A New Histogram-based Visualization Tool for Analyzing Anomaly Detection Algorithm Performance

Abstract

Performance visualization of anomaly detection algorithms is an essential aspect of anomaly and intrusion detection systems. It allows analysts to highlight trends and outliers in anomaly detection models results to gain intuitive understanding of detection models. This work presents a new way of visualizing anomaly detection algorithm results using a histogram. The approach presented in this work provides a better understanding of detection algorithms performance by revealing the exact proportions of true positives, true negatives, false positives, and false negatives of detection algorithms. The histogram-based approach was used to visualize the prediction confidence and performances of anomaly detection algorithms on multiple datasets to provide insights into the strengths and weaknesses of these algorithms on different aspects of the datasets. The proposed approach is compared with previous histogram-based visualization methods that rely on only positive and negative anomaly scores. The results show that the proposed method provides a better meaning of detection algorithms performances. Finally, this work presents further results that show how the proposed method can be applied to performance visualization and analysis of supervised machine learning techniques involving binary classification of imbalanced datasets.