Okuqukethwe[Fihla][Bonisa]
- 1. Chaza umehluko phakathi kokufunda komshini, ubuhlakani bokwenziwa, nokufunda okujulile.
- 2. Sicela uchaze izinhlobo ezahlukene zokufunda komshini.
- 3. Kuyini ukuchema nokuhluka kokuhwebelana?
- 4. Ama-algorithms wokufunda ngomshini aguquke kakhulu ngokuhamba kwesikhathi. Umuntu uyikhetha kanjani i-algorithm efanele azoyisebenzisa uma enikezwe isethi yedatha?
- 5. I-covariance nokuhlobana kwehluke kanjani?
- 6. Ekufundeni komshini, kusho ukuthini ukuhlanganisa?
- 7. Iyiphi i-algorithm oyithandayo yokufunda ngomshini?
- 8. Ukwehla Komugqa Ekufundeni Ngomshini: Kuyini?
- 9. Chaza umehluko phakathi kwe-KNN kanye ne-k-means clustering.
- 10. Kusho ukuthini “ukuchema” kuwena?
- 11. Iyini ngempela i-Bayes' Theorem?
- 12. Kumodeli Yokufunda Ngomshini, yini 'Isethi Yokuqeqesha' kanye 'Nesethi Yokuhlola'?
- 13. Iyini i-hypothesis ekuFundeni ngomshini?
- 14. Kusho ukuthini ukufakwa ngokweqile komshini wokufunda, futhi kungavinjelwa kanjani?
- 15. Bayini ngempela abafundi bezigaba baseNaive Bayes?
- 16. Kusho ukuthini Imisebenzi Yezindleko kanye Nemisebenzi Yokulahlekelwa?
- 17. Yini ehlukanisa imodeli ekhiqizayo kunemodeli ebandlululayo?
- 18. Chaza ukuhluka phakathi kwamaphutha oHlobo I kanye noHlobo II.
- 19. Ekufundeni komshini, iyiphi indlela yokufunda ye-Ensemble?
- 20. Ayini ngempela amamodeli we-parametric? Nikeza isibonelo.
- 21. Chaza ukuhlunga ngokubambisana. Kanye nokuhlunga okusekelwe kokuqukethwe?
- 22. Usho ukuthini ngempela ngochungechunge Lwesikhathi?
- 23. Chaza ukuhluka phakathi kwe-Gradient Boosting kanye ne-Random Forest algorithms.
- 24. Kungani udinga i-matrix yokudideka? Kwenzenjani?
- 25. Kuyini ngempela ukuhlaziywa kwengxenye yesimiso?
- 26. Kungani ukuzungezisa kwengxenye kubaluleke kakhulu ku-PCA (ukuhlaziywa kwengxenye eyinhloko)?
- 27. Ukujwayela nokujwayelekile kwehluka kanjani komunye nomunye?
- 28. Ukujwayezwa kanye nokumiswa kuhluke kanjani kokunye?
- 29. Kusho ukuthini ngempela ukuthi “variance inflation factor”?
- 30. Ngokusekelwe kusayizi wesethi yokuqeqeshwa, ungamkhetha kanjani umuntu ohlukanisa isigaba?
- 31. Iyiphi i-algorithm ekufundeni komshini ebizwa ngokuthi "umfundi oyivila" futhi ngani?
- 32. Yini i-ROC Curve ne-AUC?
- 33. Ayini ama-hyperparameter? Yini ebenza bahluke kumapharamitha wemodeli?
- 34. Kusho ukuthini i-F1 Score, khumbula, kanye nokunemba?
- 35. Kuyini ngempela ukuqinisekiswa okuphambene?
- 36. Ake sithi uthole ukuthi imodeli yakho inokwehluka okukhulu. Iyiphi i-algorithm, ngokubona kwakho, efaneleke kakhulu ukusingatha lesi simo?
- 37. Yini ehlukanisa ukuhlehla kwe-Ridge nokuhlehla kwe-Lasso?
- 38. Yikuphi okubaluleke kakhulu: ukusebenza kwemodeli noma ukunemba kwemodeli? Iyiphi futhi kungani uzoyithanda?
- 39. Ungayiphatha kanjani idathasethi enokungalingani?
- 40. Ungahlukanisa kanjani phakathi kwe-boosting ne-bagging?
- 41. Chaza umehluko phakathi kokufunda ngokufunda nokudonsela phansi.
