Glossary

Every term introduced in the course, in one place.

A B C D E F G H I J K L M N O P R S T U V X

A

accuracy

The fraction of samples whose predicted class equals the true class. Easy to understand, but on imbalanced data a useless classifier can score very high, so always report it with a baseline and with precision and recall.

Introduced in 2.5.

See also: class imbalance, precision, recall

axis

One dimension of an array. For a 2-D array axis 0 indexes rows and axis 1 indexes columns, so X.mean(axis=0) gives the mean of every column (one value per feature) and X.mean(axis=1) the mean of every row.

Introduced in 1.2.

See also: shape

B

baseline

A deliberately simple reference model, such as always predicting the mean of the training targets or the most common class. If your model cannot beat the baseline on held-out data, it has learned nothing useful. In scikit-learn, DummyRegressor and DummyClassifier build baselines.

Introduced in 2.3.

See also: generalisation, test set

bias–variance trade-off

The tension between two sources of error. Bias is the error from a model that is too simple to capture the true pattern; variance is the error from a model so flexible that it follows the noise in a particular training set. Increasing model complexity lowers bias but raises variance.

Introduced in 2.4.

See also: overfitting, underfitting

boolean mask

An array of booleans with the same shape as the data, usually produced by a comparison such as x > 0, and used inside square brackets to select just the elements where it is True. Masks are how you filter data without loops in NumPy and pandas.

Introduced in 1.2.

See also: vectorisation

box plot

A chart showing the median (line), the middle 50% of values (box, from the 25th to the 75th percentile), whiskers to the most extreme non-outlying points and individual outliers beyond them. Violin plots show the same plus the full shape of the distribution.

Introduced in 1.4.

See also: outlier

broadcasting

The set of rules NumPy uses to combine arrays of different but compatible shapes, for example subtracting a vector of column means from every row of a matrix without writing a loop.

Introduced in 1.2.

See also: vectorisation

C

calibration

A classifier is well calibrated when its predicted probabilities match observed frequencies, so that among samples predicted with probability 0.8, about 80% are actually positive. Good ranking (AUC) does not guarantee good calibration.

Introduced in 2.5.

See also: logistic regression

categorical data

Data whose values are labels from a fixed set (species, island, sex) rather than measurements. pandas has a dedicated category dtype for them; models need them encoded as numbers, for example with one-hot encoding.

Introduced in 1.3.

See also: one-hot encoding

class imbalance

A situation where one class is far more common than another, for example 1% active compounds in a screen. Accuracy then rewards predicting the majority class; precision, recall, PR curves and class weights are the tools that cope with it.

Introduced in 2.5.

See also: precision, recall

classification

A supervised learning task where the target is a category, such as toxic/non-toxic or one of several cell types.

Introduced in 2.1.

See also: regression, label

coefficient

The weight multiplying one feature in a linear model. Its size depends on the feature’s units and scale, and when features are correlated the coefficients can swap importance between them, so interpret them cautiously.

Introduced in 2.2.

See also: parameter

coefficient of determination (R²)

\(R^2 = 1 - \frac{\sum_i (y_i - \hat{y}_i)^2}{\sum_i (y_i - \bar{y})^2}\). A value of 1 means perfect predictions, 0 means no better than predicting the mean, and negative values are possible on test data when the model is worse than the mean baseline.

Introduced in 2.3.

See also: mean absolute error (MAE), root mean squared error (RMSE), baseline

colour-blind-safe palette

A palette whose colours differ in lightness and hue in ways that remain distinguishable for readers with colour-vision deficiency (about 8% of men and 0.5% of women). The Okabe–Ito palette and viridis are standard choices; red–green pairs are the classic mistake.

Introduced in 1.4.

ColumnTransformer

A scikit-learn transformer that routes columns to different preprocessing steps, for example scaling the numeric columns and one-hot encoding the categorical ones, then concatenates the results. It is the standard first step of a pipeline on mixed data.

