Ekufundeni komshini, akukuliqiniso njalo ukuthi ukunemba okuphezulu kuyinjongo yokugcina, ikakhulukazi lapho usebenzelana namasethi wedatha ambaled.
Isibonelo, ake kube nokuhlolwa kwezokwelapha, okuyi-95% enembile ekuboneni iziguli ezinempilo kepha zehluleke ukukhomba amacala amaningi wesifo sangempela. Ukunemba kwayo okuphezulu, nokho, kufihla ubuthakathaka obukhulu. Kulapha lapho amaphuzu we-F1 afakazela usizo.
Kungakho i-F1 Score inikeza ukubaluleka ngokulinganayo ekuqondeni (amaphesenti ezinto ezikhethiwe ezifanele) futhi akhumbule (iphesenti lezinto ezikhethiwe) ukwenza amamodeli asebenze ngokushesha ngisho nakwindaba yokukhetha idatha.
Yini amaphuzu we-F1 ekufundeni komshini?
I-F1 Score yisilinganiso sokusebenza esidumile esisetshenziswa kaningi ekufundeni komshini futhi silinganisa umkhondo wokunemba futhi sikhumbule ndawonye. Kuzuzisa imisebenzi yokuhlukaniswa ngedatha engalingani ngoba ukunemba kungadukisa.
Isikolo se-F1 sinikeza isilinganiso esifanele sokusebenza kwemodeli, engavumi amanga amanga noma amanga amanga, njengoba esebenza ngokunemba futhi akhumbula; Kokubili ama-positives anqatshelwe ngokungalungile kanye neminyene eyamukelwe ngokungalungile.
Ukuqonda izisekelo: ukunemba, ukunemba, nokukhumbula
1. Ukunemba
Incazelo: Ukunemba kukala ukunemba okuphelele kwemodeli ngokubala isilinganiso sokubonwa okucatshangelwe kahle (zombili izinto eziyiqiniso kanye nezinto ezimbi zeqiniso) kwinani lokubonwayo.
Ifomula:
Ukuqonda nqo = (TP + TN) / (TP + TN + FP + FN)
- I-TP: ama-positives ayiqiniso
- I-TN: Amandla weqiniso
- I-FP: ama-positive angamanga
- FN: Amanga amanga
Lapho ukunemba kuwusizo:
- Kuhle lapho i-dataset inezinkinga ezilinganiselwe futhi zamanga kanye nezingathobi zinemiphumela efanayo.
- Okuvamile ezinkingeni zokuhlukaniswa kwenhloso ejwayelekile lapho idatha isatshalaliswa ngokulinganayo phakathi kwamakilasi.
Ukulinganiselwa:
- Kungadukisa ama-datasets ama-mulalced.
Isibonelo: Ku-Dataset lapho ama-95% amasampula angesigaba esisodwa, abikezela wonke amasampula njengoba leso sigaba sinikeza ukunemba okungu-95%, kepha imodeli ayifundanga lutho oluzayo. - Ayihlukanise phakathi kwezinhlobo zamaphutha (ama-positive angamanga vs amanga amanga).
2. Ukunemba
Incazelo: Ukunemba kuyingxenye yokuqagelwa okuhle okubikezelwe kahle kumaphoyisenti ajwayelekile abikezelwe. Kusitshela ukuthi mangaki amacala amahle abikezelwe ayenempilo.
Ifomula:
Ukwenza kahle = TP / (TP + FP)
Ukuchazwa okunembile:
Kuzo zonke izimo ukuthi imodeli ihlukaniswa njengento enhle, ingakanani okuhle ngempela? Ukucaciswa okuphezulu kusho ukuthi zimbalwa ama-positives angamanga.
Lapho kufundisa kahle izindaba:
- Lapho izindleko zecebo elingamanga ziphezulu.
- Izibonelo:
- Ukutholwa kogaxekile nge-imeyili: Asifuni ama-imeyili abalulekile (okungewona ogaxekile) abhalwe njengogaxekile.
- Ukutholwa Kokukhwabanisa: Gwema ukuhlaba umkhosi ukuthengiselana okuningi okusemthethweni.
3. Khumbula (ukuzwela noma isilinganiso esifanele)
Incazelo: Khumbula ukuthi ingxenye yamacala angempela aboniswe yimodeli kahle.
Ifomula:
Khumbula = Tp / (tp + fn)
Ukuchazwa okunembile:
Kuzo zonke izimo ezinhle zangempela, imodeli ayithola ngempumelelo? Ukukhunjulwa okuphezulu kusho ukuthi bambalwa ababi.
Lapho kukhumbula ukugxeka:
- Lapho icala eliqondile linemiphumela emibi.
