Abstract
Multivariate time series prediction has applications in a wide variety ofdomains and is considered to be a very challenging task, especially when thevariables have correlations and exhibit complex temporal patterns, such asseasonality and trend. Many existing methods suffer from strong statisticalassumptions, numerical issues with high dimensionality, manual featureengineering efforts, and scalability. In this work, we present a novel deeplearning architecture, known as Temporal Tensor Transformation Network, whichtransforms the original multivariate time series into a higher order of tensorthrough the proposed Temporal-Slicing Stack Transformation. This yields a newrepresentation of the original multivariate time series, which enables theconvolution kernel to extract complex and non-linear features as well asvariable interactional signals from a relatively large temporal region.Experimental results show that Temporal Tensor Transformation Networkoutperforms several state-of-the-art methods on window-based predictions acrossvarious tasks. The proposed architecture also demonstrates robust predictionperformance through an extensive sensitivity analysis.