Data Science Glossary: 100+ Terms in Plain Language 📖

A searchable glossary of statistics, machine learning, deep learning, data, evaluation, MLOps, NLP and computer vision terms, each explained in one or two plain sentences with an example.

Jargon is the first wall every learner hits. This glossary explains each term in one or two plain sentences plus a short example. Search it, filter by topic, or pull a random “term of the day” 🎲.

Tip🎯 How to use it

• 🔎 Type in the search box to filter as you type
• 🏷️ Pick a topic to narrow the list
• 🎲 Press “Surprise me” to learn one new term a day
• 📜 The full A–Z list is printed below the search tool, so Ctrl+F works too

Note📚 The Resources hub

🧭 Overview: where to find resources · 🛤️ Learning paths · 📝 Cheat sheets · 📖 Glossary · 🗄️ 20 open data sources · 📄 Paper notes (coming soon)

🔎 Search the glossary

import {ms, msStyle} from "/assets/js/ms-plot.js"
cats = ["All topics", ...Array.from(new Set(glossary.map(d => d.cat))).sort()]
viewof q = Inputs.text({placeholder: "Type a term, e.g. drift, recall, embedding", label: "🔎 Search"})
viewof cat = Inputs.select(cats, {label: "🏷️ Topic"})

hits = glossary.filter(d => (cat === "All topics" || d.cat === cat) && (q === "" || (d.term + " " + d.def + " " + d.ex).toLowerCase().includes(q.toLowerCase())))
htl.html`<div style="font-size:.9rem;opacity:.75;margin-bottom:8px">${hits.length} of ${glossary.length} terms</div>
<div style="display:grid;grid-template-columns:repeat(auto-fill,minmax(300px,1fr));gap:12px;max-height:620px;overflow:auto;padding-right:4px">
${hits.map(d => htl.html`<div style="border:1px solid rgba(127,127,127,.25);border-radius:14px;padding:12px 14px">
  <div style="display:flex;justify-content:space-between;gap:8px;align-items:baseline"><b style="font-size:1.02rem">${d.term}</b><span style="font-size:.75rem;opacity:.7;white-space:nowrap">${d.cat}</span></div>
  <div style="font-size:.92rem;margin-top:4px">${d.def}</div>
  <div style="font-size:.85rem;margin-top:6px;opacity:.75">💡 ${d.ex}</div></div>`)}
</div>`

🎲 Term of the day

viewof roll = Inputs.button("🎲 Surprise me")

{
  roll;
  const d = glossary[Math.floor(Math.random() * glossary.length)];
  return htl.html`<div class="ms-readout"><span>${d.cat}</span><b>${d.term}</b><span>${d.def} <br>💡 ${d.ex}</span></div>`;
}

🏷️ Terms by topic

Plot.plot({
  ...msStyle, height: 280, width: 760, marginLeft: 190,
  x: {label: "Number of terms →", grid: true}, y: {label: null},
  marks: [
    Plot.barX(glossary, Plot.groupY({x: "count"}, {y: "cat", fill: ms.indigo, rx: 4, sort: {y: "-x"}})),
    Plot.text(glossary, Plot.groupY({x: "count", text: "count"}, {y: "cat", dx: 12, fill: "currentColor"}))
  ]
})

🧠 How the big ideas connect

flowchart LR
  D["🧹 Data<br/>features, labels"] --> S["📊 Statistics<br/>sampling, inference"]
  D --> M["🤖 Machine learning<br/>supervised, unsupervised"]
  S --> M
  M --> DL["🧠 Deep learning<br/>neural networks"]
  DL --> NLP["💬 NLP & GenAI<br/>transformers, LLMs"]
  DL --> CV["👁️ Computer vision<br/>CNNs, detectors"]
  M --> E["📏 Evaluation<br/>metrics, validation"]
  E --> OPS["🚀 Deployment & MLOps<br/>serving, monitoring"]

📜 Full A–Z list

A

A/B test · 📊 Statistics
A controlled experiment that randomly splits users between two versions and compares an outcome. Example: Half of visitors see a new checkout button; compare conversion rates.

Accuracy · 📏 Evaluation
The share of predictions that are correct. Misleading when one class is rare. Example: 99% accuracy on fraud data can mean the model never flags fraud.

Activation function · 🧠 Deep learning
A non-linear function applied to a neuron’s output so networks can learn curves, not just lines. Example: ReLU returns max(0, x).

Anomaly detection · 🤖 Machine learning
Finding data points that look very different from the rest. Example: Flagging a server whose CPU pattern suddenly changes.

API · 🚀 Deployment & MLOps
A defined way for programs to talk to each other, often over HTTP. Example: POST a payment as JSON, get back a fraud score.

