Assessing Deep Generative Models on Time Series Network Data

To achieve zero touch automation in next generation wireless networks through artificial intelligence (AI), large amounts of training data is required. This training data is publicly unavailable and is a major hindrance in research on AI applications to wireless communication. One solution is using limited real data to generate synthetic data that can be used in lieu of real data. Generative Adversarial Networks (GAN) have been used successfully for this purpose. In this paper, we choose two publicly available GAN - based models and one deep learning - based auto-regressive model. We then compare their performance at generating synthetic time-series wireless network traffic data. We also assess the impact of data scarcity on the generated data quality by varying the level of data available to the models for training. Moreover, in order to assess the usefulness of this generated data, we compare the performance of a gradient boosting regressor trained solely on generated data, real data, and a mix of both at forecasting network traffic. Our experiments show that the GANs perform better than the auto-regressive approach in each aspect considered in this work and forecasting models trained to predict network load based on data generated by these GANs yield error rates comparable to models trained on real data. Finally, augmenting small amounts of real data with generated data leads to minor performance gains in some cases.