Inferring change points in unlabelled time series data collected from the network diagnosis tool

Detecting significant statistical changes in time series data, such as change points and anomalies, is crucial for various applications, including computer network performance monitoring. Despite the availability of many detection algorithms, applying these techniques to real-world data remains a challenging topic due to their distinct effectiveness in different domains. This study focuses on identifying change points and anomalies in throughput and latency time series data from residential networks, emphasizing online methods. We evaluate well-established methods like Shewhart, EWMA, and CUSUM, which are simple to implement, and identify their limitations in real-world scenarios. We propose simple modifications to these classical methods to enhance their effectiveness when applied to data from network measurements. Furthermore, we introduce a new and flexible method, based on the concept of weighted voting. It is designed to detect change points while providing useful information to assess confidence in the results. Our methods were evaluated on two datasets: one we collected using the NDT protocol in Brazil and another from the publicly available Shao Dataset, which includes labeled time series of latency. We discuss the limitations of traditional methods, the effectiveness of our proposed approaches, and how to apply those for real-time network quality monitoring.