{"licence":{"name":"CC BY-SA 4.0","spdx":"CC-BY-SA-4.0","url":"https://creativecommons.org/licenses/by-sa/4.0/","attribution":"Atlas, a bilingual technical dictionary (https://cmaintz.github.io/tech-atlas/)"},"id":"ai/precision","url":{"en":"https://cmaintz.github.io/tech-atlas/en/terms/ai/precision/","da":"https://cmaintz.github.io/tech-atlas/da/terms/ai/precision/"},"term":{"en":"Precision","da":"Præcision (precision)"},"aka":{"en":["positive predictive value"],"da":["positiv prædiktiv værdi"]},"domain":["ai"],"cluster":"evaluation","layer":"theory","status":"current","summary":{"en":"Of all the cases a model flags as positive, the share that really are; in security terms, how many alerts were real.","da":"Af alle de tilfælde, en model markerer som positive, andelen der virkelig er det - i sikkerhedssprog, hvor mange alarmer der var ægte."},"body":{"formal":{"en":"The number of true positives divided by all cases the model marked positive (true positives plus false positives), taken from a confusion matrix for one class.","da":"Antallet af sande positive delt med alle tilfælde, modellen markerede som positive (sande positive plus falske positive), hentet fra en forvekslingsmatrix for én klasse."},"plain":{"en":"Like a mushroom picker whose basket holds only good mushrooms; every one they chose was safe, even if they walked past plenty of others.","da":"Som en svampeplukker, hvis kurv kun rummer gode svampe - hver eneste, de valgte, var spiselig, selv om de gik forbi masser af andre."},"inPractice":{"en":"At a small engineering firm, the office manager who looks after IT sees the phishing filter hold back 200 emails in a month; 180 really are phishing, so its precision is 90%, and she releases the other 20 by hand.","da":"I en mindre ingeniørvirksomhed ser kontorchefen, der også passer IT, at phishingfilteret holder 200 mails tilbage på en måned; 180 er reelt phishing, så præcisionen er 90 %, og hun frigiver selv de øvrige 20."},"whyItMatters":{"en":"Low precision buries staff in wrong alerts and teaches them to ignore the tool, so it is the number to raise when the cost of a wrong flag is high.","da":"Lav præcision begraver medarbejderne i forkerte alarmer og lærer dem at ignorere værktøjet, så det er tallet, man skal hæve, når en forkert markering koster meget."}},"deepDive":{"en":"Precision, or positive predictive value, is TP / (TP + FP): conditional on the model saying \"positive\", the probability that it is right. It is undefined when the model makes no positive predictions, a case scikit-learn handles with its zero_division parameter. False negatives do not enter the formula at all, so a model that flags only its single most certain case can score 100% precision while missing nearly everything; precision is meaningless without recall or a fixed operating point beside it. In ranking and retrieval it is usually reported at a cut-off, precision@k, and averaged over recall levels as average precision (AP); object detection reports mean AP over classes, with COCO additionally averaging over intersection-over-union thresholds from 0.5 to 0.95.\n\nUnlike recall, precision depends directly on prevalence. By Bayes' rule, PPV = (sensitivity × prevalence) / (sensitivity × prevalence + (1 − specificity) × (1 − prevalence)). A detector with 99% sensitivity and 99% specificity applied where only 0.1% of events are malicious has a precision of about 9%: roughly ten false alarms for every true one. This base-rate effect is the arithmetic behind alert fatigue in security operations and false-positive screening results in medicine, and it means precision measured on a balanced or enriched test set overstates what will be seen in production unless it is reweighted to the real class mix.\n\nPrecision is controlled through the decision threshold. Raising the threshold typically raises precision, but not strictly monotonically (it can dip when a few high-scoring negatives remain), whereas recall can only stay the same or fall. The precision-recall curve traces this trade-off; under heavy class imbalance it is more informative than the ROC curve, because the false positive rate stays deceptively small when negatives vastly outnumber positives (Davis & Goadrich, 2006; Saito & Rehmsmeier, 2015). Operational requirements are therefore often phrased as \"maximum recall subject to precision of at least X\".