- Izibonelo:
- Ukuxilongwa kwezokwelapha: Ukuphuthelwa yisifo (i-fappscy analyticslse negagali) kungabulala.
- Izinhlelo zokuphepha: Ukwehluleka ukuthola umuntu ongenayo noma usongo.
Ukunemba nokukhumbula ukuhlinzeka ngokuqonda okujulile kokusebenza kwemodeli, ikakhulukazi lapho ukunemba kukodwa akwanele. Ukuhweba kwabo kuvame ukuphathwa kusetshenziswa amamaki e-F1, esizohlola ngokulandelayo.
I-Concesion Matrix: Isisekelo samamethrikhi
A Ukudideka Matrix Ithuluzi eliyisisekelo ekufundeni komshini elibona ngephepha ukusebenza kwemodeli yokuhlukaniswa ngokuqhathanisa amalebula abikezelwe ngokumelene namalebula wangempela. Ihlukanisa izibikezelo zibe yimiphumela emine ehlukile.
| Wabikezela okuhle | Wabikezela kabi | |
| Okulungile | Iqiniso (TP) | Okubi okungekuhle (FN) |
| Okulungile | Amanga (FP) | Okubi Ngeqiniso (TN) |
Ukuqonda izingxenye
- Iqiniso (TP): Wabikezela kahle izimo ezinhle.
- Okubi Ngeqiniso (TN): Wabikezela kahle izimo ezingezinhle.
- Amanga (FP): Kubikezelwe ngokungalungile njengokwenza okuhle lapho kungekuhle.
- Okubi okungekuhle (FN): Kubikezelwe ngokungalungile njenge-negative uma okuhle.
Lezi zingxenye zibalulekile ekubaleni ama-metric ahlukahlukene asebenza kahle:
Ukubala amamethrikhi asemqoka
- Ukuqonda nqo: Kukala ukunemba okuphelele kwemodeli.
Uhlobo: Ukunemba = (TP + TN) / (TP + TN + FP + FN) - Ukwenza kahle: Kubonisa ukunemba kokubikezela okuyithemba.
Uhlobo: Ukucacisa = TP / (TP + FP) - Khumbula (ukuzwela): Izilinganiso amandla wemodeli yokubona zonke izimo ezinhle.
Uhlobo: Khumbula = tp / (tp + fn) - I-F1 Score: I-Harmonic isho ukunemba nokukhumbula, ukulinganisa lokhu okubili.
Uhlobo: I-F1 Score = 2 * (Ukucacisa * Ukukhumbula) / (Ukuqonda + Ukukhumbula)
Lawa ma-metric ababaliwe we-matrix yokudideka anika amandla ukusebenza kwamamodeli ahlukahlukene wokuhlukaniswa ukuze ahlolwe futhi alungiswe maqondana nenhloso esesandleni.
I-F1 Score: I-harmoninic kusho ukunemba nokukhumbula
Ukuchazwa kanye nefomula:
Isikolo se-F1 yi-F1 Scorey F1 yokunemba nokukhumbula. Inika inani elilodwa le-Model (noma elibi) imodeli le selokhu libheka zombili izinzuzo zamanga nezingahambi kahle.
Kungani kusho ukuthini ukuthi i-harmonic:
I-harmonic kusho ukuthi isetshenziswa esikhundleni se-arithmetic asho ukuthi inani elilinganisiwe linikeza isisindo esiphakeme kuya kokuncane okubili (ukunemba noma ukukhumbula). Lokhu kuqinisekisa ukuthi uma omunye wabo ephansi, amaphuzu e-F1 azothinteka kakhulu, agcizelele ukubaluleka okulinganayo kwezindlela ezimbili.
Uhla lwe-F1 Score:
- 0 kuya ku-1: Amanani we-F1 amabanga avela ku-0 (okubi kakhulu) kuya ku-1 (okuhle).
- 1: Ukunemba okuhle nokukhumbula.
- 0: Noma yikuphi ukunemba noma ukukhumbula kungu-0, okukhombisa ukusebenza kabi.
Ukubalwa kwesibonelo:
Unikezwe i-matrix yokudideka nge:
- I-TP = 50, Fp = 10, Fn = 5
- Ukwenza kahle = 5050 + 10 = 0.833 Frac {50} {50 + 10} = 0.83350 + 1050 = 0.833
- Khumbula = 5050 + 5 = 0.909 Frac {50} {50 + 5} = 0.90950 + 550 = 0.909
Ngakho-ke, lapho ubala amaphuzu we-F1 ngokusho kwefomula engenhla, isikolo se-F1 sizoba ngu-0.869. Kusezingeni elifanele ngoba linokulinganisela okuhle kakhulu phakathi kokunemba nokukhumbula.
