GraphFounder: Interpretable Founder Archetypes via Heterogeneous Graph Neural Networks

GraphFounder models founder credentials as a typed attribute graph, connecting founders to university prestige tiers, company sizes, industries and roles, then clusters the resulting graph neural network embeddings into interpretable archetypes.

On a 9,000-founder benchmark it produces archetypes with a 29.7 percentage-point success-rate spread, 6.9× that of tabular clustering, separating subgroups whose success rates run from 1.9% to 31.6%.

The point is not only accuracy but interpretability: each archetype is readable as a credential pattern, so the output is something an investment team can reason about rather than a score with no explanation attached.

Reproducibility

The repository is the reproducible implementation behind the ICDM 2026 Applied Track paper. Every result table and figure is committed and regenerates from cached model outputs with a single command, with no GPU and no retraining:

python reproduce.py --from-cache --all

The dataset is read only for the train/validation/test split and the labels. Deterministic steps (k-means on the committed embedding, bootstrap confidence intervals, χ², plotting) re-run on the cached artifacts to reproduce the paper’s tables and figures exactly.