Comparative Analysis of the Impact of Arrival Rates on the Performance of Distance-based Streaming Outlier Detection Algorithms

Abstract

Outlier detection in data streams comes with many challenges. Among these challenges is the variable arrival rate of streams. When data packets are sent across an unreliable network, the data sending process is interrupted due to temporary loss of signal and later all of the data is tried to send at once as signals resume, resulting in data point drop, leading to faulty outlier detection. However, which algorithm performs the best in such cases remained a question until now. This research studies the impact of the arrival rate, varying queue capacity sizes, and slide sizes on the performance of state-of-the-art outlier detection algorithms for data streams. Our experiments show that using a bounded queue for incoming data points and allowing data drop has an average detrimental impact on the F-1 score, which is 100% for NETS, 99.78% for Thresh-Leap, 99.69% for Micro-Cluster, 67.5% for Exact Storm, and 0.422% for DUE. The number of outliers lost is 0% for Thresh-Leap, -0.33% for NETS, 0% for Micro-Cluster, 39.2% for Exact Storm and 38.6% for DUE, observed on default parameters.