Glossary

118 terms from ML Basics, in plain words, each with the chapter that teaches it.

Search the glossary

A

Accuracy: The share of predictions that are right. It can mislead when one class is rare: predicting "nobody cancels" at Kettle & Crane is right about 90% of the time.
Average precision: A single number that summarizes the precision-recall curve. A random model scores about the positive rate, roughly 0.10 for the cancel data, so it suits very rare positives.

B

Baseline: The score of a simple model that learns nothing useful, such as always predicting the most common class. A real model must beat it to be worth using.
best_estimator_: Attributes of a fitted grid or randomized search. They hold the winning pipeline refit on all training rows, the winning settings, and their mean CV score.
Bootstrap sample: Rows drawn at random with repeats, as many as the training set has. Each tree in a random forest trains on its own bootstrap sample.
Break-even threshold: The chance of canceling at which a call's expected saving equals its cost: $10 ÷ $36 ≈ 0.278. If probabilities are calibrated, flag everyone at or above it.

C

Calibrated: Describes a model whose predicted probabilities match reality: of subscribers given 30%, about 30% really cancel. Priya's churn model is calibrated, so its scores can be read as honest chances.
calibration_curve: A function that sorts rows into groups by predicted risk and returns each group's real rate and average prediction, to check whether probabilities are honest.
category dtype: A pandas column type with a fixed set of possible values. Histogram gradient boosting can use these columns directly, without one-hot encoding.
class: One of the categories a classifier predicts, such as canceled or stayed.
class_weight: A model setting that makes rare-class rows count more during training. It distorts predicted probabilities and rarely improves ranking, so choosing a threshold is usually clearer.
Classification: Supervised learning where the target is a category, such as whether a Kettle & Crane subscriber will cancel (yes or no).
classification_report: A scikit-learn function that prints precision, recall, F1 and support for each class, plus macro and weighted averages.
Clustering: Unsupervised learning that groups similar rows without any labels, for example finding types of subscriber.
Coefficient: The weight a linear model multiplies a feature by. In Sam's café model, each extra seat adds about 0.46 lb because its coefficient is 0.46.
ColumnTransformer: A scikit-learn tool that sends different groups of columns through different preparation steps and glues the results side by side. Unlisted columns are dropped unless you set remainder="passthrough", which keeps them unchanged.
Confusion matrix: A table counting the four outcomes of a classifier: true and false positives and true and false negatives. Rows are the truth and columns are the predictions.
ConvergenceWarning: A warning that a model's step-by-step fitting stopped before settling. Scale the features or raise max_iter, which defaults to 100 for logistic regression.
Cross-validation (k-fold): Splits training rows into k folds, then trains and scores k times so each fold is the test part once. You report the mean score and its spread.
cross_val_predict: A scikit-learn function that returns a prediction for every training row from a model that never saw that row. It fits a copy of the model on all folds but one, predicts the held-out fold, and repeats, so the predictions can be used to choose a threshold without touching the test set.
cross_val_score: Cross-validates a model and returns one score per fold, using the metric named in scoring=.
cross_validate: Like cross_val_score but returns a dict with several metrics, timings and, optionally, training scores, which helps you spot overfitting.

D

Data leakage: Information getting into training that won't be available when you really predict. It makes a model look great in testing and fail in use.
Decision tree: A model that predicts by asking a series of yes/no questions about the features, like a flowchart. Each final box is a leaf with a cancel rate.
Distance: How far apart two rows are, treating each row as a point and measuring the straight-line distance between feature values. k-nearest neighbors uses it to find similar subscribers.
DummyClassifier: A scikit-learn model that ignores the features and makes simple guesses, such as the most common class or the mean. It gives the baseline a real model must beat.

E

Early stopping: Holding back part of the training data and stopping boosting once the score on it hasn't improved for a set number of rounds.
Ensemble: A model that combines many models, such as the hundreds of trees in a random forest.
Estimator: Any scikit-learn object you create with settings and then fit to data. Models and pipelines are estimators.
Example: One row of past data used for training or testing, such as one subscriber with a known outcome.

F

F1: One number that combines precision and recall as their harmonic mean. It stays low if either one is low.
False positive / negative: A false positive is a subscriber flagged but who really stayed (a wasted call). A false negative is a canceler the model didn't flag.
False positive rate: The share of real negatives that were flagged: FP ÷ (FP + TN). It is the horizontal axis of the ROC curve.
Feature: An input column the model uses to predict, such as tenure, plan or skipped deliveries.
Feedback loop: When a model's predictions change the data it later learns from. Subscribers Priya calls and keeps show up as "stayed", which can teach a retrained model the wrong lesson.
fit: The scikit-learn method that learns from data. Transformers then use transform to change any data the same way, and fit_transform does both at once.
Fold: One of the k equal parts the rows are cut into for cross-validation.

