Hello! I am Bonan Yang, a Computer Science Ph.D. candidate at Southern Illinois University Carbondale, advised by Dr. Gunes Ercal. I specialize in complex network analysis, knowledge graph systems, and graph machine learning. My work spans representation learning on large-scale graphs, semantic reasoning over knowledge bases, and building graph infrastructure for real-world applications. Lately I have been doing research on the structure between enterprise data and LLMs. I don't call it knowledge graphs or ontology here, because I don't think those alone solve the problem. They are some of the possible ways. See my article.
Graph modeling of enterprise data, graph representation learning,
knowledge graph systems, scalable graph infrastructure.
Specializations: Graph Modeling of Enterprise Data, Entity Resolution, Ontology Engineering (OWL, RDF, SHACL, SPARQL), LLM Agent Evaluation, Knowledge Graphs, Graph Neural Networks (GCN, GraphSAGE, GAT, Graph Autoencoders, Spatio-temporal GNNs), Complex Network Analysis, NLP & LLM, GraphRAG, Vector Databases, End-to-End ML Deployment
Languages & Databases: Python, SQL, SPARQL, Cypher, C/C++, R, DuckDB, Neo4j, Databricks, Milvus, Pinecone, HDFS
Platforms & Tools: Databricks, rdflib, owlrl, pySHACL, Streamlit, Linux, Docker, AWS, GCP, Apache Spark, PyTorch, PyG, Transformers, vLLM/SGLang, CUDA, Git, NetworkX, Gephi
Oct 2026: Presented at the IIN conference on what course graph research means for reskilling and workforce development.
Sep 2026 – present: Cross-institutional course prerequisite networks: aligning course data from 44 U.S. public universities to separate dependencies inherent to the knowledge from conventions set by each school.
Jun 2026: Poster at NetSci 2026 (Boston) with early results on cross-institutional prerequisite networks: how often the same prerequisite relation recurs across schools, and which dependency chains hold up everywhere.
Apr – Nov 2026: Graph Algorithm Engineer intern at LakeFusion (Plano, TX), graph modeling of enterprise data. Found that an LLM answering questions over enterprise tables can only guess at the structure; now trying to compute the structure directly from the tables, without an LLM, and hand it to the LLM afterwards.
Mar 2026: ASEE IL/IN 2026 Best Paper Award for a paper on curriculum–industry knowledge graph alignment.
Jan 2026: Released HPO Clinical Knowledge Graph — HPO-KG built from Human Phenotype Ontology for rare disease diagnosis. 19934 phenotype nodes, 13012 disease nodes, 200000+ edges.
Ontology from Structured Data without LLM (2026)
CSV tables in, reasoned tables out. Keys, references, hierarchies and duplicate records computed by SQL; what the data cannot settle is asked of a person; reasoning by owlrl, pySHACL and SPARQL. Tested on hundreds of randomly generated schemas and five simulated industries with planted ground truth. Three-way comparison on Databricks Genie: the reasoning layer holds, the hand-off to the agent does not. Article
Cross-Institutional Course Prerequisite Networks (2026)
Course catalogs from 44 U.S. public universities, 50K+ courses aligned into 3348 cross-school courses. One question: when a course requires a prerequisite, is that because the knowledge needs it, or because this school has always arranged it that way? The two are mixed together today and nobody has separated them. Only after that can you talk about which transfer credits to accept, which parts of a curriculum can be changed and which cannot, and what someone switching fields actually has to make up.
Large-scale Wikipedia Knowledge Graph (2025 – 2026)
Processed 18.8M page nodes and 747M link edges with HDFS and Spark. Generated entity embeddings with FastText, indexed in Milvus for semantic search. Supporting downstream GraphRAG and QA systems.
Course Graph Visualization Platform (2023 – Present)
Full-stack web app (Flask + Neo4j + Cytoscape.js) for visualizing course prerequisite networks. Deployed with Docker, Nginx, and SSL. Live at coursegraphs.com
GNN-Based Molecular Toxicity Prediction (2025)
Implemented GCN and GAT in PyTorch/PyG for multi-label toxicity prediction on Tox21 dataset (7.8K compounds). Achieved 0.81 ROC-AUC under severe class imbalance (1:22).
3D Object Quality Assessment via Graph Learning (2024 – Present)
Transformed 3D printing point clouds into graphs. Designed Graph Autoencoder with contrastive learning for similarity retrieval and deviation localization in personalized manufacturing.
Curriculum-Industry Skill Gap Analysis (2025 – Present)
Built knowledge graph integrating O*NET skill ontology with STEM course syllabi (462 courses, 3556 skills, 770 jobs). Applied semantic encoding and multi-hop reasoning for course-skill-job alignment. Providing data-driven insights for Illinois State Board of Education.
