Detecting High-Potential SMEs with Heterogeneous Graph Neural Networks

Yijiashun Qi; Hanzhe Guo; Yijiazhen Qi

cs.LGFeb 23, 2026

Detecting High-Potential SMEs with Heterogeneous Graph Neural Networks

Authors: Yijiashun Qi, Hanzhe Guo, Yijiazhen Qi

Organizations: University of Michigan, USA · The University of Hong Kong, Hong Kong, China

Abstract

Small and Medium Enterprises (SMEs) constitute 99.9% of U.S. businesses and generate 44% of economic activity, yet systematically identifying high-potential SMEs remains an open challenge. We introduce SME-HGT, a Heterogeneous Graph Transformer framework that predicts which SBIR Phase I awardees will advance to Phase II funding using exclusively public data. We construct a heterogeneous graph with 32,268 company nodes, 124 research topic nodes, and 13 government agency nodes connected by approximately 99,000 edges across three semantic relation types. SME-HGT achieves an AUPRC of 0.621 0.003 on a temporally-split test set, outperforming an MLP baseline (0.590 0.002) and R-GCN (0.608 0.013) across five random seeds. At a screening depth of 100 companies, SME-HGT attains 89.6% precision with a 2.14 lift over random selection. Our temporal evaluation protocol prevents information leakage, and our reliance on public data ensures reproducibility. These results demonstrate that relational structure among firms, research topics, and funding agencies provides meaningful signal for SME potential assessment, with implications for policymakers and early-stage investors.

Figures & tables

Node Type	Count	Features
Company	32,268	390 ∗
Topic	124	2
FundingAgency	13	3
Edge Types
(Company, operates_in , Topic)
(Company, awarded_by , FundingAgency)

Table 1. Heterogeneous Graph Schema. ∗ 7 tabular + 383 text embedding dims.

Panel of Figure 1 — Figure 1. SME-HGT architecture. Input features are projected per node type to a shared dimension, then refined through three HGT layers with residual connections. Only company node embeddings are passed to the classifier. A vertical flowchart showing five blocks connected by arrows from bottom to top: Heterogeneous Graph input, Type-Specific Linear Projection, three repeated HGT Layers with Residual connections and LayerNorm and Dropout, MLP Classifier for Company nodes, and Phase II Prediction output.

Model	AUPRC	AUROC	F1
XGBoost	0.629 $\pm$ .003	0.652 $\pm$ .003	0.590 $\pm$ .000
MLP	0.591 $\pm$ .025	0.634 $\pm$ .014	0.593 $\pm$ .001
R-GCN	0.598 $\pm$ .006	0.648 $\pm$ .006	0.593 $\pm$ .002
SME-HGT	0.606 $\pm$ .005	0.646 $\pm$ .003	0.591 $\pm$ .001

Table 2. Test Set Results (mean

\pm

std over 10 seeds)

	Precision@ $K$			Lift@ $K$
Model	@100	@500	@1K	@100	@500	@1K
XGBoost	.924	.711	.578	2.21	1.70	1.38
MLP	.845	.668	.550	2.02	1.60	1.32
R-GCN	.826	.673	.572	1.98	1.61	1.37
HGT	.834	.689	.576	2.00	1.65	1.38

Table 3. Ranking Metrics on Test Set (mean over 10 seeds)

Removed Feature	AUPRC	$\Delta$ vs. Full
None (full model)	0.603 $\pm$ .002	—
$-$ log total Phase I	0.597 $\pm$ .002	$-$ 0.006
$-$ log Phase I count	0.577 $\pm$ .003	$-$ 0.026
$-$ agency diversity	0.604 $\pm$ .002	$+$ 0.001
$-$ years active	0.607 $\pm$ .002	$+$ 0.005
$-$ log avg Phase I size	0.601 $\pm$ .002	$-$ 0.002

Table 4. Feature Ablation (HGT, AUPRC, mean

\pm

std over 5 seeds). Larger

\Delta

= more important feature.

Removed Edge Type	AUPRC	$\Delta$ vs. Full
None (full model)	0.603 $\pm$ .002	—
$-$ operates_in	0.596 $\pm$ .003	$-$ 0.007
$-$ awarded_by	0.594 $\pm$ .005	$-$ 0.009
$-$ co_topic	0.615 $\pm$ .002	$+$ 0.012
$-$ co_agency	0.605 $\pm$ .005	$+$ 0.002
$-$ co_state	0.601 $\pm$ .006	$-$ 0.002

Table 5. Edge-Type Ablation (HGT, AUPRC, mean

\pm

std over 5 seeds).

