By Wei Cai, Cox Communications, Inc.; Mohsin Afridi, Cox Cpmmunications, Inc
The question of generating accurate forecasts in a long memory (or long range) process has attracted much attention with telecom traffic data as it is crucial to formulate capacity planning and budget allocation in a cost-effective manner.
However, there has been a growing awareness of a variety of difficulties to implement long-range forecasting using telecom time series data. Firstly, insufficiency of telecom traffic data has posed challenges to effectively execute sophisticated statistical models and machine learning (ML) models [1]. Furthermore, irregular patterns in network time series data made conventional outlier detection method difficult to detect, which might introduce noise in the forecast, and hence greatly affect forecast accuracy. Lastly, the predictions made using the state-of-the-art statistical models or highly supervised machine learning models tend to experience error propagation and lose accuracy as the prediction time horizon expands. It is less likely for those models to correctly extrapolate the special characteristics of network time series data, particularly when small-scale historical data is presented as the learning dataset.
In order to address the issues highlighted earlier, literature has introduced a relatively recent approach that involves using a Generative Adversarial Network (GAN) architecture to generate a soft representation for both the short- and long-term dependencies in the time series. The GAN architecture was initially proposed by Goodfellow et al [2]. Originally, GANs were primarily designed for processing picture data. Since their introduction, substantial progress has been achieved in expanding their capabilities, and they are extensively employed in various tasks such as text generation, audio signal generation, spectral data generation, tabular data generation, and time series data generation [3][4][5].
Nonetheless, as far as we are aware, GAN has been focused less on temporal time series data. Consequently, there has not been much research done on how to use GAN to improve long-range forecasting.
In this paper, we evaluate the performance of GAN in time series forecasting and propose a hybrid forecasting strategy of incorporating GAN into a long horizon forecasting process using telecom traffic data.