Time-series forecasting contributes crucial information to industrial and institutional decision-making with multivariate time-series input. Although various models have been developed to facilitate the forecasting process, they make inconsistent forecasts. Thus, it is critical to select the model appropriately. The existing selection methods based on the error measures fail to reveal deep insights into the model’s performance, such as the identification of salient features and the impact of temporal factors (e.g., periods). This paper introduces mTSeer, an interactive system for the exploration, explanation, and evaluation of multivariate time-series forecasting models. Our system integrates a set of algorithms to steer the process, and rich interactions and visualization designs to help interpret the differences between models in both model and instance level. We demonstrate the effectiveness of mTSeer through three case studies with two domain experts on real-world data, qualitative interviews with the two experts, and quantitative evaluation of the three case studies.
https://doi.org/10.1145/3411764.3445083
The ACM CHI Conference on Human Factors in Computing Systems (https://chi2021.acm.org/)