Glossary and Quick Reference

This glossary is meant to be used while reading, not memorized in one sitting.

Term	Meaning in everyday language
R	Programming language that executes the instructions
RStudio	Editor and work environment for R
Quarto	Document/report/book tool combining prose, code, and output
Code	Written instructions for a computer
Console	Where R executes a command and displays output
Script	Saved text file containing R commands
Object	A named value or collection stored in R
Assignment (`<-`)	Save a value under an object name
Function	Named instructions, often accepting inputs and returning output
Argument	Input supplied inside function parentheses
Vector	A sequence of values of compatible type
Data frame / tibble	Table of rows and columns
Row / observation	One unit described by the columns
Column / variable	One measured characteristic
Factor	Category with defined levels
`NA`	Missing or unavailable value, NOT automatically zero
CSV	Comma-separated text file representing a table
Pipe (\|>)	Feed one operation’s result into the next
Predictor / feature	Input supplied to a model
Target / label / outcome	Correct answer the supervised model tries to learn
Unlabeled	Ground-truth answer not established
Split	Partition observations into distinct modeling roles
Train	Data on which an algorithm learns
Validation / CV	Development data used to compare choices
Test / holdout	Data reserved for final independent evaluation
Stratification	Split while approximately preserving class ratios
Patient-level split	Keep one patient’s encounters in only one split
Leakage	Letting information unavailable at prediction time or from assessment data influence fitting
Imputation	Filling in missing predictors by an explicit rule
Recipe	A set of preprocessing operations learned from the training data
Workflow	Bundle of a preprocessing recipe and model
Regression	Predict a continuous value
Classification	Predict a category or its probability
Logistic regression	Classification model estimating binary-outcome probabilities
Decision tree	Model that uses learned branching questions
Random forest	Ensemble of multiple randomized decision trees
Baseline	Simple comparator that a complex model should aim to improve
Hyperparameter	Model setting chosen by analyst or tuning algorithm
K-fold CV	Repeated train/assess process partitioning training data into k folds
Out-of-fold prediction	Prediction on a training row by a model not fitted on that row
TP / FP / FN / TN	True positive / false positive / false negative / true negative
Sensitivity / recall	TP / (TP + FN)
Specificity	TN / (TN + FP)
Precision / PPV	TP / (TP + FP)
Accuracy	(TP + TN) / total
ROC AUC	Threshold-independent ranking summary across sensitivity and false-positive rates
PR AUC	Precision-recall ranking summary
Calibration	Agreement between predicted probabilities and actual frequencies
Generalization	How well the model works on new relevant observations
Overfitting	Learning quirks of training/development data instead of patterns that hold up
Adaptive overfitting	Improving and choosing models repeatedly using feedback from the same development data
Confounding	Outside variable influences association between predictor and outcome
Selection bias	Records included are not representative of target population
PU learning	Learning from confirmed positives plus unlabeled examples, not assumed negatives
PCA	Principal component analysis, a way to summarize variation
K-means	Clustering algorithm grouping similar observations
PHI	Protected health information, requiring approved privacy/security handling

Quick reference: frequent R syntax

x <- 10                       # Save value
c(1, 2, NA)                  # Vector
is.na(x)                     # Check whether a value is missing
mean(c(1, NA), na.rm = TRUE)  # Mean of observed entries
df$age                       # Column named age
df[1, ]                      # First row
df[, "age"]                  # Age column
x == 10                      # Compare equality
x >= 10                      # Greater than or equal
library(readr)               # Load installed package
read_csv("data/file.csv")    # Read a CSV
set.seed(500)                # Make random sampling reproducible
help("mean")                 # Open help
sessionInfo()                # Record R and package details

Where to go next

After the required capstone, revisit the data governance appendix. Then continue with Part III, which starts at supervised regression, or jump to unsupervised learning in Part IV.

Extra ML methods vocabulary (Chapters 19–31)

Term	Plain-language meaning
Paradigm versus task	Paradigm specifies what feedback is available; task specifies the type of output
Multiclass	Exactly one of three or more categories per observation
Multilabel	Several labels can all be true simultaneously
Ordinal outcome	Categories with a meaningful ordering
Hyperparameter	Model setting chosen through the development/tuning process
Penalty / regularization	Discourages overly complex fitted model coefficients
Ridge / LASSO / elastic net	Common L2, L1, or mixed regression penalties
kNN	Makes predictions from nearby labeled training examples
LDA / QDA	Discriminant classifiers with shared or separate class covariances
Naive Bayes	Probabilistic classifier using a conditional independence approximation
SVM	Margin-based classification/regression, optionally using kernels
Boosting	Sequentially combines simpler models, often trees
GAM	Generalized additive model, using smooth functions of predictors
Cluster centroid	Average location of the members of a cluster
Medoid	Real observed point representing a cluster
Density-based clustering	Groups locally dense regions and may mark noise
Silhouette width	Measures how well a point matches its group compared with other groups
Gaussian mixture	Model that uses several probability components with soft membership
SVD	Matrix factorization connected with PCA
NMF	Nonnegative matrix factorization into additive components
t-SNE / UMAP	Nonlinear embeddings mainly used for visualization
Association rule	Pattern such as A → B describing co-occurrence, not causation
Support	Fraction of all transactions that contain an item set
Confidence	Fraction containing B among transactions containing A
Lift	Confidence(A→B) divided by prevalence(B)
Isolation forest / LOF	Methods for finding unusual patterns
Right censoring	Event time is unknown after the end of observation
Kaplan–Meier / Cox	Survival-curve estimator / proportional-hazards model
Rolling-origin validation	Fit on past periods, test on a later window, then advance
Semi-supervised	Learn from a combination of labeled and unlabeled examples
Weak supervision	Learn from noisy/incomplete heuristic labels
Self-supervised	Use information from input data to create a training signal
Transfer learning	Adapt knowledge learned in a different context
Reinforcement learning	Learn a policy for actions based on rewards
Adaptive overfitting	Repeated feedback-driven changes can fit the same development data
FWER	Probability of making any false rejection across a specified test family

A glossary definition is not an endorsement of a method for real patients. Read the corresponding chapters and governance appendix before using these ideas in a research proposal.