ECOD

Algorithm introduction

The Empirical Cumulative Outlier Detection (ECOD) algorithm is an unsupervised anomaly detection algorithm. The core idea is to calculate the probability of outliers by using the experience cumulative distribution function (ECDF) to estimate the joint cumulative distribution function of the data (calculating the outlier score for each sample by combining the outlier scores of different dimensions of the same sample). The algorithm was inspired by the fact that outliers are usually rare events that appear at the tail of a distribution.

Use scenario

Anomaly detection for , can , with low time and spatial complexity. The processing can be accelerated by parallel computation.

Algorithm principle

The algorithm first calculates the left-tail ECDF and right-tail ECDF of each dimension (index), then calculates the skewness of each dimension, and finally performs probabilistic aggregation based on this. After traversing each data point, Oleft-only is obtained by left-tailed ECDF aggregation, and Oright-only is obtained by right-tailed ECDF aggregation. Oauto is calculated according to skewness, and the maximum values of the three are taken as the outliers of the data points, and the threshold can be set to screen out the outliers.

** Paper Links ** : <https://www.researchgate.net/publication/357552688_ECOD_Unsupervised_Outlier_Detection_Using_Empirical_Cumulative_Distr ibution_Functions>