Jaejun Lee

Jaejun Lee

Korea Advanced Institute of Science and Technology
1 previous affiliation
38Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Jaejun Lee focuses on graph machine learning, graph neural networks, and knowledge graph representation. His early work explored multi-agent control systems and edge-deployed machine listening before transitioning to structural and relational reasoning on complex graphs. His current research investigates inductive reasoning, hyper-relational knowledge graphs, and generative models for fact synthesis.

Research Trajectory

The early trajectory of this research explored decentralized architectures for resilient multi-agent coordination. In 2016, this included Fuzzy Formation Controlling Phugoid Model-Based Multi-Agent Systems, which presented a robust control strategy for coordinating autonomous agents, such as gliders, that follow phugoid flight models. Subsequently, the focus shifted toward privacy-preserving machine listening via edge feature extraction. The development of Howl: A Deployed, Open-Source Wake Word Detection System in 2020 introduced a production-ready, browser-based toolkit that leveraged JavaScript for local inference, powering the Firefox Voice extension without relying on cloud processing.

By 2022, the research focus transitioned squarely into graph representation learning. Participation in Semantic Grasping Via a Knowledge Graph of Robotic Manipulation: A Graph Representation Learning Approach involved the creation of "roboKG," a knowledge graph designed to reason about object materials, tasks, and appropriate gripper configurations to enable more sophisticated robotic manipulation.

This foundational interest in knowledge graphs expanded significantly in 2023, particularly within the research direction of Inductive Reasoning for Dynamic Knowledge Graphs. To address graphs where facts contain extra context like dates or locations, he introduced the HyNT framework in Representation Learning on Hyper-Relational and Numeric Knowledge Graphs with Transformers. This dual-transformer architecture was designed to predict missing entities as well as specific numeric values. Parallel work addressed multimodal integration in VISTA: Visual-Textual Knowledge Graph Representation Learning. During the same year, a major contribution to the areas of Scalable Subgraph Reasoning for Inductive Knowledge Graph Completion and Topology Aware Inductive Reasoning in Knowledge Graphs was published as InGram: Inductive Knowledge Graph Embedding via Relation Graphs. This framework generated embeddings for entirely new entities and relations at inference time by constructing a relation graph to capture semantic affinities between different types of links.

Moving into theoretical foundations, work in the broader area of Structural Preservation and Stability in Graph Aggregation led to PAC-Bayesian Generalization Bounds for Knowledge Graph Representation Learning in 2024. This research introduced the ReED (Relation-aware Encoder-Decoder) framework to theoretically demonstrate why common empirical techniques, such as weight normalization and mean aggregation, are essential for a model's ability to generalize to unseen data.

In 2025, the focus on the research direction of Inductive Relation Prediction via Structural Patterns culminated in Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs. This paper proposed MAYPL, a purely structure-driven message-passing framework that computes representations based solely on how entities and relations are connected, enabling inductive reasoning on entirely new elements. Collaborative theoretical work continued with Stability and Generalization Capability of Subgraph Reasoning Models for Inductive Knowledge Graph Completion, which contributed to the area of Inductive Relation Reasoning in Knowledge Graphs. This study introduced the Relational Tree Mover’s Distance (RTMD) to prove that a model's generalization capability is directly proportional to its structural stability in response to small input changes.

Most recently, the trajectory has advanced into generative paradigms. The 2026 paper Generative Representation Learning on Hyper-relational Knowledge Graphs via Masked Discrete Diffusion introduced KREPE, a framework that moves beyond traditional link prediction to generate entirely new, valid facts from scratch using contextual message passing and masked discrete diffusion.

Current Directions

The most recent papers in the evidence address the theoretical limits and generative capabilities of hyper-relational knowledge graphs. Current work investigates fact generation paradigms, moving beyond traditional "fill-in-the-blank" link prediction to synthesize entirely new facts from highly incomplete queries using masked discrete diffusion. Additionally, ongoing research explores the theoretical foundations of subgraph reasoning, specifically how structural stability—measured by metrics like Relational Tree Mover’s Distance—dictates a graph neural network's ability to generalize to unseen entities and relations in inductive settings.

Profile updated

See something wrong? Report a mistake