Tensor Time-Series Forecasting and Anomaly Detection with Augmented Causality
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Time Series, Forecasting, Anomaly Detection
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: In time series, variables often exhibit high-dimensional characteristics, and relationships between variables tend to be intricate, encompassing aspects such as non-linearity and time-dependency. Understanding the interaction of variables and comprehending the distribution of their values can significantly enhance the effectiveness of time series data analysis tasks, such as forecasting and anomaly detection. Hence, in this paper, we start from the tensor time series, which can encode higher dimensional information than classic multivariate time series, and aim to discover and leverage their fine-grained time-dependent causal relations to contribute to a more accurate analysis. To this end, we first form an augmented Granger Causality model, named TBN-Granger Causality, which adds time-respecting Bayesian Networks to the time-lagged Neural Granger Causality through a bi-level optimization, such that the overlooking of instantaneous effects in typical causal time series analysis can be addressed. Then, we propose an end-to-end deep generative model, named TacSas, which takes the historical tensor time series, outputs the future tensor time series, and detects possible anomalies, by leveraging the TBN-Granger Causality in the history. Moreover, we show TacSas not only can capture the ground-truth causality but also can be applied when the ground-truth causal structures are hardly available, to help forecasting and anomaly detection. For evaluations, besides synthetic benchmark data, we have four datasets from the climate domain benchmark database ERA5 as the real-world tensor time series for forecasting. Moreover, we extend ERA5 with the extreme weather database NOAA for testing anomaly detection accuracy. We show the effectiveness of TacSas in different time series analysis tasks by comparing with causal baselines, forecasting baselines, and anomaly detection baselines.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7927
Loading