Dataiku ML Practitioner flashcards

31 cards. Read the term and the definition here. The deck at the top of this page is optional.

Visual ML Lab

A Dataiku workspace where a dataset is split, preprocessed, and trained into one model per experiment, with results, metrics, and interpretation shown in the interface.

Train/test split

A step that divides rows into a training portion used to fit the model and a held-out test portion used to estimate generalization, with configurable ratio and seed.

AUC

Area under the ROC curve, ranging 0 to 1 and threshold-independent; 0.5 means random ranking, and higher values mean better separation of positive from negative classes.

Precision formula

Precision equals true positives divided by true positives plus false positives, measuring how many predicted positives are actually positive; useful when false positives are costly.

Recall formula

Recall equals true positives divided by true positives plus false negatives, measuring how many actual positives were found; useful when missing positives is costly.

F1 score

The harmonic mean of precision and recall, computed as 2 times precision times recall divided by precision plus recall; a single balanced metric for imbalanced classification.

Hyperparameter search

Grid or random search over a defined parameter space, scored with cross-validation folds, to select algorithm settings that generalize rather than settings that fit one split.

Feature handling

Per-feature or per-type configuration of encoding for categoricals, imputation for missing values, and rescaling such as standardization or min-max before the algorithm sees the data.

Variable importance

A model interpretation readout ranking features by their contribution, often permutation or impurity based, showing which inputs most affect predictions on the evaluated data.

Partial dependence plot

A chart showing how the model's predicted outcome changes as one feature varies while other features are held at their observed distribution, revealing direction and shape of effect.

Saved model

A versioned, deployable artifact produced from a trained experiment, usable in scoring recipes or API nodes and refreshable through scheduled retraining on new data.

Regression metrics

RMSE, MAE, R2, and MAPE summarize numeric prediction error; RMSE penalizes large errors more than MAE, and R2 reports the share of variance explained.

Cross-validation

Repeatedly splitting training data into folds, fitting on all but one fold and scoring on the held-out fold, then averaging scores to reduce variance of the performance estimate.

Decision threshold

The probability cutoff that converts a model score into a predicted class; moving it trades precision against recall without retraining the underlying model.

Class imbalance handling

Techniques such as class weighting, resampling, or threshold adjustment applied when one class is rare, because accuracy alone misleads and precision/recall or AUC are more informative.

Overfitting

A model fitting noise in training data, shown by strong training scores but weak test or cross-validation scores; remedies include simpler models, regularization, and more data.

Univariate statistics

Per-column summary values including count, missing count, distinct count, min, max, mean, median, standard deviation, and quartiles, shown in a dataset's Statistics tab.

Histogram binning

Grouping numeric values into intervals to display distribution shape; the chosen bin count changes apparent skew and modality, so reporting should state the binning used.

Box plot whiskers

Lines extending from quartiles, conventionally to 1.5 times the interquartile range; points beyond the whiskers are flagged as potential outliers.

Interquartile range

The difference between the third quartile and the first quartile, covering the middle half of values and serving as the scale for box-plot outlier whiskers.

Frequency chart

A categorical-column chart listing top values with counts and their share of rows, exposing dominant categories, rare categories, and cardinality problems.

Distinct count

The number of unique values in a column; high cardinality complicates encoding and can signal an identifier column rather than a useful feature.

Missing count

The number of rows with no value in a column; paired with its share of total rows, it indicates whether imputation or exclusion is warranted before modeling.

Pearson correlation

A pairwise measure of linear association between numeric columns ranging -1 to 1; sensitive to outliers and unable to capture non-linear relationships or causation.

Correlation matrix

A grid of pairwise correlation values across numeric columns, used to spot strongly related features that may be redundant or that suggest multicollinearity in models.

Cross statistics

A bivariate view comparing a numeric metric across groups defined by a categorical column, reporting group means and counts for pre-model comparison.

p-value

The probability of observing a result at least as extreme as the data under the null hypothesis; small values suggest the observed difference is unlikely by chance alone.

Null hypothesis

The default assumption that there is no difference between groups or no association between variables, which a statistical test either rejects or fails to reject.

Chi-square test

A test comparing observed versus expected counts in categorical cross-tabulations, returning a p-value for whether the two categorical variables are associated.

T-test

A test comparing means between two groups, returning a p-value; it assumes roughly normal distributions and is commonly used in group comparison worksheets.

Sample effect on statistics

Any sample or filter active before the Statistics tab changes computed values; confirming the active sample ensures reported statistics describe the intended population.