A Novel Outlier Detection Method for Multivariate Data
Yahya Almardeny,Noureddine Boujnah,Frances Cleary
DOI: https://doi.org/10.1109/tkde.2020.3036524
IF: 9.235
2020-01-01
IEEE Transactions on Knowledge and Data Engineering
Abstract:Detecting anomalous objects from given data has a broad range of real-world applications. Although there is a rich number of outlier detection algorithms, most of them involve hidden assumptions and restrictions. This paper proposes a novel, yet effective outlier learning algorithm that is based on decomposing the full attributes space into different combinations of subspaces, in which the 3D-vectors, representing the data points per 3D-subspace, are rotated about the geometric median, using Rodrigues rotation formula, to construct the overall outlying score. The proposed approach is parameter-free, requires no distribution assumptions and easy to implement. Extensive experimental study and comparison are conducted on both synthetic and real-world datasets with six popular outlier detection algorithms, each from different category. The comparison is evaluated based on the precision @s, average precision, rank power, AUC ROC and time complexity metrics. The results show that the performance of the proposed method is competitive and promising.
computer science, information systems, artificial intelligence,engineering, electrical & electronic