Frequency-aware Time Series Forecasting, Anomaly Detection, Classification and Granger Causality

• Published in: 2022 14th International Conference on COMmunication Systems & NETworkS (COMSNETS) pp. 217 - 221
• Main Authors: Banerjee, Serene, Martin, Raul R, Pardo, Abel
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.01.2022
• Subjects: Anomaly Detection; Classification; Feature extraction; Filter banks; Forecasting; Granger Causality; High frequency; Key Performance Indicators; Machine learning; Multivariate Time Series Analysis; Passive Intermodulation; Rendering (computer graphics); Root Cause Analysis; Time series analysis; Time-frequency analysis
• ISSN: 2155-2509
• DOI: 10.1109/COMSNETS53615.2022.9668359

Automated methods for forecasting and anomaly detection on time series data is very essential for LTE, 5G and upcoming networks. However, due to spectral bias, at finer time granularities, it is difficult for machine learning models to give reasonable outputs as they tend to approximate the low spectral frequencies only. Inspired by the application of Fourier Feature Mapping [3] to preserve high frequency details for rendering applications, we demonstrate the use of the same for four relevant applications in RAN, including, but not limited to: (a) time-series forecasting, (b) anomaly detection, (c) auto-encoding of Radio environments, and (d) neural-network based Granger causality detection. Our proposed approach beats state-of-the-art methods in time series forecasting by consistently having a mean average error (MAE) of less than 0.2. We validate our proposed approach on customer data collected from LTE/5G networks. As Fourier Feature Mapping is a bank of filters, the approach extends well to multi-frequency multi-variate time series as well.