Choose a pruning strategy

explain/explainf expose four independent choices. This guide is practical guidance on picking values; see How scoring and pruning fit together for why these axes exist and what each is actually doing.

Order: HeuristicOrder vs GreedyForward

explain(ds, model)                          # default: HeuristicOrder, built from scoring automatically
explain(ds, model; order=GreedyForward())   # no scoring-based order; pure stepwise search

• HeuristicOrder (the default whenever order is left as nothing) uses the importance scores ShapleyExplainer already computed, sorting candidates and adding them via bisection – O(log n) evaluations of the objective. Use this unless you have a specific reason not to; it's the faster option and reuses work already done during scoring.
• GreedyForward ignores any precomputed score and, at each step, actually tries every remaining candidate and keeps the best one – O(n) evaluations per step. Slower, but doesn't depend on the scoring strategy's importance estimates being accurate. Worth trying if HeuristicOrder's result looks worse than expected, or if you're using a scoring strategy whose scores you don't fully trust for ordering.

Granularity: levelbylevel

explain(ds, model; levelbylevel=true)    # default: one hierarchy depth at a time
explain(ds, model; levelbylevel=false)   # the whole tree at once

levelbylevel=true is generally faster in practice for hierarchical samples (bags of bags, nested objects) since each level's search operates over a much smaller candidate set than the whole tree at once. Set it to false if you want the search to consider trade-offs across levels simultaneously rather than committing to shallower levels before deeper ones are decided.

Post-passes: random_removal and finetune

explain(ds, model; random_removal=true, finetune=true)   # both on (default)

Both default to true and both only ever make the result smaller or equal – they never undo a valid explanation, only look for redundancy the main search missed. There's rarely a reason to turn either off; doing so trades a small amount of speed for a possibly-larger (but still valid) explanation.

• random_removal: repeatedly tries removing currently-kept items in random order, keeping the removal if the objective still holds.
• finetune: a small local-search pass (alternating batched add/remove) that catches local-optimum artifacts the single-pass search can leave behind.

A reasonable starting point

The defaults (HeuristicOrder, levelbylevel=true, both post-passes on) are a reasonable starting point for most samples. Reach for GreedyForward or levelbylevel=false only if you have a specific reason to suspect the default search order or granularity is producing a poor result for your data.