Woojin Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Woojin Lee has a total of 46 co-authors. Here are the top co-authors by number of shared publications:
- Jaewook Lee
Seoul National University
4 shared publications - Hoki Kim
Department of Industrial Security, Chung-Ang University, Seoul, South Korea
3 shared publications - Youngdoo Son
Department of Industrial and Systems Engineering, Dongguk University-Seoul, Seoul, Republic of Korea
3 shared publications Daehun Yoo
2 shared publications- Hyeonjin Kim
Department of Computer Science and Artificial Intelligence, Dongguk University, Seoul 04620, Republic of Korea
2 shared publications Jaehwan Lee
Dongguk Univ Seoul, Dept Ind & Syst Engn, Seoul 04620, South Korea
2 shared publications
Latest Publications
- Open Access
- 2 September 2026
- Published by Elsevier in SSRN Electronic Journal
Leakage-aware spatial and temporal validation of flood susceptibility models on Jeju Island
Study Region: Jeju Island, Republic of Korea: a 1,833 km² basaltic shield volcano whose permeable lava sequence keeps the incised radial channels (geoncheon) dry between storms and routes runoff to a densely settled coastal fringe, while orographic rainfall peaks inland.Study Focus: We tested whether flood-susceptibility skill survives evaluation designs respecting the mapped observation units. A 398,878-cell inventory from 1,795 inundation polygons spanning seven flood years (2006–2019) was audited for leakage and partitioned by event and polygon before pixel extraction. Frequency ratio, regularised logistic regression, random forest and gradient-boosted trees were compared on identical folds under a random pixel split, grouped spatial blocks, leave-one-event-out validation and a 2019 temporal holdout, with uncertainty resampled from polygons and blocks.New Hydrological Insights for the Region: The random split shares 94.3% of polygons and all seven years; the area under the receiver operating characteristic curve (AUC) falls monotonically from 0.9019–0.9906 to 0.8000–0.8795, 0.7304–0.7819 and 0.7115–0.7711, inverting the model ordering. The flood label’s autocorrelation range, 19.1 km, exceeds any block a 73 by 41 km island can accommodate, so grouped skill only bounds geographic transfer from above. Orographic rainfall is spatially decoupled from mapped inundation: co-varying with lithology and valley depth, its negative attribution marks the interior-to-coast contrast, not reduced flood generation. Watershed-basin membership identifies location, not a catchment control. The maps are relative screening scores, not transferable predictions.
- 2 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLeakage-aware spatial and temporal validation of flood susceptibility models on Jeju Island
Study Region: Jeju Island, Republic of Korea: a 1,833 km² basaltic shield volcano whose permeable lava sequence keeps the incised radial channels (geoncheon) dry between storms and routes runoff to a densely settled coastal fringe, while orographic rainfall peaks inland.Study Focus: We tested whether flood-susceptibility skill survives evaluation designs respecting the mapped observation units. A 398,878-cell inventory from 1,795 inundation polygons spanning seven flood years (2006–2019) was audited for leakage and partitioned by event and polygon before pixel extraction. Frequency ratio, regularised logistic regression, random forest and gradient-boosted trees were compared on identical folds under a random pixel split, grouped spatial blocks, leave-one-event-out validation and a 2019 temporal holdout, with uncertainty resampled from polygons and blocks.New Hydrological Insights for the Region: The random split shares 94.3% of polygons and all seven years; the area under the receiver operating characteristic curve (AUC) falls monotonically from 0.9019–0.9906 to 0.8000–0.8795, 0.7304–0.7819 and 0.7115–0.7711, inverting the model ordering. The flood label’s autocorrelation range, 19.1 km, exceeds any block a 73 by 41 km island can accommodate, so grouped skill only bounds geographic transfer from above. Orographic rainfall is spatially decoupled from mapped inundation: co-varying with lithology and valley depth, its negative attribution marks the interior-to-coast contrast, not reduced flood generation. Watershed-basin membership identifies location, not a catchment control. The maps are relative screening scores, not transferable predictions.
