Riba:
- Ability don zaɓar da fassara rarrabuwa da ma'auni na koma baya (matrix ruɗani, daidaito / tunawa / F1, MAE / RMSE / R²) bisa ga manufar aikin
- Ikon gano koyo ta hanyar kwatanta horo da maki gwaji da samun ingantaccen aiki tare da tabbatarwa
- Ability don kimanta ko kowane samfurin yana ƙara ƙimar gaske ta hanyar kwatanta shi da tushe
Yana da sauƙi don saita samfurin; Yana da wuya a auna gaskiya ko yana aiki da gaske. Batun wannan rukunin shine ƙwarewar masanin kimiyyar bayanai mafi mahimmanci: kimanta samfurin tare da ma'auni masu dacewa, cimma ingantaccen aikin tsinkaya, da kama mafi girman tarko: koyo ( haddace). AI yana ƙididdigewa, fassara da kwatanta ma'auni; amma ya rage ga ɗan adam ya yanke shawarar wane awo ne daidai don kasuwancin ku da ko samfurin ya “isa”. Ƙungiya da ke kallon ma'auni mara kyau na iya kuskuren samfurin mara kyau na tsawon watanni a matsayin "nasara."
Me yasa daidaito bai isa ba: rikicewar matrix
Ma'auni na farko da ke zuwa a hankali a cikin rarrabuwa shine daidaito (daidaita - jimillar madaidaicin tsinkaya). Amma kamar yadda muka gani a cikin Raka'a 6, daidaito yana yaudarar bayanai marasa daidaituwa. Mafi kyawun farawa shine matrix ɗin ruɗani - tebur da ke raba tsinkaya zuwa bins huɗu:
- Gaskiya tabbatacce (TP): Mun kira karya abin da gaske karya ne. (mai kyau)
- Gaskiya mara kyau (TN): Mun ce da gaske tsabta. (mai kyau)
- Ƙarya (FP): Mun yi kuskure mun ce mai tsabta karya ne. (ƙarar ƙararrawa)
- Ƙarya mara kyau (FN): Mun rasa karya kuma mun kira shi mai tsabta. (Rashin barazana)
Ma'auni biyu masu mahimmanci sun samo asali daga waɗannan akwatuna huɗu. Precision: "Nawa ne abin da na kira karya a zahiri karya ne?" - mahimmanci idan farashin ƙararrawa na ƙarya ya yi yawa. Hankali (tunawa a cikin Ingilishi): "Nawa nawa na gaske karya ne?" - mahimmanci lokacin rasa barazanar yana da tsada. Sau da yawa su biyun suna yin rashin jituwa da juna: idan kun runtse bakin kofa kuma ku ƙara cewa “karya”, tuna yana ƙaruwa amma daidaito yana raguwa. Makin F1 shine madaidaicin matsakaici (ma'anar jituwa) na waɗannan biyun.
Hankali: Amsar tambayar "Wane awo ne mai mahimmanci" yanke shawara ne na kasuwanci, ba na fasaha ba. Rasa shari'ar (ƙaramar tunawa) a cikin gwajin cutar kansa yana da illa; Yana da ban haushi don aika imel mai mahimmanci zuwa spam (ƙananan madaidaicin) a cikin tace spam. Kudin yana ƙayyade ma'auni.
Ma'aunin koma baya
Idan fitarwar lambobi ne, ana amfani da ma'auni daban-daban. MAE (Ma'anar Cikakkun Kuskure): nawa kididdigar ta karkata daga ainihin matsakaita, a cikin raka'a ɗaya (misali "mun yi kuskure ta 4,200 TL akan matsakaita"). RMSE (Kuskuren Ma'anar Madaidaicin Tushen): yana azabtar da manyan kurakurai da tsanani kuma yana kula da masu fita waje. R² (R-squared - coefficient of determination): nawa daga cikin sauye-sauye a cikin manufa samfurin ya bayyana (kusa da 1 yana da kyau, 0 yana da mummunan kamar tsinkaya ma'anar, korau ya fi muni).
Babban tarko mai ban tsoro: overiling
Ƙarfafawa - inda samfurin ke haddace bayanan horo amma ya gaza akan sababbin bayanai - shine kuskuren da ya fi kowa a kimiyyar bayanai. Alamar ta bayyana a fili: samfurin yana da kyau a kan tsarin horo (98%), mara kyau a kan gwajin gwajin (72%). Samfurin ya haddace amo na wannan samfurin, ba ainihin tsari a cikin bayanan ba. Har ila yau, akwai akasin haka: rashin dacewa - samfurin ba shi da kyau a cikin horo da gwaji saboda yana da sauƙi don kama tsarin.
