Skip to content

MLI in Studio

Studio's Explain surface combines the Python library with backend orchestration, saved evaluation evidence and optional AI annotations. Its compute presets are MLIComputeConfig values, separate from the library's ProfilingConfig. A minimal Python profile does not reproduce Studio's preset artifact set.

In short

  • Four built-in depths: Shallow, Standard, Standard + applyable business rules (preselected in the app) and Deep. Deeper presets add PDP, interactions, a residual surrogate, stability and an independent rule model.
  • Profiling reads at most 20,000 rows, sampled from the experiment's recorded row roles.
  • SHAP, archetypes and model rules can describe a readable base model rather than the deployed ensemble. Read the provenance before comparing views.
  • A finished compute job does not mean every view exists. A missing value is not zero, and an unavailable check is not a pass.
  • AI-written names and summaries are added after the computation and never change the numbers.

Compute depth

The app's compute picker preselects Standard + applyable business rules, and so does the explain step of a full rerun. A compute request that names no preset runs Standard. Built-in presets request the following work; availability depends on the model, task, data and whether each stage succeeds.

Requested work Shallow Standard Standard + applyable business rules Deep
Native profiling when supported; EDA Yes Yes Yes Yes
Extracted model rules No Yes Yes Yes
ALE No Yes Yes Yes
Permutation importance No 1,000-row sample, 2 repeats Same as Standard 5,000-row sample, 10 repeats
Prediction surrogate and error tree No Yes Yes Yes
PDP No No No Top 10 features, 40-point grid, up to 5,000 rows
Residual surrogate and SHAP interactions No No No Yes
Archetype stability bootstrap Off Off Off 20 resamples
Independent applyable rule model No No 15 Optuna trials 40 Optuna trials

The independent rule model is classification-only: a separate compact LightGBM fits the raw attributes against the true target. Its rules are not a surrogate of the deployed model. The slow all-class rule engine is off in every built-in preset; a custom configuration can request it for binary tasks.

The default profiling row cap is 20,000. Supervised row selection uses the recorded train/validation/test roles and a seeded sample; rule holdout rows are kept separate from mining rows. Performance uses saved holdout scores when available, rather than substituting metrics from sampled profiling rows. An imported run without that evidence can therefore lack Performance data.

The default archetype space is Model leaves; the compute picker also offers SHAP. Models without usable per-sample leaves use SHAP space. For expensive scorers, the backend can reduce prediction-bound analysis budgets and record the reason; Deep or explicitly requested full-depth work retains the requested budget. Runtime and coverage are not guaranteed by the preset name.

Which model does a view explain?

The deployed pipeline produces predictions and model-agnostic feature effects. Native SHAP, archetypes and extracted model rules can describe a selected introspectable member instead. For an ensemble, these are different models: a member's drivers need not explain the deployed ensemble's exact probability. Read the model provenance and fidelity information when comparing those views.

The backend tries native analysis when an appropriate estimator is available, and can fall back to monitor_predictions when native computation is unavailable or fails. The resulting backend is recorded as native, lite_model or lite_predictions. Lite output does not supply exact native SHAP or tree-leaf archetypes of the opaque scorer.

Availability and experiment kinds

A completed compute job does not mean every artifact exists. A surface may have been omitted by the preset, unsupported for the task, absent from saved evidence, empty after filtering, or failed during computation. Read its status and explanation; a missing value is not zero, and an unavailable fairness or drift assessment is not a pass.

Forecasting, clustering, anomaly detection and association runs have dedicated lenses. Ranking and recommendation use task-specific compute paths. They do not all have binary-classification calibration, fairness, rule mining or SHAP. Forecasting tree explanations require a tree backend; other forecast backends show their interpretability limitation. Unsupervised runs do not have supervised holdout evaluation or classification label-noise scans.

Narration enriches computed results

The computed profile contains no AI-written narrative. Studio can generate explain.narration and explain.fields annotations after computation according to workspace policy, or on an explicit request. Rule and class explanations are requested separately. The host stores annotations with provenance and reattaches permitted text when loading the profile.

Studio's Landscape treemap works from computed archetypes without narration. Names and narrative groups enrich it; archetypes not assigned to a narrative group appear under Ungrouped archetypes. A profile with no surviving archetypes has no treemap. The Python result.plot_landscape() helper has a different contract: it requires attached narration with groups and archetype data.

Narrative regeneration has its own progress and failure state. Existing text can remain visible while replacement text is being generated; its presence does not mean that the new job completed successfully. Numeric evidence remains readable independently of narrative generation.

See AI-written narration, Evaluation & trust and Interactive visualizations for the library contracts.

Next