Extended Random Histogram Forest for Unsupervised Anomaly Detection

Abstract

Anomaly detection is a fundamental task in the field of unsupervised machine learning, aimed at identifying instances that significantly deviate from other input data. This problem has various applications, including identifying defective products in industries, detecting network intrusions, medical diagnostics, and many other cases. Despite extensive research conducted in this field, a solution with satisfactory performance under all conditions and types of data has not yet been achieved. One effective unsupervised method is the Random Histogram Forest (RHF) approach, which utilizes a probabilistic approach. Based on the evaluation metric of average precision of the area under the precision-recall curve (AP), this approach has shown better performance compared to other methods in terms of AP. However, due to its inherent mechanism, this approach also has limitations. In this article, these limitations are examined, and solutions and methods are proposed to address these constraints, and an extension approach called Extended Random Histogram Forest (ERHF) is introduced for unsupervised anomaly detection. ERHF utilizes data projection and dimensionality reduction methods to enhance the compatibility between the data and the internal mechanism of RHF. Instead of randomly selecting splitting points in the process of constructing random trees, the Skewness score is used to choose more targeted splitting points. Additionally, multidimensional hyperplanes with random slopes are used for data partitioning instead of pointwise splits. Furthermore, the mechanism for calculating the anomaly score of instances is modified to enable the use of subsampling. The performance evaluation of the proposed ERHF method using various metrics (AP, AUC, Recall, F1) on ODDS datasets demonstrates that ERHF significantly outperforms the RHF method.