Ukuqhathanisa amamethrikhi: ukusebenzisa i-F1 amaphuzu ngokunemba
Ungasebenzisa nini i-F1 Score?
- Imininingwane emangalisa:
Kufanelekile ukusebenzisa amaphuzu we-F1 lapho amakilasi ebumbana kudathabhethi (ukutholwa kwenkohliso, ukuxilongwa kwezifo). Ezimweni ezinjalo, ukunemba kuyinkomba, njengemodeli engahle ibe nokunemba okuphezulu ngenxa yokuhlukanisa iningi lemininingwane yeklasi eliningi kungaba nokunemba kwemininingwane yeklasi elincane.
- Ukunciphisa zombili inani lezinhloso zangempela kanye nezinto ezimbi zeqiniso
I-F1 Score ilungele kakhulu lapho ubungozi obunamandla bemiphelo yamanga, ebizwa nangokuthi amaphutha we-Type i, namaphutha amanga, eyaziwa nangokuthi amaphutha e-Type II, kuyabiza. Isibonelo, ukuthi amacala angemuhle amanga noma amanga acishe afanelwe ngokulinganayo ekuhlolweni kwezokwelapha noma ekutholeni ogaxekile.
Kanjani u-F1 amaphuzu wokulinganisa futhi ukhumbula:
I-F1 Score is 'ilungelo' lokulinganisa, ukuhlanganisa ukunemba (Mangaki ala macala akhonjwe kahle) futhi akhumbule (Mangaki abikezelwa ngokunembile njengezimo ezinhle).
Lokhu kungenxa yokuthi lapho enye yezilinganiso iphansi, amaphuzu we-F1 anciphisa leli nani, ngakho-ke imodeli igcina isilinganiso esihle.
Lokhu kunjalo ikakhulukazi kulezo zinkinga lapho kungasebenzi kahle khona ukuba nokusebenza okungajulile kuzo zombili izinhloso, futhi lokhu kungabonakala emikhakheni eminingi edingekayo.
Sebenzisa amacala lapho kukhethwa khona amaphuzu we-F1:
1. Ukuxilongwa kwezokwelapha
Ngokuthile okufana nomdlavuza, sifuna isivivinyo okungenzeka ukuthi siphuthelwe yisiguli somdlavuza kodwa ngeke sikwazi ukuveza kahle umuntu ophilile njengomuntu omuhle. Ngokwezinga elithile, amaphuzu we-F1 asiza ukugcina zombili izinhlobo zamaphutha lapho zisetshenziswa.
2. Ukutholwa Kokukhwabanisa
Ngokucutshungulwa kwezezimali, amamodeli wokutholwa kwenkohliso kumele athole noma akhombe ukuthengiselana okukhohlisayo (ukukhumbula okuphezulu) ngenkathi ekhomba kanyekanye futhi abhale inani lezeqisi lokuthengiselana kwangempela njengenkohliso (ukunemba okuphezulu). Isikolo se-F1 siqinisekisa le esele.
Ukunemba kunini ukunemba kwanele?
- Imininingwane enokulinganisela
Ngokuqondile, lapho amakilasi esethi yedatha enokulinganisela, ukunemba imvamisa kuyizinga elifanele ukukala ukusebenza kwemodeli njengoba kulindeleke ukuthi kube nemodeli enhle ukuba ikhiphe izibikezelo ezinengqondo zazo zombili lezi zifundo.
- Umthelela ophansi we-positives yamanga / engabi nalutho
Amazinga aphezulu we-positives yamanga nalabo abanganaki kungenzeka abe yinkinga enkulu kwezinye izimo, okwenza ukunemba isilinganiso esihle semodeli.
Ukubekezela
I-F1 Score kufanele isetshenziswe lapho imininingwane ingalingani, ukutholwa kwamanga nokuthola okungamanga nokuthola okungokoqobo kubalulekile, futhi ezindaweni ezinobungozi obukhulu njengokuxilongwa kwezokwelapha, ukutholwa kwenkohliso, njll.
Sebenzisa ukunemba lapho amakilasi enokulinganisela, futhi amanga amanga kanye namaphutha akuyona inkinga enkulu enomphumela wokuhlola.
Njengoba amaphuzu we-F1 ebheka ukunemba nokukhumbula, kungaba lula kwimisebenzi lapho izindleko zamaphutha zingabaluleka khona.
Ukutolika amaphuzu we-F1 ekusebenzeni
Yini eyakha amaphuzu “amahle” F1?
Amanani we-F1 Score ahlukahluka ngokwesimo kanye nesigaba kuhlelo lokusebenza oluthile.