AUC (ROC AUC) · 📏 Evaluation
The probability that the model ranks a random positive above a random negative. 0.5 is random, 1.0 is perfect. Example: A churn model with AUC 0.82.

B

Backpropagation · 🧠 Deep learning
The algorithm that computes how much each weight contributed to the error, so gradient descent can update it. Example: Errors flow backward from the output layer to the first layer.

Bagging · 🤖 Machine learning
Training many models on random resamples of the data and averaging them to reduce variance. Example: A random forest is bagged decision trees plus random feature choice.

Batch inference · 🚀 Deployment & MLOps
Scoring many records at once on a schedule instead of one request at a time. Example: Nightly churn scores for every customer.

Bayes’ theorem · 📊 Statistics
A rule for updating the probability of a belief after seeing evidence. Example: Updating the chance of disease after a positive test.

Bias (statistical) · 📊 Statistics
A systematic difference between an estimate and the true value. Example: Surveying only app users overestimates tech comfort.

Bias-variance trade-off · 🤖 Machine learning
Simple models miss patterns (bias); complex models chase noise (variance). Good models balance both. Example: A straight line underfits; a degree-20 polynomial overfits.

Bootstrap · 📊 Statistics
Estimating uncertainty by resampling the data with replacement many times. Example: 1,000 resamples give a confidence interval for the median.

Bounding box · 👁️ Computer vision
A rectangle that marks where an object is in an image. Example: x, y, width, height around each car.

C

Canary release · 🚀 Deployment & MLOps
Sending a small share of traffic to a new version before rolling it out fully. Example: 5% of payments go to model v15 first.

Categorical variable · 🧹 Data
A variable whose values are labels, not numbers. Example: Plan type: basic, plus, premium.

Central limit theorem · 📊 Statistics
Averages of many independent samples are approximately normal, whatever the original distribution. Example: Daily average order values look bell-shaped.

Classification · 🤖 Machine learning
Predicting a category. Example: Spam or not spam.

Clustering · 🤖 Machine learning
Grouping similar items without labels. Example: k-means groups customers by behaviour.

CNN · 🧠 Deep learning
A convolutional neural network: layers of small filters that detect patterns in images. Example: Early layers find edges, later layers find faces.

Concept drift · 🚀 Deployment & MLOps
When the relationship between inputs and the target changes over time. Example: Fraudsters switch to small test payments.

Confidence interval · 📊 Statistics
A range that would contain the true value in a stated share (e.g. 95%) of repeated studies. Example: Conversion 4.1% ± 0.3%.

Confusion matrix · 📏 Evaluation
A table of true positives, false positives, true negatives and false negatives. Example: Shows which mistakes a classifier makes.

Container · 🚀 Deployment & MLOps
A packaged app with its libraries that runs the same everywhere. Example: A Docker image of a FastAPI model service.

Correlation · 📊 Statistics
A number from -1 to 1 measuring how strongly two variables move together in a straight line. Example: Height and weight: r ≈ 0.7. Correlation is not causation.

Cross-validation · 📏 Evaluation
Splitting data into k folds and training k times, each time testing on a different fold. Example: 5-fold CV gives five scores; report their mean.

D

Data drift · 🚀 Deployment & MLOps
When the distribution of input data changes compared with training data. Example: Average payment size doubles after a new market launch.

Data leakage · 🧹 Data
When information unavailable at prediction time sneaks into training, inflating scores. Example: Using ‘account closed date’ to predict churn.

Data pipeline · 🧹 Data
An automated chain of steps that moves and transforms data. Example: Extract from API, clean, load to a warehouse.

Decision tree · 🤖 Machine learning
A model that makes predictions by asking a sequence of yes/no questions. Example: Is tenure < 6 months? Is plan month-to-month?

Deep learning · 🧠 Deep learning
Machine learning with neural networks that have many layers. Example: Image recognition, speech, language models.

Dimensionality reduction · 🤖 Machine learning
Compressing many features into fewer while keeping most information. Example: PCA turns 50 survey questions into 3 factors.

Dropout · 🧠 Deep learning
Randomly switching off neurons during training to reduce overfitting. Example: Dropout rate 0.3 in a dense layer.

E

Embedding · 💬 NLP & GenAI
A list of numbers that represents a word, sentence, image or item so that similar things are close together. Example: ‘king’ and ‘queen’ have nearby embeddings.

Ensemble · 🤖 Machine learning
Combining several models to get better predictions than any single one. Example: Averaging a forest, a boosted model and a linear model.

Epoch · 🧠 Deep learning
One full pass of the training data through the model. Example: Train for 20 epochs.

ETL / ELT · 🧹 Data
Extract, transform, load: moving data from sources into a store. ELT loads first and transforms inside the warehouse. Example: Daily sales from the shop database to BigQuery.

