Indlela yokuqinisekisa ehlanganisiwe kuzibalo

Indlela Yokuqinisekisa Ehlanganisiwe ku-Izibalo

Kuzibalo kanye nesayensi yedatha, enye yezinselelo ezinkulu ukuqinisekisa ukuthi imodeli ayigcini nje ngokwenza kahle kudatha eqeqeshwe kuyo, kodwa futhi isebenza kahle kudatha entsha, engabonwanga ngaphambili. Le nkinga ivame ukubizwa ngokuthi ukujwayelekile. Yilapho ukuqinisekiswa okuphambene kungena khona: indlela yokuhlola imodeli eyenzelwe ukukala ukusebenza kwemodeli ngokulinganayo nangokungaguquguquki kunokuhlola okukodwa kusetshenziswa isethi yedatha eyodwa.

Kungani Kudingeka Ukuqinisekiswa Okuphambene?

Uma sakha imodeli yokubikezela—isibonelo, imodeli yokubuyela emuva ukubikezela amanani endlu noma imodeli yokuhlukanisa ukuze sithole ugaxekile—sivame ukuhlukanisa idatha ibe izingxenye ezimbili: isethi yokuqeqesha kanye nesethi yokuhlola. Imodeli iqeqeshwe kudatha yokuqeqesha bese ihlolwa kudatha yokuhlola. Le ndlela ilula, kodwa inobuthakathaka: imiphumela yokuhlola ingancika kakhulu endleleni idatha ehlukaniswa ngayo. Uma idatha yokuhlola “ilula,” ukusebenza kubonakala kuphezulu; uma idatha yokuhlola “inzima,” ukusebenza kubonakala kuphansi.

Ukuqinisekiswa okuphambene kunciphisa ukuthembela kusethi yedatha eyodwa ngokwenza izinqubo eziningi zokuqeqesha nokuhlola kumasethi edatha ahlukene bese kubalwa isilinganiso semiphumela. Lokhu kuphumela ekulinganisweni kokusebenza okumelela kakhulu izimo zangempela.

Imiqondo Eyisisekelo Yokuqinisekiswa Kwesiphambano

Ingqikithi yokuqinisekiswa okuphambene ukuhlukanisa idatha ibe izingxenye eziningana (ama-fold). Ekuphindaphindeni ngakunye, amanye ama-fold asetshenziselwa ukuqeqesha imodeli, futhi i-fold eyodwa isetshenziselwa ukuhlola imodeli. Le nqubo iphindaphindwa kuze kube yilapho yonke i-fold isisetshenziswe njengedatha yokuhlola. Amaphuzu okuhlola avela ku-fold ngayinye abe esehlanganiswa (ngokuvamile ne-mean futhi ngezinye izikhathi kanye nokuphambuka okujwayelekile) ukuze kunikezwe umbono jikelele wokusebenza kwemodeli.

Isibonelo, ekuqinisekisweni kwe-k-fold cross nge-k=5, idatha ihlukaniswe ngama-fold ama-5. I-iteration yokuqala: phinda u-1 njengokuhlolwa, phinda u-2–5 njengokuqeqeshwa. I-iteration yesibili: phinda u-2 njengokuhlolwa, njalo njalo kuze kufike ku-5.

FUNDA  Izibalo ocwaningweni lwekhwalithi

Izinhlobo Ezivamile Zokuqinisekiswa Kwesiphambano

1. Ukuqinisekiswa Kokubamba (Ukuhlukaniswa Kokuhlolwa Kwesitimela)
Nakuba ngobuchwepheshe kungeyona "indlela ephindaphindwayo" yokuqinisekisa, indlela yokubamba ivame ukubhekwa njengesinyathelo sokuqinisekisa esiyisisekelo. Idatha ihlukaniswa kanye, isibonelo, ukuqeqeshwa okungu-80% kanye nokuhlolwa okungu-20%. Inzuzo ukuthi iyashesha futhi ilula, kodwa ububi buwukuhlukahluka okukhulu kwemiphumela ngoba incike ekuhlukanisweni okukodwa.

Le ndlela ivame ukusetshenziswa uma idatha inkulu kakhulu, kangangokuthi ngisho nokwahlukana okukodwa kumelele ngokwanele.

2. Ukuqinisekiswa kwe-K-Fold Cross
Lolu uhlobo oludumile kakhulu lokuqinisekiswa okuphambene. Ipharamitha ka-k ivame ukukhethwa njengo-5 noma u-10 ngoba ibhekwa njengokulinganisela izindleko zokubala kanye nekhwalithi yokulinganisa.

