How accurate and how costly each extraction strategy is, measured against exact ground truth on seeded synthetic corpora and on real dumps. Recall and precision are over (subject class, property, object class or datatype) patterns. Estimates should have their 95 % bounds (±2·RSE before voidx 4b) hold the truth about 95 % of the time. An exact count that differs from truth at all is a defect.
How each engine took each source in this run. A diverged engine stored the source differently from its canonical form (expected → stored counts); a failed one did not load it. Either way its cells were skipped. "Detected as" is what voidx took the engine for. "Union view" is the truth an engine is scored against: bag (a triple in two graphs counts twice) or set (the RDF merge), as the engine declares it merges graphs, by its union default graph or FROM. "Stores differently" is the datatype rewrites an engine was measured to make on what it stores: its datatype patterns cannot match truth, so its exact-count defects there are the engine's storage, not voidx's. They still count: exact means exact.
HTTP requests sent (log scale), retries included, against the share of true patterns found. Up and to the left is better.
Per strategy that produced estimates. The dashed line is the 95 % target.
Relative error of estimates against the non-blank truth: median and 95th percentile over each run's estimates, then the median over repetitions. Exact counts have no error: any difference is a defect.
Each corpus harvested 16 times at 1/16 and 16 times at 1/256, pinned to sampling, every estimate scored against its non-blank truth. 4b is what voidx reports (the conditional MLE for a pattern, exact 95 % bounds, clamped by the counts the run knows); 4b-uncapped is the same point and bounds before the caps, so the estimator's own bias shows apart from the caps' (where a cap equals the truth, capping biases the point low); pre-4b is h / π with ±2·RSE from the same hits. Bias and RMSE are of the relative error; coverage is the share of intervals holding the truth, with its range over slices; detection is the share of (true target, slice) pairs with an estimate. The size groups leave out literal objects.
Every repetition, with robustness counts from the outcomes voidx recorded. Select a run to see its patterns.
Select a run above.
Select a run above.
One chart per mode, the baselines beside voidx's: each point is a (source, engine, faults) group's median HTTP requests (log scale) against its recall. voidx's points are coloured by tier; a baseline's take its mode's colour, with the tier as its outline. Every chart shares the same axes. Live endpoints have no truth, so they sit on a cost-only strip below, coloured by mode.