Estimating Latent Traffic Demand From QoS Degradation Using a Time Series Causal Inference Approach

Quality of Service (QoS) degradation often results in Quality of Experience (QoE) degradation, leading users to abandon applications or causing applications to reduce content size to avoid congestion and further QoE degradation. In both cases, traffic will be reduced. For efficient traffic engineering or network design, it is essential to estimate traffic before the reduction, termed latent traffic demand. If network provisioning is based solely on observed traffic, which is typically lower due to QoS degradation, it can lead to under-provisioning. This underestimation may cause a cycle in which improved QoS leads to increased traffic, resulting in congestion and subsequent QoS degradation. Therefore, accurately estimating latent traffic demand helps in designing networks that can handle actual user demands, ensuring better QoS and user satisfaction. We adopt a causal inference approach, treating the impact as a causality from QoS degradation to traffic reduction. Existing causal inference techniques typically assume acyclicity, implying a one-way relationship between cause and effect variables. However, in our case, QoS degradation is primarily caused by increased traffic and subsequent congestion, indicating a two-way relationship. We focus on the fact that these two causal relationships operate on different time scales; thus, by constructing a time series causal inference structure with appropriate temporal granularity, we aim to separate these two relationships into two one-way causal relationships. Additionally, to reduce the bias and variance of naive time series causal inference techniques, we propose a source-separated causal inference method. We evaluate our method through various simulations using both synthesized and real traffic data, confirming the possibility of an accurate estimation.