- 19 August 2026
- Published by Elsevier in Research in International Business and Finance
- Vol. 91
Who provides liquidity in retail-dominated markets? Evidence from Korea
- 19 August 2026
- Published by Elsevier in Research in International Business and Finance
- Vol. 91
Who provides liquidity in retail-dominated markets? Evidence from Korea
- Open Access
- 5 June 2026
- Published by Elsevier in SSRN Electronic Journal
Who Provides Liquidity in Retail-Dominated Markets? Evidence from Korea
In the less liquid segment of the Korean equity market, foreign net trading exhibits the return patterns associated with risk-averse liquidity provision, a role that has been documented for retail investors in the U.S. setting. Using eleven years of weekly investor-level trading data, we show that net buying by foreigners predicts positive future returns at short horizons, with the magnitude of the abnormal return rising monotonically across illiquidity terciles. The corresponding patterns for institutional and individual investor flows differ in sign and statistical strength, and the foreign predictability in the less liquid segment is robust to alternative specifications of the trading-imbalance measure, the illiquidity proxy, the asset-pricing benchmark, the sample of stocks, and the procedure used to rank stocks by NIT. Three indirect microstructure tests on the information ordering across investor types, the contemporaneous price impact of trading, and the horizon decay of the long-short alpha are each consistent with the liquidity-provision interpretation. The pattern is consistent with a structural feature of the Korean market in which domestic individuals account for the majority of trading volume without displaying the contrarian or short-horizon predictive content of U.S. retail flow.
- 5 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessWho Provides Liquidity in Retail-Dominated Markets? Evidence from Korea
In the less liquid segment of the Korean equity market, foreign net trading exhibits the return patterns associated with risk-averse liquidity provision, a role that has been documented for retail investors in the U.S. setting. Using eleven years of weekly investor-level trading data, we show that net buying by foreigners predicts positive future returns at short horizons, with the magnitude of the abnormal return rising monotonically across illiquidity terciles. The corresponding patterns for institutional and individual investor flows differ in sign and statistical strength, and the foreign predictability in the less liquid segment is robust to alternative specifications of the trading-imbalance measure, the illiquidity proxy, the asset-pricing benchmark, the sample of stocks, and the procedure used to rank stocks by NIT. Three indirect microstructure tests on the information ordering across investor types, the contemporaneous price impact of trading, and the horizon decay of the long-short alpha are each consistent with the liquidity-provision interpretation. The pattern is consistent with a structural feature of the Korean market in which domestic individuals account for the majority of trading volume without displaying the contrarian or short-horizon predictive content of U.S. retail flow.
- Open Access
- 26 May 2026
- Published by Elsevier in SSRN Electronic Journal
SeMoBridge: A Semantic Modality Bridge for Intra-Modal Alignment in CLIP
While Contrastive Language-Image Pretraining (CLIP) excels at zero-shot tasks by aligning image and text embeddings, its performance in few-shot classification is hindered by a critical limitation: intra-modal misalignment. This issue, caused by a persistent modality gap and CLIP's exclusively inter-modal training objective, leaves the embedding spaces uncalibrated, making direct image-to-image comparisons unreliable. Existing methods attempt to address this by refining similarity logits or by computationally expensive per-sample optimization.To overcome these challenges, we introduce SeMoBridge, a lightweight yet powerful approach that directly addresses the misalignment. Our method maps images into the text modality, while keeping their semantic content intact through what we call a Semantic Modality Bridge. SeMoBridge is closed-form and can optionally be trained through multi-modal supervision, combining image and text-alignment losses to optimize the projection. Experiments show that the trained version, SeMoBridge-T, requires only a fraction of the training time while overall outperforming other methods, particularly in low-data scenarios (1, 2, and 4 shots).
- 26 May 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessSeMoBridge: A Semantic Modality Bridge for Intra-Modal Alignment in CLIP
While Contrastive Language-Image Pretraining (CLIP) excels at zero-shot tasks by aligning image and text embeddings, its performance in few-shot classification is hindered by a critical limitation: intra-modal misalignment. This issue, caused by a persistent modality gap and CLIP's exclusively inter-modal training objective, leaves the embedding spaces uncalibrated, making direct image-to-image comparisons unreliable. Existing methods attempt to address this by refining similarity logits or by computationally expensive per-sample optimization.To overcome these challenges, we introduce SeMoBridge, a lightweight yet powerful approach that directly addresses the misalignment. Our method maps images into the text modality, while keeping their semantic content intact through what we call a Semantic Modality Bridge. SeMoBridge is closed-form and can optionally be trained through multi-modal supervision, combining image and text-alignment losses to optimize the projection. Experiments show that the trained version, SeMoBridge-T, requires only a fraction of the training time while overall outperforming other methods, particularly in low-data scenarios (1, 2, and 4 shots).
- Open Access
- 22 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 82166-82179
Exploring the Effect of Multi-Step Ascent in Sharpness-Aware Minimization
Recently, Sharpness-Aware Minimization (SAM) has shown state-of-the-art performance by seeking flat minima. To minimize the maximum loss within a neighborhood in the parameter space, SAM uses an ascent step, which perturbs the weights along the direction of gradient ascent with a given radius. While single-step or multi-step can be taken during ascent steps, previous studies have shown that multi-step ascent SAM rarely improves generalization performance. However, this phenomenon is particularly interesting because the multi-step ascent is expected to provide a better approximation of the maximum neighborhood loss. Therefore, in this paper, we analyze the effect of the number of ascent steps and investigate the difference between both single-step ascent SAM and multi-step ascent SAM. We identify the effect of the number of ascents on SAM optimization and reveal that single-step ascent SAM and multi-step ascent SAM exhibit distinct loss landscapes. Based on these observations, we finally suggest a simple modification that can mitigate the suboptimal generalization of multi-step ascent SAM.