G

Google Colab: A free service that runs Python notebooks in your browser, so you don't have to install anything.
Gradient boosting: Builds small trees one after another, each trained to correct the errors left so far, then adds them up. Often the most accurate model for tables of data.
GridSearchCV: Tries every combination of the hyperparameter values you list, cross-validates each, and keeps the best. It then refits the winner on all training rows.
Group leakage: Leakage when rows from the same group, such as one café's weeks, land on both sides of a split. Keep each group on one side with GroupShuffleSplit or GroupKFold.
GroupKFold: A cross-validation splitter that keeps all rows of one group, such as one café's weeks, in the same fold. Pass groups= when you use it.

H

Harmonic mean: An average of rates that stays low if either one is low. F1 uses it to combine precision and recall.
HistGradientBoostingClassifier: scikit-learn's fast gradient boosting model. It first sorts each feature into at most 255 bins, accepts missing values, and can use category columns directly.
Hyperparameter: A setting you choose before fitting, such as n_neighbors or max_depth. Parameters, by contrast, are what the model learns.

I

Imbalanced classes: When one class is much rarer than the other. Only about 10% of Kettle & Crane subscribers cancel, so accuracy hides almost everything.
Imputation: Filling missing values with an estimate, such as the median learned from the training rows.
Interaction: When one feature's effect depends on another. In the café data, temperature matters about twice as much for cafés with a drive-through, and trees can find this on their own.
Intercept: A linear model's starting value, the prediction when every feature is zero.

J

joblib: The library, installed with scikit-learn, used to save a fitted model to a file with dump and read it back with load.

K

k-nearest neighbors: A model that predicts for a new row by finding the k most similar training rows and letting them vote. Training just stores the examples.
KFold: Splitter objects that control how cross-validation folds are cut. Use StratifiedKFold for classification and TimeSeriesSplit for time-ordered rows, which always trains on the past and tests on the future.

L

Label: The known answer for a training example, for instance whether a subscriber canceled in September.
Leaf: A final box of a decision tree. It holds the training rows that ended up there, and the prediction is their cancel rate.
Learning curve: A plot of training and validation scores against the number of training rows. It shows whether more data would help.
Learning rate: In gradient boosting, how much of each new tree's correction is added. Smaller steps learn more carefully but need more trees.
Linear regression: Predicts a number as a weighted sum of the features plus an intercept, as Sam does to forecast each café's weekly order.
Log-odds: The logarithm of the odds p ÷ (1 − p). It is the scale of logistic regression's weighted sum, where each step of 1 multiplies the odds by about 2.7.
Logistic regression: A classifier that computes a weighted sum and squeezes it through an S-shaped curve so the result is a probability between 0 and 1. Despite its name, it classifies.
loguniform: Range generators from scipy.stats for randomized search. randint draws whole numbers; loguniform draws decimals spread evenly on a log scale.

M

Machine learning: Building models that learn patterns from past examples with known answers, instead of writing the rules by hand.
MAE: Mean absolute error: the average size of the errors, in the target's units. A café MAE of 4.5 means predictions are typically off by about 4.5 lb.
make_pipeline: Chains steps and a model into a pipeline, naming each step after its lower-case class (such as standardscaler). Use named_steps to reach a step.
max_depth: Hyperparameters that rein in trees. max_depth is the most questions in a row; min_samples_leaf is the fewest training rows a leaf may hold; max_leaf_nodes caps the leaves per tree.
max_iter: In logistic regression, the maximum number of fitting steps (default 100). In histogram gradient boosting, the number of trees (rounds) to build.
Model: A function learned from data that takes the facts about a new row, such as a subscriber, and returns a prediction.
Model card: A short document describing a model's purpose, data, performance, limits and owner for the people who will use or review it.

N

n_estimators: The number of trees in a random forest. More trees mostly make results steadier rather than more accurate.
NumPy array: A grid of numbers that works much like a DataFrame without labels. scikit-learn returns predictions as arrays, and array[:, 1] means every row, column 1.

O

One-hot encoding: Turns a category column into one 0/1 column per category, so "plan" becomes plan_weekly, plan_biweekly and plan_monthly without implying an order. With handle_unknown="ignore", a category never seen in training becomes all zeros instead of an error.
Out-of-fold prediction: A prediction for a row made by a model that was not trained on that row, because the row was in the fold held out during cross-validation. These predictions give an honest estimate of how the model behaves on unseen data, so they can be used to choose a threshold without using the test set.
Overfitting: Learning the training data's noise instead of the general pattern. The training score is far above the validation score.

