Hankel Imputation for Robust Network Throughput Forecasting

Abstract

Network operators rely on analysis of data of monitoring applications to support strategic planning and ensure Quality of Service (QoS). Among network performance metrics, throughput is particularly critical, as it reflects the actual data successfully delivered between endpoints despite external interference. However, throughput time series are often affected by missing values due to measurement-test failures, the retention of incomplete or failed test records in datasets, and general server-side issues , which compromise data quality and hinder downstream tasks such as traffic forecasting. To address this data quality challenge, we propose HAIL (Hankel Adaptive Imputation via Latent Space), a novel algorithm designed to reconstruct incomplete throughput series using the available throughput measurements and the intrinsic temporal information. HAIL combines Hankel matrix embedding with latent space imputation to capture and preserve the underlying temporal structure. In addition to evaluating HAIL’s imputation accuracy, we assessed the downstream utility of the reconstructed datasets by using them to train GRU and LSTM forecasting models and measuring the resulting forecasting performance. The experimental results demonstrate that the HAIL algorithm restores missing throughput patterns, outperforming the baseline methods at missingness rates of 10%, 15%, and 20%, and achieving a slightly lower RMSE in the forecasting results obtained with the datasets imputed by HAIL. Experimental results demonstrate that HAIL effectively restores missing throughput patterns, maintaining temporal consistency and improving the reliability of network performance forecasting.