Model	Edge Mode	AUPRC	AUROC
SME-HGT	Multi	0.606 $\pm$ .003	0.647 $\pm$ .003
SME-HGT	Single	0.599 $\pm$ .004	0.645 $\pm$ .004
R-GCN	Multi	0.607 $\pm$ .007	0.653 $\pm$ .005
R-GCN	Single	0.594 $\pm$ .010	0.650 $\pm$ .007
MLP	Multi	0.557 $\pm$ .070	0.608 $\pm$ .051
MLP	Single	0.557 $\pm$ .070	0.608 $\pm$ .051

Table 6. Multi-Edge vs. Single-Edge (AUPRC, mean

\pm

std over 5 seeds).

Model	Early	Current	Late
XGBoost	0.732 $\pm$ .001	0.628 $\pm$ .005	0.548 $\pm$ .005
MLP	0.660 $\pm$ .078	0.557 $\pm$ .070	0.433 $\pm$ .086
R-GCN	0.676 $\pm$ .059	0.607 $\pm$ .007	0.483 $\pm$ .080
SME-HGT	0.717 $\pm$ .004	0.606 $\pm$ .003	0.542 $\pm$ .005

Table 7. Temporal Robustness (AUPRC, mean

\pm

std over 5 seeds). Test periods: Early (2018–2020), Current (2020–2022), Late (2021–2023).

Coverage-Aware Web Crawling for Domain-Specific Supplier Discovery via a Web--Knowledge--Web Pipeline
Feb 27, 2026Yijiashun Qi, Yijiazhen Qi, Tanmay WaghSupply Chain Management KG Construction
University of Michigan Ann Arbor, USA · The University of Hong Kong Hong Kong, China · Santa Clara University Santa Clara, USA
Integrated Multivariate Segmentation Tree for Heterogeneous Credit Data Analysis in Small- and Medium-Sized Enterprises
Aug 30, 2025Lu Han, Xiuying WangFinancial Services Interpretable ML
School of Management Science and Engineering, Central University of Finance and Economics, Beijing 100081, China · School of Computer Science, University of Sydney, NSW 2006, Australia
Aethel: A Reproducible Graph-Retrieval Framework for Multi-Hop Financial Diligence
Jul 20, 2026Krish SapruCoreference Resolution Graph-Based Retrieval

Feb 27, 2026J/K move · Enter open · S save

Coverage-Aware Web Crawling for Domain-Specific Supplier Discovery via a Web--Knowledge--Web Pipeline

Yijiashun Qi, Yijiazhen Qi, Tanmay Wagh

University of Michigan Ann Arbor, USA · The University of Hong Kong Hong Kong, China · Santa Clara University Santa Clara, USA

Supply Chain Management KG Construction

Open paper AskPDF arXiv

Identifying the full landscape of small and medium-sized enterprises (SMEs) in specialized industry sectors is critical for supply-chain resilience, yet existing business databases suffer from substantial coverage gaps -- particularly for sub-tier suppliers and firms in emerging niche markets. We propose a \textbf{Web--Knowledge--Web (W

\to

\to

W)} pipeline that iteratively (1)~crawls domain-specific web sources to discover candidate supplier entities, (2)~extracts and consolidates structured knowledge into a heterogeneous knowledge graph using domain-adapted few-shot LLM prompting, and (3)~uses the knowledge graph's topology and coverage signals to guide subsequent crawling toward under-represented regions of the supplier space. To quantify discovery completeness, we introduce a coverage estimation framework inspired by ecological species-richness estimators (Chao1, ACE) adapted for web-entity populations. Experiments on the semiconductor equipment manufacturing sector (NAICS 333242) demonstrate that the W

\to

\to

W pipeline achieves the highest precision (0.165) and F1 (0.123) among all methods while using only 144 pages -- 32% fewer than the 213-page baseline budget -- building a knowledge graph of 664 entities and 542 relations with 100% relation type-consistency.

Abstract

Explore similar work

Coverage-Aware Web Crawling for Domain-Specific Supplier Discovery via a Web--Knowledge--Web Pipeline

Integrated Multivariate Segmentation Tree for Heterogeneous Credit Data Analysis in Small- and Medium-Sized Enterprises

Aethel: A Reproducible Graph-Retrieval Framework for Multi-Hop Financial Diligence