Introduced in 2.6.

See also: pipeline, one-hot encoding

conda / mamba

conda is a package and environment manager; mamba is a faster drop-in replacement. Both read environment.yml, create an isolated environment and install libraries (including compiled ones such as RDKit later in the course) from the conda-forge channel.

Introduced in 0.2.

See also: virtual environment

confusion matrix

A table that cross-tabulates true classes against predicted classes. For a binary problem it has four cells: true positives, false positives, false negatives and true negatives. Precision, recall and accuracy are all read off it.

Introduced in 2.5.

See also: precision, recall

correlation coefficient

Pearson’s \(r\), a number between \(-1\) and \(1\) measuring the strength of the linear relationship between two variables. Zero means no linear relationship, not no relationship. A heatmap of all pairwise correlations is a standard EDA figure.

Introduced in 1.4.

See also: scatter plot

cross-validation

A resampling method that splits the data into \(k\) folds, trains on \(k-1\) of them and evaluates on the remaining fold, rotating until each fold has served as the evaluation set. The \(k\) scores are averaged. It gives a more stable estimate than a single split and is the standard way to choose hyperparameters.

Introduced in 2.4.

See also: validation set, hyperparameter

D

data leakage

Any way in which information that would not be available at prediction time, or information from the test set, influences training. Typical causes are preprocessing before splitting, duplicated samples across the split, and features derived from the target. Leakage makes test scores unrealistically good.

Introduced in 2.3.

See also: train/test split, test set

DataFrame

The central pandas data structure, a two-dimensional table with labelled rows (the index) and labelled columns, where each column has its own data type.

Introduced in 1.3.

See also: tidy data

decision boundary

The set of points in feature space where a classifier’s prediction changes from one class to another. For logistic regression it is a straight line (hyperplane); for kNN it is irregular and follows the training points.

Introduced in 2.5.

See also: classification

decision threshold

The cut-off applied to a predicted probability to decide the class, 0.5 by default. Moving it trades precision against recall, which is what ROC and precision–recall curves display across all thresholds.

Introduced in 2.5.

See also: ROC curve and AUC, precision–recall curve

distribution shift

A systematic difference between the data a model was trained on and the data it is applied to, for example new chemical scaffolds, a different hospital or a different sequencing batch. Models typically degrade under shift, and random splits do not reveal it.

Introduced in 2.1.

See also: generalisation, scaffold split

dtype

The data type of the elements of a NumPy array or pandas column, for example float64, int64, bool or str. Mixing types in one array forces a common type (often object), which is slow and a sign that something is wrong.

Introduced in 1.2.

See also: NumPy array

duplicate

A row that repeats another one, exactly or in all the ways that matter (the same molecule under two names, the same patient measured twice). Duplicates inflate apparent dataset size, and when copies fall on both sides of a train/test split they leak information.

Introduced in 1.5.

See also: data leakage

E

estimator

scikit-learn’s name for any object that learns from data through a fit method. Models add predict and score; preprocessors add transform. Because all estimators share this interface, they can be swapped and combined in pipelines.

Introduced in 2.2.

See also: pipeline

exploratory data analysis (EDA)

The first stage of any data project: summarising and plotting the data to understand distributions, relationships, missing values, outliers and duplicates before any model is fitted.

Introduced in 1.5.

See also: tidy data

F

f-string

A formatted string literal, written f"...", in which expressions inside braces are evaluated and inserted. Format specifiers control the output, for example f"{rmse:.3f}" prints three decimals.

Introduced in 1.1.

F1 score

\(F_1 = 2 \cdot \frac{\mathrm{precision} \cdot \mathrm{recall}}{\mathrm{precision} + \mathrm{recall}}\), a single number summarising the trade-off between precision and recall. It is high only when both are high.

Introduced in 2.5.

See also: precision, recall

feature

A measurable property of a sample used as input to a model, such as molecular weight or the number of hydrogen-bond donors. The features of all samples form the matrix \(X\) with one row per sample and one column per feature.

