The Artificial Immune Ecosystem: A Bio-Inspired Meta-Algorithm for Boosting Time Series Anomaly Detection with Expert Input

摘要

One of the challenges in machine learning, especially in the Big Data era, is to obtain labeled data sets. Indeed, the difficulty of labeling large amounts of data had lead to an increasing reliance on unsupervised classifiers, such as deep autoencoders. In this paper, we study the problem of involving a human expert in the training of a classifier instead of using labeled data. We use anomaly detection in network monitoring as a field of application. We demonstrate how using crude, already existing monitoring software as a heuristic to choose which points to label can boost the classification rate with respect to both the monitoring software and the classifier trained on a fully labeled data set, with a very low computational cost. We introduce the Artificial Immune Ecosystem meta-algorithm as a generic framework integrating the expert, the heuristic and the classifier.