A Mobile Network Performance Evaluation Method Based on Multivariate Time Series Clustering with Auto-Encoder

Abstract

With the extensive use of smartphones, the amount of collected mobile data and types of mobile network performance evaluation indicators are rapidly expanding. This has resulted in a challenge for the mobile network providers of how to assess the mobile network performance quickly and precisely. This paper proposes a mobile network performance evaluation method based on multivariate time series clustering with auto-encoder. In this method, firstly, we remove a part of indicators with high redundancy by combining Pearson's correlation coefficient and lightgbm. Secondly, we consider the changes in indicators over time as multivariate time series, the distance matrices of which are measured by fastDTW and auto-encoder. Finally, the indicators data under different periods are clustered into three categories by using k-medoids on matrices. We apply this method to the indicator data provided by a mobile network provider and discover that the clusters refer to different levels of mobile network performance according to the mobile network provider' KPI standards.