Generative Representation Learning in Recurrent Neural Networks for Causal Timeseries Forecasting

Feed-forward deep neural networks (DNNs) are the state of the art in timeseries forecasting. A particularly significant scenario is the causal one: when an arbitrary subset of variables of a given multivariate timeseries is specified as forecasting target, with the remaining ones (exogenous variables) <i>causing</i> the target at each time instance. Then, the goal is to predict a temporal window of future target values, given a window of historical exogenous values. To this end, this article proposes a novel deep recurrent neural architecture, called generative-regressing recurrent neural network (GRRNN), which surpasses competing ones in causal forecasting evaluation metrics, by smartly combining generative learning and regression. During training, the generative module learns to synthesize historical target timeseries from historical exogenous inputs via conditional adversarial learning, thus internally encoding the input timeseries into semantically meaningful features. During a forward pass, these features are passed over as input to the regression module, which outputs the actual future target forecasts in a sequence-to-sequence fashion. Thus, the task of timeseries generation is synergistically combined with the task of timeseries forecasting, under an end-to-end multitask training setting. Methodologically, GRRNN contributes a novel augmentation of pure supervised learning, tailored to causal timeseries forecasting, which essentially forces the generative module to transform the historical exogenous timeseries to a more appropriate representation, before feeding it as input to the actual forecasting regressor. Extensive experimental evaluation on relevant public datasets obtained from disparate fields, ranging from air pollution data to sentiment analysis of social media posts, confirms that GRRNN achieves top performance in multistep long-term forecasting.