seq2graph: Discovering Dynamic Dependencies from Multivariate Time Series with Multi-level Attention

• DOI: 10.48550/arxiv.1812.04448
• Publication Date: 2018-01-01
• Authors: Xuan-Hong Dang; Syed Yousaf Shah; Petros Zerfos
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Statistics - Machine Learning; 0202 electrical engineering, electronic engineering, information engineering; Machine Learning (stat.ML); 02 engineering and technology; Machine Learning (cs.LG)

Discovering temporal lagged and inter-dependencies in multivariate time series data is an important task. However, many real-world applications, such as commercial cloud management, manufacturing predictive maintenance, portfolios performance analysis, dependencies can be non-linear time-variant, which makes it more challenging to extract through traditional methods Granger causality or clustering. In this work, we present a novel deep learning model that uses multiple layers of customized gated recurrent units (GRUs) for discovering both behaviors well inter-timeseries the form directed weighted graphs. We introduce key component Dual-purpose neural network decodes information domain discover within each series, encodes them into set vectors which, collected from all informative inputs inter-dependencies. Though discovery two types are separated at different hierarchical levels, they tightly connected jointly trained end-to-end manner. With joint training, one type dependency immediately impacts other one, leading overall accurate discovery. empirically test our on synthetic exact (non-linear) known. also evaluate its (i) monitoring provider, exhibit highly dynamic, non-linear, volatile behavior and, (ii) sensor plant. further show how approach able capture these via intuitive interpretable graphs use generate forecasts.