Package: splitGraph 0.3.0

What this vignette is for | Build a split_spec to work with | Adapter 1 — base R: leave-one-group-out folds | Honoring block variables | Adapter 2 — rsample::group_vfold_cv() | Adapter 3 — rsample::rolling_origin() | Going across language boundaries via JSON | Letting the spec pick the adapter | When you need a custom adapter

Derive and serialize in R | Read in Python | Verify the round-trip | Drive scikit-learn | Why this matters

Cluster-style relations: site, region, platform, assay | Pairwise relation: genetic relatedness | Pairwise relation: spatial proximity | Deriving on a subset is leakage-safe | Thresholds are inputs, not modeling

Why splitGraph exists | A realistic toy dataset | Fast path: graph_from_metadata() | Ingest metadata and build typed nodes and edges | Assemble the dependency graph | Visualize the typed structure | Validate before you split | When you really do need to relax a check | Query the graph to inspect hidden structure | Derive split constraints from the graph | Batch constraints | Time constraints | Composite constraints | Cluster-style relations: site, region, platform, assay | Site | Region | Platform and assay | Pairwise relations: relatedness and spatial proximity | Time ordering can come from precedence edges alone | Translate the constraint into a split specification | Summarize the leakage picture in one object | Downstream handoff | Persisting the handoff | Case studies | Case study 1: repeated subjects in a longitudinal cohort | Case study 2: a subject reused across studies | Case study 3: partially observed technical metadata | Case study 4: choosing a defensible split strategy | When splitGraph is useful | What splitGraph is not for | Takeaway