- Isiphetho
Amabhizinisi asebenzisa ubuchwepheshe obusezingeni eliphezulu, obufana nobuhlakani bokwenziwa (AI) nokufunda ngomshini, ukukhulisa ukufinyeleleka kolwazi namasevisi kubantu ngabanye.
Lobu buchwepheshe bamukelwa yizimboni ezihlukahlukene, okuhlanganisa amabhange, ezezimali, ukuthengisa, ukukhiqiza, nokunakekelwa kwezempilo.
Enye yezindima ezifunwa kakhulu zenhlangano esebenzisa i-AI ingeyososayensi bedatha, onjiniyela bobuhlakani bokwenziwa, onjiniyela bokufunda ngemishini, nabahlaziyi bedatha.
Lokhu okuthunyelwe kuzokuholela ezinhlobonhlobo ze ukufunda imishini imibuzo yenhlolokhono, kusukela kokuyisisekelo kuye kokuyinkimbinkimbi, ukukusiza ukuthi ulungele noma yimiphi imibuzo ongabuzwa lapho ufuna umsebenzi owufanele.
1. Chaza umehluko phakathi kokufunda komshini, ubuhlakani bokwenziwa, nokufunda okujulile.
Ubuhlakani bokwenziwa busebenzisa izinhlobonhlobo zokufunda komshini nezindlela zokufunda ezijulile ezivumela amasistimu wekhompiyutha ukuthi enze imisebenzi esebenzisa ubuhlakani obunjengobomuntu ngomqondo nemithetho.
Ukufunda ngomshini kusebenzisa izibalo ezihlukahlukene kanye nezindlela Zokufunda Okujulile ukuze imishini ikwazi ukufunda ekusebenzeni kwayo kwangaphambili futhi ibe nobuchule bokwenza imisebenzi ethile iyodwa ngaphandle kokugadwa umuntu.
I-Deep Learning iqoqo lama-algorithms avumela isofthiwe ukuthi ifunde kuyona futhi yenze imisebenzi ehlukahlukene yezohwebo, njengokuqaphela izwi nesithombe.
Amasistimu aveza izendlalelo eziningi amanethiwekhi we-neural inani elikhulu ledatha yokufunda bayakwazi ukufunda ngokujulile.
2. Sicela uchaze izinhlobo ezahlukene zokufunda komshini.
Ukufunda ngomshini kukhona ngezinhlobo ezintathu ezihlukene ngobubanzi:
- Ukufunda Okugadiwe: Imodeli idala izibikezelo noma izahlulelo isebenzisa idatha enelebula noma yomlando ekufundeni komshini ogadiwe. Amasethi edatha amakiwe noma alebulwe ukuze kwandiswe incazelo yawo abizwa ngokuthi idatha enelebula.
- Ukufunda okungagadiwe: Asinayo idatha enelebula yokufunda engagadiwe. Kudatha engenayo, imodeli ingathola amaphethini, okungavamile, nokuhlobana.
- Ukuqinisa Ukufunda: Imodeli ingakwazi funda ngokusebenzisa ukuqinisa ukufunda kanye nemivuzo eyayithola ngokuziphatha kwayo kwangaphambili.
3. Kuyini ukuchema nokuhluka kokuhwebelana?
Ukufaka i-overfitting kuwumphumela wokuchema, okuyizinga imodeli elingana ngalo nedatha. Ukuchema kubangelwa ukucabanga okungalungile noma okulula kakhulu kuwe i-algorithm yokufunda yomshini.
Ukwehluka kusho amaphutha abangwe ubunkimbinkimbi ku-algorithm yakho ye-ML, okukhiqiza ukuzwela kumazinga amakhulu okuhluka kudatha yokuqeqeshwa kanye nokufakwa ngokweqile.
Umehluko ukuthi imodeli ihlukahluka kangakanani kuye kokufakwayo.
Ngamanye amazwi, amamodeli ayisisekelo achemile ngokwedlulele kodwa azinzile (ukwehluka okuphansi). Ukufaka i-overfitting kuyinkinga ngamamodeli ayinkimbinkimbi, yize noma kunjalo ethwebula iqiniso lemodeli (ukuchema okuphansi).
Ukuze uvimbele kokubili ukuhluka okuphezulu nokuchema okuphezulu, ukuhwebelana phakathi kokuchema nokuhluka kuyadingeka ukuze kuncishiswe amaphutha angcono kakhulu.
