Please use this identifier to cite or link to this item:
Scopus Web of Science® Altmetric
Type: Conference paper
Title: Anomaly detection via neighbourhood contrast
Author: Chen, B.
Ting, K.M.
Chin, T.J.
Citation: Lecture Notes in Artificial Intelligence, 2020 / Lauw, H.W., Wong, R.C.W., Ntoulas, A., Lim, E.P., Ng, S.K., Pan, S.J. (ed./s), vol.12085, pp.647-659
Publisher: Springer
Publisher Place: Cham, Switzerland
Issue Date: 2020
Series/Report no.: Lecture Notes in Computer Science; 12085
ISBN: 3030474356
ISSN: 0302-9743
Conference Name: Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) (11 May 2020 - 14 May 2020 : Singapore)
Editor: Lauw, H.W.
Wong, R.C.W.
Ntoulas, A.
Lim, E.P.
Ng, S.K.
Pan, S.J.
Statement of
Bo Chen, Kai Ming Ting, and Tat-Jun Chin
Abstract: Relative scores such as Local Outlying Factor and mass ratio have been shown to be better scores than global scores in detecting anomalies. While this is true, our analysis reveals for the first time that these relative scores have a key shortcoming: anomalies have greatly different relative scores if they are located in different regions where the curvatures of the density surface are very different. As a result, the lowscore anomalies could be ranked lower than some normal points. This revelation motivates (i) a new score called Neighbourhood Contrast (NC) which produces approximately the same high scores for all anomalies, regardless of varying curvatures of the density surface in different regions; and (ii) an anomaly detection method based on NC. Our experiments show that the proposed method which employs the new score significantly outperforms methods using the aforementioned relative scores on benchmark datasets.
Keywords: Neighbourhood Contrast; Anomaly detection; Relative scores
Rights: © Springer Nature Switzerland AG 2020
DOI: 10.1007/978-3-030-47436-2_49
Published version:
Appears in Collections:Computer Science publications

Files in This Item:
There are no files associated with this item.

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.