Introduced in 1.2.

See also: label, molecular descriptor

feature scaling

Transforming each feature so that they have comparable ranges, most often by standardising to zero mean and unit standard deviation (StandardScaler). Distance-based models such as kNN and regularised models depend on it.

Introduced in 2.6.

See also: k-nearest neighbours (kNN), pipeline

G

generalisation

The ability of a model to perform well on data it has not seen during training. Generalisation, not training performance, is the real goal of machine learning.

Introduced in 2.1.

See also: overfitting, test set

Google Colab

Google Colaboratory, a free hosted Jupyter notebook service that runs in the browser. Every lab has an Open in Colab button. Files you create there disappear when the session ends unless you save them to Drive.

Introduced in 0.2.

See also: Jupyter notebook

gradient

For a function of several variables, the vector of its partial derivatives. It points in the direction in which the function increases fastest, which is why optimisation steps go the opposite way.

Introduced in A.

See also: gradient descent

gradient descent

An iterative optimisation method that repeatedly moves the parameters a small step in the direction that decreases the loss fastest, the negative gradient. The step size is the learning rate.

Introduced in 2.2.

See also: loss function, gradient

grid search

Systematically evaluating every combination of candidate hyperparameter values with cross-validation and keeping the best. scikit-learn’s GridSearchCV does this and then refits the best model on all the training data.

Introduced in 2.4.

See also: cross-validation, hyperparameter

group split

Any train/test or cross-validation split that assigns whole groups of related samples, such as all measurements from one patient or all molecules with one scaffold, entirely to one side. scikit-learn’s GroupKFold implements it.

Introduced in 2.6.

See also: scaffold split, data leakage

groupby

The pandas operation that splits a DataFrame into groups sharing a key (for example species), applies an aggregation such as mean or count to each group, and combines the results into a new table. The pattern is often called split-apply-combine.

Introduced in 1.3.

See also: DataFrame

H

histogram

A chart that divides the range of a variable into bins and draws a bar whose height is the number of values in each bin. The first plot to make for any numeric column.

Introduced in 1.4.

See also: kernel density estimate (KDE)

hyperparameter

A setting of a learning algorithm that is not learned from the data but chosen beforehand, such as the regularisation strength of ridge regression or the number of neighbours in kNN. Hyperparameters are tuned with validation data or cross-validation, never with the test set.

Introduced in 2.4.

See also: cross-validation, regularisation

I

i.i.d.

Random variables \(X_1, \ldots, X_n\) are independent and identically distributed if they all have the same distribution and are mutually independent. Most results about averages (the law of large numbers, the central limit theorem) assume it, and a random train/test split assumes that training and test samples are i.i.d. draws from the same population. Duplicates, closely related molecules and batch effects break the assumption, which is why later chapters use scaffold and group splits.

Introduced in A.

See also: train/test split, data leakage

imputation

Replacing missing values with estimates, for example the column median (SimpleImputer). It must be fitted on the training data only and applied inside a pipeline, otherwise test information leaks into training.

Introduced in 2.6.

See also: missing value (NaN), pipeline

index (pandas)

The labels attached to the rows of a DataFrame or Series. By default they are integers starting at 0, but they can be anything (compound names, dates). loc selects by these labels, iloc by position.

Introduced in 1.3.

See also: loc and iloc

J

joblib

A utility library for saving Python objects efficiently. joblib.dump(model, "model.joblib") persists a fitted scikit-learn pipeline and joblib.load restores it. Record the library versions alongside the file: a model saved with one scikit-learn version may not load with another.

Introduced in 2.6.

Jupyter notebook

A document made of cells that can hold code, its output (text, tables, figures) or formatted text. Notebooks run in JupyterLab, VS Code or Google Colab. The course labs are notebooks; the episodes are rendered from the same kind of cells.

Introduced in 0.2.