Izinzuzo:
- Ukusebenzisa idatha ngendlela ephumelela kakhulu (idatha ngayinye iba yingxenye yesitimela nokuhlolwa).
– Izilinganiso zokusebenza zizinzile kakhulu kunokungabambeki.

Ukushoda:
– Kuthatha isikhathi eside ngoba kuqeqesha izikhathi zemodeli k.
– Uma idatha inkulu kakhulu noma imodeli iyinkimbinkimbi kakhulu, izindleko zokubala zingaba phezulu.

3. Ukuqinisekiswa kwe-K-Fold Cross okune-Stratified
Ngezinkinga zokuhlela, ikakhulukazi uma amakilasi engalingani (isb., 90% negative, 10% positive), i-k-fold ejwayelekile ingakhiqiza ama-fold ane-skew class distributions. I-k-fold ehlanganisiwe iqinisekisa ukuthi isilinganiso samakilasi ku-fold ngayinye sicishe sifane nesilinganiso samakilasi kudatha yokuqala.

Lokhu kubaluleke kakhulu ekuhloleni amamodeli okuthola izifo, ukukhwabanisa, noma ezinye izimo lapho isigaba sabantu abambalwa sincane.

4. Ukuqinisekiswa Kokunqamula Kokushiya-One-Out (LOOCV)
Ku-LOOCV, inani lama-fold lilingana nenani ledatha (k = n). Lokhu kusho ukuthi ku-iteration ngayinye, ukubonwa okukodwa kuphela okuba idatha yokuhlola, kanti okunye kuba idatha yokuqeqesha.

Izinzuzo:
– Cishe yonke idatha isetshenziselwa ukuqeqeshwa ekuphindaphindweni ngakunye, ngakho-ke ukubandlulula kokulinganisa kungaba kuncane.

Ukushoda:
– Kubiza kakhulu ngokwezibalo kumasethi edatha amakhulu.
– Ukwehluka kwesilinganiso kungaba phezulu kwezinye izinhlobo zezinkinga ngoba isethi yokuhlola iyiphuzu elilodwa kuphela ngokuphindaphinda.

I-LOOCV ivame ukusetshenziswa uma kunedatha encane kakhulu, isibonelo ucwaningo olunesampula encane.

FUNDA  Izibalo zesayensi yezemvelo

5. Ukuqinisekiswa Kwesiphambano Esiphindaphindwayo se-K-Fold
Le ndlela iphinda i-k-fold izikhathi eziningi ngezabelo ezahlukene (ezingahleliwe) zokugoqa. Umgomo ukunciphisa ukuncika esabelweni sokugoqa esisodwa nokukhiqiza izilinganiso ezizinzile kakhudlwana.

Isibonelo, “ukuphindaphinda ka-10 izikhathi ezi-3” kusho ukuqhuba izikhathi ezi-3 ezi-10 (ukuqeqeshwa nokuhlolwa okungu-30).

6. Ukuqinisekiswa Kochungechunge Lwesikhathi Esinqamulelayo
Kudatha yochungechunge lwesikhathi, ukuqinisekiswa okuvamile okuphambeneyo akufanele ngoba "kungavuza ikusasa" enkambisweni yokuqeqesha. Ochungechungeni lwesikhathi, ukuhleleka kwesikhathi kumele kugcinwe. Ngakho-ke, izindlela ezinjengalezi:
– Ifasitela eligingqikayo/elishelelayo: qeqesha esikhathini sokuqala bese uvivinya esikhathini esilandelayo, bese ifasitela liyashintsha.
– Iwindi lokwandisa: idatha yokuqeqesha iyanda ngokuhamba kwesikhathi, bese ihlolwa esikhathini esilandelayo.

Le ndlela ifanelekela ukubikezela ukuthengisa kwanyanga zonke, amanani esitoko, noma izinzwa zesikhathi sangempela.

Amamethrikhi Okuhlola Ekuqinisekisweni Okuphambene

Ukuqinisekiswa okuphambene kumane kuwuhlaka lokuhlola; izilinganiso ezisetshenziswayo zincike ohlotsheni lwenkinga:
– Ukuhlehla: MSE, RMSE, MAE, R-squared.
– Ukuhlukaniswa: ukunemba, ukunemba, ukukhumbula, amaphuzu e-F1, i-ROC-AUC.
– Ukuhlukaniswa okungalingani: ROC-AUC, PR-AUC (ukukhumbula ngokunemba), ukunemba okulinganiselayo.

Imiphumela yokuqinisekiswa okuphambene ngokuvamile ibikwa njengesilinganiso se-metric kanye nokuphambuka okujwayelekile (isb., ukunemba 0,89 ± 0,03). Ukuphambuka okujwayelekile kusiza ukuqonda ukuzinza kwemodeli.

