Sangyeop Kim
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Sangyeop Kim has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Sungzoon Cho
Department of Industrial Engineering, Seoul National University, Seoul, Republic of Korea
3 shared publications Cholhwan Jung
Department of Industrial Engineering, Seoul National University, Gwanak-gu, Seoul, Republic of Korea
1 shared publication- Eojin Kim
Department of Industrial Engineering, Seoul National University, Gwanak-gu, Seoul, Republic of Korea
1 shared publication - Hangyeul Lee
Department of Industrial Engineering, Seoul National University, Seoul, Republic of Korea
1 shared publication - Hyunjong Kim
Department of Industrial Engineering, Seoul National University, Gwanak-gu, Seoul, Republic of Korea
1 shared publication - Jaewon Jung
Department of Mathematics, Korea University, Seoul, Republic of Korea
1 shared publication
Latest Publications
- 25 October 2026
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 77
Attention divergence via process-step masked self-supervised learning for semiconductor anomaly detection
- 25 October 2026
- Published by Elsevier in Advanced Engineering Informatics
- Vol. 77
Attention divergence via process-step masked self-supervised learning for semiconductor anomaly detection
- Open Access
- 22 April 2026
- Published by Elsevier in SSRN Electronic Journal
Confidence-guided dual-adapter fusion for multimodal aspect-based sentiment analysis
Multimodal Aspect-Based Sentiment Classification (MASC) aims to determine the sentiment polarity toward a specific aspect term given an image-text pair. While most existing methods assume that the accompanying image provides complementary cues for sentiment understanding, this assumption is frequently violated in real-world social media data, where images are often noisy or irrelevant to the target aspect. To address this challenge, we propose COnfidence-guided Dual-Adapter fusion (CODA), a novel framework that reframes MASC as an adaptive multi-view fusion problem. CODA constructs complementary unimodal and multimodal views by independently training low-rank adaptation adapters on a shared frozen multimodal large language model backbone. The predictions from both views are calibrated via temperature scaling, and a meta-feature vector encoding the calibrated probabilities, confidence scores, predictive entropy, and cross-view discrepancy is fed into a meta-fusion classifier that adaptively integrates the two views at the instance level. Experiments on three benchmark datasets demonstrate that CODA achieves state-of-the-art performance, with particularly substantial gains under domain shift. Additional experiments and in-depth analyses confirm the importance of confidence calibration, the complementary roles of meta-feature components, and the robustness and generalizability of CODA across different backbones and in training-free settings. The code is available at https://github.com/hjkim811/CODA.
- 22 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessConfidence-guided dual-adapter fusion for multimodal aspect-based sentiment analysis
Multimodal Aspect-Based Sentiment Classification (MASC) aims to determine the sentiment polarity toward a specific aspect term given an image-text pair. While most existing methods assume that the accompanying image provides complementary cues for sentiment understanding, this assumption is frequently violated in real-world social media data, where images are often noisy or irrelevant to the target aspect. To address this challenge, we propose COnfidence-guided Dual-Adapter fusion (CODA), a novel framework that reframes MASC as an adaptive multi-view fusion problem. CODA constructs complementary unimodal and multimodal views by independently training low-rank adaptation adapters on a shared frozen multimodal large language model backbone. The predictions from both views are calibrated via temperature scaling, and a meta-feature vector encoding the calibrated probabilities, confidence scores, predictive entropy, and cross-view discrepancy is fed into a meta-fusion classifier that adaptively integrates the two views at the instance level. Experiments on three benchmark datasets demonstrate that CODA achieves state-of-the-art performance, with particularly substantial gains under domain shift. Additional experiments and in-depth analyses confirm the importance of confidence calibration, the complementary roles of meta-feature components, and the robustness and generalizability of CODA across different backbones and in training-free settings. The code is available at https://github.com/hjkim811/CODA.
