Njira yotsimikizira ziwerengero

Njira Yotsimikizira Mtanda mu Ziwerengero

Mu ziwerengero ndi sayansi ya deta, vuto lalikulu ndikuwonetsetsa kuti chitsanzo sichimangogwira bwino pa deta yomwe idaphunzitsidwa, komanso chimagwira bwino pa deta yatsopano, yomwe sinali yodziwika kale. Vutoli nthawi zambiri limatchedwa kufalikira. Apa ndi pomwe kutsimikizika kwa cross-validation kumabwera: njira yowunikira chitsanzo yopangidwira kuyeza magwiridwe antchito a chitsanzo moyenera komanso mosasintha kuposa kuwunika kamodzi pogwiritsa ntchito deta imodzi.

Nchifukwa chiyani Kutsimikizika kwa Mtanda Kukufunika?

Tikapanga chitsanzo cholosera—mwachitsanzo, chitsanzo chowongolera mitengo ya nyumba kapena chitsanzo chogawa magulu kuti tipeze sipamu—nthawi zambiri timagawa deta m'magawo awiri: gulu lophunzitsira ndi gulu loyesera. Chitsanzocho chimaphunzitsidwa pa deta yophunzitsira kenako n’kuyesedwa pa deta yoyesera. Njirayi ndi yosavuta, koma ili ndi vuto: zotsatira zowunikira zitha kudalira kwambiri momwe deta imagawidwira. Ngati deta yoyesera ndi "yosavuta," magwiridwe antchito amawoneka okwera; ngati deta yoyesera ndi "yovuta," magwiridwe antchito amawoneka otsika.

Kutsimikizira deta m'njira zosiyanasiyana kumachepetsa kudalira deta imodzi mwa kuchita maphunziro ambiri ndi kuyesa deta zosiyanasiyana kenako n’kuwerengera zotsatira zake. Izi zimapangitsa kuti pakhale kuyerekezera magwiridwe antchito komwe kumayimira kwambiri momwe zinthu zilili padziko lapansi.

Malingaliro Oyambira a Kutsimikizira Mtanda

Chofunika kwambiri pakutsimikizira deta ndikugawa deta m'magawo angapo (ma folds). Pa nthawi iliyonse yobwerezabwereza, ma folds ena amagwiritsidwa ntchito kuphunzitsa chitsanzocho, ndipo fold imodzi imagwiritsidwa ntchito kuyesa chitsanzocho. Njirayi imabwerezedwa mpaka fold iliyonse itagwiritsidwa ntchito ngati deta yoyesera. Zigoli zowunikira kuchokera pa nthawi iliyonse zimaphatikizidwa (nthawi zambiri ndi mean ndipo nthawi zina ngakhale standard deviation) kuti zipereke chithunzithunzi cha momwe chitsanzocho chikuyendera.

Mwachitsanzo, mu k-fold cross validation yokhala ndi k=5, deta imagawidwa m'ma fold 5. Kubwereza koyamba: pindani 1 ngati mayeso, pindani 2–5 ngati maphunziro. Kubwereza kwachiwiri: pindani 2 ngati mayeso, ndi zina zotero mpaka pindani 5.

WERENGANI  Ziwerengero mu kafukufuku wa qualitative

Mitundu Yodziwika ya Kutsimikizika kwa Mtanda

1. Kutsimikizira Kuti Munthu Ali ndi Chilema (Kugawanika kwa Mayeso a Sitima)
Ngakhale kuti njira yotsimikizira deta si "yobwerezabwereza", nthawi zambiri imaonedwa ngati njira yoyambira yotsimikizira deta. Deta imagawidwa kamodzi, mwachitsanzo, maphunziro a 80% ndi mayeso a 20%. Ubwino wake ndi wakuti ndi yachangu komanso yosavuta, koma vuto lake ndi kusiyana kwakukulu kwa zotsatira chifukwa imadalira kugawa kamodzi.

Njira imeneyi nthawi zambiri imagwiritsidwa ntchito pamene deta ili yaikulu kwambiri, kotero kuti ngakhale gawo limodzi limakhala lokwanira.

2. Kutsimikizira kwa K-Fold Cross
Iyi ndi njira yotchuka kwambiri yotsimikizira cross-validation. K parameter nthawi zambiri imasankhidwa ngati 5 kapena 10 chifukwa imaonedwa kuti imagwirizanitsa mtengo wa makompyuta ndi khalidwe loyerekeza.

