Youngkwan Son | Yonsei University

About Me

I am a Ph.D. candidate in Industrial Engineering (Smart Cities) at Yonsei University, advised by Prof. Wooju Kim, with expected completion in February 2027. Trained as an architectural engineer, I work on Spatial AI for the built environment.

My research develops auditable methods that turn heterogeneous floor plans into topology-centered building knowledge graphs, and grounds language-based spatial reasoning in explicit graph evidence. My planned postdoctoral research extends this foundation toward topology-aware multimodal representations that connect floor plans, indoor 3D observations, building knowledge graphs, and language for verifiable spatial reasoning.

Before my Ph.D., I co-founded NEOSPECTRA and served as its CTO, leading R&D in remote sensing and 3D computer vision.

Research Interests

Spatial Intelligence for the Built Environment: Spatial AI that represents and reasons about buildings through ontology-grounded building knowledge graphs and auditable, graph-grounded spatial reasoning with LLMs (GraphRAG)
3D Scene Understanding: Spatial AI that perceives and structures 3D spaces through multimodal 3D scene understanding, ontology-grounded 3D scene graphs, topology-aware representations, and cross-modal 3D scene alignment
Explainable and Trustworthy AI: quantitative evaluation of XAI methods for vision and language models, attribution-guided LLM data augmentation, and reliability of LLM-based text analysis

Publications

Under Review

An Attribution-Aware Framework for Constructing and Validating Topology-Centered Building Knowledge Graphs from Heterogeneous Floor Plans

Y. Son and W. Kim*

Under Review · Advanced Engineering Informatics, 2026

An attribution-aware G0–G1–G2 RDF pipeline that builds topology-centered building knowledge graphs from heterogeneous floor plans, evaluated on 22,479 plans with SHACL validation, negative-control audits, and coverage-aware competency questions.
Stage-Aware GraphRAG for Evidence-Grounded Question Answering over Topology-Centered Building Knowledge Graphs

Y. Son and W. Kim*

Under Review · Automation in Construction, 2026

A GraphRAG framework that keeps G0–G1–G2 provenance, separates proximity from navigability, and returns the supporting triples and relation paths behind every answer so spatial reasoning can be verified.
Evaluating Predictive Performance and Pixel-Level Feature Consistency in Remote Sensing Super-Resolution via Explainable AI

Y. Son

Under Review · IEEE Transactions on Geoscience and Remote Sensing, 2026

Measures how deep-learning super-resolution of aerial imagery across ground sample distances affects downstream tasks, and uses XAI to check that models attend to the same pixel-level features as in native high-resolution data.
Confusion-Targeted Data Augmentation With Large Language Models for Few-Shot Text Classification

Y. Son, S. Kwon, J. Shin, and W. Kim*

Under Review · IEEE Access, 2026

CTDA turns out-of-fold classifier errors into targets for LLM-generated training data, with pairwise relabeling and target-consistency filtering. Across 12 dataset–shot–backbone settings on Banking77, HWU64, and TREC-50, it reaches the highest aggregate accuracy (75.48%) and macro-F1 (74.68%) among the compared augmentation methods.
Fine-Tuned Large Language Models for Theory-Aligned Social Science Content Analysis: Evidence From Collaborative Learning Discourse

Y. Shin, Y. Son*, H. Lee, and M. Kim

Under Review · Social Science Computer Review, 2026

An LLM-based coding pipeline for socially shared metacognitive regulation, compared with human coders on agreement, performance, efficiency, and error patterns.
AI-Assisted Collaborative Problem-Solving Discourse Coding for Learning Analytics: XAI-Informed Review Evidence

Y. Shin, Y. Son*, and Y. Im

Under Review · Education and Information Technologies, 2026

A BERT-based, XAI-informed framework for automated collaborative problem-solving discourse coding.