- Open Access
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 4889-4894
Upcycling Candidate Tokens of Large Language Models for Query Expansion
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 4889-4894
Open AccessUpcycling Candidate Tokens of Large Language Models for Query Expansion
- Open Access
- 10 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 192403-192415
Fashion Image Retrieval With Vision–Language Model Guided Fine-Grained Textual Attributes and Cross-Domain Contrastive Optimization
Fashion image retrieval (FIR) enables consumers to discover products using visual queries, however current methods struggle with fine-grained fashion differences and cross-domain gaps between user photos and professional catalogs. We propose a novel two-stage framework that combines vision-language model guided fine-grained textual attributes with weighted contrastive optimization to address these fundamental limitations. The first stage employs Sigmoid Loss for Language Image Pre-training (SigLIP) with systematic twelve-attribute fashion annotations using binary classification to accommodate multi-positive scenarios where single garments can be validly described by multiple attributes. This approach naturally handles the semantic richness of fashion items, where multiple descriptions may accurately characterize individual garments. The second stage introduces Weighted Multi Normalized Temperature-Scaled Cross-Entropy (NT-Xent) loss that strategically prioritizes street-to-shop matching while eliminating complex negative sampling requirements, enabling effective cross-domain learning between casual user photography and professional catalog imagery. Extensive experiments on Street2Shop and DeepFashion datasets demonstrate that our framework outperforms existing methods in terms of retrieval accuracy and zero-shot classification performance. Comprehensive ablation studies confirm the effectiveness of each proposed component in capturing detailed fashion characteristics and enabling practical cross-domain matching. This paper contributes a comprehensive framework addressing fashion retrieval challenges, novel application of SigLIP with specialized contrastive learning, and practical improvements for real-world applications.
- 10 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 192403-192415
Open AccessFashion Image Retrieval With Vision–Language Model Guided Fine-Grained Textual Attributes and Cross-Domain Contrastive Optimization
Fashion image retrieval (FIR) enables consumers to discover products using visual queries, however current methods struggle with fine-grained fashion differences and cross-domain gaps between user photos and professional catalogs. We propose a novel two-stage framework that combines vision-language model guided fine-grained textual attributes with weighted contrastive optimization to address these fundamental limitations. The first stage employs Sigmoid Loss for Language Image Pre-training (SigLIP) with systematic twelve-attribute fashion annotations using binary classification to accommodate multi-positive scenarios where single garments can be validly described by multiple attributes. This approach naturally handles the semantic richness of fashion items, where multiple descriptions may accurately characterize individual garments. The second stage introduces Weighted Multi Normalized Temperature-Scaled Cross-Entropy (NT-Xent) loss that strategically prioritizes street-to-shop matching while eliminating complex negative sampling requirements, enabling effective cross-domain learning between casual user photography and professional catalog imagery. Extensive experiments on Street2Shop and DeepFashion datasets demonstrate that our framework outperforms existing methods in terms of retrieval accuracy and zero-shot classification performance. Comprehensive ablation studies confirm the effectiveness of each proposed component in capturing detailed fashion characteristics and enabling practical cross-domain matching. This paper contributes a comprehensive framework addressing fashion retrieval challenges, novel application of SigLIP with specialized contrastive learning, and practical improvements for real-world applications.
- 13 September 2025
- Published by Elsevier in Information Fusion
- Vol. 127
Fusing sequential recommender and ad-hoc planner for multi-faceted preference understanding in conversational recommendation
- 13 September 2025
- Published by Elsevier in Information Fusion
- Vol. 127
Fusing sequential recommender and ad-hoc planner for multi-faceted preference understanding in conversational recommendation
- 31 July 2025
- Published by Elsevier in Information Processing & Management
- Vol. 63 (1)
Human-guided collective LLM intelligence for strategic planning via two-stage information retrieval
- 31 July 2025
- Published by Elsevier in Information Processing & Management
- Vol. 63 (1)
Human-guided collective LLM intelligence for strategic planning via two-stage information retrieval