4. Ama-algorithms wokufunda ngomshini aguquke kakhulu ngokuhamba kwesikhathi. Umuntu uyikhetha kanjani i-algorithm efanele azoyisebenzisa uma enikezwe isethi yedatha?
Indlela yokufunda yomshini okufanele isetshenziswe incike kuphela ohlotsheni lwedatha kudathasethi ethile.
Uma idatha iwumugqa, kusetshenziswa ukuhlehla komugqa. Indlela yokufaka izikhwama izosebenza kangcono uma idatha ikhombisa ukungahambisani nomugqa. Singasebenzisa izihlahla zesinqumo noma i-SVM uma idatha kufanele ihlolwe noma ihunyushwe ngezinjongo zokuthengisa.
Amanethiwekhi e-Neural angase abe usizo ukuthola impendulo enembile uma idathasethi ihlanganisa izithombe, amavidiyo, nomsindo.
Ukukhethwa kwe-algorithm yesimo esithile noma ukuqoqwa kwedatha akukwazi ukwenziwa ngesilinganiso esisodwa nje.
Ngenhloso yokuthuthukisa indlela efanelekile, kufanele siqale sihlole idatha sisebenzisa ukuhlaziywa kwedatha yokuhlola (EDA) futhi siqonde umgomo wokusebenzisa idathasethi.
5. I-covariance nokuhlobana kwehluke kanjani?
I-Covariance ihlola ukuthi okuguquguqukayo okubili kuxhumene kanjani nokuthi umuntu angashintsha kanjani ekuphenduleni izinguquko komunye.
Uma umphumela uhle, ubonisa ukuthi kukhona ukuxhumana okuqondile phakathi kwezinto eziguquguqukayo nokuthi umuntu angakhuphuka noma ehle ngokunyuka noma ukuncipha kokuguquguquka kwesisekelo, kucatshangwa ukuthi zonke ezinye izimo zihlala zingashintshi.
Ukuhlobana kukala isixhumanisi phakathi kokuguquguquka okungahleliwe okubili futhi kunamanani amathathu kuphela ahlukene: 1, 0, kanye -1.
6. Ekufundeni komshini, kusho ukuthini ukuhlanganisa?
Izindlela zokufunda ezingagadiwe eziqoqa amaphuzu edatha ndawonye zibizwa ngokuthi ukuhlanganisa. Ngeqoqo lamaphuzu edatha, indlela yokuhlanganisa ingasetshenziswa.
Ungaqoqa wonke amaphuzu edatha ngokwemisebenzi yawo usebenzisa leli su.
Izici nezimfanelo zamaphoyinti edatha awela esigabeni esifanayo ziyefana, kuyilapho lawo amaphuzu edatha awela emaqenjini ahlukene ahlukile.
Le ndlela ingasetshenziswa ukuhlaziya idatha yezibalo.
7. Iyiphi i-algorithm oyithandayo yokufunda ngomshini?
Unethuba lokubonisa okuncamelayo kanye namakhono ahlukile kulo mbuzo, kanye nolwazi lwakho olubanzi lwamasu amaningi okufunda emishini.
Nawa ama-algorithms wokufunda womshini ambalwa ajwayelekile ongacabanga ngawo:
- Ukuhlehla komugqa
- Ukuhlehla kwezinto
- AmaNaive Bayes
- Izihlahla zesinqumo
- K kusho
- I-algorithm yehlathi engahleliwe
- Umakhelwane oseduze no-K (KNN)
8. Ukwehla Komugqa Ekufundeni Ngomshini: Kuyini?
I-algorithm yokufunda komshini ogadiwe ukuhlehla komugqa.
Isetshenziswa ekuhlaziyeni okubikezelwayo ukuze kutholwe ukuxhumana komugqa phakathi kweziguquguquko ezincikile nezizimele.
Isibalo sokuhlehla komugqa simi kanje:
Y = A + BX
lapho:
- Okokufaka noma okuhlukile okuzimele kubizwa ngokuthi u-X.
- Okuhluka okuncikile noma okukhiphayo ngu-Y.
- I-coefficient ka-X ingu-b, futhi ukunqamuka kwayo kungu-a.
9. Chaza umehluko phakathi kwe-KNN kanye ne-k-means clustering.
Umehluko oyinhloko ukuthi i-KNN (indlela yokuhlukanisa, ukufunda okugadiwe) idinga amaphuzu anelebuli kanti u-k-means awafuni (i-algorithm yokuhlanganisa, ukufunda okungagadiwe).
