Riba:
- Ikon gane nau'ikan yabo bayanai (manufa, lokaci, aiwatarwa, jeren rukuni) da kuma tambayar makin 'mayi kyau ya zama gaskiya' azaman ƙararrawa.
- Ability don hana yayyo tare da farkon rabuwar saitin gwajin, bututun bututu da daidaitaccen rabo (tsari / rukuni)
- Ability don yin bincike reproducible tare da kafaffen tsaba, version iko da kau da manual matakai
Akwai kurakurai guda biyu waɗanda suka ɓata mafi yawan ƙoƙari a cikin ilimin kimiyyar bayanai, kuma duka biyun sun kasance masu ɓarna saboda suna haifar da bala'i daidai lokacin da komai "da alama yana da kyau." Na farko shine zubewar bayanai: samfurin yana aiki sosai akan saitin gwajin amma ya fashe a samarwa. Na biyu shi ne irreproducibility: kuna gudanar da bincike bayan watanni shida kuma ku sami sakamako daban. An sadaukar da wannan rukunin don sani da kuma nisantar waɗannan ramuka biyu a zurfafan su. AI na iya ƙara haɗarin duka biyu (yana haifar da sauri, yana ba da shawarar ɓoyayyiyar ɓoye, yana sauƙaƙa muku ɗaukar matakan hannu) amma kuma yana iya rage su idan aka yi amfani da su daidai. Bambancin yana cikin horo.
Tushen bayanai: samfurin clairvoyant
Zubar da bayanai shine lokacin da samfurin ya ga bayanai yayin horo wanda ba zai samu ba a lokacin ainihin tsinkaya. Samfurin "mai cuta" tare da wannan bayanin, yana kallo mai kyau akan saitin gwajin, amma faduwa a samarwa ba tare da wannan bayanin ba. Alamar yabo kusan koyaushe iri ɗaya ce: yayi kyau a zama gaskiya. Kafin ku yi murna lokacin da kuka ga daidaito 99%, yakamata ku nemi leaks.
Babban nau'ikan yabo sune:
1. Goal leakage: Wani fasali shine sakamakon burin. A cikin hasashen "an soke", ginshiƙan "kwanakin sokewa" ko "adadin dawowa" shine sakamakon manufa; Za a cika su ne kawai lokacin da sakamakon ya bayyana.
2. Leak Time: Kawo bayanai na gaba zuwa abubuwan da suka gabata. Lokacin ƙididdige "matsakaicin kwanaki 30 na ƙarshe", haɗa da kwanakin bayan ranar hasashen, ko raba jerin lokaci ba da gangan ba.
3. Pre-processing leakage: Koyo canje-canje kamar sikelin, cika, coding daga duk bayanai kafin horo / gwaji bangare. Matsakaicin bayanan gwaji yana tsoma baki tare da horo.
4. Kwafi/ruwan layuka: Layukan na mutum ɗaya suna nan a cikin horo da gwaji (ziyara guda biyu na majiyyaci ɗaya a cikin nau'i daban-daban). Samfurin yana haddace mutum.
Nau'in leak
Yadda ake haihuwa
Yadda ake hana
zubewar manufa
Shagon wanda shine sakamakon manufa
"Shin ina da shi a lokacin tsinkaya" gwajin
lokacin zubarwa
Kawo gaba ga abin da ya gabata
Rarraba na zamani, sarrafa taga
Preprocessing yabo
Juyawa da aka riga aka raba
Bututu, dacewa daga horo kawai
Zubewar layin rukuni
Raka'a ɗaya a cikin saiti biyu
Raba ta rukuni (GroupKFold)
Iyakar horo don hana zubewa
Magani na gama gari na kowane nau'in leken asiri ya gangara zuwa jumla ɗaya: Ware saitin gwajin da wuri-wuri don kwaikwayi ainihin nan gaba, kuma kada ku "koyar da" komai. A aikace, wannan yana nufin: na farko raba, sannan ku koyi duk canje-canje kawai daga horo kuma kuyi amfani da su a cikin bututu (tsarin da ke tattara duk matakai a cikin sarkar guda). Ga kowane fasali, yi tambayar "Shin ina da wannan bayanin a lokacin hasashen?" Idan akwai lokaci, raba shi bisa ga tsarin lokaci; Idan ɗayan ɗayan yana maimaituwa, raba ta rukuni.
