한국시뮬레이션학회논문지 (Journal of the Korea Society for Simulation)

한국시뮬레이션학회 (The Korea Society for Simulation)
권호 표지
DOI QR코드

DOI QR Code

트랜스포머 기반 다변량 시계열 분류 모델에 대한 XAI 기법 비교

Comparison of XAI Methods for Transformer based Multivariate Time Series Classification Model

  • 투고 : 2025.08.05
  • 심사 : 2025.09.09
  • 발행 : 2025.09.30
PDF 다운로드
⟨ 이전 논문 다음 논문 ⟩

초록

다변량 시계열 분류 문제는 많은 응용 분야에서 연구되고 있다. 인공지능 기반의 많은 모델이 최근 개발되었고 그 성능이 향상되고 있다. 현재 개발되고 있는 많은 인공지능 모델은 블랙박스 형태로 모델의 출력에 대한 설명가능성이 요구되고 있다. 의학, 금융, 군사 분야 등의 주요 의사결정 분야에서 이러한 요구사항이 증가되고 있으며, 이러한 요구에 따라 다양한 인공지능 모델에 대하여 설명가능한 인공지능 기술(XAI)들이 제시되었다. 본 연구에서는 탄도 미사일의 궤적을 다변량 시계열 데이터로 간주하고, 궤적을 입력으로 받아 탄도 미사일의 탄종을 분류하는 트랜스포머 기반의 모델에 대하여 XAI 기법들을 적용하였고 그 결과를 비교, 분석하였다. 트랜스포머 기반의 다변량 시계열 분류 모델에 적합한 XAI 기법으로 LIME 분석과 SHAP 분석, 기울기 기반(Gradient Based) 분석 방법 등을 적용하고 각 방법의 분석 결과를 비교하였다. 각 방법을 적용하여 모델에 사용된 여러 특징 변수 중에서 가장 중요한 변수를 식별하였고, 시계열 데이터의 특성을 고려하여 어느 구간에서 어떤 특징 변수가 중요한지를 추가 식별하였다. 또한 각 기법들에 대하여 검증 결과를 비교하였다. 다양한 XAI 기법을 적용하고 그 결과를 비교하여 모델의 설명 가능성을 제시할 수 있었으며, 이는 모델의 분류 결과에 대한 신뢰도를 높일 수 있을 뿐 아니라 향후 모델의 성능 향상에도 기여할 수 있을 것이다.

Multivariate time series classification problems are being studied in many application fields. Many AI-based models have been developed recently, and their performances are improving. Many AI models being developed require the explainability of the model output in the form of a black box. This requirement is increasing in major decision-making fields such as medicine, finance, and military, and various AI technologies (XAI) that can explain AI models have been proposed in response to this demand. In this study, we considered the trajectory of a ballistic missile as a multivariate time series data, applied XAI techniques to a transformer-based model that receives the trajectory as input and classifies the type of ballistic missile. We applied LIME, SHAP, and gradient-based analysis methods as XAI techniques suitable for the transformer-based multivariate time series classification model, and compared the analysis results of each method. By applying each method, we identified the most important variable among the various feature variables used in the model, and additionally identified which feature variable is important in which section by considering the characteristics of the time series data. We also compared the verification results for each technique. By applying various XAI techniques and comparing the results, we were able to present the explainability of the model, which will not only increase the reliability of the model's classification results, but also contribute to improving the model's performance in the future.

키워드

과제정보

본 논문은 2024년 정부(방위사업청)의 재원으로 국방과학연구소의 지원을 받아 수행된 연구입니다(UG243001TD).

