Recall
Recall is the fraction of relevant, positive, or target outcomes that a system successfully identifies, retrieves, or captures.
Grokking Machine Learning
A classification metric equal to the proportion of positive-labeled data points that are correctly predicted as positive.
- How it is used
- Evaluates classification models when missed positive cases matter more than accuracy alone reveals.
- How deep it goes
- Defines recall through true positives and false negatives, calculates it from confusion matrices, and extends it to threshold trade-offs rather than model construction.
- Examples
- Uses medical and spam-detection confusion matrices, plus exercises on recommendations, diagnosis, pedestrian detection, and voice assistants.
- What you can do after
- Evaluate/understand classifiers by calculating recall and deciding when missed positives deserve emphasis.
Starting Data Analytics with Generative AI and Python
The proportion of actual positive cases correctly classified as positive, equivalent to sensitivity.
- How it is used
- Measures how many positive reviews a sentiment classifier finds when its predictions are checked against review-score labels.
- How deep it goes
- Defines recall through true positives and false negatives, then uses confusion-matrix counts and method comparisons rather than deriving statistical theory.
- Examples
- Compares keyword search and ChatGPT 4 API classifications on 500 reviews, then contrasts additional pretrained sentiment-analysis methods.
- What you can do after
- Evaluate/understand how often a positive-review classifier misses positives and weigh improvements against specificity, response-parsing limitations, and cost.
Read how Starting Data Analytics with Generative AI and Python explains recall →
Natural Language Processing in Action, Second Edition
Recall is the proportion of relevant items that a retrieval or classification system successfully identifies.
- How it is used
- Measures whether search systems and classifiers find the relevant documents or positive cases they are meant to identify.
- How deep it goes
- Defines recall through missed positives and works through classifier counts, without developing a general evaluation procedure.
- Examples
- Uses toxic-comment predictions, capitalization-sensitive queries, and a big-cat search that misses documents naming individual species.
- What you can do after
- Diagnose/debug missed positive cases with a scikit-learn confusion matrix and weigh higher recall against lower precision.
Read how Natural Language Processing in Action, Second Edition explains recall →
Machine Learning System Design
The proportion of actual positive cases that a model correctly predicts as positive.
- How it is used
- Measures how many actual positive cases a classifier catches and helps assess whether a search shortlist contains clicked photos.
- How deep it goes
- Defines recall through true positives and false negatives, then compares simulated results; stops short of deriving a training objective.
- Examples
- Uses fraud-detector simulations with unequal false alarms and a photo-search check of clicked results in the top 1,000.
- What you can do after
- Evaluate/understand it by comparing missed positives with false alarms and interpreting Recall@1000 as evidence for a possible retrieval stage.
Effective Conversational AI
Recall is the proportion of relevant or correct cases that a classifier successfully identifies.
- How it is used
- Measures missed correct intents in chatbot classification and missed relevant documents in retrieval.
- How deep it goes
- Defines classifier recall with true positives, false negatives, and a worked calculation; distinguishes document retrieval recall without a retrieval calculation.
- Examples
- Works through human-agent intent recognition and compares login-intent blind tests before and after adding representative training utterances.
- What you can do after
- Diagnose/debug with it by locating missed intents in blind tests, adding representative examples, and checking whether recall improves.
Delivering Machine Learning Projects: From design to deployment
Recall is the proportion of relevant positive instances that a predictive classification system correctly identifies.
- How it is used
- Frames recall as a proposed measure for churn and demand predictions and as one criterion in model testing and selection.
- How deep it goes
- Treats recall as a named performance measure alongside precision and F1, without a worked calculation or threshold-setting procedure.
- Examples
- Uses customer-churn and demand prediction hypotheses as concrete cases where recall and precision will be measured.
- What you can do after
- Evaluate/understand it by considering recall alongside other test results and documenting its role in model selection.
Read how Delivering Machine Learning Projects: From design to deployment explains recall →
Statistics Slam Dunk
Sensitivity is the proportion of actual positive cases that a classification procedure correctly identifies.
- How it is used
- Evaluates how effectively a classification model identifies actual positive outcomes.
- How deep it goes
- Defines the metric, gives its formula, computes it from a confusion matrix, and contrasts it with specificity.
- Examples
- Uses NBA postseason predictions with 40 true positives, 10 false negatives, and an observed sensitivity of 0.80.
- What you can do after
- Evaluate logistic-regression predictions by calculating recall from true-positive and false-negative counts.
Knowledge Graphs and LLMs in Action
Recall is the proportion of relevant positive instances that are successfully identified.
- How it is used
- Evaluates how many existing positive entities, relations, or graph nodes a model successfully identifies.
- How deep it goes
- Defines recall operationally, contrasts it with precision and F1-score, and applies it through validation plots and class averaging.
- Examples
- Uses knowledge-graph entity and relation extraction plus AML licit-versus-illicit node classification as evaluation examples.
- What you can do after
- Evaluate model coverage with per-class recall and compare it across models, epochs, and averaging strategies.
