Detecting High-Potential SMEs with Heterogeneous Graph Neural Networks

Small and Medium Enterprises (SMEs) constitute the vast majority of businesses globally and generate a significant portion 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 early-stage (Phase I) innovation grant 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 funding agency nodes connected by approximately 99,000 edges across five semantic relation types. Against baselines including MLP and R-GCN, XGBoost achieves the strongest overall performance (AUPRC 0.629 ± 0.003, AUROC 0.652 ± 0.003), consistent with the documented strength of tree-based methods on tabular data. However, graph-based methods substantially outperform the non-graph MLP baseline (SME-HGT AUROC 0.646 ± 0.003 vs. MLP 0.634 ± 0.014), confirming that relational structure provides meaningful signal for neural prediction. Feature ablation identifies Phase I award count as the dominant predictive feature (ΔAUPRC = −0.026 when removed); temporal robustness analysis across three non-overlapping time windows reveals substantial concept drift (10–18 pp AUPRC decline), motivating periodic retraining. At a screening depth of 100 companies, XGBoost attains 92.4% precision (2.21 × lift), while SME-HGT achieves 83.4% (2.00 × lift). 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 bodies provides meaningful signal for SME potential assessment, with implications for economic policymakers and early-stage investors globally.