Zochulukira:
- Kugwiritsa ntchito deta moyenera (deta iliyonse imakhala gawo la maphunziro ndi mayeso).
- Ziwerengero za magwiridwe antchito ndizokhazikika kuposa zosakhalitsa.

Kukonzekera:
- Zimatenga nthawi yayitali chifukwa zimaphunzitsa chitsanzo cha k nthawi.
- Ngati deta ndi yayikulu kwambiri kapena chitsanzocho ndi chovuta kwambiri, ndalama zowerengera zitha kukhala zokwera.

3. Kutsimikizika kwa K-Fold Cross Yokhazikika
Pa mavuto okhudza magulu, makamaka ngati maguluwo ali ndi kusalinganika (monga 90% negative, 10% positive), k-fold yokhazikika imatha kupanga ma fold okhala ndi magawano a kalasi osokonekera. K-fold yolumikizidwa imatsimikizira kuti gawo la makalasi mu fold iliyonse ndi lofanana ndi gawo la makalasi mu deta yoyambirira.

Izi ndizofunikira kwambiri poyesa njira zopezera matenda, chinyengo, kapena milandu ina pomwe gulu laling'ono ndi laling'ono.

4. Kutsimikizira Kusiya Chimodzi-Chimodzi (LOOCV)
Mu LOOCV, chiwerengero cha ma fold ndi chofanana ndi kuchuluka kwa deta (k = n). Izi zikutanthauza kuti mu nthawi iliyonse, kuwona kamodzi kokha kumakhala deta yoyesera, pomwe kotsalako kumakhala deta yophunzitsira.

Zochulukira:
- Pafupifupi deta yonse imagwiritsidwa ntchito pophunzitsa nthawi iliyonse, kotero kuyerekezera kungakhale kochepa.

Kukonzekera:
- Ndi okwera mtengo kwambiri pa ma data akuluakulu.
- Kusiyana kwa kuyerekezera kungakhale kwakukulu pamavuto ena chifukwa mayeso amakhala ndi mfundo imodzi yokha pa nthawi iliyonse.

LOOCV imagwiritsidwa ntchito nthawi zambiri pamene pali deta yochepa kwambiri, mwachitsanzo kafukufuku wokhala ndi chitsanzo chochepa.

WERENGANI  Ziwerengero mu sayansi ya zachilengedwe

5. Kutsimikizira Kobwerezabwereza kwa K-Fold Cross
Njirayi imabwereza k-fold kangapo ndi ma assignments osiyanasiyana (osasinthika). Cholinga chake ndi kuchepetsa kudalira pa assignment imodzi yofold ndikupereka ziwerengero zokhazikika.

Mwachitsanzo, "kubwerezabwereza katatu ka 10" kumatanthauza kuchita maulendo 10 katatu (maphunziro ndi kuwunika kokwanira 30).

6. Kutsimikizira kwa Nthawi Yotsatizana
Pa deta ya nthawi, kutsimikizira kwachizolowezi sikoyenera chifukwa kumatha "kulowetsa tsogolo" mu njira yophunzitsira. Mu nthawi, dongosolo la nthawi liyenera kusungidwa. Chifukwa chake, njira monga:
- Zenera lozungulira/lotsetsereka: phunzitsani mu nthawi yoyamba kenako yesani mu nthawi yotsatira, kenako zenera limasuntha.
- Kukulitsa nthawi: deta yophunzitsira imawonjezeka pakapita nthawi, kenako imayesedwa munthawi yotsatira.

Njirayi ndi yofunikira pa kuneneratu malonda pamwezi, mitengo ya masheya, kapena masensa a nthawi yeniyeni.

Mayeso Owunikira mu Kutsimikizira Kosiyanasiyana

Kutsimikizira kosiyanasiyana ndi njira yowunikira yokha; miyezo yomwe imagwiritsidwa ntchito imadalira mtundu wa vuto:
- Kubwerera m'mbuyo: MSE, RMSE, MAE, R-squared.
- Kugawa: kulondola, kulondola, kukumbukira, F1-score, ROC-AUC.
- Kusalinganika kwa magulu: ROC-AUC, PR-AUC (kukumbukira molondola), kulondola kolinganizika.

Zotsatira zotsimikizirana nthawi zambiri zimanenedwa ngati mean metric ndi standard deviation (monga kulondola 0,89 ± 0,03). Standard deviation imathandiza kumvetsetsa kukhazikika kwa chitsanzocho.

