NegGraphBench v1.0: A Benchmark for Relational-Exclusion Path Queries

NegGraphBench v1.0 is a benchmark for relational-exclusion path queries: shortest-path queries in which an exclusion policy forbids the answer path from selecting more than a bounded number of vertices from any relation component (the same institution, the same beneficial owner, the same evidence channel). Such queries arise in conflict-of-interest-aware referral, financial-forensics routing, and evidence-independence analysis. The benchmark exposes both realistic feasible and infeasible regimes: on the natural policies, roughly half of the real queries have no admissible answer, so proving infeasibility is part of the workload. It is the companion benchmark of the paper "GuardedPath: Structure-Aware Processing of Relational-Exclusion Path Queries" (under submission to PVLDB). Reproduction code, experiment runners, and analysis scripts are maintained at https://github.com/zxf-work/cnga (results reported at commit b183ab1). FILES neggraphbench-graphs-v1.0.tar.xz (32 MB): all built workload graphs in a plain-text format (edges.txt: "u v w" per line; memberships.txt: vertex-component pairs; meta.json: provenance; entity id maps back to OpenAlex), plus the build and crawl scripts and a README documenting formats, construction, and pinned query conventions. neggraphbench-openalex-raw-v1.0.tar.xz (163 MB): the raw OpenAlex works records (JSON lines) from which the Scholar graphs are built, comprising API crawls plus a complete filter of the official OpenAlex S3 snapshot (subfield 1702, years 2010-2020). This is the only non-re-derivable layer: the OpenAlex corpus changes daily, so a fresh extraction cannot reproduce these graphs bit-for-bit. MANIFEST-v1.0.sha256: SHA-256 checksums for both archives and every packaged file. WORKLOADS Scholar: OpenAlex coauthorship, database-community slice; 96,668 authors, 233,791 edges, 6,939 institution components (policy: at most one author per institution). Scholar-4x: OpenAlex AI subfield, 2010-2016; 352,956 authors, 863,044 edges, 14,640 components. Scholar-10x: OpenAlex AI subfield, 2010-2020, snapshot-complete; 913,073 authors, 2,770,257 edges, 23,809 components. STRING: human protein-protein interaction workload from STRING v12; 19,622 proteins, 236,930 edges, 1,393 evidence-channel components (per-channel union-find over score-thresholded links; four channels). Finance: LDBC FinBench SF1 transfers; 200,336 accounts, 614,043 directed edges, with three ownership policies of increasing closure (direct owner; 25-percent-control closure; unrestricted closure). Layered synthetic: generated by the engine with an analytic birthday-bound difficulty knob; generator parameters live in the experiment scripts and are not materialized here. LICENSES OpenAlex-derived data is CC0. STRING-derived data is CC BY 4.0 (please cite Szklarczyk et al., STRING v12). FinBench-derived data follows the LDBC data license. Scripts are Apache 2.0. This record as a whole is published under CC BY 4.0.