Created by W.Langdon from gp-bibliography.bib Revision:1.4824
We address this problem, proposing ensemble anomaly detection techniques that perform well in many applications, with four major contributions: using bootstrapping to better detect anomalies on multiple subsamples, sequential application of diverse detection algorithms, a novel adaptive sampling and learning algorithm in which the anomalies are iteratively examined, and improving the random forest algorithms for detecting anomalies in streaming data.
We design and evaluate multiple ensemble strategies using score normalization, rank aggregation and majority voting, to combine the results from six well-known base algorithms. We propose a bootstrapping algorithm in which anomalies are evaluated from multiple subsets of the data. Results show that our independent ensemble performs better than the base algorithms, and using bootstrapping achieves competitive quality and faster runtime compared with existing works.
We develop new sequential ensemble algorithms in which the second algorithm performs anomaly detection based on the first algorithm's outputs; best results are obtained by combining algorithms that are substantially different. We propose a novel adaptive sampling algorithm which uses the score output of the base algorithm to determine the hard-to-detect examples, and iteratively resamples more points from such examples in a complete unsupervised context.
On streaming datasets, we analyse the impact of parameters used in random trees, and propose new algorithms that work well with high-dimensional data, improving performance without increasing the number of trees or their heights. We show that further improvements can be obtained with an Evolutionary Algorithm.
Second supervisor: Chilukuri Mohan",
Genetic Programming entries for Zhiruo Zhao