Graph Deep Factors for Probabilistic Time-series Forecasting

Effective time-series forecasting methods are of significant importance to solve a broad spectrum research problems. Deep probabilistic techniques have recently been proposed for modeling large collections time-series. However, these explicitly assume either complete independence (local model) or dependence (global between in the collection. This corresponds two extreme cases where every is disconnected from other collection likewise, that related resulting completely connected graph. In this work, we propose deep hybrid graph-based framework called Graph Factors (GraphDF) goes beyond extremes by allowing nodes and their be others an arbitrary fashion. GraphDF consists relational global local model. particular, model learns complex non-linear patterns globally using structure graph improve both accuracy computational efficiency. Similarly, instead independently, not only considers its individual but also The experiments demonstrate effectiveness compared state-of-the-art terms accuracy, runtime, scalability. Our case study reveals can successfully generate cloud usage forecasts opportunistically schedule workloads increase cluster utilization 47.5% on average. Furthermore, target addressing common nature many applications provided streaming version; however, most fail leverage newly incoming values result worse performance over time. article, online incremental learning forecasting. theoretically proven lower time space complexity. universally applied machine learning-based methods.