<(From left) Graduate School of Artificial Intelligence Professor Kim Jae-cheol, Dr. Lee Chan-woo, Park Yong-jin, doctoral candidate, and Professor Choi Jae-sik>
A research team led by Professor Choi Jae-sik of KAIST's Kim Jae-chul AI Graduate School announced that, in collaboration with Kakao Bank, they have developed an accelerated explanation technology that can explain the basis for decisions made by artificial intelligence (AI) models in real time. The results of this research significantly increase the possibility of practical application of Explainable Artificial Intelligence (XAI) technology in fields that require real-time decision-making, such as financial services, by achieving an average processing speed of 8.5 times and up to 11 times faster than existing explanation algorithms for AI model predictions.
In the financial sector, clear explanations of decisions made by AI systems are essential. Particularly in services that directly concern customer rights, such as loan screening and anomaly detection, regulatory requirements to transparently present the basis for AI model decisions are becoming increasingly strict. However, traditional explainable artificial intelligence (XAI) technologies require iterative calculations of hundreds to thousands of baselines to generate accurate explanations, resulting in huge computational costs. This has been the main factor limiting the application of XAI technology in real-time service environments.
To address this issue, Professor Choi's research team developed the ABSQR (Amortized Baseline Selection via Rank-Revealing QR) framework to speed up explanation algorithms. ABSQR noticed that the value function matrix generated during the AI model explanation process has a low-rank structure. Introduced a method to select only a few important baselines from the hundreds available. This significantly reduced the computational complexity, which was previously proportional to the number of baselines, to be proportional only to the number of selected significant baselines, maximizing computational efficiency while maintaining explanatory accuracy.
Specifically, ABSQR operates in two stages. The first stage uses singular value decomposition (SVD) and Rank-Revealing QR decomposition techniques to systematically select significant baselines. Unlike existing random sampling methods, this is a deterministic selection method that aims to preserve information recovery, ensuring explanatory accuracy while significantly reducing computation. In the second stage, we introduce an amortization inference mechanism. It reuses the baseline precomputed weights through cluster-based search, allowing the system to provide an explanation of the model's predicted results in a real-time service environment without repeatedly evaluating the model. The research team verified the superiority of ABSQR through experiments on various real-world datasets. Tests on standard datasets across five domains: finance, marketing, and demographics showed that ABSQR achieved an average of 8.5 times faster processing speed than existing explanation algorithms using all baselines, with a maximum speed improvement of more than 11 times. Furthermore, we minimized the decrease in explanatory accuracy due to speed acceleration and maintained up to 93.5% explanatory accuracy compared to the baseline algorithm. This level is sufficient to meet the explanatory quality required in actual applications.
< ABSQR フレームワークの概要。 (1) ベースライン選択ステージでは、価値関数行列の低ランク構造を利用して少数の主要なベースラインのみを選択します。(2) 高速検索ステージでは、クラスターに基づいて事前に計算されたベースライン重み係数を再利用します。これにより、ベースラインの数に比例していた計算の複雑さが、選択された主要なベースラインの数にのみ比例するように劇的に軽減されます。 >
Kakao Bank officials said, “We will continue to work tirelessly on research and development to improve the reliability and convenience of financial services and introduce innovative financial technology that our customers can experience.'' KAIST co-lead authors Chanwoo Lee and Youngjin Park explained the importance of this study: “We have demonstrated that this methodology can solve important acceleration problems for real-time applications in the financial sector and provide users with the reasoning behind learning model decisions in real time.”Furthermore, “this study provides new insights into unnecessary computational components and the selection of critical baselines in explanation algorithms, contributing substantially to improving the efficiency of explanation techniques.” This research was co-authored by Lee Chang-woo and Park Young-jin, doctoral candidates at KAIST's Kim Jae-chul Graduate School of Artificial Intelligence, and Lee Hyung-eun and Yoo Ye-eun, researchers at the Kakao Bank Institute of Financial Technology, and was presented on November 12 at CIKM 2025 (ACM International Conference on Information and Knowledge Management), the world's most prestigious academic conference in the field of information and knowledge management. *Paper title: Amortized Baseline Selection via Rank-Revealing QR for Efficient Model Preparation
*Author information:
*Author information: DOI: https://doi.org/10.1145/3746252.3761036
- Co-lead authors: Chanwoo Lee (KAIST Kim Jae-cheol AI Graduate School), Youngjin Park (KAIST Kim Jae-cheol AI Graduate School), Hyogeun Lee (KakaoBank), Yeeun Yoo (KakaoBank)
- Co-authors: Daehee Han (Kakao Bank), Junho Choi (KAIST Kim Jae-cheol AI Graduate School), Kunhyung Kim (KAIST Kim Jae-cheol AI Graduate School)
- Corresponding authors: Nari Kim (KAIST Kim Jae-cheol AI Graduate School), Jae-sik Choi (KAIST Kim Jae-cheol AI Graduate School)
The results of this research were carried out through Kakao Bank's industry-academia research project “Advanced Research on Explainable Artificial Intelligence Algorithms in the Financial Field'' and the Ministry of Science, Technology and Communication/Institute of Information and Communication Technology Planning and Evaluation (IITP) supported project “Development of explainable artificial intelligence technology that provides explainability through plug-and-play and verification of providing explanations to AI systems''.
/Open to the public. This material from the original organization/author may be of a contemporary nature and has been edited for clarity, style, and length. Mirage.News does not take any institutional position or position, and all views, positions, and conclusions expressed herein are those of the authors alone. Read the full text here.