Tsanaki: Mafi haɗari al'amari na zub da jini shi ne ya gabatar da kansa a matsayin nasara. Wani mummunan samfurin zai haifar da mummunan sakamako kuma za a lura da shi; Samfurin da aka leka yana aiki da kyau, yana faranta wa kowa rai, kuma an saka shi cikin samarwa - anan ne rushewar ta fara. Shi ya sa sakamakon "mai kyau sosai" ke haifar da ƙararrawa, ba bikin ba.
Reproducibility: samun sakamako iri ɗaya sau biyu
Maimaituwa shine ikon samun sakamako iri ɗaya lokacin da kuka sake gudanar da bincike a wani lokaci, akan wata na'ura. Idan ba tare da wannan ba, bincikenku na faruwa ne, ba kimiyya ba. Babban dalilai da mafita waɗanda ke lalata haifuwa:
Matakai na hannu: Canza tantanin halitta da hannu a cikin Excel, gyara taswira da hannu. Magani: sami kowane mataki a lamba.
Rashin Kafaffen bazuwar: Horon samfuri, samfuri, rarrabuwa sun haɗa da bazuwar. Magani: gyara nau'in bazuwar (ƙimar farko na janareta bazuwar) (random_state=42).
Sigar tana canzawa: Sakamakon na iya canzawa lokacin da sigar ɗakin karatu ta canza. Magani: gyara abubuwan dogaro (requirements.txt, fayil ɗin yanayi).
Babu ajiyar rikodi: Ba a bayyana wace bayanai ba, wace lamba, wacce siga aka yi amfani da ita. Magani: sarrafa sigar (Git - tsarin da ke adana duk nau'ikan lambar) da sigar bayanai.
"Yana aiki akan injina kawai": Magani: rubuta muhalli, yi amfani da kwantena (Docker) idan zai yiwu.
uku mini lokuta
Case 1 - Zubar da manufa. Wani bincike na kiwon lafiya ya nuna shafi na "maganin zubar da ruwa" a cikin tsinkayar "ko za a sake shigar da majiyyaci." An cika wannan shafi ne kawai bayan an sallami mai haƙuri. Samfurin ya ba da 96%, a cikin samarwa 61%. Aikin mako 8 shara ne. Darasi: tambayi kowace sifa "shin yana nan a lokacin tsinkaya?"
Case 2 - Preprocessing yabo. Ƙungiya ɗaya ta daidaita duk bayanan sannan ta raba su. Ma'anar bayanan gwajin ya shiga cikin ƙima. CV ci 89%, ainihin samarwa 76%. Nasarar karya ta ɓace lokacin da na ƙaura zuwa Pipeline kuma na koyi game da sauye-sauye daga horo kawai. Darasi: raba farko, canzawa daga baya.
Hali na 3 - Rashin haɓakawa. Wani manazarci ya so sabunta taswirar da ya gabatar wa gudanarwa bayan watanni uku amma ya kasa tuna yadda ya samar da ita; matakai da yawa an yi su da hannu a cikin Excel. Sakamakon bai yi tasiri ba kuma amana ta girgiza. Darasi: babu matakai na hannu, komai yana cikin lamba da Git.
Samfura huɗu masu kwafi
1) Duban lek:
Matsayinku: leak inspector. Manufa: "churn" (0/1), kwanan wata hasashen hasashen: rikodin_date. Zan ba ku wannan jerin fasali. Ga KOWANE fasali: (a) sakamakon makasudin ne, (b) Shin yana samuwa a gare ni a lokacin tsinkaya, (c) Tagar lokaci ta ƙunshi nan gaba? Alama shi a matsayin "marasa lafiya/shakku/leak" kuma rubuta dalili. Siffofin: [jeri]
2) Bututun da ba ya zube:
Saita bututun sklearn: jirgin ƙasa na farko / gwaji (stratified, iri = 42), sannan ya dace da duk abubuwan da aka riga aka tsara (impute, scale, encode) a cikin bututun daga horo KAWAI. Bayyana dalilin da yasa lambar ba ta da ruwa, wane mataki aka koya a ina.
3) Lambar lissafin sake fasalin:
Ina so in mayar da bincike na. Ba da shawarar lamba/tsarin da ke ƙara: (1) iri mai ƙarfi don duk bazuwar, (2) nau'ikan laburare da aka yi amfani da su, (3) alamar kwanan wata/version don bayanai da fitarwa. Har ila yau, ba ni lissafin bincike don tabbatar da cewa babu matakai na hannu.
