Minimal Rare-Pattern-Based Outlier Detection Method for Data Streams by Considering Anti-monotonic Constraints

Abstract

In the collected associated data streams, some potential outliers are often fixed with the normal data instances, thus, it is necessary to accurately detect the outliers to improve the reliability of the data streams. In real life, people are more concerned about whether some outliers existed in the small scale data instances that satisfy their constraints, rather than in the huge entire datasets. However, the existing association-based outlier detection methods were proposed to detect the outliers from the entire data streams, thus, the time consumption is very long. To content with the existence of the constraints, this paper proposes an efficient constrained minimal rare pattern-based outlier detection method for data streams, namely AMCMRP-Outlier, to process the succinct and convertible anti-monotonic constraints. In the pattern mining phase, the matrix structure is used to quickly mine the minimal rare patterns that satisfy the constraints, thus providing the pattern basis for the outlier detection. In the outlier detection phase, two deviation indices are defined to measure the deviation degree of each transaction, and then the transactions having large deviation degrees are determined as the outliers. Finally, extensive experiments on one synthetic dataset and two public datasets verify that the AMCMRP-Outlier method can accurately detect the outliers with less time cost.

This work was partly supported by National Natural Science Foundation of China (NSFC grant number: U1836116), and the project of Jiangsu provincial Six Talent Peaks (Grant number XYDXXJS-016).