Completing Ecosystem Knowledge Graphs: Few-Shot Learning with Calibrated Rules and the Governance of Innovative Start-Ups

Ecosystem knowledge graphs make young firms' governance observable, but registers record its rarest relations unevenly. We build LUCE-KG, a graph of 15,493 Italian innovative start-ups and SMEs, and propose CRML, a few-shot learner whose predictions are explained by rules with confidences shrunk toward pooled rates by an empirical-Bayes estimator whose risk advantage we prove. On four graphs, shrinkage cuts the error of one-shot confidences by 47–76 per cent; on LUCE-KG, CRML recovers rare-office holders already in the firm as well as retrained embeddings, though no method finds new ones, gaining from observed links, not the ontology. Panel, clustering and explained machine-learning analyses of 11,764 firm-years show that specialised governance offices are associated with higher profitability across firms, not within them, and larger boards with lower. A firm-level classifier trained on link scores beats naive baselines (AUC 0.88 vs 0.81). Completion helps when trained for the firm-level quantity the analysis needs.