See also: kernel, Google Colab

K

k-nearest neighbours (kNN)

A model that predicts a sample by looking up the \(k\) training samples closest to it in feature space and averaging their labels (regression) or taking a vote (classification). It has no training phase and is sensitive to feature scaling.

Introduced in 2.5.

See also: feature scaling

kernel

The background Python process that executes notebook cells. It keeps all variables in memory between cells, which is why running cells out of order can confuse you: restart the kernel and run from the top when in doubt.

Introduced in 0.2.

See also: Jupyter notebook

kernel density estimate (KDE)

A smooth curve estimating the distribution of a variable, obtained by placing a small bump (kernel) at every data point and summing them. It is a smoothed histogram; the bandwidth controls how smooth.

Introduced in 1.4.

See also: histogram

L

label

The target value associated with each sample, which a supervised model learns to predict. Continuous labels give a regression problem; categorical labels give a classification problem. The labels of all samples form the vector \(y\).

Introduced in 2.1.

See also: feature, regression, classification

lasso regression

Linear regression whose loss adds \(\alpha \sum_j |\theta_j|\). Unlike ridge, the absolute-value penalty drives some coefficients to exactly zero, so lasso performs feature selection as a side effect.

Introduced in 2.4.

See also: regularisation, ridge regression

learning curve

A plot of training and validation score against the number of training samples. A large persistent gap signals overfitting (more data would help); two low curves close together signal underfitting (a richer model would help).

Introduced in 2.4.

See also: validation curve, bias–variance trade-off

learning rate

The factor that scales each gradient-descent step. Too large and the loss oscillates or explodes; too small and training takes forever. It is the first hyperparameter of any iterative optimiser.

Introduced in 2.2.

See also: gradient descent

list comprehension

A compact expression that builds a list (or dict, set) by transforming and filtering another sequence, for example [m.upper() for m in names if m]. It replaces a three-line for-loop with an append.

Introduced in 1.1.

loc and iloc

The two indexers of a DataFrame. df.loc[rows, cols] uses labels (index values and column names, including boolean masks); df.iloc[rows, cols] uses integer positions. Using plain square brackets for anything beyond a column name is a common source of confusion.

Introduced in 1.3.

See also: index (pandas), boolean mask

log transform

Replacing a positive variable \(x\) by \(\log_{10} x\) (or \(\ln x\)). It compresses quantities that span several orders of magnitude, such as solubilities or concentrations, into a range where additive models and symmetric error measures make sense.

Introduced in 1.5.

See also: logS

logistic regression

A linear model for binary classification: it computes a weighted sum of the features and passes it through the sigmoid function to give a probability between 0 and 1. Despite the name it is a classifier.

Introduced in 2.5.

See also: classification, sigmoid function

logS

The base-10 logarithm of a compound’s aqueous solubility expressed in moles per litre. A logS of −3 means a solubility of 0.001 mol/L. Taking the logarithm makes a quantity that spans many orders of magnitude easier to model.

Introduced in 1.5.

See also: molecular descriptor

loss function

A function that measures the disagreement between predictions and true values, such as the mean squared error. Training a model means adjusting its parameters to make the loss on the training data small.

Introduced in 2.1.

See also: mean squared error (MSE), gradient descent

M

mean absolute error (MAE)

The average of the absolute differences between predictions and true values, \(\mathrm{MAE} = \frac{1}{n}\sum_i |y_i - \hat{y}_i|\). It is in the same units as the target and treats all errors proportionally.

Introduced in 2.3.

See also: root mean squared error (RMSE), coefficient of determination (R²)

mean squared error (MSE)

\(\mathrm{MSE} = \frac{1}{n}\sum_i (y_i - \hat{y}_i)^2\), the average squared difference between predictions and truth. It is the loss minimised by ordinary least-squares regression; its square root is the RMSE.

Introduced in 2.2.

