Styxis
Company
Knowledge

DecisionTree check_input=False: a validation flag cannot replace the missing-value mask

Skipping input checks requires knowing which information the skipped stage normally creates for the algorithm.

One team's validation is not another team's input preparation

A single validated flag cannot substitute for derived information needed by a downstream algorithm. Confirming dtype and shape does not establish that missing-value positions are ready. Before saving time by skipping validation, distinguish rejection checks from steps that compute information required for execution. [1, 2, 3]

Information missing from the public fit path

scikit-learn issue 35048 reports that DecisionTree fit with check_input=False can skip computation of the missing-feature mask and treat NaN as ordinary values. The source distinguishes passing a mask to private _fit from the public fit path, which does not expose the same parameter. An internal forest's use of the private path is not direct evidence that an external call is safe. [1]

Separate the validation scope in the handoff

Record missing-value inspection and missing-mask generation separately in the preprocessing acceptance table. Avoid a skip path lacking required information. State which data was checked and which derived material was created. If the public API cannot receive the necessary information, retain the normal validation path rather than generalizing private-method calls as a workaround. [1, 3]

A true flag can coexist with absent information

Suppose illustrative data contains NaN and only dtype checking has finished. The required mask state is available. Actual available material satisfies this preparation condition. An absent mask combined with a validation-complete flag must hold the operation. These are handoff-table state names, not actual scikit-learn API arguments. The fixture also does not claim that we trained trees and reproduced a performance difference. [1, 3]

Truthound can find a contradictory preparation record

ConditionalExpressionValidator in Truthound Core 3.3.2 can use condition to select a skip request with missing values and then_expr to require the preparation state. The check assumes missingness extracted from actual input and an execution-path record. Truthound does not generate a tree's mask or make private APIs safe. Passing the preparation table and establishing the validity of a trained model remain separate tasks. [1, 2, 3]

Application conditions

Prepare caller, contains_missing, mask_available and skip_input_checks from raw input and approved references, preserving row identity.

Alternatives

Retain a supported public path that performs validation. If optimizing, verify the documented input scope first.

Limitations

  • This does not claim that the public API accepts a mask or recommend private API calls as a general solution.

Evidence

Event checked
Subject checked
Modified