Docs · Step 5

Go deeper with Deep Analysis

A real, resumable AutoML search that trades your machine's idle time for a materially stronger model.

1. Start a run

  1. Open Predict and choose the outcome column you want to predict.
  2. Start a Deep Analysis run instead of (or after) a quick Predict run.

Most of CLAIR finishes an analysis in minutes. Deep Analysis is built for the opposite trade: it can run for hours, spread across however many idle windows it takes, in exchange for a search that a quick pass can't do.

2. Let it run while idle

A run is a staged, checkpointed pipeline. It searches feature selection, tries an Optuna-driven hyperparameter search across several model families (XGBoost, LightGBM, Random Forest, Neural Net, and CatBoost when available), and stacks the results into a meta-model, only keeping the stack if it actually beats the best single model. You can close CLAIR and come back; the run resumes from its last checkpoint rather than starting over.

3. Read the results

When a run finishes, you get a leaderboard and a champion model. Two honesty checks apply to what you see:

  • A champion that never clearly beat the obvious guess (the majority answer for a category, the plain average for a number) is labelled weak, with the same caution wherever its result appears: the results screen, the exported report, and a "what predicts X" chat answer.
  • A column that looks like a plain row number, or one that lines up suspiciously well with the outcome, is checked and disclosed rather than left to quietly dominate the "what drives this" ranking.

4. Pause and resume

A run that shows as paused has genuinely stopped: nothing is running in the background while it waits. Click Resume and it continues from where it left off rather than restarting the search.

What's actually happening

This is a real search, not a wrapper around a single model

Feature selection is chosen by cross-validated score, not a fixed rule. Feature engineering is purely structural (products, ratios, transforms) and never fitted on the outcome, so it can't leak the answer into the inputs. Every model family's out-of-fold predictions feed a meta-learner that's only kept if it earns its place. When you set a higher "seeds" setting, the whole pipeline repeats across multiple random seeds and averages the result, specifically to reduce the "got lucky once" effect a single run can't rule out.

Once a champion is trained, it's saved and reused: if your data changes later, CLAIR re-fits the same known-best pipeline instead of searching from scratch, and can periodically challenge it with a fresh search, promoting a challenger only if it beats the incumbent by a real margin.