Skip to content

Exact-linked public benchmarks

VertiMosaic keeps linkage categories separate so a reproducible vertical split is never presented as stronger evidence than the source data actually support.

MovieLens 1M exact multi-table linkage

vertimosaic-linked-movielens is the fully public, provider-backed multi-table linked benchmark in v0.2.0. It downloads MovieLens 1M directly from GroupLens at runtime and uses the dataset's observed identifiers:

  • UserID links users.dat to ratings.dat;
  • MovieID links ratings.dat to movies.dat;
  • the active party receives demographic attributes from users.dat;
  • the passive party receives historical-rating aggregates and genre behavior derived from the earliest 80% of each user's observed ratings;
  • the binary target is derived only from the latest 20% of that user's observed ratings: whether the late-period mean rating is at least four stars;
  • the target-defining late rating events are excluded from passive-party features;
  • train/validation/test splitting is entity-level and all preprocessing is fitted on training entities only;
  • the output records the downloaded archive SHA-256, provider/license metadata, environment, sampled peak RSS, protocol communication metadata, bootstrap uncertainty and pseudonymous test predictions.

This is genuine exact linkage across multiple public tables describing the same service users. It is still a single-service dataset, not cross-organization linkage, and it does not demonstrate PSI. MovieLens source files are not committed or redistributed by VertiMosaic; they are downloaded from GroupLens for research use at runtime.

IEEE-CIS transaction/identity

Authorized local IEEE-CIS files can be joined on the exact TransactionID intersection. This is a genuinely linked two-table transaction/identity benchmark, but the competition files are not downloaded or redistributed by VertiMosaic.

UCI Credit exact-row vertical partition

vertimosaic-linked-uci provides a public exact-link sanity benchmark using UCI Default of Credit Card Clients (dataset 350). The benchmark:

  • retrieves the source directly from UCI at runtime;
  • uses the exact same real source rows on both parties;
  • partitions disjoint raw source columns between an active and passive party;
  • fits imputation/scaling on training rows only;
  • preserves a held-out entity split;
  • writes pseudonymous test identifiers, source checksum, license metadata, metrics and a SHA-256 manifest.

This benchmark is intentionally labeled single-source exact-row vertical partitioning. It does not claim that the two feature partitions originated from independent organizations, and it is not evidence of cross-organization entity resolution.

Cross-industry four-party benchmark

The Bank/Telecom/Insurance/Retail public-source benchmark remains an externally grounded semi-synthetic linkage construction. The source datasets do not describe the same real people. That benchmark tests cross-domain modeling assumptions; it must not be conflated with MovieLens/IEEE-CIS exact linkage.