F

F1 score · 📏 Evaluation
The harmonic mean of precision and recall. Example: Useful when both false alarms and misses matter.

Feature · 🧹 Data
An input variable used by a model. Example: Tenure, monthly charge, number of support calls.

Feature engineering · 🧹 Data
Creating new, more useful inputs from raw data. Example: Turning timestamps into ‘hour of day’ and ‘is weekend’.

Feature store · 🚀 Deployment & MLOps
A system that stores and serves features consistently for training and prediction. Example: The same ‘tx_last_hour’ value used offline and online.

Fine-tuning · 💬 NLP & GenAI
Continuing to train a pre-trained model on a smaller, task-specific dataset. Example: Adapting a general language model to support tickets.

G

Gradient boosting · 🤖 Machine learning
Building trees one after another, each fixing the errors of the previous ones. Example: XGBoost, LightGBM, scikit-learn HistGradientBoosting.

Gradient descent · 🤖 Machine learning
An optimisation method that repeatedly steps weights in the direction that reduces the loss. Example: Learning rate 0.01, 1,000 steps.

H

Hallucination · 💬 NLP & GenAI
When a generative model states something fluent but false. Example: Inventing a citation that does not exist.

Hyperparameter · 🤖 Machine learning
A setting chosen before training, not learned from data. Example: Tree depth, learning rate, number of clusters.

Hypothesis test · 📊 Statistics
A procedure for judging whether data are consistent with a default assumption (the null hypothesis). Example: Is the new page’s conversion rate really higher?

I

Imbalanced data · 🧹 Data
When one class is much rarer than the others. Example: 0.2% of transactions are fraud.

Imputation · 🧹 Data
Filling in missing values. Example: Replace missing income with the median.

Inference · 🚀 Deployment & MLOps
Using a trained model to make predictions. (In statistics, it means drawing conclusions from data.) Example: The API runs inference in 20 ms.

IoU · 👁️ Computer vision
Intersection over union: how much a predicted box overlaps the true box, from 0 to 1. Example: IoU ≥ 0.5 often counts as a correct detection.

K

k-nearest neighbours · 🤖 Machine learning
Predicting by looking at the k most similar training examples. Example: A house is priced like its 5 nearest neighbours.

L

Label · 🧹 Data
The correct answer the model learns to predict. Example: ‘churned’ = 1 or 0.

Large language model (LLM) · 💬 NLP & GenAI
A very large neural network trained on text to predict the next token, able to generate and reason over language. Example: Chat assistants and coding assistants.

Latency · 🚀 Deployment & MLOps
How long one request takes. Example: p95 latency of 60 ms.

Learning rate · 🧠 Deep learning
How big a step gradient descent takes each update. Example: Too high diverges, too low crawls.

Linear regression · 🤖 Machine learning
Predicting a number as a weighted sum of inputs. Example: Price = 50,000 + 120 × square feet.

Logistic regression · 🤖 Machine learning
A linear model passed through a sigmoid to predict a probability. Example: Probability that a customer churns.

Loss function · 🤖 Machine learning
A number that measures how wrong the model is; training tries to make it small. Example: Mean squared error, cross-entropy.

M

MAE · 📏 Evaluation
Mean absolute error: average size of errors in the target’s units. Example: Forecasts are off by 12 units on average.

mAP · 👁️ Computer vision
Mean average precision: the standard accuracy score for object detectors. Example: COCO mAP 50:95 of 54.7.

MLOps · 🚀 Deployment & MLOps
Practices that automate and monitor training, testing, deployment and retraining of ML models. Example: A pipeline retrains monthly and gates releases.

Model registry · 🚀 Deployment & MLOps
A catalogue of model versions with metrics, lineage and aliases. Example: ‘payguard@champion’ points to version 15.

Monitoring · 🚀 Deployment & MLOps
Watching a live system’s health, data and quality over time. Example: Alert when PSI > 0.2.

Multicollinearity · 📊 Statistics
When input variables are strongly correlated with each other, making coefficients unstable. Example: House size in square feet and in square metres.

N

Naive Bayes · 🤖 Machine learning
A fast probabilistic classifier that assumes features are independent given the class. Example: A classic spam filter.

Neural network · 🧠 Deep learning
Layers of simple units connected by weights that together learn complex functions. Example: Input layer → hidden layers → output.

Normal distribution · 📊 Statistics
The symmetric bell curve described by a mean and standard deviation. Example: About 95% of values lie within 2 standard deviations.

Normalisation / scaling · 🧹 Data
Rescaling features to a common range or spread. Example: StandardScaler gives mean 0, standard deviation 1.

Null hypothesis · 📊 Statistics
The default assumption a test tries to find evidence against. Example: ‘The new design has no effect.’

