Dataiku ML Practitioner key terms

Dataiku's point-and-click environment for exploring a dataset, building charts, and testing statistical hypotheses without writing code. It produces reusable analyses and cards that can be shared on dashboards or fed into discussions.

A saved visualization inside a Visual Analysis that records the chosen dimensions, measures, aggregation, and chart type. Cards can be published to dashboards and refreshed against new data.

Exploration of a single variable at a time, typically through histograms, box plots, or frequency tables. It reveals distribution shape, central tendency, spread, and outliers before any modeling begins.

Examination of the relationship between two variables, using scatter plots, correlation, or grouped comparisons. It helps decide whether a feature carries signal for a target and whether transformation is needed.

A number between minus one and one summarizing the strength and direction of a linear relationship between two numeric variables. Zero indicates no linear association, but it does not prove independence or causation.

A procedure in a Visual Analysis that computes a test statistic and p-value to judge whether an observed pattern could arise by chance. Dataiku offers tests suited to numeric, categorical, and paired data.

The default assumption that no real effect or difference exists in the population. A statistical test measures how incompatible the observed sample is with this assumption before it can be rejected.

The probability of observing a result at least as extreme as the one computed, assuming the null hypothesis is true. Small values suggest the data are unlikely under that assumption, but do not measure effect size.

A pre-chosen threshold, often 0.05, against which a p-value is compared to decide whether to reject the null hypothesis. Choosing it before analysis limits opportunistic interpretation of results.

A range of plausible values for a population parameter, built so that a stated percentage of such intervals would contain the true value across repeated sampling. Width shrinks as sample size grows.

Rejecting a true null hypothesis, that is, reporting an effect that does not actually exist. Its probability equals the significance level when the test assumptions hold.

Failing to reject a false null hypothesis, meaning a real effect goes undetected. Its probability rises when samples are small or effects are weak, lowering statistical power.

The probability that a test correctly detects an effect when one truly exists. Power increases with larger samples, bigger effect sizes, and lower noise in the measurements.

A test comparing observed category counts with counts expected under independence or a specified distribution. It is commonly used to check whether two categorical variables are associated.

A cross-tabulation counting how many records fall into each combination of two categorical variables. Row or column percentages make association patterns easier to read than raw counts.

A test comparing the means of one or two groups while accounting for variability and sample size. Variants exist for one sample, independent groups, and paired measurements.

Analysis of variance, a test that compares means across three or more groups by partitioning total variability. A significant result indicates at least one group differs, not which one.

A symmetric bell-shaped distribution fully described by its mean and standard deviation. Many classical tests assume approximate normality, which large samples often satisfy through the central limit theorem.

A measure of how far values typically sit from their mean, in the same units as the data. It complements variance, which is its square and is harder to interpret directly.

The standard deviation of a statistic's sampling distribution, such as that of a sample mean. It shrinks as sample size grows and underlies confidence intervals and many test statistics.

A chart grouping numeric values into bins and showing how many records fall in each. Bin width strongly shapes the picture, so trying several widths avoids misleading conclusions.

A chart summarizing a numeric distribution through quartiles, median, and flagged extreme values. It makes comparison across groups compact and highlights skew and outliers.

A chart placing one numeric variable on each axis to reveal relationships, clusters, and unusual points. Adding a color dimension can expose group-specific patterns hidden in a single cloud.

A record whose value lies far from the bulk of the distribution. Outliers can distort means, correlations, and model coefficients, so they should be inspected rather than blindly removed.

A variable whose values are labels or groups rather than measurable quantities. Its levels may be ordered or unordered, which determines which charts and encoding steps are appropriate.

The number of records used in an analysis or test. Larger samples narrow confidence intervals and raise power, but they also make trivial differences statistically significant.

Dataiku's visual modeling environment where a candidate assembles a dataset, chooses a prediction target, trains several algorithms, and compares them. Sessions can be deployed as saved models or exported as code.

A trained model produced inside a Lab session, holding the algorithm settings, preprocessing, and learned parameters. It can be deployed to a Flow for scoring new data.

The column a supervised model is trained to predict. Its type, binary, multiclass, or numeric, determines whether the task is classification or regression and which metrics apply.

An input column used by a model to make predictions. Features may be raw, derived, or transformed, and their handling of missing values and categories affects model quality.

Dividing data into a portion used to fit a model and a held-out portion used to estimate generalization. Dataiku also supports a validation split for tuning before final test evaluation.

Repeatedly splitting training data into folds, fitting on some and validating on others, then averaging the scores. It gives a more stable performance estimate than a single split.

Fitting a model so closely to training data that it captures noise and performs poorly on new records. Symptoms include strong training scores paired with weak validation scores.

A penalty added to a model's objective to discourage excessive complexity, shrinking coefficients or limiting tree growth. It trades a little training accuracy for better generalization.

A setting chosen before training, such as tree depth, learning rate, or regularization strength, that controls how a model learns. It is tuned by search rather than learned from data.

An optimization approach that trains a model across a defined grid of hyperparameter combinations and keeps the best according to a chosen metric. Cost grows quickly with the number of values tried.

Side-by-side ranking of candidate models on the same validation data using consistent metrics. Fair comparison requires identical splits, preprocessing, and metric definitions across sessions.

A ranking of how much each input contributes to a model's predictions, computed by methods such as permutation or impurity reduction. It helps with interpretation and feature selection but can mislead on correlated inputs.

The per-column settings in a visual model that decide how each feature is treated: categorical, numeric, ignored, or forced. Wrong choices, such as treating an identifier as numeric, degrade results.

Moving a trained visual model out of the Lab so it scores new data in the Flow or via an API endpoint. Deployed models can be retrained on refreshed data without rebuilding the analysis.