Kutsimikizira kwa Cross for Model Selection ndi Parameter Tuning

Chimodzi mwa zinthu zazikulu zomwe zimagwiritsidwa ntchito potsimikizira zinthu zosiyanasiyana ndi kusankha chitsanzo ndi kusintha kwa hyperparameter. Mwachitsanzo:
– Kusankha k mu k-NN.
- Sankhani kuzama kwakukulu mu mtengo wosankha.
- Dziwani magawo a regularization mu ridge/lasso regression.
- Dziwani C ndi gamma mu SVM.

Mwa njira yabwino, njira yosinthira deta imachitidwa pa deta yophunzitsira pogwiritsa ntchito njira yotsimikizira, pomwe deta yomaliza yoyesera imasungidwa padera kuti iwunikenso komaliza. Izi zimaletsa "chiyembekezo chopitirira muyeso" chifukwa cha chitsanzocho kuyikidwa mopitirira muyeso pa deta yowunikira.

Njira yokhwima kwambiri imatchedwa nested cross-validation, yomwe ndi cross-validation mkati mwa cross-validation: loop yakunja ndi yowunikira, loop yamkati ndi yokonza. Izi ndizodziwika bwino mu kafukufuku chifukwa zimapereka ziwerengero zosakondera za magwiridwe antchito.

WERENGANI  Kufunika kwa ziwerengero mu sayansi

Ubwino ndi Zofooka za Cross Validation

Ubwino waukulu:
1. Imapereka ziyerekezo zokhazikika za magwiridwe antchito kuposa kugawa kamodzi.
2. Gwiritsani ntchito deta moyenera, makamaka pamene detayo ndi yaying'ono.
3. Zimathandiza kusankha chitsanzo chofala kwambiri komanso kuchepetsa chiopsezo chowonjezera.

Keterbatasan:
1. Ndalama zowerengera zimawonjezeka pamene maphunziro akubwerezedwa kangapo.
2. Kutaya deta kumatha kuchitikabe ngati kukonza deta sikunachitike bwino.
3. Pa deta yamagulu (monga deta ya odwala yomwe ili ndi zolemba zingapo), njira yapadera imafunika, monga gulu la k-fold, kuti munthu m'modzi asawonekere mu sitima ndikuyesa nthawi imodzi.

Makhalidwe Abwino Ogwiritsira Ntchito Kutsimikizira Mtanda

Kuti kuwunika kukhale kovomerezeka, mfundo zingapo zofunika ziyenera kutsatiridwa:
- Chitani zinthu zokonzekera (kukonza zinthu, kuyika zinthu, kusankha zinthu) mkati mwa fold iliyonse, osati kamodzi kokha pa deta yonse. Kupanda kutero, chidziwitso chochokera mu fold yoyesera chikhoza kutuluka mu fold yoyendera.
- Gwiritsani ntchito k-fold yolumikizidwa m'magulu pogwiritsa ntchito magulu osalinganika.
- Gwiritsani ntchito njira yapadera yopezera deta ya nthawi kuti dongosolo lisaphwanyidwe.
- Ikani pambali seti yomaliza yoyesera ngati cholinga chanu ndikuwunika momwe chitsanzocho chikuyendera musanagwiritse ntchito.

Kutseka

Kutsimikizira kosiyanasiyana ndi chida chofunikira kwambiri mu ziwerengero zogwiritsidwa ntchito komanso kuphunzira kwa makina poyesa magwiridwe antchito a chitsanzo mwachilungamo komanso mwamphamvu. Pogwiritsa ntchito kugawana deta mobwerezabwereza, kutsimikizira kosiyanasiyana kumathandiza kuchepetsa tsankho lomwe limayambitsidwa ndi kusankha kugawa kwa mayeso, kuzindikira kukwanira kwambiri, ndikuthandizira kusankha chitsanzo ndi kusintha kwa hyperparameter. Ngakhale mtengo wa makompyuta uli wokwera, maubwino nthawi zambiri amakhala ofunika, makamaka pamene deta ndi yaying'ono kapena pamene zisankho zochokera ku zotsatira za chitsanzo zimakhala ndi zotsatirapo zazikulu. Mwa kusankha mtundu woyenera wa kutsimikizira kosiyanasiyana ndikukhazikitsa njira zabwino, titha kupanga mitundu yodalirika kwambiri yomwe ili yokonzeka kugwiritsidwa ntchito pa data yeniyeni.

Siyani ndemanga