KPI Synthetic Estimation for Handling Missing Metrics and Updating Cellular Datasets

With the deployment of Fifth Generation (5G) cellular networks, the development and application of Artificial Intelligence (AI) and Machine Learning (ML) techniques has been expanding continuously. Such techniques enhance the reliability, efficiency, capacity, and self-healing capabilities of cellular networks, having as requisite the availability of enough up-to-date training data, especially in the case of AI-based techniques. For this reason, mobile operators and service providers are actively working to generate updated datasets to feed and update all these mechanisms. In this paradigm, the lifecycle management of datasets becomes key. Older well-known data sets are the reference for many carefully tuned AI systems in use, and the effort put into the creation of such datasets is lost with the inclusion of new key network metrics. In this article, a data imputation technique based on Transfer Learning (TL) and Variational AutoEncoders (VAEs) for updating reference cellular datasets is presented. Specifically, this method leverages previous labeled data sets that have been used in the training of complex AI systems by extracting knowledge from their multivariate time series composed of different metrics, such as Key-Performance Indicators (KPIs), as well as from new updated datasets that contain new metrics created to improve the outcome of such complex AI systems. In this way, older data sets that contain real metrics are updated with the new synthetic KPIs, expanding the available reference data.