- 22 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 82166-82179
Open AccessExploring the Effect of Multi-Step Ascent in Sharpness-Aware Minimization
Recently, Sharpness-Aware Minimization (SAM) has shown state-of-the-art performance by seeking flat minima. To minimize the maximum loss within a neighborhood in the parameter space, SAM uses an ascent step, which perturbs the weights along the direction of gradient ascent with a given radius. While single-step or multi-step can be taken during ascent steps, previous studies have shown that multi-step ascent SAM rarely improves generalization performance. However, this phenomenon is particularly interesting because the multi-step ascent is expected to provide a better approximation of the maximum neighborhood loss. Therefore, in this paper, we analyze the effect of the number of ascent steps and investigate the difference between both single-step ascent SAM and multi-step ascent SAM. We identify the effect of the number of ascents on SAM optimization and reveal that single-step ascent SAM and multi-step ascent SAM exhibit distinct loss landscapes. Based on these observations, we finally suggest a simple modification that can mitigate the suboptimal generalization of multi-step ascent SAM.
- Open Access
- 14 April 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (8) , 3813
Explainable Graph-Based Golf Swing Analysis Integrating Club and Body Keypoints for Ball Flight Outcome Prediction
We propose a graph-based, explainable golf swing analysis framework that integrates human body keypoints with golf club keypoints to predict ball flight outcomes. We collected 321 driver swing sequences from six amateur golfers in a controlled studio setting, synchronizing monocular swing videos with TrackMan-derived ball trajectory measurements. Using a unified spatial–temporal graph that jointly models body joints and golf club keypoints, we trained graph neural networks (ST-GCN and STGAT) to perform three prediction tasks: Spin Axis and Launch Direction (classification) and Ball Speed (regression). Model performance was evaluated using AUC and accuracy for classification, and R2 and RMSE for regression. STGAT achieved the best overall performance, reaching an AUC of 0.9188 and an accuracy of 78.33% for Spin Axis classification, an AUC of 0.7599 and an accuracy of 69.81% for Launch Direction classification, and an R2 of 0.6925 with an RMSE of 6.4020 for Ball Speed prediction, outperforming traditional machine learning baselines. Finally, we applied Integrated Gradients to quantify the importance of both body and club keypoints across swing phases, enabling interpretable, phase-specific feedback to support individualized swing refinement.
- 14 April 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (8) , 3813
Open AccessExplainable Graph-Based Golf Swing Analysis Integrating Club and Body Keypoints for Ball Flight Outcome Prediction
We propose a graph-based, explainable golf swing analysis framework that integrates human body keypoints with golf club keypoints to predict ball flight outcomes. We collected 321 driver swing sequences from six amateur golfers in a controlled studio setting, synchronizing monocular swing videos with TrackMan-derived ball trajectory measurements. Using a unified spatial–temporal graph that jointly models body joints and golf club keypoints, we trained graph neural networks (ST-GCN and STGAT) to perform three prediction tasks: Spin Axis and Launch Direction (classification) and Ball Speed (regression). Model performance was evaluated using AUC and accuracy for classification, and R2 and RMSE for regression. STGAT achieved the best overall performance, reaching an AUC of 0.9188 and an accuracy of 78.33% for Spin Axis classification, an AUC of 0.7599 and an accuracy of 69.81% for Launch Direction classification, and an R2 of 0.6925 with an RMSE of 6.4020 for Ball Speed prediction, outperforming traditional machine learning baselines. Finally, we applied Integrated Gradients to quantify the importance of both body and club keypoints across swing phases, enabling interpretable, phase-specific feedback to support individualized swing refinement.