4) Bangaren rukuni (leke naúrar guda ɗaya):
A cikin bayanan abokin ciniki_id iri ɗaya yana cikin layuka da yawa. Yi rarrabuwa (GroupKFold koGroupShuffleSplit, rukuni = abokin ciniki_id) wanda ke hana abokin ciniki iri ɗaya kasancewa cikin horo da gwaji. Haɗa lamba don tabbatar da cewa babu abokin ciniki a cikin saitin biyu bayan an raba.
Rauni mai ƙarfi / Ƙarfi mai ƙarfi
Rawanin faɗakarwa:
Samfurina ya dawo daidai 98%, shin ba haka bane? Inganta lambar.
Bikin 98% yana ɓoye ɓoyayyiyar. Kafin ingantawa, yakamata a yi tambaya ko wannan maki na gaske ne ko a'a.
Ƙarfi mai ƙarfi:
Matsayinku: leak inspector. Samfurin nawa yana dawo da daidaiton kashi 98% akan saitin gwajin, wanda yayi kama da "mai kyau ya zama gaskiya" a gareni. Bincika: (1) duk wani fasali ne sakamakon abin da aka yi niyya, (2) ana yin jujjuyawar da aka yi kafin rabuwa, (3) raka'a iri ɗaya ce a cikin saiti biyu, (4) shin akwai lokacin da za a iya ɓata lokaci. Jera duk abubuwan da ake tuhuma; Mayar da hankali kan gano ɗigon ruwa, ba daidaita maki ba.
Anan, babban maki ana ɗaukar shi azaman alamar tambaya, ba don bikin ba.
Kuskuren gama gari
- Bikin sakamakon "mai kyau sosai". Maki-mai kyau-zuwa-gaskiya faɗakarwa ce, ba nasara ba.
- Koyon canji daga duk bayanai kafin rarraba. Mafi yawan zubewa; Raba farko da bututu.
- Rarraba jerin lokaci ba da gangan ba. Samfurin yana ganin makomar gaba; Rarraba na lokaci-lokaci dole ne.
- Barin raka'a ɗaya cikin saiti biyu. Misali yana haddace mutum; Raba ta rukuni.
- Ba shiga da hannu da rubutu a cikin lambar ba. Binciken ya zama wanda ba zai iya haifar da shi ba; komai yakamata ya kasance a cikin code da Git.
Tukwici: Rubuta jimla guda biyu "albarin girmamawa" a farkon aikinku: "Ban taɓa gwajin da aka saita ta kowace hanya ba kafin in gan shi a samarwa. Kowane mataki yana cikin lambar kuma an gyara iri." Idan ba za ku iya sanya hannu kan waɗannan jimloli biyu da gaskiya ba, har yanzu sakamakonku bai kasance abin dogaro ba tukuna.
A takaice
Fitar bayanai da rashin sakewa su ne kurakuran shiru guda biyu mafi tsada a kimiyyar bayanai. Leakage shine hangen nesa na samfurin na gaba kuma yana gabatar da kansa a matsayin nasarar karya; Magani shine raba saitin gwajin da wuri, koyi canje-canje kawai daga horo (bututun bututu), tambayi kowane fasalin tambayar "Shin ina da shi a lokacin tsinkaya" kuma kuyi daidaitaccen rarrabuwa (tsari / rukuni). Maimaituwa yana iya samun sakamako iri ɗaya sau biyu; Maganin sa shine cire matakai da hannu, sanya iri, daskare sigar, da adana komai a Git. AI na iya ƙarawa ko rage waɗannan haɗari; Horon ku ne ke ƙayyade.
Aikin aikace-aikace
Ɗauki jerin fasalulluka na ƙirar da kuka gina (ko na hasashe) kuma ku yi kowane fasalin tambayar "shin ina da wannan bayanin a lokacin tsinkaya?" a rubuce; Nemo aƙalla ɗan takara guda ɗaya. Sa'an nan kuma cika jerin abubuwan bincike don yin nazarin binciken ku: an gyara iri, akwai matakan hannu, akwai nau'ikan rajista, suna cikin Git. Gyara gazawar.
jerin abubuwan dubawa
- [ ] Na tambayi makin "mafi kyau a zama gaskiya" azaman faɗakarwa?
- [ ] Shin na koyi duk canje-canje bayan an raba, daga horo kawai?
- [ ] Na raba bisa ga tsarin lokaci/rukuni (na zamani/RukuninKFold)?
- [ ] Na yi duk abin da bazuwar maimaituwa tare da tsayayyen iri?
- [ ] Na cire matakan jagora kuma na adana komai a cikin lamba da sarrafa sigar?