Journal Articles

Explainer-Agnostic Quantitative Evaluation Framework for XAI Techniques in Generative Language Models

S. Kwon, S. Kim, Y. Kim, J. Yum, B. Kim, Y. Son, and W. Kim*

Journal of Intelligence and Information Systems, 31(4), 477–492, 2025

An explainer-agnostic framework for quantitatively evaluating XAI techniques in generative language models.

DOI

Work in Progress

LogicGraph3D: Bidirectional Neural-Logic Learning for Ontology-Grounded Indoor 3D Panoptic Scene Graphs

Y. Son

Working Paper · Targeting IEEE Robotics and Automation Letters

A neural-symbolic framework that jointly predicts 3D object instances and spatial relations, aligns scene graphs with an ontology, and uses logical feedback to refine predictions, with planned evaluation on 3RScan/3DSSG, ScanNet200, and ScanNet++.

* Corresponding author

Research Experience

Auditable Building Knowledge Graphs and Spatial GraphRAGMar. 2023 – Present

Doctoral Research, Yonsei University

Built an end-to-end pipeline that constructs and validates topology-centered building knowledge graphs from 22,479 heterogeneous floor plans, and designed a stage-aware GraphRAG framework with a 15,937-instance evidence-grounded spatial QA benchmark (BOT, RDF, SHACL, SPARQL, provenance tracking).

Explainable AI for Computer Vision and Large Language ModelsJan. 2025 – Present

Research Project, Yonsei University

Applied and evaluated XAI methods (SHAP, LIME, Grad-CAM) for vision and NLP models and LLM-specific techniques (SPEX, MExGen, ContextCite); contributed to an evaluation framework that received the KIISS 2025 Excellent Paper Award.

Legal Knowledge Graph Construction and Ontology Modeling (AIHUB)Sep. 2023 – Feb. 2024

Research Project, Yonsei University

Designed a domain ontology for legal precedent data and built a knowledge graph model with SBERT, TransE, and RotatE.

3D Reconstruction and Spatial Computer Vision PipelinesDec. 2021 – Feb. 2023

Corporate R&D, NEOSPECTRA Co., Ltd.

Co-developed a patented pipeline that removes transient objects from multi-view aerial imagery before 3D reconstruction, meshed indoor point clouds with Poisson surface reconstruction, and built segmentation, inpainting, image-quality classification, and super-resolution models for aerial imagery.

Solar Irradiance and Photovoltaic Power ForecastingDec. 2020 – Feb. 2021

Consulting Project, Architectural Environment Planning Lab, Chung-Ang University

Forecast solar irradiance and PV power generation with LSTM networks for renewable energy use in buildings.

Education

Ph.D. Candidate in Industrial Engineering (Smart Cities)Mar. 2023 – Present

Yonsei University · Advisor: Prof. Wooju Kim · Expected Feb. 2027

Dissertation (working title): Staged Knowledge Graphs of Residential Floor Plans: Construction, Validation, and Effects on LLM-Based Question Answering

M.S. in Digital AnalyticsMar. 2019 – Aug. 2020

Yonsei University Graduate School (now College of Artificial Intelligence) · Advisor: Prof. Wooju Kim

Thesis: Comparative study of I/O structure, learning method and model type for predicting performance of data center equipment via deep learning

B.S. in Architectural EngineeringMar. 2009 – Feb. 2015

Yonsei University

Professional Experience

Co-Founder & CTO, AI Team LeadNov. 2021 – May 2023

NEOSPECTRA Co., Ltd.

Led corporate R&D in remote sensing and computer vision for spatial data analysis.

MIRAE ISE Co., Ltd. (Mirae Structural Engineering)Mar. 2015 – Aug. 2016

Architectural engineering firm

Honors & Patent

Excellent Paper Award, KIISS Fall Conference 2025Nov. 2025

BK21 Research Fellowship, Ministry of Education, KoreaMar. 2025 – Present

National Smart City Innovative Talent Fellowship, MOLIT, KoreaJan. 2024 – Dec. 2025

Patent (Granted): Method and System for 3D Model Generation Using 2D ImagesJan. 2023

Korean Intellectual Property Office, No. 10-2495954 · Co-inventor