Hanyoyi don magance overlearning: sauƙaƙa samfurin, ƙarin bayanai, daidaitawa - birki na lissafi wanda ke kiyaye ƙirar daga yin rikitarwa - kuma mafi mahimmanci, tabbatar da giciye.
Ƙaddamar da ketare: kar a amince da aiki ɗaya
Kwas ɗin horo/gwaji ɗaya na iya zama sa'a ko rashin sa'a; Kuna iya samun manyan maki don saitin gwajin kawai saboda samfurin yana da sauƙi. Tabbatarwa (CV a takaice) yana magance wannan. Mafi yawan nau'i na k-ninka CV: an raba bayanai zuwa sassan k (misali 5); A kowane zagaye, bangare guda yana gwadawa sauran kuma horo ne; wannan mirgine sau 5 kuma an daidaita maki 5. Don haka za ku ga cewa wasan kwaikwayon ya dogara ne akan matsakaicin rarrabuwar kawuna, ba raba sa'a ɗaya ba. Rarraba maki kuma yana ba da labari: makiyoyi masu saurin canzawa (85% akan bene ɗaya, 62% akan ɗayan) yana nuna cewa ƙirar ba ta da ƙarfi.
Ba a yi amfani da k-ninka na al'ada ba a cikin jerin lokaci (yana zazzage gaba); Madadin haka, kuna yin “sarkar gaba” (tsaga jerin lokaci): koyaushe horo tare da abubuwan da suka gabata da gwada gaba.
awo
Nau'in matsala
Yaushe abin yake?
Daidaito
Rabewa
Idan azuzuwan sun daidaita
Daidaitawa
Rabewa
Ƙararrawar ƙarya yana da tsada
Hankali (tunawa)
Rabewa
Idan yana da tsada a rasa
F1
Rabewa
Idan ana buƙatar ma'auni
MAE
koma baya
Kuskuren fassara
RMSE
koma baya
Idan manyan kurakurai suna da mahimmanci
R²
koma baya
ikon bayani
uku mini lokuta
Case na 1 - Haƙiƙa na babban daidaito. Ɗayan samfurin zamba ya nuna daidaito 99.2% kuma ƙungiyar ta yi bikin. Duban matrix ɗin rikicewa, ainihin: ƙimar karya a cikin bayanan shine 0.8%; Samfurin ya kama kusan babu karya, kawai yana cewa "duk bayyane". Tunawa ya kasance 6%. Darasi: duba awo, ba daidaito ba.
Hali na 2 - Latent overlearing. Ƙungiya ɗaya ta isar da haɓaka XGBoost zuwa 97% akan tsarin horo. Saitin gwajin ya kasance 69%, amma babu wanda ya duba. Samfurin ya rushe a samarwa. Idan da an yi haƙƙin giciye, babban bambanci tsakanin folding (overlearning) zai kasance a bayyane daga farko. Darasi: aminta da CV da makin gwaji, ba makin ilimi ba.
Case 3 - Lucky raba. Wani manazarci ya yi farin cikin samun kashi 88% cikin rarrabuwa guda. Lokacin da abokin aikinsa ya yi CV mai ninki 5, maki sun kasance 88%, 71%, 83%, 64%, 79% - matsakaicin shine 77%, amma yana da rauni sosai. Samfurin ba shi da kwanciyar hankali; Wurin guda ɗaya ya kasance mai ɓatarwa. Darasi: Dubi ma'anar CV da rarrabawa, ba bin mutum ɗaya ba.
Samfura huɗu masu kwafi
1) Cikakken rahoton rabe-rabe:
Na horar da samfuri da saitin gwaji. Ƙirƙirar: matrix ruɗani, daidaito, tunowa, F1 (na kowane aji) da ƙidayar tallafi. Ina magana, amma ya rage nawa: Zan gaya muku wane awo ne mai mahimmanci. Lura cewa daidaito na iya zama ɓata tare da bayanan da ba daidai ba.
2) Ikon Ilmantarwa:
Buga makin ƙira na a cikin TARBIYYA da TEST saitin gefe da gefe. Yi ƙididdige bambanci tsakanin su kuma yi gargaɗi saboda babban bambanci alama ce ta koyo. Idan bambancin ya yi girma, bayar da shawarar daidaitawa ko sauƙaƙawa.
