Precision asks whether what was found is right. Recall asks what was missed.

Precision is the fraction of retrieved items that are relevant. It asks whether what came back was right. It says nothing about what did not come back.

Recall is the fraction of all relevant items in the collection that were actually retrieved. It asks what was missed. It is the harder of the two to measure, because it requires knowing what exists.

The F-measure combines precision and recall into one score, using their weighted harmonic mean. It exists because the two pull against each other. Reporting one number hides which of the two was traded away.

Accuracy is the overall proportion of correct decisions. In retrieval it is nearly useless, because almost everything in a collection is irrelevant to any query. A system that returns nothing scores well.

A precision-recall curve plots precision against recall across operating points. It shows the trade-off rather than one position on it. It is the honest form of the report.

Precision at K measures the proportion of relevant items within the top K results. It reflects what a user actually sees. For a result page, K is roughly ten and in practice closer to three.

Recall at K measures how many of all relevant items appear within the top K. It combines the question of completeness with the constraint of a short list. It is where most systems fail quietly.

The true positive rate is recall under another name, used when plotting against the false positive rate. The duplicate vocabulary comes from two fields solving the same problem. Knowing they are the same prevents a good deal of confusion.

The false positive rate is the proportion of non-relevant items wrongly retrieved. It is the cost side of precision. It matters most where a wrong result is expensive rather than merely annoying.

Inverse precision evaluates precision on the reversed problem, where the negative class is treated as positive. It exposes behaviour that the ordinary measure hides. It is rarely reported.

Inverse recall is recall computed for the reversed classification. Together with inverse precision it completes the picture. Reporting only the forward pair is a choice, not a necessity.

Prevalence is the proportion of actually relevant items in the collection. It determines what a given precision score is worth. The same precision means very different things at one percent and at fifty.

The Matthews correlation coefficient summarises performance while correcting for chance. It is robust where accuracy is misleading. It is the single number worth reporting when classes are unbalanced.

Cohen's kappa discounts the agreement that would occur by chance alone. It is used both for classifier performance and for agreement between judges. Its second use is the more important one here.

A ROC curve plots true positive rate against false positive rate across thresholds. It shows how a system behaves as its threshold moves. It is less informative than a precision-recall curve when relevant items are rare.

Informedness is a chance-corrected renormalisation of recall. It tells you how much better than guessing the system is at finding what exists. It is the recall-side counterpart to markedness.

Markedness is the chance-corrected renormalisation of precision. It tells you how much better than guessing a positive result is. Together with informedness it gives a symmetrical account.

It defines the measures and their relations. It leaves out every measure designed for realistic conditions, the entire question of where the relevance judgments come from, the fact that they are never complete, the cost of different errors, and the institutions that produced the whole framework.

R-precision measures precision at a cutoff equal to the number of relevant documents for that query. The cutoff adapts to the query instead of being fixed. Standard evaluation uses static cutoffs that suit some queries and not others.

The F-beta score weights precision and recall differently according to which matters more. Evaluation defaults to the balanced version without asking whether balance is right. For most commercial questions it is not.

Rank-biased precision models a user who continues down the list with some probability at each step. It replaces a hard cutoff with browsing behaviour. Metrics assuming a fixed cutoff assume a user who does not exist.

Bpref is designed for the case where relevance judgments are incomplete. Evaluation assumes complete binary judgments for every candidate. Complete judgments exist for no collection of any size.

PR-AUC is the area under the precision-recall curve, expressed as a single number. Analyses plot the curve without integrating it. The scalar is what allows systems to be compared.

The break-even point is where precision equals recall. It isolates one interpretable position on the curve. Reports give the whole curve and never name it.

Interpolated precision smooths the jagged steps of a raw precision curve. It makes curves comparable across queries. Raw scores are recorded without it.

Fallout is the proportion of non-relevant items in the collection that were retrieved. It looks at the corpus rather than at the result set. Metrics focused on what was returned cannot see it.

Miss rate is the proportion of relevant items that were not retrieved. It is recall stated as a failure. Stating it that way changes how it is received in a meeting.

The micro F1-score aggregates over all instances, so large categories dominate. It answers how the system does overall. Reporting an unlabelled F-measure hides which aggregation was used.

The macro F1-score averages across categories with equal weight, so small categories count as much as large ones. It answers how the system does across the range. Micro and macro can point in opposite directions.

Subtopic recall measures how many distinct information needs behind a query were covered. It treats a query as having several possible meanings. Standard metrics treat relevance as a single binary match.

