Deep Representation Learning for Cluster-Level Time Series Forecasting †

In today’s data-driven world, time series forecasting is an intensively investigated temporal data mining technique. In practice, there is a range of forecasting techniques that have been proven to be efficient at capturing different aspects of an input. For instance, classic linear forecasting models such as seasonal autoregressive integrated moving average (S-ARIMA) models are known to capture the trends and seasonality evident in temporal datasets. In contrast, neural-network-based forecasting approaches are known to be best at capturing nonlinearity. Despite such differences, most forecasting techniques inherently assume that models are fitted using a single input. In practice, there are often cases where we cannot deploy forecasting models in this manner. For instance, in most wireless communication traffic forecasting problems, temporal datasets are defined by taking samples from hundreds of base stations. Moreover, the base stations are expected to have spatial correlation due to user mobility, land use, settlement patterns, etc. Thus, in such cases, it is often advised that forecasting should be approached using clusters that group the base stations based on their traffic patterns. However, when this approach is used, the quality of the cluster centroids and the overall cluster formation process is expected to have a significant impact on the performance of forecasting models. In this paper, we show the effectiveness of representation learning for cluster formation and cluster centroid definition, which in turn improves the quality of cluster-level forecasting. We demonstrate this concept using data traffics collected from 729 wireless base stations. In general, based on the experimental results, the representation learning approach outperforms cluster-level forecasting models based on classical clustering techniques such as K-means and dynamic time warping barycenter averaging K-means (DBA K-means).