The development of 5G network technology has given rise to a large number of Configuration and Optimization Parameters (COPs). Adjusting COPs can help to achieve optimal Key Performance Indicators (KPIs) that have an impact on the network, hence improve user Quality of Service (QoS). Configuring parameters manually to be optimal is unrealistic. Based on data correlation, machine learning (ML) models are investigated to learn KPIs behavior with COPs adjustment. However, there are spurious associations between the COP and KPI generated by intermediate variables in realistic scenarios. Traditional ML models are insufficient to give solutions. Thus, we combine the causal inference to estimate the COP adjustment effect. Specifically, we designed a model named Dynamic Causal Deconfounder Network (DCDN). DCDN learns complex and long-term dependencies among indicators through the self-attention mechanism. And from the perspective of causal inference, we reduce spurious associations based on the idea of adversarial for predicting the KPI value accurately. Extensive experiments on the COP-KPI dataset collected from real-world cellular networks validate the effectiveness of the model.
DCDN: Estimating Handover Parameter Adjusting Effect with Causal Inference
10.10.2023
1981323 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Causal inference framework for generalizable safety effect estimates
Online Contents | 2017
|ArXiv | 2022
|Europäisches Patentamt | 2024
|