[4I3-GS-11] AI and Society:

[4I3-GS-11-01] Interpretable Patch-based Time Series Forecasting

〇Hiroki Tomioka¹, Genta Yoshimura¹ (1. Mitsubishi Electric Corporation)

Keywords:time series forecasting, neural network, interpretability, explainability, XAI

Numerous neural network-based time series forecasting methods have been proposed for high-accuracy predictions,
but their complexity often makes it difficult for humans to intuitively understand the rationale behind the predictions and the inference process.
In this study, we propose a highly accurate and interpretable neural network-based time series forecasting method that divides the time series into subsequences (patches) for each variable and calculates the prediction as the sum of contributions from each patch.
This approach allows us to decompose the prediction into the contributions of specific subsequences of each variable, including covariates.
Evaluation on datasets for univariate time series, and multivariate time series forecasting with covariates demonstrate that the proposed method achieves high predictive accuracy comparable to recent forecasting methods while enhancing interpretability.