P

Parameter: A value the model learns during fitting, such as a coefficient. Names ending in an underscore, like mean_, are things learned in fit.
parse_dates: A read_csv option that turns date text into real dates so you can compare them, for example to split weeks by a cutoff.
Permutation importance: Shuffles one column at a time on held-out rows and measures how much the score drops. A big drop means the model relied on that column.
Pickle: Python's format for saving objects, used by joblib. Loading a pickle can run arbitrary code, so only load files you trust.
Pipeline: A chain of preparation steps and a model that acts as one estimator. It fits every step on training data only, so preparation can't leak.
Positive class: The class you are looking for, here "canceled". It is not meant as a good thing.
Precision: TP ÷ (TP + FP): of the subscribers flagged, the share that really canceled. It tells Priya how many of her calls are worth making.
Precision-recall curve: A plot of precision against recall at every threshold. It shows the cost of wasted flags that ROC AUC can hide when positives are rare.
Predicted probability: The model's estimated chance of each class, one column per class. predict_proba returns them, and predict picks the class with the higher one.
Preprocessing leakage: Leakage from fitting preparation steps, such as filling gaps with the average, on rows that include test rows. Putting them in a pipeline prevents it.
Proxy: A feature that stands in for another, often sensitive, one. Zip code can track income or race, so leaving race out doesn't keep its effect out.

R

Random forest: Hundreds of varied decision trees whose predictions are averaged. It is more accurate and stable than one tree.
random_state: A setting that fixes randomness, such as the shuffle in a split, so you get identical results every run. Any whole number works.
RandomizedSearchCV: Tries a fixed number of random hyperparameter combinations drawn from ranges you give, cross-validating each. It is usually as good as a full grid in less time.
Recall (true positive rate): TP ÷ (TP + FN): of the subscribers who really canceled, the share the model flagged.
Regression: Supervised learning where the target is a number, such as how many pounds a café will order next week.
Regularization: A penalty on large coefficients that keeps a model simpler and guards against overfitting. In logistic regression, C controls it, and smaller C means more regularization.
Residual (error): The true value minus the prediction for one row.
RMSE: Root mean squared error: square the errors, average them, take the square root. Like MAE, but big misses count extra.
ROC curve, ROC AUC: The ROC curve plots true positive rate against false positive rate at every threshold. Its area (AUC) is the chance a random canceler outranks a random stayer: 0.5 is guessing and 1.0 is perfect.
R²: The share of variation explained, compared with always predicting the mean. 1 is perfect, 0 is no better than the mean, and below 0 is worse.

S

scikit-learn (sklearn): Python's main free, open-source machine learning library, imported as sklearn.
SciPy: A math library that scikit-learn builds on, installed alongside it.
SimpleImputer: A transformer that fills gaps with the mean, median, most frequent value or a constant, learned from training rows. With add_indicator=True it also adds a 0/1 column marking which rows had a gap.
skops: A library that loads saved models without running arbitrary code, suggested for files from outside sources.
Sparse matrix: A memory-saving table format that stores only the non-zero values. OneHotEncoder returns one by default; sparse_output=False gives an ordinary array.
Standardizing: Rescaling each column to mean 0 and standard deviation 1 (z-scores), so no column dominates just because of its units. StandardScaler does it.
Step-parameter names: To set a hyperparameter of a step inside a pipeline, write the step's name, two underscores, then the setting, as in histgradientboostingclassifier__max_depth.
Stratified: Keeping class shares equal across splits or folds, so each part has about 10% cancelers. In train_test_split you pass stratify=y.
Supervised learning: Learning from labeled examples, as if a teacher had marked the answers. Classification and regression are its two kinds.
Support: In a classification report, the number of true rows of each class.

T

Target: The thing the model predicts, such as whether a subscriber canceled. By convention it is stored in y.
Target leakage: A feature that is a consequence of the target rather than a clue to it. The exit survey is sent only after someone cancels, so it gives the answer away.
Test set: Rows held back for one final, honest check at the very end. Changing the model because of its test score makes the test set stop being honest.
Threshold: The probability at or above which a row is flagged. 0.5 is the default but only a choice. Lowering it raises recall and lowers precision, and it should be set by cost.
Tie (within the fold spread): When two models' CV scores differ by less than their fold-to-fold spread, treat them as tied and prefer the simpler one.
Time-based split: For data collected over time, train on earlier rows and test on later ones, instead of shuffling. A random split would let the model learn from the future.
train_test_split: A scikit-learn function that shuffles rows and returns X_train, X_test, y_train and y_test. test_size sets the held-back share.
Training (fitting): Finding the patterns in past examples, that is, learning a model's parameters from data.
Training set: The rows the model learns from.
Transformer: An object that learns something from training data with fit, then changes any data the same way with transform. Scalers, encoders and imputers are transformers.
True positive / negative: A true positive is a subscriber flagged who really canceled. A true negative is one not flagged who really stayed.

U

Underfitting: Being too simple to capture the pattern. Training and validation scores are both low and close together.
Unsupervised learning: Finding structure in data, such as groups, without any labels.

V

Validation curve: A plot of training and CV scores across values of one hyperparameter. It shows where the model underfits, where it overfits and where the sweet spot is.
Validation set: Rows used to make choices, such as which model or which k, so the test set stays untouched for the final check.