Change detection method using cluster transition probability
Shoko Takahashi, Kei Takeshita, Kazuhisa Yamagishi, Masataka Masuda · 2022
AI models such as anomaly detection are being incorporated into various systems, and it is essential to update such models in response to changes in the external environment. When trying to automate the determination of the update timing of the AI model, one promising solution is considered to be the use of a change detection method. However, constraints such as stationarity and iid imposed on target data by existing methods often become barriers to application to time-series data in telecommunications, which often show periodic fluctuations. In this paper, we propose a new change detection method that solves the problems of existing methods, focusing on time-series data showing daily and weekly fluctuations. The proposed method is based on clustering that imposes no restrictions on target data and detects changes by tracking cluster transitions and calculating the distance between cluster transition distributions for the past and current period. The effectiveness of the proposed method is shown by using actual telecommunications data with tens of millions of user connections to network devices.