- Isikolo esiphakeme se-F1 (0.8-1.0): Isho izimo ezinhle zemodeli ephathelene nokunemba nokukhumbula inani lemodeli.
- Isikolo esilinganiselayo se-F1 (0.6-0.8): Ngokuqiniseka nokuncoma kahle ukusebenza okungcono, kepha kunikeza izincomo ezibonisa isikhala esanele esidinga ukumbozwa.
- Isikolo esiphansi se-F1 (<0.6): Isiginali ebuthakathaka ekhombisa ukuthi kuningi okufanele kuthuthukiswe kumodeli.
Kwesinye isikhathi, njengokuthola ukuxilonga noma ukuphendula amacala okukhwabanisa, ngisho nesikolo se-F1 metric singaba phezulu kakhulu noma okulinganiselayo, futhi izikolo eziphakeme zincaniswa.
Usebenzisa i-F1 Score yokukhethwa kwemodeli kanye nokuhleleka
Isikolo se-F1 sisebenza ku:
- Ukuqhathanisa amamodeli: Inikeza isilinganiso esilinganayo nesifanele sokuhlola, ikakhulukazi uma kuqhathaniswa nezimo zokungalingani kweklasi.
- I-Hyperparameter Tuning: Lokhu kungenziwa ngokushintsha amanani azenzakalelayo epharamitha eyodwa ukukhulisa isilinganiso se-F1 semodeli.
- Ukulungiswa komkhawulo: Imikhawulo eguquguqukayo yezinqumo ezihlukile ze-CPU ingasetshenziswa ukulawula ukunemba nosayizi wolwazi olufanele olusethiwe futhi ngenxa yalokho, kwandise amaphuzu we-F1.
Isibonelo, singasebenzisa ukuqinisekiswa kokuqinisekiswa ukuze sisebenze kahle ama-hyperparameters ukuthola amaphuzu aphezulu we-F1, noma usebenzise amasu wokusesha angahleliwe noma wegridi.
Ama-macro, ama-micro, futhi anesisindo ama-F1 izikolo zezinkinga ze-multi-Class
Ekuhlukanisweni kwamakilasi amaningi, izindlela zokulinganisa zisetshenziselwa ukufaka amaphuzu we-F1 kuwo wonke amakilasi amaningi:
- Isikolo seMacro F1: Kuqala kukala isikolo se-F1 ekilasini ngalinye bese kuthatha isilinganiso sezikolo. Njengoba kubhubhisa wonke amakilasi kungakhathalekile ukuthi kwenzeka kaningi kangakanani, lokhu kubaphatha ngokulinganayo.
- Isikolo se-Micro F1: Kuhlanganisa imiphumela etholwe kuwo wonke amakilasi ukuthola amaphuzu aphakathi we-F1. Lokhu ngokuqinisekile kubeka amakilasi ajwayelekile esikalini esiphakeme kunamanye amakilasi abakhona abakhona abaphansi.
- Isisindo se-F1 Score: Isilinganiso sesikolo se-F1 seklasi ngalinye sibalwa kusetshenziswa ifomula F1 = 2 (Ukucacisa X Ukukhumbula) / (Ukuqonda kwe-X Khumbula) ekilasini ngalinye, ngesisindo esingeziwe se-positives ethile yangempela. Lokhu kubhema ukungalingani kweklasingo ngokunikeza izinsimbi ezengeziwe kumakilasi anabantu abaningi ku-dataset.
Ukukhethwa kwendlela yokulinganisa kususelwa kumazinga wesicelo esithile nohlobo lwedatha esetshenzisiwe.
Ukugcina
Le khasi I-F1 Score Ingabe i-metric ebalulekile ekufundeni komshini, ikakhulukazi lapho usebenzelana nemininingwane emibi noma lapho ama-positives amanga aphathe imiphumela ebalulekile. Amandla awo okulinganisa ukunemba nokukhumbuza kwenza ukuthi kubaluleke kakhulu ekuhlolweni kwezokwelapha kanye nokutholwa kwenkohliso.
Uhlelo lwe-Mit IDS Science Science kanye noHlelo Lokufunda lomshini lunikeza ukuqeqeshwa okuphelele kochwepheshe ukujulisa ukuqonda kwabo ama-metric kanye nezicelo zabo.
Le nkambo ye-inthanethi eyi-12-isonto, yathuthukiswa yi-Mit Faculty, ihlanganisa izihloko ezibalulekile kufaka phakathi ukuhlaziya okubikezelwa, ukuhlolwa kwamamodeli, kanye nezifundo zecala langempela, kuhlobanisa abahlanganyeli ngamakhono okwenza amakhono, izinqumo eziqhutshwa idatha.