See also: root mean squared error (RMSE), loss function

melt and pivot

Two reshaping operations. melt converts a wide table with one column per measurement into a long one with a variable and a value column, which is what seaborn expects. pivot and pivot_table go the other way, spreading one column’s categories into separate columns.

Introduced in 1.3.

See also: tidy data

merge (join)

Combining two DataFrames by matching rows on one or more key columns, as in a database join. how="inner" keeps only matching rows; how="left" keeps every row of the left table.

Introduced in 1.3.

See also: DataFrame

missing value (NaN)

A cell with no recorded value, represented as NaN (not a number) or <NA> in pandas. Count them with df.isna().sum(). Options are to drop the rows, drop the column or impute a replacement; scikit-learn models refuse to fit with NaNs present.

Introduced in 1.3.

See also: imputation

model

A function \(f(x;\theta)\) that turns a feature vector \(x\) into a prediction, with parameters \(\theta\) that are adjusted (trained) to make its predictions match known labels. Linear regression, logistic regression, kNN and neural networks are all models in this sense.

Introduced in 2.1.

See also: parameter, loss function

molecular descriptor

A numerical property computed from a molecular structure, for example molecular weight, number of rings or topological polar surface area. Descriptors are the simplest way to turn molecules into feature vectors; Chapter 5 computes them with RDKit.

Introduced in 1.5.

See also: feature, SMILES

N

nested cross-validation

An outer cross-validation loop for estimating performance with an inner loop, run on each outer training fold, for choosing hyperparameters. It is the honest way to report the performance of a tuned model when you cannot afford a separate test set.

Introduced in 2.4.

See also: cross-validation

normal equation

The linear system \(X^\top X \theta = X^\top y\), obtained by setting the gradient of the mean squared error to zero; its solutions are exactly the least-squares fits. When the columns of \(X\) are linearly independent, \(X^\top X\) is invertible and the unique solution is \(\hat\theta = (X^\top X)^{-1} X^\top y\). In code, solve the system rather than inverting the matrix: np.linalg.solve(X.T @ X, X.T @ y).

Introduced in 2.2.

See also: gradient descent, mean squared error (MSE)

NumPy array

NumPy’s core data structure, ndarray: a grid of elements that all share one data type, with a shape such as (1128, 5). Arithmetic on arrays is applied element-wise in compiled code, which is what makes scientific Python fast.

Introduced in 1.2.

See also: dtype, shape, vectorisation

O

one-hot encoding

Representing a categorical feature with one binary column per category, so that a model that only understands numbers can use it. scikit-learn’s OneHotEncoder does this inside a pipeline.

Introduced in 2.6.

See also: pipeline

outlier

A value that lies far from the bulk of the data. Outliers can be measurement errors, data-entry mistakes, or genuine rare cases, and only domain knowledge can tell which. Look at them before deciding whether to remove them.

Introduced in 1.5.

See also: exploratory data analysis (EDA)

overfitting

When a model fits its training data so closely, noise included, that it performs worse on new data. Symptoms are a large gap between training and test performance. Remedies include more data, simpler models, regularisation and cross-validated model selection.

Introduced in 2.4.

See also: underfitting, regularisation, generalisation

P

parameter

A value inside a model that is learned from the data, such as the slope and intercept of a line. Distinguish from hyperparameters, which are chosen before training.

Introduced in 2.1.

See also: hyperparameter

parity plot

A scatter plot with the true value on one axis and the predicted value on the other. Perfect predictions fall on the diagonal line \(y = x\). It shows at a glance where a model is biased, where it saturates and which points are outliers.

Introduced in 2.3.

See also: residual

pathlib

The standard-library module whose Path class represents file-system paths. It joins paths with /, checks existence with .exists() and reads files with .read_text(), and it handles Windows and Unix separators for you.

Introduced in 1.1.

pipeline

A scikit-learn object that chains preprocessing steps (scaling, imputation, encoding) with a final model so that they are fitted together on the training data and applied together to new data. Pipelines prevent data leakage and make cross-validation honest.

