Editable Temporal Graph Neural Network for Interactive Time Series Forecasting
Abstract: Time-series data often contain multiple dependent variables that evolve with persistent relationships. Modeling such relational information with graphs can improve the forecasting accuracy significantly, while providing explainable insights to the human users on the underlying dependencies. Conventional graph temporal models propose learning such relational information from the data, which might not capture the correct relations. Ideally, it is desirable to provide mechanisms for users to modify the learned relationships before using the models for forecasting, especially in the high-stakes decision making scenarios. In this paper, we propose a novel model, Editable Temporal Graph Neural Network (ETGNN), and an editing algorithm that allows users to make edits on the learned graphs before obtaining forecasts, in a superior way compared to the alternatives. A key novelty of ETGNN model is the use of global embeddings in both graph learning and time series forecasting, which allows fast editing the forecasts in a local region without affecting unrelated forecasts. We show that after editing, ETGNN can achieve close-to-perfect accuracy for latent graph prediction (with the F1-score of larger than 0.9) and reduce the time series forecasting error by 10\%$\sim$40\% over multiple datasets, with much lower editing cost compared to retraining. Overall, ETGNN provides a convenient mechanism to incorporate user feedback to improve the accuracy and explainablity of complex forecasting tasks.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Artem_Babenko1
Submission Number: 1112
Loading