Ukuqinisekiswa Okuphambene Kokukhethwa Kwemodeli Nokulungiswa Kwepharamitha

Enye yezindlela eziyinhloko zokusebenzisa ukuqinisekiswa okuphambene ukukhetha imodeli kanye nokulungisa i-hyperparameter. Isibonelo:
– Ukukhetha u-k ku-k-NN.
– Khetha ukujula okuphezulu kakhulu esihlahleni sesinqumo.
– Nquma amapharamitha okuqondisa kabusha ku-ridge/lasso regression.
- Thola i-C kanye ne-gamma ku-SVM.

Ngomkhuba omuhle, inqubo yokulungisa yenziwa kudatha yokuqeqesha kusetshenziswa ukuqinisekiswa okuphambene, kuyilapho idatha yokuhlola yokugcina igcinwa yodwa ukuze kuhlolwe kokugcina. Lokhu kuvimbela "ukulindela ngokweqile" ngenxa yokuthi imodeli ifakwe ngokweqile kudatha yokuhlola.

Indlela eqinile ibizwa ngokuthi i-nested cross-validation, okuyi-cross-validation ngaphakathi kwe-cross-validation: i-loop yangaphandle iwukuhlola, i-loop yangaphakathi iwukulungisa. Lokhu kudumile ocwaningweni ngoba kunikeza izilinganiso zokusebenza ezingachemile.

FUNDA  Ukubaluleka kwezibalo kwisayensi

Izinzuzo kanye nemikhawulo yokuqinisekiswa kwesiphambano

Izinzuzo eziyinhloko:
1. Inikeza izilinganiso zokusebenza ezizinzile kakhulu kunesigaba esisodwa.
2. Sebenzisa idatha kahle, ikakhulukazi uma isethi yedatha incane.
3. Kusiza ukukhetha imodeli evamile futhi kunciphisa ingozi yokufaka ngokweqile.

Imikhawulo:
1. Izindleko zokubala ziyanda njengoba ukuqeqeshwa kuphindaphindwa kaningi.
2. Ukuvuza kwedatha kusengenzeka uma ukucubungula kusengaphambili kungenziwanga kahle.
3. Ngemininingwane eqoqwe ngamaqoqo (isibonelo, idatha yesiguli enamarekhodi amaningana), kudingeka indlela ekhethekile, njenge-group k-fold, ukuze umuntu oyedwa angaveli esitimeleni futhi avivinye ngesikhathi esifanayo.

Imikhuba Emihle Ekusebenziseni Ukuqinisekiswa Okuphambene

Ukuze ukuhlolwa kube semthethweni, kumele kulandelwe izimiso eziningana ezibalulekile:
– Yenza ukucubungula kwangaphambili (ukulungisa, ukufaka, ukukhetha izici) ngaphakathi kokugoqwa ngakunye, hhayi kanye kuphela kwedatha yonke. Ngaphandle kwalokho, ulwazi oluvela ekugoqweni kokuhlola lungavuza lungene ekugoqweni kwesitimela.
– Sebenzisa i-k-fold ehlukaniswe ngezigaba ukuze uhlukanise ngezigaba ezingalingani.
– Sebenzisa uhlelo olukhethekile lwedatha yochungechunge lwesikhathi ukuze i-oda lingaphulwa.
– Beka eceleni isethi yokugcina yokuhlola uma umgomo wakho uwukuhlola ukusebenza kokugcina kwemodeli ngaphambi kokusetshenziswa.

I-Penutup

Ukuqinisekiswa okuphambene kuyithuluzi eliyisisekelo kuzibalo ezisetshenzisiwe kanye nokufunda komshini kokuhlola ukusebenza kwemodeli ngokulingana nangokuqinile. Ngokusebenzisa ukwabelana ngedatha okuphindaphindiwe, ukuqinisekiswa okuphambene kusiza ukunciphisa ukubandlulula okubangelwa ukukhethwa kokuhlukaniswa kwesitimela, kuthola ukulingana ngokweqile, futhi kusekela ukukhethwa kwemodeli kanye nokulungiswa kwe-hyperparameter. Ngenkathi izindleko zokubala ziphakeme, izinzuzo zivame ukuba zifanele, ikakhulukazi lapho isethi yedatha incane noma lapho izinqumo ezisekelwe emiphumeleni yemodeli zinemiphumela ebalulekile. Ngokukhetha uhlobo olufanele lokuqinisekiswa okuphambene nokusebenzisa imikhuba emihle kakhulu, singakha amamodeli athembekile alungele ukusetshenziswa kudatha yangempela.

Shiya amazwana