Basic research on interpretability and causality in modeling time-dependent phenomena

2023 Fiscal Year Final Research Report

• Project/Area Number: 22K21278
• Research Category: Grant-in-Aid for Research Activity Start-up
• Allocation Type: Multi-year Fund
• Review Section: 1001:Information science, computer engineering, and related fields
• Research Institution: Kobe University
• Principal Investigator: Komatsu Mizuka 神戸大学, システム情報学研究科, 助教 (80856766)
• Project Period (FY): 2022-08-31 – 2024-03-31
• Keywords: SciML / 深層学習 / 応用代数 / 微分方程式 / 時系列データ

Outline of Final Research Achievements

In recent years, research utilizing techniques from both machine learning and computational science has been conducted in the interdisciplinary field. However, there is room for advancement particularly in the interpretability of models and treatment of causality in phenomena within this domain. Based on this, in this study, several approaches were developed with a focus on these aspects of modeling aimed at acquiring scientific insights into observed time-series data.
In broad terms, two researches were conducted as follows:
parameter estimation methods based on algebraic techniques for the case where the governing equations of the phenomenon are at least partially known and deep learning-based approaches are effective when the equations governing the phenomena are unknown.

Free Research Field

応用代数，モデリング

Academic Significance and Societal Importance of the Research Achievements

時系列データから現象に関する知見の獲得を目指す場合，用いられるモデルとして，支配方程式のような解釈性の高いモデルと，深層学習ベースの解釈性の低いモデルがある．前者に関して，モデルの構造によっては，データからパラメータが一意に定まらず，解析結果の信憑性が担保されないという問題がある．本研究では，これを回避すべく代数に基づく推定手法を提案した．後者に関して，時系列グラフ等を導入し解釈性を向上した．