Handling Imbalanced 5G and Beyond Network Tabular Data Using Conditional Generative Models

Data-driven machine learning based approaches have been playing increasingly important roles in 5G and beyond (B5G) network management and optimisation. One of the major challenges is that the existence of class imbalance in networking datasets can greatly bias the classifier towards a majority classification. This discrepancy can pose a serious problem for AI-based models, which require large and diverse amounts of data to learn patterns and generate classifications. The generation of synthetic data is a process that has evolved over time from the application of statistical models to a more machine learning-centric approaches. In this research work we present the development, implementation, evaluation and comparison of four generative models for tabular data on a B5G network management dataset. These models have been previously optimised according to certain evaluation metrics of the generated synthetic data. The dataset presents a problem of imbalance between its classes, which is improved by using generative models to enrich it with synthetic data. The results show that machine learning based generative models obtain more accurate data than traditional statistical models, and are much faster in terms of conditional data sampling.