Hanyar Tabbatar da Hadin Kai a Kididdiga
A fannin kididdiga da kimiyyar bayanai, ɗaya daga cikin manyan ƙalubalen shine tabbatar da cewa samfurin ba wai kawai yana aiki da kyau akan bayanan da aka horar da shi a kai ba, har ma yana aiki da kyau akan sabbin bayanai, waɗanda ba a gani a baya. Wannan matsalar galibi ana kiranta da janareta. Nan ne ake samun ingantaccen tsari: hanyar kimanta samfurin da aka tsara don auna aikin samfurin da adalci da daidaito fiye da kimantawa ɗaya ta amfani da saitin bayanai guda ɗaya.
Me yasa ake buƙatar Tabbatar da Haɗin gwiwa?
Idan muka gina samfurin hasashen - misali, samfurin komawa baya don annabta farashin gidaje ko samfurin rarrabuwa don gano spam - yawanci muna raba bayanan zuwa sassa biyu: saitin horo da saitin gwaji. Ana horar da samfurin akan bayanan horo sannan a kimanta shi akan bayanan gwaji. Wannan hanyar tana da sauƙi, amma tana da matsala: sakamakon kimantawa na iya dogara sosai akan yadda aka raba bayanan. Idan bayanan gwajin suka zama "mai sauƙi," aiki yana bayyana sama; idan bayanan gwajin suka zama "mai wahala," aiki yana bayyana ƙasa.
Tabbatar da bayanai tsakanin mutane yana rage dogaro da saitin bayanai guda ɗaya ta hanyar yin horo da gwaje-gwaje da yawa akan saitin bayanai daban-daban sannan a daidaita sakamakon. Wannan yana haifar da kimanta aiki wanda ya fi wakiltar yanayin duniya na gaske.
Ka'idoji na Asali na Tabbatar da Giciye
Ma'anar tantancewa tsakanin bayanai shine raba bayanai zuwa sassa da dama (naɗewa). A kowane naɗewa, ana amfani da wasu naɗewa don horar da samfurin, kuma ana amfani da naɗewa ɗaya don gwada samfurin. Ana maimaita wannan tsari har sai an yi amfani da kowane naɗewa azaman bayanan gwaji. Sannan ana haɗa sakamakon kimantawa daga kowane naɗewa (yawanci tare da matsakaicin wani lokacin kuma daidaitaccen karkacewa) don samar da taƙaitaccen bayani game da aikin samfurin.
Misali, a cikin tabbatar da giciye na k-fold tare da k=5, an raba bayanan zuwa ninki 5. Na farko: ninka 1 a matsayin gwaji, ninka 2-5 a matsayin horo. Na biyu: ninka 2 a matsayin gwaji, da sauransu har zuwa ninki 5.
Nau'ikan Tabbatar da Giciye da Aka Fi Sani
1. Tabbatar da Holdout (Raba-Gwajin Jirgin Ƙasa)
Duk da cewa a zahiri ba a "maimaita" tantancewa ta hanyar giciye ba, hanyar riƙewa galibi ana ɗaukarta a matsayin matakin tabbatarwa na asali. Ana raba bayanan sau ɗaya, misali, horo 80% da gwaji 20%. Fa'idar ita ce yana da sauri kuma mai sauƙi, amma rashin amfani shine babban bambancin sakamakon saboda yana dogara ne akan rabawa ɗaya.
Yawanci ana amfani da wannan hanyar ne idan bayanai suka yi yawa, don haka ko da sashe ɗaya ya isa ya wakilci.
2. Tabbatar da Giciye na K-Fold
Wannan ita ce hanyar da ta fi shahara wajen tantancewa. Sau da yawa ana zaɓar ma'aunin k a matsayin 5 ko 10 saboda ana ɗaukarsa a matsayin daidaita farashin lissafi da ingancin kimantawa.
Kelebihan:
- Amfani da bayanai yadda ya kamata (kowane bayanai ya zama wani ɓangare na horo da gwaji).
– Kimantawar aiki ta fi kwanciyar hankali fiye da riƙewa.
