Towards a scalable anomaly detection with pseudo-optimal hyperparameters

Jellis Vanhoeyveld and David Martens

Working Papers from University of Antwerp, Faculty of Business and Economics

Abstract: Anomaly detection (AD) involves the detection of instances that show a conduct that differs from a well de ned notion of normal behaviour. This apparently simple task still poses two open issues that are addressed in this work. (1) Each AD technique typically requires a number of hyperparameters to be specified. As labels are presumed unknown, supervised tuning approaches are ruled out. The community developed heuristics that are either too computationally involved and/or are tailored to a single type of AD method. Benchmarking studies comparing several AD methods therefore rely on an arbitrary parameter choice or adopt standard statistics (mean, min, max) to summarize the performances of a range of parameter settings. We consider the latter methodology as inappropriate and develop a heuristic to tune the hyperparameters of any AD algorithm. We choose the parameter combination that, on `average', achieves the highest performance on an external labelled data repository of AD tasks. The Wilcoxon-signed rank test reveals this setting to significantly outperform a random choice. (2) AD techniques can be designed with a focus on predictive performance and neglect scalability issues, limiting their applicability. We present the fixed-width anomaly detection (FWAD) method that conducts a preliminary data compression with fixed-width clustering and applies the original AD technique to the resulting cluster centres. The latter retain information of the original dataset in the form of local density estimates. Our results indicate that AD is significantly slower than FWAD if the former has a superlinear complexity in the number of instances N. Substantial bene ts are obtained for datasets of low-dimensionality with N large. Often, no significant AUC differences are observed between FWAD and AD, though we noted a drop in AUC for FWAD on a (small) majority of datasets. Apparently, the local density estimates hold enough information to allow for high-compression levels.

Keywords: Anomaly detection; Unsupervised hyperparameter tuning; Fixed-width clustering; Data compression (search for similar items in EconPapers)
Pages: 61 pages
Date: 2018-10
References: View references in EconPapers View complete reference list from CitEc
Citations:

Downloads: (external link)
https://repository.uantwerpen.be/docman/irua/450120/154651.pdf (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:ant:wpaper:2018012

Access Statistics for this paper

More papers in Working Papers from University of Antwerp, Faculty of Business and Economics Contact information at EDIRC.
Bibliographic data for series maintained by Joeri Nys ().