Series Saliency: Temporal Interpretation for Multivariate Time Series Forecasting

TL;DR

This paper introduces the series saliency framework for multivariate time series forecasting, providing temporal and feature interpretability while enhancing forecast accuracy using saliency maps and data augmentation.

Contribution

It proposes a novel temporal interpretation method that considers both feature and time dimensions, applicable to various deep learning models, and improves forecasting accuracy.

Findings

Effective temporal interpretation demonstrated on real datasets.

Framework enhances forecast accuracy through data augmentation.

Saliency maps reveal important time points and features.

Abstract

Time series forecasting is an important yet challenging task. Though deep learning methods have recently been developed to give superior forecasting results, it is crucial to improve the interpretability of time series models. Previous interpretation methods, including the methods for general neural networks and attention-based methods, mainly consider the interpretation in the feature dimension while ignoring the crucial temporal dimension. In this paper, we present the series saliency framework for temporal interpretation for multivariate time series forecasting, which considers the forecasting interpretation in both feature and temporal dimensions. By extracting the "series images" from the sliding windows of the time series, we apply the saliency map segmentation following the smallest destroying region principle. The series saliency framework can be employed to any well-defined deep learning models and works as a data augmentation to get more accurate forecasts. Experimental results on several real datasets demonstrate that our framework generates temporal interpretations for the time series forecasting task while produces accurate time series forecast.