3) Tabbataccen ƙetare:
Yi tabbataccen giciye mai ninki 5 (don rarrabuwa, rarrabawa). Buga makin, ma'ana da daidaitaccen karkacewar kowane ninka. Idan maki sun kasance marasa ƙarfi (high std), nuna cewa ƙirar ba ta da ƙarfi. Yi amfani da bututun bututu domin ana iya koyan sauye-sauye daga sashin horo a kowane Layer.
4) Kwatancen tushe:
Sanya ma'auni na ƙirar gefe da gefe tare da tushen tushen DummyClassifier. Shin samfurin yana taka rawar gani sosai? Idan bai wuce ba, bayyana a sarari cewa ƙirar ba ta ƙara kowane ƙimar gaske ba.
Rauni mai ƙarfi / Ƙarfi mai ƙarfi
Rawanin faɗakarwa:
Shin samfurina yana da kyau?
"mai kyau" ba a bayyana shi ba; Ba a bayyana wanne awo ba, wane kofa, wane tushe. AI na iya ba da lamba daidai guda ɗaya kuma ya ɓata.
Ƙarfi mai ƙarfi:
Matsayinku: mataimakin kima. Rarraba, bayanan da ba su daidaita (12% tabbatacce). fifiko: tuno (ba a rasa tabbataccen abu ba). Aiki: (1) matrix ruɗani, (2) madaidaicin / tuna/F1, (3) ma'anar ma'anar CV mai ninki 5 da std, (4) DummyClassifier kwatankwacin tushe. Mayar da hankali kan tunawa da bambancin tushe, ba daidaito ba. Yi sharhi, amma na yanke shawara ta ƙarshe.
Anan, fifikon awo, CV da yanayin asali a bayyane suke.
Kuskuren gama gari
- Amincewa da daidaito a cikin bayanan da ba daidai ba. 99% daidaito bazai iya kama 'yan tsiraru ba kwata-kwata; Dubi matrix ruɗani.
- Kallon karatun ku kawai. Babban ilimi, ƙarancin jarrabawa shine overlearning; Koyaushe kwatanta maki biyu.
- Dogaro da ɗaki guda ɗaya. Rarraba sa'a bata ce; Dubi ma'ana da rarraba tare da tabbatarwa.
- Ba kwatanta da asali ba. Ba za ku iya cewa "mai kyau" ba tare da sanin ko samfurin ya doke mai tsinkayar wawa ba.
- Amfani da k-ninka na al'ada akan jerin lokaci. Yana zubar da gaba; Ana buƙatar raba jerin lokaci.
Tukwici: Rubuta lambobi uku kusa da kowane samfuri: makin horo, matsakaicin tabbataccen giciye, da makin asali. Wannan nau'in ukun yana bayyanawa a kallo overlearing (horo >> CV) da rashin ƙima (CV ≈ asali).
A takaice
Gina samfurin yana da sauƙi, amma kimanta shi da gaskiya yana da wuyar gaske. A cikin rarrabuwa, matrix ɗin ruɗani, daidaito, da tunawa sun fi ƙarin bayyani fiye da daidaito; Kudin aikin yana ƙayyade wanda yake da mahimmanci. Ana amfani da MAE, RMSE da R² a cikin koma baya. Mafi girman tarko shine koyo: koyaushe kwatanta horo da maki gwaji. Dogaro ga ma'ana da rarraba ƙetare tabbatarwa, ba guda ɗaya ba; Kwatanta komai da tushe. AI yana ƙididdige waɗannan duka, amma ya rage ga ɗan adam ya yanke shawarar wane awo zai yi amfani da shi.
Aikin aikace-aikace
Cire matrix ruɗani, daidaito, tunawa da F1 don ƙirar ƙira; Sa'an nan kuma yi 5-ninka ƙetare-validation kuma duba rarraba maki a cikin folds. A ƙarshe, rubuta ƙimar horo, matsakaicin CV, da maƙiyan tushe gefe da gefe. Yi sharhi a cikin jumla ɗaya ko samfurin ku yana koyo kuma yana wuce tushe.
jerin abubuwan dubawa
- [ ] Na kalli ainihin ma'auni na aikin (daidaici / tunawa / F1 da dai sauransu) maimakon daidaito?
- [ ] Na bincika matrix ɗin ruɗani?
- [ ] Shin na kwatanta horo da maki na gwaji kuma na duba don koyo?
- [ ] Na kalli ma'ana da rarrabawa tare da tabbatarwa?
- [ ] Na kwatanta samfurin zuwa tushen tushe?