fnctId=thesis,fnctNo=358
[심성현] Distributed Lag Transformer based on Time-Variable-Aware Learning for Explainable Multivariate Time Series Forecasting
- 작성자
- scsc연구센터
- 저자
- Younghiw Kim, Sunghyun Sim, Dohee Kim, Joongrock Kim
- 발행사항
- 발행일
- 20260713
- 저널명
- IEEE Access
- 국문초록
- 영문초록
- Time series data is a key element of big data analytics, commonly found in domains such as
finance, healthcare, climate forecasting, and transportation. In large-scale real-world settings, such data is
often high-dimensional and multivariate, requiring advanced forecasting methods that are both accurate and
interpretable. Although Transformer-based models have achieved strong performance in multivariate time
series forecasting (MTSF), their lack of explainability limits their use in critical applications. To address
this limitation, we propose the Distributed Lag Transformer (DLFormer), a novel Transformer architecture
for explainable and scalable MTSF. DLFormer integrates distributed lag embedding and time-variableaware learning (TVAL) to structurally model both local and global temporal dependencies and explicitly
capture the influence of past variables on future outcomes. Experiments on ten benchmark and real-world
datasets demonstrate that DLFormer achieves competitive predictive accuracy, yielding MSE reductions of
1.41% on average across all settings and 2.66% in short-term forecasting relative to three top-performing
baselines, including iTransformer, NLinear, and TFT. Beyond predictive accuracy, we evaluate the quality
of explainability through global attention-based analysis, perturbation-based faithfulness evaluation, and
domain-oriented case studies. These analyses show that DLFormer provides more robust, faithful, and valid
explanations of temporal-variable dependencies than competing models. Overall, these results suggest that
DLFormer effectively bridges the gap between predictive performance and explainability. The source code
is publicly available at: http://github.com/kYounghwi/DLFormer
- 일반텍스트
- 첨부파일