Example-Based Robust DB-Outlier Detection for High Dimensional Data

Abstract

This paper presents a method of outlier detection to identify exceptional objects that match user intentions in high dimensional datasets. Outlier detection is a crucial element of many applications like financial analysis and fraud detection. Scholars have made numerous investigations, but the results show that current methods fail to directly discover outliers from high dimensional datasets due to the curse of dimensionality. Beyond that, many algorithms require several decisive parameters to be predefined. Such vital parameters are considerably difficult to determine without identifying datasets beforehand. To address these problems, we take an Example-Based approach and examine behaviors of projections of the outlier examples in a dataset. An example-based approach is promising, since users are probably able to provide a few outlier examples to suggest what they want to detect. An important point is that the method should be robust, even if user-provided examples include noises or inconsistencies. Our proposed method is based on the notion of DB- (Distance-Based) Outliers. Experiments demonstrate that our proposed method is effective and efficient on both synthetic and real datasets and can tolerate noise examples.