Skip to main content
Version: Next

EDA & Statistics

EDA worksheets are interactive exploratory-data-analysis surfaces on a dataset — univariate summaries, correlation matrices, scatter-and-fit, and a library of statistical test cards. Unlike a recipe, a worksheet computes on demand and is not part of the build graph; it is where an analyst pokes at a dataset before deciding what to model.

Worksheets persist per dataset (honeyframe.dataset_worksheets) and read through the governed statistics API, so PII masking applies to every value a worksheet shows.

Univariate & bivariate

  • Univariate — per-column summaries: count, nulls, distinct, min/max, mean, quantiles, and a distribution sketch.
  • Correlation matrix — pairwise correlation across the numeric columns you select.
  • Scatter + fit (scatter_fit) — a scatter of two columns with a fitted trend line.

Sampling

Statistics run over a sample so exploration stays fast on large tables. The sampling picker offers first, random, or last; the worksheet notes which sample a result was computed on.

Statistical test cards

Beyond summaries, a worksheet offers named statistical tests, each rendered as a card with its statistic, p-value (where applicable), and a plot:

CardUse
Chi-squaredIndependence between two categorical columns.
Welch's t-testDifference in means between two groups (unequal variance).
Kolmogorov–SmirnovWhether two samples come from the same distribution.
CDFEmpirical cumulative distribution of a column.
Fit distributionFit a column to a candidate distribution and report goodness-of-fit.
ACFAutocorrelation of a time-ordered series at increasing lags.
Mann–KendallMonotonic trend test for a time series.
QuantilesQuantile table / plot for a numeric column.

API

The statistics surface is served by the /api/statistics router (univariate, correlation matrix, scatter_fit, and the test cards). Worksheet state is saved so a worksheet reopens where you left it.

Relation to profiling and recipes

  • The generate_statistics Prepare processor / recipe materializes per-column profiling (fill rate, null/distinct counts, min/max) as a dataset you can build on and schedule — use it when you want profiling in the pipeline rather than an ad-hoc worksheet.
  • The Catalog column profile (median, p25/p75, IQR, sample stddev) is the cached, at-a-glance version surfaced on the asset's Profile tab.

Use worksheets to explore, generate_statistics to persist, and the catalog profile for a quick governed snapshot.

See also