Introduced in 2.6.

See also: estimator, data leakage

polynomial features

Features created by raising the originals to powers and multiplying them together (\(x^2\), \(x_1 x_2\), …). A linear model on polynomial features fits curved relationships; raising the degree too far is the classic way to overfit.

Introduced in 2.4.

See also: overfitting

precision

\(\mathrm{precision} = \frac{TP}{TP + FP}\), the fraction of predicted positives that are truly positive. High precision means few false alarms.

Introduced in 2.5.

See also: recall, confusion matrix

precision–recall curve

A plot of precision against recall as the decision threshold varies. Unlike the ROC curve it ignores true negatives, so it is the better picture when the positive class is rare, as in a screening campaign.

Introduced in 2.5.

See also: ROC curve and AUC, class imbalance

R

random number generator

NumPy’s Generator object, created with np.random.default_rng(seed), which produces pseudo-random numbers. Fixing the seed makes every run identical. Create one generator at the top of a script and pass it around rather than calling the legacy np.random.* functions.

Introduced in 1.2.

See also: random state

random state

A number (the seed) that initialises a pseudo-random number generator so that a random operation, such as a train/test split or a shuffled cross-validation, produces exactly the same result every time it is run. In scikit-learn the argument is random_state; in NumPy you create a generator with np.random.default_rng(seed).

Introduced in 1.2.

See also: train/test split

recall

\(\mathrm{recall} = \frac{TP}{TP + FN}\), the fraction of actual positives that the model identifies. Also called sensitivity. High recall means few misses.

Introduced in 2.5.

See also: precision, confusion matrix

regression

A supervised learning task where the target is a continuous number, such as a solubility, a binding affinity or a body mass.

Introduced in 2.1.

See also: classification, label

regularisation

Adding a penalty term to the loss function that discourages large or many coefficients, trading a little training accuracy for better generalisation. Ridge regression penalises squared coefficients; Lasso penalises absolute values and drives some coefficients to exactly zero.

Introduced in 2.4.

See also: overfitting, hyperparameter

residual

The difference between an observed value and the model’s prediction, \(r_i = y_i - \hat{y}_i\). Plotting residuals against predictions or against features reveals systematic errors that a single summary metric hides.

Introduced in 2.3.

See also: parity plot

ridge regression

Linear regression whose loss adds \(\alpha \sum_j \theta_j^2\), the squared size of the coefficients. The penalty shrinks all coefficients towards zero, stabilising the fit when features are correlated or numerous. \(\alpha\) is a hyperparameter.

Introduced in 2.4.

See also: regularisation, lasso regression

ROC curve and AUC

The receiver operating characteristic curve plots the true-positive rate against the false-positive rate as the decision threshold varies. The area under it (AUC) summarises ranking quality: 1 is perfect, 0.5 is random.

Introduced in 2.5.

See also: precision, recall

root mean squared error (RMSE)

The square root of the mean of the squared errors, \(\mathrm{RMSE} = \sqrt{\frac{1}{n}\sum_i (y_i - \hat{y}_i)^2}\). Also in target units, but because errors are squared first, a few large mistakes raise RMSE much more than they raise MAE.

Introduced in 2.3.

See also: mean absolute error (MAE), coefficient of determination (R²)

S

scaffold split

A way of splitting a molecular dataset by core structure (scaffold) rather than at random, so that the test set contains scaffolds the model has never seen. It gives a more realistic estimate of performance on new chemistry. Introduced properly in Part II.

Introduced in 2.6.

See also: distribution shift, group split

scatter plot

A chart with one point per sample positioned by two numeric variables, optionally coloured by a third (hue). It shows relationships, clusters and outliers that summary statistics hide.

Introduced in 1.4.

See also: correlation coefficient

Series

pandas’ one-dimensional labelled array. Selecting one column of a DataFrame gives a Series; it has an index, a name and a single dtype.

Introduced in 1.3.