참고문헌

  1. Ali, A., Schnake, T., Eberle, O., Montavon, G., Müller, K. R., & Wolf, L. (2022, June). XAI for transformers: Better explanations through conservative propagation. In International conference on machine learning (pp. 435-451). PMLR..
  2. Atanasova, P. (2024). A diagnostic study of explain ability techniques for text classification. In Accountable and Explainable Methods for Complex Reasoning over Text (pp. 155-187). Cham: Springer Nature Switzerland.
  3. Chefer, H., Gur, S., & Wolf, L. (2021). Generic attention model explain ability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 397-406)
  4. Crothers, E., Viktor, H., & Japkowicz, N. (2024, September). Robust Infidelity: When Faithfulness Measures on Masked Language Models Are Misleading. In International Conference on Machine Learning, Optimization, and Data Science (pp. 133-147). Cham: Springer Nature Switzerland.
  5. Hsieh, T. Y., Wang, S., Sun, Y., & Honavar, V. (2021, March). Explainable multivariate time series classification: a deep neural network which learns to attend to important variables as well as time intervals. In Proceedings of the 14th ACM international conference on web search and data mining (pp. 607-615)
  6. Ismail, A. A., Gunady, M., CorradaBravo, H., & Feizi, S. (2020). Benchmarking deep learning interpretability in time series predictions. Advances in neural information processing systems, 33, 6441-6452.
  7. Jahin, M. A., Shovon, M. S. H., Mridha, M. F., Islam, M. R., & Watanobe, Y. (2024). A hybrid transformer and attention based recurrent neural network for robust and interpretable sentiment analysis of tweets. Scientific Reports, 14(1), 24882. https://doi.org/10.1038/s41598-024-76079-5
  8. Jutte, A., Ahmed, F., Linssen, J., & van Keulen, M. (2025). C-SHAP for time series: An approach to high-level temporal explanations. arXivpreprint arXiv:2504.11159.
  9. Lee, D., Choi, Y., Kim, J., Kim, D., Kim, S., & Park, S. J. (2025). Development of Ballistic Missile Classification method based on Dynamic Information. Journal of the Korea Society for Simulation, 34(1), 33-40. https://doi.org/10.9709/JKSS.2025.34.1.033
  10. Lee, M. H. (2024). Towards gradient-based time-series explanations through a spatiotemporal attention network. arXivpreprint arXiv:2405.17444.
  11. Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. Advances in neural information processing systems, 30.
  12. Mujkanovic, F., Doskoč, V., Schirneck, M., Schäfer, P., & Friedrich, T. (2020). timeXplain--A Framework for Explaining the Predictions of Time Series Classifiers. arXivpreprint arXiv:2007.07606.
  13. Nwakanma, C. I., Nkoro, E. C., Ahakonye, L. A. C., Anyanwu, G. O., Njoku, J. N., & Kim, D. S. (2023). Explainable Activity Detection and Classification: SHAP vs LIME. Proceedings of the KICS Conference, 1816-1817.
  14. Remmer, E. (2022). Explain ability methods for transformer-based artificial neural networks: A comparative analysis(Master's thesis, KTH Royal Institute of Technology)
  15. Ribeiro, M. T., Singh, S., & Guestrin, C. (2016, August). "Why should itrust you?" Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (pp. 1135-1144).
  16. Saluja, R., Malhi, A., Knapič, S., Främling, K., & Cavdar, C. (2021). Towards a rigorous evaluation of explainabilityfor multivariate time series. arXivpreprint arXiv:2104.04075.
  17. Samek, W., Binder, A., Montavon, G., Lapuschkin, S., & Müller, K. R. (2016). Evaluating the visualization of what a deep neural network has learned. IEEE transactions on neural networks and learning systems, 28(11), 2660-2673. https://doi.org/10.1109/TNNLS.2016.2599820
  18. Schlegel, U., Arnout, H., El-Assady, M., Oelke, D., & Keim, D. A. (2019, October). Towards a rigorous evaluation of XAI methods on time series. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW) (pp. 4197-4201). IEEE.
  19. Schlegel, U., Vo, D. L., Keim, D. A., & Seebacher, D. (2021, September). Ts-mule: Local interpretable model-agnostic explanations for time series forecast models. In Joint European conference on machine learning and knowledge discovery in databases (pp. 5-14). Cham: Springer International Publishing.
  20. Serramazza, D. I., Nguyen, T. T., Le Nguyen, T., & Ifrim, G. (2023, September). Evaluating explanation methods for multivariate time series classification. In International Workshop on Advanced Analytics and Learning on Temporal Data (pp. 159-175). Cham: Springer Nature Switzerland.
  21. Serramazza, D. I., Nguyen, T. L., & Ifrim, G. (2024, August). Improving the evaluation and action ability of explanation methods for multivariate time series classification. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (pp. 177-195). Cham: Springer Nature Switzerland.
  22. Šimić, I., Sabol, V., & Veas, E. (2021). XAI methods for neural time series classification: A brief review. arXivpreprint arXiv:2108.08009.
  23. Simonyan, K., Vedaldi, A., & Zisserman, A. (2013). Deep inside convolutional networks: Visualising image classification models and saliency maps. arXivpreprint arXiv:1312.6034.
  24. Škrlj, B., Džeroski, S., Lavrač, N., & Petković, M. (2020). Feature importance estimation with selfattention networks. In ECAI 2020 (pp. 1491-1498). IOS Press.
  25. Sundararajan, M., Taly, A., & Yan, Q. (2017, July). Axiomatic attribution for deep networks. In International conference on machine learning (pp. 3319-3328). PMLR.
  26. Theissler, A., Spinnato, F., Schlegel, U., & Guidotti, R. (2022). Explainable AI for time series classification: a review, taxonomy and research directions. Ieee Access, 10, 100700-100724. https://doi.org/10.1109/ACCESS.2022.3207765
  27. Van Zyl, C., Ye, X., & Naidoo, R. (2024). Harnessing explain able artificial intelligence for feature selection in time series energy forecasting: A comparative analysis of Grad-CAM and SHAP. Applied Energy, 353, 122079.
  28. Veerappa, M., Anneken, M., Burkart, N., & Huber, M. F. (2022). Validation of XAI explanations for multivariate time series classification in the maritime domain. Journal of Computational Science, 58, 101539. https://doi.org/10.1016/j.jocs.2021.101539
  29. Wang, Y., Zhang, T., Guo, X., & Shen, Z. (2024). Gradient based feature attribution in explainable ai: A technical review. arXivpreprint arXiv:2403.10415.
  30. Yeh, C. K., Hsieh, C. Y., Suggala, A., Inouye, D. I., & Ravikumar, P. K. (2019). On the (in) fidelity and sensitivity of explanations. Advances in neural information processing systems, 32.
  31. Zenário, P. A. G. (2023). Explainable artificial intelligence for multivariate time series regression models(Master's thesis, Universidadede Lisboa)