Who Inherits the Field? A Leakage-Audited Benchmark for Intellectual Inheritance in Doctoral Training
Data for the paper "Who Inherits the Field? A Leakage-Audited Benchmark for Intellectual Inheritance in Doctoral Training" (submitted to the KDD 2027 Datasets and Benchmarks Track). The benchmark asks whether a doctoral student still works in the advisor's early research area ten years later, framed as forward prediction under a testable time contract: every input feature is observable by the student's fifth career year, and the label is read fifteen years out. It spans five scientific genealogies built from one 2026 snapshot under one byte-identical protocol: economics, mathematics, physics, chemistry, and neuroscience, 330,282 raw student-advisor pairs distilled into 68,235 modeled pairs (one student per pair). The funnel is dominated by two requirements: both members of a pair must resolve to OpenAlex authors, and the label must be observable by 2026. FINDINGS THE DATA SUPPORTS A strong tabular model on early features sets the prediction ceiling in every discipline; four graph neural network architectures fail to beat it in nineteen of twenty discipline-architecture comparisons, so the suite serves as a negative control for graph learning, with the one surviving exception (a relational graph network in chemistry) flagged and audited. The true advisor beats a cohort-matched placebo in all five disciplines, while the mechanism differs by field. An initial aggregation had reported seven ceiling crossings; three corrections restoring the pre-registered protocol removed six, and that evaluation audit ships with the artifact. CONTENTS (zenodo_archive.zip, 34 files) - data/clean_dataset_{econ,math,neuro,physics,chemistry}.parquet: the five frozen modeling tables, one row per student-advisor pair.- data/pairs_resolved_{econ,math,neuro,physics,chemistry}.parquet: resolved pairs with OpenAlex author ids.- data/outcomes_{econ,math,neuro,physics,chemistry}.parquet: outcome tables for the mechanism analyses.- data/SHA256SUMS: hash manifest; every table is pinned and verified by the code before any experiment runs.- data/funnel_*.json, dataset_summary_*.json, coauthorship_*.json, network_modularity.json: attrition funnels, build summaries, and descriptive statistics reported in the paper.- datasheet/DATASHEET.md: datasheet following Gebru et al.- reproduction/reproduce_assertions.py: 26 assertion checks against the frozen tables. USE Code, documentation, and the full reproduction pipeline are in the public repository: https://github.com/Xinke-Li/who-inherits-the-field Download this archive, place the parquet files under data/, and verify with sha256sum -c data/SHA256SUMS. The four certificates (y-scrambling placebo, advisor-disjoint split, advisor placebo, student-only control) rerun from the frozen tables on a CPU. Rebuilding the tables from scratch requires an OpenAlex API key supplied through the environment; reproducing the paper's experiments from the frozen tables does not. LICENSES AND CITATION Derived tables are released under CC BY 4.0, following the Academic Family Tree genealogy terms (David 2020, doi:10.5281/zenodo.4247419). Underlying publication records come from OpenAlex (CC0); code and datasheet are MIT. If you use this data, please also cite OpenAlex: Priem, J., Piwowar, H., and Orr, R. (2022). OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv:2205.01833. The dataset describes public scholarly records and must not be used to rank or evaluate individual researchers.