCOD against other state-of-art approaches.