O

One-hot encoding · 🧹 Data
Turning a category into several 0/1 columns. Example: Plan → is_basic, is_plus, is_premium.

Outlier · 🧹 Data
A value far from most others. Example: A $50,000 order in a shop where most are $40.

Overfitting · 🤖 Machine learning
When a model memorises training data and performs poorly on new data. Example: Training accuracy 99%, test accuracy 70%.

P

p-value · 📊 Statistics
The probability of seeing results at least this extreme if the null hypothesis were true. Example: p = 0.03 is unusual under ‘no effect’. It is not the chance the null is true.

PCA · 🤖 Machine learning
Principal component analysis: finds new axes that capture the most variance. Example: Plot 100 features in 2D.

Precision · 📏 Evaluation
Of the items flagged positive, the share that really are positive. Example: Of 100 flagged payments, 80 were fraud: precision 0.8.

Prompt · 💬 NLP & GenAI
The input text or instructions given to a generative model. Example: ‘Summarise this ticket in one sentence.’

PSI · 🚀 Deployment & MLOps
Population stability index: a score of how much a feature’s distribution has shifted. Example: Above 0.2 usually needs attention.

R

RAG · 💬 NLP & GenAI
Retrieval-augmented generation: fetch relevant documents, then let an LLM answer using them. Example: A help bot that cites your own docs.

Random forest · 🤖 Machine learning
Many decision trees trained on random samples and features, averaged together. Example: A strong default for tabular data.

Recall · 📏 Evaluation
Of all real positives, the share the model found. Example: Caught 90 of 100 frauds: recall 0.9.

Regression · 🤖 Machine learning
Predicting a number. Example: Next week’s sales.

Regularisation · 🤖 Machine learning
Adding a penalty for complexity to reduce overfitting. Example: L2 (ridge) shrinks coefficients; L1 (lasso) can zero them.

Reinforcement learning · 🤖 Machine learning
Learning by trial and error to maximise a reward. Example: A robot learns to walk; a model learns from human feedback.

RMSE · 📏 Evaluation
Root mean squared error: like MAE but punishes large errors more. Example: RMSE of 15 units.

S

Sampling · 📊 Statistics
Selecting a subset of a population to study. Example: Survey 1,000 of 1 million customers at random.

Segmentation (image) · 👁️ Computer vision
Labelling every pixel with a class or object. Example: Masks for road, car, pedestrian.

Sigmoid · 🧠 Deep learning
An S-shaped function that squeezes any number into 0–1. Example: Turns a score into a probability.

Standard deviation · 📊 Statistics
The typical distance of values from the mean. Example: Delivery times: mean 3 days, SD 1 day.

Stratified split · 📏 Evaluation
A train/test split that keeps class proportions the same in each part. Example: Both splits keep 15% churners.

Supervised learning · 🤖 Machine learning
Learning from examples that have correct answers (labels). Example: Past emails labelled spam / not spam.

T

Test set · 📏 Evaluation
Data held back until the end to estimate real-world performance. Example: Never tune on it.

Time series · 🧹 Data
Data points recorded in time order. Example: Daily website visits.

Token · 💬 NLP & GenAI
A chunk of text (word or word piece) that a language model reads and writes. Example: ‘unbelievable’ → ‘un’, ‘believ’, ‘able’.

Train/validation/test split · 📏 Evaluation
Separate data for learning, tuning and final evaluation. Example: 70% / 15% / 15%.

Transfer learning · 🧠 Deep learning
Reusing a model trained on one task as the starting point for another. Example: ImageNet CNN fine-tuned to spot defects.

Transformer · 💬 NLP & GenAI
A neural architecture built on attention, the basis of modern language and vision models. Example: BERT, GPT, ViT, DETR.

Type I / Type II error · 📊 Statistics
Type I: a false alarm (rejecting a true null). Type II: a miss (failing to reject a false null). Example: Approving a useless drug vs missing a useful one.

U

Underfitting · 🤖 Machine learning
When a model is too simple to capture the pattern. Example: A straight line through a curved relationship.

Unsupervised learning · 🤖 Machine learning
Finding structure in data without labels. Example: Clustering, dimensionality reduction.

V

Validation set · 📏 Evaluation
Data used to choose hyperparameters and compare models. Example: Pick the tree depth with the best validation AUC.

W

Weight · 🧠 Deep learning
A learned number that scales an input’s influence on a neuron or model output. Example: In y = 3x + 2, the weight is 3.

Y

YOLO · 👁️ Computer vision
‘You Only Look Once’: a family of fast single-pass object detectors. Example: Real-time detection on a phone or camera.

Note📝 Missing a term?

Use the contact page to suggest one. The glossary grows with every new article.