SplitCheck

The results of check_splits().

Usage

Source

SplitCheck(
    values,
    schema,
    columns,
    label,
    near,
    threshold,
    ngram,
    num_perm,
    drift_columns,
)

Attributes

splits

The number of rows of each split, by name.

overlap

A dictionary for each pair of splits (an earlier split first and a later one second), with the number of rows of the later split that are also in the earlier one (exact) and, with near=True, the number of other rows of the later split that are nearly the same as one in the earlier split (near).

labels

The check of the labels (or None without label=): the counts of the labels in each split, the labels of the first split that each other split doesn’t have (missing), those of each other split that the first doesn’t have (unseen), and the PSI of the shares of the labels of each other split, compared with the first.

drift

A dictionary for each column of drift_columns= and each split after the first, with the PSI and (for a numeric column) the KS statistic and its p-value, compared with the first split, and the level of the change ("stable", "moderate", "major", or None when there are too few values).

issues

The problems found, as messages: overlapping rows, missing or unseen labels, and major changes in distribution (a PSI of at least 0.2). A moderate change isn’t an issue, but it’s shown in the report.