\n\nThe word collides with other meanings. In metrology (ISO 5725) precision means the closeness of repeated measurements to each other, a property of variance rather than of correctness; in numerics it refers to floating-point formats such as FP16 or BF16 used for model weights. In retrieval-augmented generation, \"context precision\" measures how much of the retrieved material is relevant, a retrieval metric distinct from the precision of the final answer.","da":"Præcision, også kaldet positiv prædiktiv værdi, er TP / (TP + FP): givet at modellen siger \"positiv\", sandsynligheden for, at den har ret. Den er udefineret, når modellen ikke laver nogen positive forudsigelser, hvilket scikit-learn håndterer med parameteren zero_division. Falske negative indgår slet ikke i formlen, så en model, der kun markerer sit ene sikreste tilfælde, kan opnå 100 % præcision og samtidig overse næsten alt; præcision er meningsløs uden genkaldelse eller et fast arbejdspunkt ved siden af. Ved rangordning og søgning angives den som regel ved en grænse, precision@k, og som gennemsnit over genkaldelsesniveauer som average precision (AP); objektdetektion rapporterer mean AP over klasser, og COCO tager desuden gennemsnittet over intersection-over-union-tærskler fra 0,5 til 0,95.\n\nModsat genkaldelse afhænger præcision direkte af forekomsten. Efter Bayes' regel er PPV = (sensitivitet × forekomst) / (sensitivitet × forekomst + (1 − specificitet) × (1 − forekomst)). En detektor med 99 % sensitivitet og 99 % specificitet, der bruges, hvor kun 0,1 % af hændelserne er ondsindede, har en præcision på cirka 9 %: omkring ti falske alarmer for hver ægte. Denne basisrate-effekt er regnestykket bag alarmtræthed i sikkerhedsdrift og falsk positive screeningsresultater i sundhedsvæsenet, og den betyder, at præcision målt på et balanceret eller beriget testsæt overvurderer, hvad man vil se i drift, medmindre den omvægtes til den reelle klassefordeling.\n\nPræcision styres via beslutningstærsklen. En højere tærskel giver typisk højere præcision, men ikke strengt monotont - den kan falde, når nogle få højt scorede negative er tilbage - mens genkaldelse kun kan forblive uændret eller falde. Præcision-genkaldelse-kurven viser afvejningen; ved stærk klasseubalance er den mere informativ end ROC-kurven, fordi den falsk positive rate forbliver vildledende lav, når negative tilfælde langt overstiger positive (Davis & Goadrich, 2006; Saito & Rehmsmeier, 2015). Driftskrav formuleres derfor ofte som \"højest mulig genkaldelse under forudsætning af mindst X præcision\".\n\nOrdet støder sammen med andre betydninger. I metrologien (ISO 5725) betyder præcision, hvor tæt gentagne målinger ligger på hinanden, altså en egenskab ved variansen og ikke ved korrektheden; i numerik henviser det til flydende-komma-formater som FP16 eller BF16, der bruges til modelvægte. I retrieval-augmented generation måler \"context precision\", hvor stor en del af det hentede materiale der er relevant, et retrieval-mål, der er noget andet end præcisionen af det endelige svar."},"edges":[{"type":"requires","to":"ai/confusion-matrix","confidence":"high","strength":"normal"},{"type":"requires","to":"ai/classification","confidence":"high","strength":"normal"},{"type":"requires","to":"security/false-positive","confidence":"high","strength":"normal"},{"type":"part-of","to":"ai/model-evaluation","confidence":"high","strength":"normal"},{"type":"contrasts-with","to":"ai/recall","why":{"en":"Precision asks how many of the flagged cases were right; recall asks how many of the real cases were found. Raising one usually lowers the other.","da":"Præcision spørger, hvor mange af de markerede tilfælde der var rigtige; genkaldelse spørger, hvor mange af de ægte tilfælde der blev fundet. Hæver man den ene, falder den anden som regel."},"confidence":"high","strength":"primary"}],"depth":3,"sources":[{"title":"Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft (§4.9, Evaluation: Precision, Recall, F-measure)","url":"https://web.stanford.edu/~jurafsky/slp3/","tier":"textbook"},{"title":"Fawcett (2006), An introduction to ROC analysis","url":"https://doi.org/10.1016/j.patrec.2005.10.010","tier":"reference","publisher":"Pattern Recognition Letters"},{"title":"Goodfellow, Bengio & Courville, Deep Learning (ch. 11.1)","url":"https://www.deeplearningbook.org/","tier":"textbook","publisher":"MIT Press"},{"title":"scikit-learn User Guide, Metrics and scoring (classification metrics)","url":"https://scikit-learn.org/stable/modules/model_evaluation.html","tier":"official-doc","publisher":"scikit-learn"}],"draft":true}