- 21 February 2026
- Published by Springer Nature in Computational Economics
- p. 1-31
Large Language Models as Financial Analysts: Sector-Aware Reasoning for Investment Decisions
- 21 February 2026
- Published by Springer Nature in Computational Economics
- p. 1-31
Large Language Models as Financial Analysts: Sector-Aware Reasoning for Investment Decisions
- Open Access
- 2 February 2026
- Published by MDPI AG in Journal of Theoretical and Applied Electronic Commerce Research
- Vol. 21 (2) , 45
Disconfirmation Dynamics in Service Recovery: Insights from Online Customer Reviews
This paper analyzes the effects of service recovery and the double-deviation phenomenon on customer satisfaction using an analysis of online reviews. With the application of Word2Vec and AFINN sentiment analysis to 4793 Skytrax review comments, this paper bridges the gap between conceptual theories about customer behavior and big data analysis. Findings from this research indicate that although overall types of service recovery increase customer satisfaction, explanation is the most preferable yet effective type of recovery. Most importantly, this analysis shows that service recovery disconfirmation has been found to be a highly important moderator. When service efforts involve intangible actions, such as an apology or explanation, that fail to satisfy customers, they tend to perform worse than no recovery at all, thus supporting the double-deviation phenomenon. Compensation, however, has been found to be an efficient type of recovery since its impacts are not reduced much due to service recovery disconfirmation. These insights provide service providers with critical guidance on prioritizing transparent communication and avoiding poorly executed intangible recoveries.
- 2 February 2026
- Published by MDPI AG in Journal of Theoretical and Applied Electronic Commerce Research
- Vol. 21 (2) , 45
Open AccessDisconfirmation Dynamics in Service Recovery: Insights from Online Customer Reviews
This paper analyzes the effects of service recovery and the double-deviation phenomenon on customer satisfaction using an analysis of online reviews. With the application of Word2Vec and AFINN sentiment analysis to 4793 Skytrax review comments, this paper bridges the gap between conceptual theories about customer behavior and big data analysis. Findings from this research indicate that although overall types of service recovery increase customer satisfaction, explanation is the most preferable yet effective type of recovery. Most importantly, this analysis shows that service recovery disconfirmation has been found to be a highly important moderator. When service efforts involve intangible actions, such as an apology or explanation, that fail to satisfy customers, they tend to perform worse than no recovery at all, thus supporting the double-deviation phenomenon. Compensation, however, has been found to be an efficient type of recovery since its impacts are not reduced much due to service recovery disconfirmation. These insights provide service providers with critical guidance on prioritizing transparent communication and avoiding poorly executed intangible recoveries.
- Open Access
- 16 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 71053-71065
LM-CLIP: Adapting Positive Asymmetric Loss for Long-Tailed Multi-Label Classification
Accurate multi-label image classification is essential for real-world applications, especially in scenarios with long-tailed class distributions, where some classes appear frequently while others are rare. This imbalance often leads to biased models that struggle to accurately recognize underrepresented classes. Existing methods either trade off performance between head and tail classes or rely on image captions, limiting adaptability. To address these limitations, we propose LM-CLIP, a novel framework built around a unified loss function. Our Balanced Asymmetric Loss (BAL) extends traditional asymmetric loss by emphasizing the gradients of rare positive samples where the model is uncertain, mitigating bias toward dominant classes. This is complemented by a contrastive loss that pushes negative samples further from the decision boundary, creating a more optimal embedding space even in long-tailed scenarios. These loss functions together ensure balanced performance across all classes. Our framework is built on pre-trained models utilizing textual and visual features from millions of image-text pairs. Furthermore, we incorporate a dynamic sampling strategy that prioritizes rare classes based on their occurrence, which ensures effective training without compromising overall performance. Experiments conducted on VOC-MLT and COCO-MLT benchmarks demonstrate the effectiveness of our approach, achieving +4.66% and +8.14% improvements in mean Average Precision (mAP) over state-of-the-art methods. Our code is publicly available at https://github.com/damilab/lm-clip.
- 16 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 71053-71065
Open AccessLM-CLIP: Adapting Positive Asymmetric Loss for Long-Tailed Multi-Label Classification
Accurate multi-label image classification is essential for real-world applications, especially in scenarios with long-tailed class distributions, where some classes appear frequently while others are rare. This imbalance often leads to biased models that struggle to accurately recognize underrepresented classes. Existing methods either trade off performance between head and tail classes or rely on image captions, limiting adaptability. To address these limitations, we propose LM-CLIP, a novel framework built around a unified loss function. Our Balanced Asymmetric Loss (BAL) extends traditional asymmetric loss by emphasizing the gradients of rare positive samples where the model is uncertain, mitigating bias toward dominant classes. This is complemented by a contrastive loss that pushes negative samples further from the decision boundary, creating a more optimal embedding space even in long-tailed scenarios. These loss functions together ensure balanced performance across all classes. Our framework is built on pre-trained models utilizing textual and visual features from millions of image-text pairs. Furthermore, we incorporate a dynamic sampling strategy that prioritizes rare classes based on their occurrence, which ensures effective training without compromising overall performance. Experiments conducted on VOC-MLT and COCO-MLT benchmarks demonstrate the effectiveness of our approach, achieving +4.66% and +8.14% improvements in mean Average Precision (mAP) over state-of-the-art methods. Our code is publicly available at https://github.com/damilab/lm-clip.