Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpreting Time Series Transformer Models and Sensitivity Analysis of Population Age Groups to COVID-19 Infections

Md Khairul Islam, Tyler Valentine, Timothy Joowon Sue, Ayush Karmacharya, Luke Neil Benham, Zhengguang Wang, Kingsley Kim, Judy Fox

TL;DR

The paper tackles the challenge of interpreting transformer-based time-series models applied to multivariate, multi-horizon COVID-19 forecasting at the county level. It presents a model-agnostic, perturbation-based local interpretation framework and benchmarks eight interpretation methods, demonstrating a strong performance of FEDformer for COVID-19 predictions and showing that age-group sensitivity can be both evaluated with ground truth and forecasted over horizons. The authors extend the approach to Electricity and Traffic domains to establish generalizability across time-series tasks. They also introduce a novel age-group sensitivity evaluation using CDC data and demonstrate that past infection data and specific age groups drive predictions, with practical implications for real-time decision-making and public health analysis. Overall, the work provides a comprehensive, transferable methodology for explaining complex time-series transformers in dynamic, real-world datasets.

Abstract

Interpreting deep learning time series models is crucial in understanding the model's behavior and learning patterns from raw data for real-time decision-making. However, the complexity inherent in transformer-based time series models poses challenges in explaining the impact of individual features on predictions. In this study, we leverage recent local interpretation methods to interpret state-of-the-art time series models. To use real-world datasets, we collected three years of daily case data for 3,142 US counties. Firstly, we compare six transformer-based models and choose the best prediction model for COVID-19 infection. Using 13 input features from the last two weeks, we can predict the cases for the next two weeks. Secondly, we present an innovative way to evaluate the prediction sensitivity to 8 population age groups over highly dynamic multivariate infection data. Thirdly, we compare our proposed perturbation-based interpretation method with related work, including a total of eight local interpretation methods. Finally, we apply our framework to traffic and electricity datasets, demonstrating that our approach is generic and can be applied to other time-series domains.

Interpreting Time Series Transformer Models and Sensitivity Analysis of Population Age Groups to COVID-19 Infections

TL;DR

The paper tackles the challenge of interpreting transformer-based time-series models applied to multivariate, multi-horizon COVID-19 forecasting at the county level. It presents a model-agnostic, perturbation-based local interpretation framework and benchmarks eight interpretation methods, demonstrating a strong performance of FEDformer for COVID-19 predictions and showing that age-group sensitivity can be both evaluated with ground truth and forecasted over horizons. The authors extend the approach to Electricity and Traffic domains to establish generalizability across time-series tasks. They also introduce a novel age-group sensitivity evaluation using CDC data and demonstrate that past infection data and specific age groups drive predictions, with practical implications for real-time decision-making and public health analysis. Overall, the work provides a comprehensive, transferable methodology for explaining complex time-series transformers in dynamic, real-world datasets.

Abstract

Interpreting deep learning time series models is crucial in understanding the model's behavior and learning patterns from raw data for real-time decision-making. However, the complexity inherent in transformer-based time series models poses challenges in explaining the impact of individual features on predictions. In this study, we leverage recent local interpretation methods to interpret state-of-the-art time series models. To use real-world datasets, we collected three years of daily case data for 3,142 US counties. Firstly, we compare six transformer-based models and choose the best prediction model for COVID-19 infection. Using 13 input features from the last two weeks, we can predict the cases for the next two weeks. Secondly, we present an innovative way to evaluate the prediction sensitivity to 8 population age groups over highly dynamic multivariate infection data. Thirdly, we compare our proposed perturbation-based interpretation method with related work, including a total of eight local interpretation methods. Finally, we apply our framework to traffic and electricity datasets, demonstrating that our approach is generic and can be applied to other time-series domains.
Paper Structure (25 sections, 4 equations, 4 figures, 9 tables)

This paper contains 25 sections, 4 equations, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Background Terminologies
  4. Related Work
  5. Time Series Interpretation
  6. Interpreting COVID-19 Infection
  7. Problem Statement
  8. Methodology for Local Interpretation of Time-Series
  9. Interpretation Methods
  10. Datasets
  11. Experiment Setup
  12. COVID-19 Data Preprocessing
  13. Models and Parameters
  14. Implementation Library
  15. Prediction Results
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Test predictions comparison with ground truth aggregated over all counties.
  • Figure 2: Weekly COVID-19 cases weekly_cases for each of the eight age subgroups over the study period .
  • Figure 3: Predicted age sensitivity on the extended dataset with weekly COVID-19 cases by age groups as ground truth.
  • Figure 4: Interpreting different feature attribution for the Los Angeles, California county from our test set using the FEDformer model and Feature Ablation method.