Kekurangan:
– Yana ɗaukar lokaci mai tsawo domin yana horar da lokutan samfurin k.
– Idan bayanai sun yi yawa sosai ko kuma samfurin yana da rikitarwa sosai, farashin lissafi na iya zama mai yawa.
3. Tabbatar da Tsarin Giciye na K-Fold
Ga matsalolin rarrabuwa, musamman idan azuzuwan ba su da daidaito (misali, kashi 90% mara kyau, kashi 10% mai kyau), k-fold na yau da kullun na iya samar da naɗewa tare da rarraba aji mara kyau. K-fold ɗin da aka daidaita yana tabbatar da cewa rabon azuzuwan a kowane naɗewa yayi daidai da rabon azuzuwan a cikin bayanan asali.
Wannan yana da matuƙar muhimmanci musamman wajen tantance samfuran gano cututtuka, zamba, ko wasu lokuta inda ƙananan rukuni ba su da yawa.
4. Tabbatar da Giciye na Bar-One-Out (LOOCV)
A cikin LOOCV, adadin ninki yana daidai da adadin bayanai (k = n). Wannan yana nufin cewa a cikin kowane maimaitawa, lura ɗaya kawai ya zama bayanan gwaji, yayin da sauran ya zama bayanan horo.
Kelebihan:
- Kusan dukkan bayanai ana amfani da su don horo a kowane maimaitawa, don haka nuna bambanci na ƙiyasin na iya zama ƙarami.
Kekurangan:
– Yana da tsada sosai ga manyan bayanai.
– Bambancin ƙiyasi na iya zama babba a wasu nau'ikan matsaloli saboda saitin gwajin maki ɗaya ne kawai a kowane maimaitawa.
Sau da yawa ana amfani da LOOCV idan babu bayanai sosai, misali bincike tare da ƙaramin girman samfuri.
5. Tabbatar da Giciyen K-Fold Mai Maimaitawa
Wannan hanyar tana maimaita k-fold sau da yawa tare da ayyuka daban-daban (bazuwar) nadaukar nauyin. Manufar ita ce rage dogaro da aikin nadaukar nauyin guda ɗaya da kuma samar da kimantawa mafi daidaito.
Misali, "maimaita sau 10 sau 3" yana nufin yin gudu sau 10 sau 3 (jimillar horo da kimantawa 30).
6. Tabbatar da Tsarin Lokaci
Don bayanai game da jerin lokaci, tantancewa ta al'ada ba ta dace ba domin tana iya "ba da damar yin amfani da makomar" cikin tsarin horo. A cikin jerin lokaci, dole ne a kiyaye tsarin lokaci. Saboda haka, hanyoyin kamar:
– Tagar birgima/zamiya: horo a farkon lokacin sannan a gwada a cikin lokaci na gaba, sannan taga ta canza.
- Faɗaɗa taga: bayanan horo suna ƙaruwa akan lokaci, sannan a gwada su a cikin lokaci na gaba.
Wannan hanyar ta dace da hasashen tallace-tallace na wata-wata, farashin hannun jari, ko na'urori masu auna lokaci-lokaci.
Ma'aunin Kimantawa a Tabbatarwa Tsakanin
Tabbatar da daidaito tsarin kimantawa ne kawai; ma'aunin da aka yi amfani da shi ya dogara da nau'in matsalar:
– Komawa baya: MSE, RMSE, MAE, R-squared.
– Rarrabawa: daidaito, daidaito, tunawa, maki F1, ROC-AUC.
– Rarrabawa mara daidaito: ROC-AUC, PR-AUC (tunawa daidai), daidaito mai daidaito.
Sakamakon tantancewa tsakanin lambobi yawanci ana bayar da rahoton su a matsayin matsakaicin ma'auni da karkacewar misali (misali, daidaito 0,89 ± 0,03). karkacewar misali tana taimakawa wajen fahimtar daidaiton samfurin.
