PG-SB: Benchmarking Schema Discovery for Property Graphs

PG-SB unifies ten datasets (five real and five synthetic) spanning the domains of social networks, neuroscience connectomes, biomedicine, finance/leaks, communications, stream analytics, and internet measurements. For each dataset, we provide the Neo4j dataset dump, the ground-truth schema, and the corresponding type patterns observed in the data, capturing the structural variability of label and property co-occurrence. The benchmark includes a configurable noise injection framework that:(i) randomly removes 0–40% of node and edge properties, and(ii) varies label availability across three settings: 100% (all labels retained), 50% (half retained), and 0% (no labels),resulting in a total of 150 test cases. This Zenodo release includes the datasets (Neo4j dumps), the ground-truth schemas, the extracted type patterns, and the scripts used for computing the evaluation metrics. All dataset resources are provided as Neo4j database dumps, named:{dataset_name}_neo4j_{neo4j-version}.dump. The full reproducible benchmark pipeline — including noise injection, and experimental execution — is available on GitHub: https://github.com/sophisid/PG-SB