Enhancement of the Local Outlier Factor Algorithm for Anomaly Detection in Time Series

EasyChair Preprint 14238

Abstract

A plethora of methodological approaches have been proferred in recent years within the topical field of time series anomaly detection. Despite many advancements in adjacent fields, the well-known local outlier factor algorithm has stood the test of time in this domain, with recent comparative studies indicating that it remains competitive with newer approaches in benchmark tests. In this paper, we enhance this algorithm by leveraging ensembling techniques and GPU (graphics processing unit) acceleration. To the best of our knowledge, our ensembling approach establishes a new state-of-the-art accuracy of 66.8% in respect of the well-known UCR Time Series Anomaly Detection benchmark, while our GPU implementation was approximately eleven times faster than our CPU (central processing unit) baseline.

Keyphrases: Knowledge Discovery and Data Mining, Local outlier factor algorithm, anomaly detection, machine learning, matrix profile, time series

@booklet{EasyChair:14238,
  author    = {Daniel Barrish and Jan van Vuuren},
  title     = {Enhancement of the Local Outlier Factor Algorithm for Anomaly Detection in Time Series},
  howpublished = {EasyChair Preprint 14238},
  year      = {EasyChair, 2024}}