Dataiku ML Practitioner cheat sheet

The Visual ML Lab creates a trainable model from a dataset; each experiment holds one model and one set of preprocessing, feature handling, and algorithm settings.

Train/test split: the Split step assigns rows to train and test sets; default is random split with a settable ratio (commonly 80/20) and an optional seed for reproducibility.

Algorithm choice drives metric defaults: classification models report accuracy, precision, recall, F1, AUC, and log loss; regression models report RMSE, MAE, R2, and MAPE.

AUC (area under ROC curve) is threshold-independent and ranges 0 to 1; 0.5 equals random guessing, so compare models by AUC before tuning decision thresholds.

Precision = TP/(TP+FP); recall = TP/(TP+FN); F1 is the harmonic mean 2PR/(P+R). Choose the metric that matches the business cost of false positives versus false negatives.

Feature handling options include categorical encoding (dummy/label), missing-value imputation, rescaling (standardization, min-max), and feature generation; these are configured per feature or by type.

Hyperparameter search uses grid search or random search over a defined parameter space, with cross-validation folds; the reported score is the cross-validated score, not the holdout score.

The model's interpretation tab exposes variable importance (permutation or impurity based) and partial dependence plots to show how each feature moves predictions.

Deployment turns a trained model into a saved model (versioned) that can be used in a scoring recipe or API node; retraining can be scheduled to refresh the saved model.

The Statistics tab on a dataset offers univariate analysis per column: count, missing count, distinct count, min, max, mean, median, standard deviation, and quartiles for numeric columns.

Histograms bin numeric values to show distribution shape; the bin count is adjustable, and changing bins changes apparent skew and modality, so state the binning when reporting.

Box plots mark median, quartiles (Q1, Q3), and whiskers; points beyond 1.5*IQR from the quartiles are conventionally flagged as outliers.

For categorical columns, the frequency/count chart gives top values and their share; the distinct-count and missing-count readouts reveal cardinality and data-quality gaps.

The correlation matrix computes pairwise Pearson correlation for numeric columns; values near +1 or -1 indicate linear association, but correlation does not establish causation and is sensitive to outliers.

The bivariate/cross-statistics view compares a metric across groups defined by a categorical column, reporting group means and counts to support comparison before modeling.

A statistical test worksheet can run t-tests, chi-square, or ANOVA-style comparisons; results include a p-value, and a small p-value indicates the observed difference is unlikely under the null hypothesis.

Sampling and filtering applied before the Statistics tab change the computed values; confirm the active sample or filter so reported statistics match the intended population.

The Visual ML Lab creates a trainable model from a dataset; each experiment holds one model and one set of preprocessing, feature handling, and algorithm settings.

Train/test split: the Split step assigns rows to train and test sets; default is random split with a settable ratio (commonly 80/20) and an optional seed for reproducibility.

Algorithm choice drives metric defaults: classification models report accuracy, precision, recall, F1, AUC, and log loss; regression models report RMSE, MAE, R2, and MAPE.

AUC (area under ROC curve) is threshold-independent and ranges 0 to 1; 0.5 equals random guessing, so compare models by AUC before tuning decision thresholds.

Precision = TP/(TP+FP); recall = TP/(TP+FN); F1 is the harmonic mean 2PR/(P+R). Choose the metric that matches the business cost of false positives versus false negatives.

Feature handling options include categorical encoding (dummy/label), missing-value imputation, rescaling (standardization, min-max), and feature generation; these are configured per feature or by type.

Hyperparameter search uses grid search or random search over a defined parameter space, with cross-validation folds; the reported score is the cross-validated score, not the holdout score.

The model's interpretation tab exposes variable importance (permutation or impurity based) and partial dependence plots to show how each feature moves predictions.

Deployment turns a trained model into a saved model (versioned) that can be used in a scoring recipe or API node; retraining can be scheduled to refresh the saved model.

The Statistics tab on a dataset offers univariate analysis per column: count, missing count, distinct count, min, max, mean, median, standard deviation, and quartiles for numeric columns.

Histograms bin numeric values to show distribution shape; the bin count is adjustable, and changing bins changes apparent skew and modality, so state the binning when reporting.

Box plots mark median, quartiles (Q1, Q3), and whiskers; points beyond 1.5*IQR from the quartiles are conventionally flagged as outliers.

For categorical columns, the frequency/count chart gives top values and their share; the distinct-count and missing-count readouts reveal cardinality and data-quality gaps.

The correlation matrix computes pairwise Pearson correlation for numeric columns; values near +1 or -1 indicate linear association, but correlation does not establish causation and is sensitive to outliers.

The bivariate/cross-statistics view compares a metric across groups defined by a categorical column, reporting group means and counts to support comparison before modeling.

A statistical test worksheet can run t-tests, chi-square, or ANOVA-style comparisons; results include a p-value, and a small p-value indicates the observed difference is unlikely under the null hypothesis.

Sampling and filtering applied before the Statistics tab change the computed values; confirm the active sample or filter so reported statistics match the intended population.