Read how Knowledge Graphs and LLMs in Action explains recall →
Deep Learning with PyTorch, Second Edition
A classification metric measuring the fraction of actual positives that are correctly identified, which can be undefined when there are no actual positives.
- How it is used
- Checks whether an imbalanced nodule classifier finds actual positives when overall percentage correct looks misleadingly high.
- How deep it goes
- Defines the ratio, illustrates threshold trade-offs, and calculates it from classification counts; does not develop a general zero-denominator handling rule.
- Examples
- Contrasts a permissive guard-dog threshold with epoch results where an all-negative nodule classifier has zero recall.
- What you can do after
- Diagnose/debug with it when high classification accuracy hides missed positives, and track it during training alongside precision.
Read how Deep Learning with PyTorch, Second Edition explains recall →
Grokking Statistics
True positive rate is the proportion of positive instances correctly classified as positive.
- How it is used
- Evaluates how well a classifier identifies relevant positive events, especially when accuracy can mislead under class imbalance.
- How deep it goes
- Defines recall through confusion-matrix counts and develops its threshold tradeoff with precision using ROC and precision-recall curves.
- Examples
- Uses wolf detection and toddler-meltdown prediction to contrast missed alarms, false alarms, and threshold choices.
- What you can do after
- Evaluate/understand how threshold changes affect missed positive cases and precision in a classification decision.
Statistics Every Programmer Needs
Sensitivity is the proportion of actual positive instances that a classifier correctly identifies as positive.
- How it is used
- Evaluates how well logistic regression identifies actual positive outcomes.
- How deep it goes
- Defines the formula, maps it to confusion-matrix counts, and demonstrates direct and scikit-learn calculations.
- Examples
- Uses raisin classification counts, a confusion matrix, Python arithmetic, and scikit-learn's recall_score().
- What you can do after
- Apply it via scikit-learn or confusion-matrix counts to evaluate binary predictions.
Read how Statistics Every Programmer Needs explains recall →
Just Use Postgres!
Recall is the proportion of relevant items that a search system retrieves from all relevant items available.
- How it is used
- Evaluates the accuracy of approximate nearest-neighbor indexes in vector similarity search.
- How deep it goes
- Defines recall, establishes exact-search behavior, and connects HNSW parameters to accuracy and resource trade-offs.
- Examples
- Uses movie-embedding searches, a full-table scan baseline, HNSW execution plans, and ef_search-adjusted result lists.
- What you can do after
- Evaluate vector-index accuracy and tune HNSW construction or search parameters to improve retrieved results.
The Art of AI Product Development
Recall is the proportion of actual positive cases that a classifier correctly identifies.
- How it is used
- Evaluates whether classifiers capture target cases and whether search systems retrieve the available relevant information.
- How deep it goes
- Defines recall, explains its precision trade-off, and applies it to classifier segments and search-document evaluation.
- Examples
- Uses user-segment classification and a search case with 15 relevant documents, of which 8 are retrieved.
- What you can do after
- Evaluate classifier and search coverage, then judge whether missed cases or false positives matter more.
Read how The Art of AI Product Development explains recall →
Generative AI in Action
Recall is the proportion of relevant or actual positive cases that a system correctly identifies.
- How it is used
- Evaluates coverage in summaries, translations, content filters, and fine-tuned model tasks.
- How deep it goes
- Defines coverage and links recall to precision and F1, but stops short of mathematical derivation and semantic quality assessment.
- Examples
- Uses reference summaries, machine translations, generated text, content filters, user feedback, expert analysis, and A/B testing.
- What you can do after
- Evaluate summaries, translations, filters, and adapted models while interpreting recall alongside precision, F1, qualitative review, and operational checks.
Financial AI in Practice
The proportion of all fraudulent cases found within a selected group of model-ranked cases, such as the top n percent.
- How it is used
- Measures detected fraud as a share of all actual fraud, including within a model-ranked review queue.
- How deep it goes
- Defines sensitivity and walks through ranked test-set thresholds and capture rates, without developing a mathematical proof.
- Examples
- Compares fraud caught, precision, and alert volumes at the top 0.1%, 0.2%, and 5% of ranked transactions.
- What you can do after
- Apply it via ranked test-set analysis to choose alert thresholds against fraud coverage and review capacity.
Data-Driven Cybersecurity
Recall is the proportion of actual instances in a class that a classification model correctly identifies.
- How it is used
- Evaluates how well cybersecurity classifiers capture actual high-risk articles or network threats.
- How deep it goes
- Interprets class-specific recall in classification reports alongside precision, F1-score, accuracy, and confusion matrices.
- Examples
- Uses cybersecurity articles, network traffic, and login activity to interpret class-specific recall.
- What you can do after
- Evaluate whether a classifier detects actual threats by reading class-specific recall alongside false negatives and other performance measures.
Ensemble Methods for Machine Learning
Recall is the fraction of actual positive examples that are correctly predicted as positive.
- How it is used
- Evaluates class-level positive detection and serves as a scoring metric for model and feature-importance analysis.
- How deep it goes
- Defines recall as a fraction and demonstrates per-class interpretation, but does not develop its calculation beyond reported scores.
