Evaluation of Local Explanation Methods for Multivariate Time Series Forecasting

Abstract

Being able to interpret a machine learning model is a crucial task in many applications of machine learning. Specifically, local interpretability is important in determining why a model makes particular predictions. Despite the recent focus on AI interpretability, there has been a lack of research in local interpretability methods for time series forecasting while the few interpretable methods that exist mainly focus on time series classification tasks. In this study, we propose two novel evaluation metrics for time series forecasting: Area Over the Perturbation Curve for Regression and Ablation Percentage Threshold. These two metrics can measure the local fidelity of local explanation models. We extend the theoretical foundation to collect experimental results on two popular datasets, Rossmann sales and electricity. Both metrics enable a comprehensive comparison of numerous local explanation models and find which metrics are more sensitive. Lastly, we provide heuristical reasoning for this analysis.

DANLIP: Deep Autoregressive Networks for Locally Interpretable Probabilistic Forecasting

“…Our preliminary analysis indicates that training a single GBR model for each dataset (i.e., similar to the other three models) leads to a significant deterioration in forecasting performance across the datasets, hence we adopt the above-explained approach. We note that above results for these datasets are largely inline with the forecasting performance values reported in previous studies (e.g., see [8,19,23]).…”

“…Post-hoc interpretability methods have been used to interpret the decisions of time series models. Mujkanovic [17] used SHAP to interpret time series classifiers, whereas Ozyegen et al [19] evaluated three post-hoc interpretability methods, including SHAP, to interpret the time series forecasting models. On the other hand, many time series forecasting methods take into account interpretability considerations in model development [3,8].…”