Intent-aware precision weights results by the distribution of likely intents behind a query. It accepts that one query means different things to different people. Evaluation models assume a single meaning.

Balanced accuracy averages performance across the positive and negative classes. It repairs accuracy under imbalance. Plain accuracy is reported instead.

Youden's J summarises a ROC curve at a single threshold. It gives one number where a curve gives a shape. Trade-offs are discussed through curves alone.

The diagnostic odds ratio combines sensitivity and specificity into one efficiency measure, borrowed from epidemiology. Retrieval studies omit it. Medicine solved several of these measurement problems first.

The Brier score measures the accuracy of probabilistic predictions rather than thresholded decisions. It rewards being right about how confident to be. Evaluation focuses on binary outputs.

Expected calibration error measures whether stated confidence matches observed frequency. A system that says ninety percent should be right nine times in ten. Classification analysis measures matches without checking calibration.

Inter-annotator agreement measures how far human judges agree on what is relevant. Evaluations treat ground truth as absolute. Where judges disagree, the ground truth is a committee decision.

Fleiss' kappa extends chance-corrected agreement to more than two annotators. Agreement metrics emphasise pairs. Real annotation is done by teams.

Krippendorff's alpha handles missing judgments and varied measurement scales. Reliability measurement assumes fixed scales and complete data. Neither holds in practice.

Pooling judges only the documents that some system retrieved, rather than the whole collection. It is how large evaluations are made feasible. It also means anything no system found is counted as irrelevant.

Incomplete judgments are the normal condition: only a fraction of candidates is ever assessed. Standard metrics assume full manual judgment. Recall computed against an incomplete set overstates itself.

The Cranfield method is the offline evaluation paradigm of fixed queries, fixed judgments and comparable systems. It is the framework that made retrieval a science. Analysis of the metrics without it omits where they come from.

Time-biased gain accounts for how long a user spends reading before moving on. It replaces rank position with elapsed time. Metrics assume inspection is instantaneous and free.

Session recall measures coverage across a whole sequence of queries rather than one. People refine and repeat. Metrics isolate single turns.

Session precision tracks accumulated precision across an interaction. It follows the user rather than the query. Evaluations measure single response sets.

Diversification rewards a result set that covers different interpretations instead of ten versions of one. Standard precision and recall reward homogeneity. A perfectly precise page of near-identical results scores well and serves badly.

Top-K search is the algorithmic problem of retrieving the best K items efficiently. Metric analysis focuses on definitions rather than on the algorithms that produce the list. The algorithm constrains what the metric can ever see.

A confusion matrix shows the raw counts from which all these ratios are derived. Analyses present the ratios without the counts. The counts reveal what the ratios conceal.

The sensitivity index d-prime measures discriminability on a continuous scale, from signal detection theory. Evaluation relies on discrete counting. The continuous view separates ability from where the threshold was set.

Criterion C measures the bias in where a decision threshold sits. Metrics assume a fixed threshold without asking whether it leans. Two systems with identical ability can differ entirely in this.

The threat score, or critical success index, comes from forecasting and handles rare events. Domain metrics stay with standard retrieval forms. Rare-event measurement has been solved elsewhere.

The G-mean is the geometric mean of sensitivity and specificity. Harmonic means dominate evaluation. The geometric version penalises imbalance between the two more sharply.

Cost-sensitive learning assigns different costs to different errors. Metrics treat false positives and false negatives as equally bad. In almost every real application they are not.

Expected utility expresses results as economic gain and loss rather than as match counts. Metrics count matches. What a search is worth is a business quantity that no standard measure represents.

Class imbalance is the condition where relevant items are a tiny minority. Prevalence appears as a marginal note. Severe imbalance invalidates most of the intuitive measures and is the normal case in retrieval.

A gold standard set is the reference against which performance is measured. Literature focuses on relevance labels without asking how a reference is constructed and by whom. Whoever builds it defines what counts as correct.

ISO 25010 is the formal standard for software product quality characteristics. Evaluation centres on statistical retrieval quality. The formal quality framework exists and is not connected to it.

TREC is the long-running evaluation programme at NIST that produced the collections, the judgments and much of the methodology. Consensus lists the metrics without naming the institution. The institution is why the metrics are comparable at all.

Precision describes the results that were shown. Recall describes what existed and was not shown, including your page. A system with high precision and unmeasured recall can be excellent and still never surface you, and no report will say so.

Everything here measures a system against a list of what somebody decided is relevant. That list is incomplete, made by judges who disagree, and pooled from what systems already found. A subject nobody has written about is not in the list, so no measure records its absence.