Tabbatar da Tsarin Zaɓe da Daidaita Sigogi
Ɗaya daga cikin manyan amfani da tabbatar da daidaito tsakanin samfuran shine zaɓin samfuri da daidaita hyperparameter. Misali:
– Zaɓar k a cikin k-NN.
– Zaɓi mafi girman zurfin da ke cikin bishiyar yanke shawara.
– Ƙayyade sigogin daidaitawa a cikin ridge/lasso regression.
- Ƙayyade C da gamma a cikin SVM.
A aikace mai kyau, ana yin tsarin daidaitawa akan bayanan horo ta amfani da tantancewa tsakanin bayanai, yayin da ake ajiye bayanan gwaji na ƙarshe daban don kimantawa ta ƙarshe. Wannan yana hana "fiye da fata" saboda an sanya samfurin fiye da kima ga bayanan kimantawa.
Wata hanya mafi tsauri ita ce ake kira nested cross-validation, wadda ita ce cross-validation a cikin cross-validation: madauki na waje don kimantawa ne, madauki na ciki don daidaitawa ne. Wannan ya shahara a cikin bincike saboda yana ba da ƙarin kimantawa na aiki mara son kai.
Fa'idodi da Iyakancewar Tabbatar da Giciye
Babban fa'idodi:
1. Yana samar da kimantawa mai ƙarfi fiye da rabo ɗaya.
2. Yi amfani da bayanai yadda ya kamata, musamman idan bayanan sun yi ƙanƙanta.
3. Yana taimakawa wajen zaɓar samfurin da ya fi dacewa da kowa kuma yana rage haɗarin yin amfani da shi fiye da kima.
Iyakoki:
1. Kudaden lissafi suna ƙaruwa yayin da ake maimaita horo sau da yawa.
2. Har yanzu ana iya samun ɓullar bayanai idan ba a yi aikin da ya dace ba kafin a fara aiki.
3. Ga bayanan da aka haɗa a rukuni (misali, bayanan marasa lafiya waɗanda ke da bayanai da yawa), ana buƙatar wata hanya ta musamman, kamar rukuni na k-fold, don kada mutum ɗaya ya bayyana a cikin jirgin kuma ya gwada a lokaci guda.
Kyawawan Ayyuka Wajen Amfani da Tabbatar da Daidaito
Domin kimantawa ta zama mai inganci, dole ne a bi wasu muhimman ƙa'idodi:
– A yi aikin sarrafawa kafin lokaci (daidaitawa, ƙididdigewa, zaɓin fasali) a cikin kowane ninka, ba sau ɗaya ba ga dukkan bayanan. In ba haka ba, bayanai daga ninka gwajin na iya zubewa cikin ninka jirgin.
- Yi amfani da nau'in k-fold mai tsari don rarrabawa tare da azuzuwan da ba su da daidaito.
– Yi amfani da tsari na musamman don bayanan jerin lokaci don kada a karya umarnin.
– Ajiye saitin gwajin ƙarshe a gefe idan burin ku shine tantance aikin ƙarshe na samfurin kafin a fara aiki.
Penutup
Tabbatar da daidaiton bayanai muhimmin kayan aiki ne a cikin kididdigar da aka yi amfani da su da kuma koyon injin don kimanta aikin samfuri cikin adalci da ƙarfi. Ta hanyar amfani da raba bayanai akai-akai, tantancewa tsakanin bayanai yana taimakawa rage son zuciya da zaɓin raba gwajin jirgin ƙasa ke haifarwa, yana gano dacewa fiye da kima, kuma yana tallafawa zaɓin samfuri da daidaita hyperparameter. Duk da cewa farashin lissafi ya fi girma, fa'idodin galibi suna da daraja, musamman lokacin da bayanai suka yi ƙanƙanta ko kuma lokacin da yanke shawara bisa ga sakamakon samfurin ke da sakamako mai mahimmanci. Ta hanyar zaɓar nau'in tantancewa tsakanin bayanai da aiwatar da mafi kyawun ayyuka, za mu iya gina samfuran da suka fi aminci waɗanda suka shirya don amfani da su akan bayanan duniya na gaske.