- Examples
- Uses AdaBoost digit classification, per-class scores, a confusion matrix, and recall-based permutation feature importance.
- What you can do after
- Apply recall through scikit-learn classification reports and compare it with other metrics when diagnosing class-specific performance.
Read how Ensemble Methods for Machine Learning explains recall →
LLMs in Production
Recall is the proportion of reference items that are correctly identified by a prediction.
- How it is used
- Uses recall within ROUGE to evaluate generated summaries against ground-truth text without requiring exact matches.
- How deep it goes
- Defines recall as reference-item coverage and applies it to ROUGE overlap, without presenting a mathematical derivation.
- Examples
- Demonstrates ROUGE-1 and ROUGE-L on two Zelda summaries, reporting recall alongside precision and F-measure.
- What you can do after
- Apply it via rouge_score to inspect how much reference-summary content a generated summary overlaps.
Introduction to Generative AI, Second Edition
Recall is the proportion of actual positive cases that a classifier correctly identifies.
- How it is used
- Uses recall to assess whether an LLM judge identifies positive cases consistently against expert annotations.
- How deep it goes
- Defines recall and applies it to classifier evaluation, while leaving the metric’s formula and implementation details implicit.
- Examples
- Uses binary expert labels for enthusiasm and positivity in generated real estate listing descriptions.
- What you can do after
- Evaluate annotation prompts and LLM judges for reliable positive-case detection against expert labels.
Read how Introduction to Generative AI, Second Edition explains recall →
Machine Learning Algorithms in Depth
The proportion of actual positive examples correctly predicted as positive.
- How it is used
- Evaluates binary classifier performance, especially the rate at which actual positive examples are detected.
- How deep it goes
- Defines recall algebraically and relates it to thresholds, confusion matrices, precision-recall curves, and F1, without deriving implementation details.
- Examples
- Uses TP and FN counts, threshold variation, and precision-recall plots as concrete interpretive examples.
- What you can do after
- Evaluate binary classifiers by calculating recall and interpreting its tradeoff with precision across thresholds.
Read how Machine Learning Algorithms in Depth explains recall →
The AI Pocket Book
A classification or retrieval metric measuring how many relevant or positive items are successfully found.
- How it is used
- Measures information-retrieval performance by counting relevant instances successfully identified.
- How deep it goes
- Defines recall with a numerical example and explains its tradeoff with precision, stopping short of detailed implementation.
- Examples
- Uses legal-case retrieval and fraud-transaction detection to illustrate finding relevant items in large pools.
- What you can do after
- Evaluate retrieval systems by judging missed relevant items alongside irrelevant results and business priorities.
Elasticsearch in Action, Second Edition
In information retrieval, recall is the proportion of all relevant documents that are retrieved.
- How it is used
- Evaluates how completely a search retrieves the relevant documents for a user query.
- How deep it goes
- Defines the metric, works through a television-search calculation, and stops short of direct query tuning.
- Examples
- Uses a 4K television search with six retrieved and three omitted relevant documents.
- What you can do after
- Evaluate search completeness and adjust queries while balancing recall against precision.
Read how Elasticsearch in Action, Second Edition explains recall →
Time Series Forecasting Using Foundation Models
Recall is the proportion of actual positive observations that a detection system correctly identifies.
- How it is used
- Evaluates how completely anomaly detectors identify labeled anomalies.
- How deep it goes
- Defines recall, implements its confusion-matrix calculation, and demonstrates its trade-off with precision and F1 score.
- Examples
- Uses TimeGPT, Chronos, Llama-3.2, and daily New York City taxi anomaly detection examples.
- What you can do after
- Evaluate anomaly detectors by measuring missed anomalies and balancing recall against false-positive costs.
Read how Time Series Forecasting Using Foundation Models explains recall →
Fabulous Adventures in Data Structures and Algorithms
Test sensitivity is the conditional probability that a medical test is positive when the tested condition is present.
- How it is used
- Measures positive-case accuracy while supporting Bayesian interpretation of medical test results.
- How deep it goes
- Defines sensitivity numerically and contrasts it with specificity, then stops short of treating sensitivity as posterior accuracy.
- Examples
- Uses Frob Syndrome, a 99.5%-accurate test, six-million-person counts, and a C# joint distribution.
- What you can do after
- Evaluate medical test results by separating sensitivity from posterior belief and accounting for false positives.
Read how Fabulous Adventures in Data Structures and Algorithms explains recall →
Related concepts
- f-scoreused by · An F-score is a classification or evaluation metric combining precision and recall via their weighted harmonic mean; F1 is the equal-weight case.
- precisioncontrasts with · Precision is the proportion of items identified as positive, selected, or retrieved that satisfy the relevant correctness or relevance criterion.
- false-negative ratecontrasts with · The false-negative rate is the proportion of actual positive cases that are incorrectly classified or identified as negative.
- receiver operating characteristicused by · A receiver operating characteristic curve plots a binary classifier’s true-positive rate against its false-positive rate as the decision threshold varies.
- similarity searchmeasures · Similarity search is a retrieval technique that finds or ranks items by comparing their representations with a query according to a similarity or distance measure.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.