See also: DataFrame, index (pandas)

shape

A tuple giving the length of an array along each of its axes. A feature matrix has shape (n_samples, n_features); a label vector has shape (n_samples,). Checking .shape is the first thing to do when code misbehaves.

Introduced in 1.2.

See also: NumPy array, axis

sigmoid function

\(\sigma(z) = \frac{1}{1 + e^{-z}}\), an S-shaped function mapping any real number to the interval \((0, 1)\). Logistic regression uses it to turn a linear score into a probability.

Introduced in 2.5.

See also: logistic regression

SMILES

Simplified Molecular-Input Line-Entry System, a compact text notation for molecular structures, for example CCO for ethanol or c1ccccc1 for benzene. In this course SMILES strings appear as plain text columns until Chapter 5, where RDKit parses them into molecules.

Introduced in 1.5.

See also: molecular descriptor

supervised learning

Learning a function from input features to known labels, using examples where both are given. Regression and classification are the two main kinds.

Introduced in 2.1.

See also: unsupervised learning, label

T

target leakage

A form of data leakage where a feature carries information about the target that would not be available at prediction time, for instance a lab value recorded only for patients already diagnosed. Models with target leakage look superb in testing and fail in use.

Introduced in 2.6.

See also: data leakage

test set

Data held back from training and used only at the end to estimate how the model will perform on new, unseen examples. Looking at the test set while choosing models or settings turns it into a validation set and makes the final estimate optimistic.

Introduced in 2.3.

See also: train/test split, validation set

tidy data

A layout convention for tables: each variable is a column, each observation is a row, and each type of observational unit is a table. Tidy tables are what plotting libraries and scikit-learn expect.

Introduced in 1.3.

See also: DataFrame

traceback

The report Python prints when an exception is not handled. The last line names the error and its message; the lines above trace the chain of function calls that led there, innermost last. Read the last line first, then find the line of your code nearest the bottom.

Introduced in 1.1.

train/test split

Dividing a dataset into a training set, which the model learns from, and a test set, which is kept back and used only once to estimate how the model performs on new data. The split should be random and reproducible (random_state), and the test set must never influence any modelling decision.

Introduced in 2.3.

See also: test set, training set, data leakage

training set

The subset of the data used to fit (train) a model. Performance measured on the training set is optimistic, because the model has already seen these examples.

Introduced in 2.3.

See also: train/test split, test set

U

underfitting

When a model is too simple or too constrained to capture the structure in the data, so it performs poorly on both the training and the test set.

Introduced in 2.4.

See also: overfitting

unsupervised learning

Learning patterns from inputs alone, without labels, for example clustering samples into groups or compressing many features into a few components (PCA).

Introduced in 2.1.

See also: supervised learning

V

validation curve

A plot of training and validation score against the value of one hyperparameter, such as polynomial degree or ridge \(\alpha\). The best setting is where the validation score peaks, not where the training score does.

Introduced in 2.4.

See also: learning curve, hyperparameter

validation set

Data held back from training and used to compare models or tune hyperparameters. Unlike the test set it may be consulted many times, which is why a separate, untouched test set is still needed for the final estimate.

Introduced in 2.4.

See also: test set, cross-validation

vectorisation

Expressing a computation as operations on whole arrays rather than element-by-element Python loops. NumPy executes vectorised operations in compiled code, typically tens to hundreds of times faster.

Introduced in 1.2.

See also: broadcasting

virtual environment

A self-contained directory holding a Python interpreter and its own installed packages. Each project gets its own, so upgrading a library for one project cannot break another. Created with python -m venv, uv venv or conda create.

Introduced in 0.2.

See also: conda / mamba

X

X / y convention

The layout scikit-learn and this course use everywhere: features live in a two-dimensional array X with one row per sample and one column per feature, and the target lives in a one-dimensional array y with one entry per sample. Capital X because it is a matrix, lower-case y because it is a vector.

Introduced in 1.2.

See also: feature, label, shape