Ungakwazi ukuhlukanisa idatha enelebula ibe iphoyinti elingalebulanga ngokusebenzisa i-K-Nearest Neighbors. Ukuhlanganisa kwe-K kusebenzisa ibanga elimaphakathi phakathi kwamaphoyinti ukuze kufundwe ukuthi uwaqoqa kanjani amaphuzu angenalebuli.
10. Kusho ukuthini “ukuchema” kuwena?
Ukuchema esigabeni sokusampula sokuhlolwa kungenxa yokunemba kwezibalo.
Iqembu elilodwa lesampula likhethwa kaningi kunamanye amaqembu esivivinyweni ngenxa yokunemba.
Uma ukuchema kokukhethwa kungavunywa, kungaholela esiphethweni esingalungile.
11. Iyini ngempela i-Bayes' Theorem?
Uma sazi amanye amathuba, singanquma amathuba sisebenzisa i-Bayes' Theorem. Inikeza amathuba angemuva esenzeko asekelwe olwazini lwangaphambilini, ngamanye amazwi.
Indlela ezwakalayo yokulinganisa amathuba anemibandela inikezwa yile theory.
Lapho uthuthukisa izinkinga zokubikezela ngokwezigaba futhi uhlanganisa imodeli ekuqeqeshweni idathasethi ekufundeni komshini, i-theorem ye-Bayes isetshenziswa (okungukuthi i-Naive Bayes, i-Bayes Optimal Classifier).
12. Kumodeli Yokufunda Ngomshini, yini 'Isethi Yokuqeqesha' kanye 'Nesethi Yokuhlola'?
Isethi yokuqeqesha:
- Isethi yokuqeqeshwa iqukethe izimo ezithunyelwa kumodeli ukuze zihlaziywe nokufunda.
- Lena idatha enelebula ezosetshenziswa ukuqeqesha imodeli.
- Ngokuvamile, u-70% wayo yonke idatha isetshenziswa njengedathasethi yokuqeqeshwa.
Isethi Yokuhlola:
- Isethi yokuhlola isetshenziselwa ukuhlola ukunemba kokukhiqizwa kwe-hypothesis yemodeli.
- Sihlola ngaphandle kwedatha enelebula bese sisebenzisa amalebula ukuze siqinisekise imiphumela.
- U-30% osele usetshenziswa njengedathasethi yokuhlola.
13. Iyini i-hypothesis ekuFundeni ngomshini?
Ukufunda Ngomshini kuvumela ukusetshenziswa kwamadathasethi akhona ukuqonda kangcono umsebenzi othile oxhumanisa okokufaka nokuphumayo. Lokhu kwaziwa ngokuthi ukulinganiselwa komsebenzi.
Kulesi simo, ukulinganisa kufanele kusetshenziswe kumsebenzi oqondiwe ongaziwa ukuze kudluliselwe konke ukubonwa ongase kucatshangwe ngokusekelwe esimweni esinikeziwe ngendlela engcono kakhulu engakhona.
Ekufundeni komshini, i-hypothesis imodeli esiza ekulinganiseni umsebenzi oqondiwe kanye nokuqedela imephu efanelekile yokufaka kuya kokuphumayo.
Ukukhethwa nokuklanywa kwama-algorithms kuvumela incazelo yesikhala semibono engase ibe khona engamelwa imodeli.
Ukuze uthole i-hypothesis eyodwa, kusetshenziswa igama elithi h (h) elincane, kodwa usonhlamvukazi u-h (H) usetshenziselwa sonke isikhala se-hypothesis esiseshwayo. Sizobuyekeza kafushane lezi zimpawu:
- I-hypothesis (h) imodeli ethile eyenza kube lula ukwenziwa kwemephu kokufakwayo kuya kokuphumayo, okungase kamuva kusetshenziselwe ukuhlola nokubikezela.
- Isethi ye-hypothesis (H) iyindawo eseshekayo yemibono engasetshenziswa ukwenza imephu okokufaka kokuphumayo. Uhlaka lwendaba, imodeli, nokucushwa kwemodeli kuyizibonelo ezimbalwa zemikhawulo ejwayelekile.
14. Kusho ukuthini ukufakwa ngokweqile komshini wokufunda, futhi kungavinjelwa kanjani?
Uma umshini uzama ukufunda kudathasethi enganele, ukugcwalisa ngokweqile kwenzeka.
Ngenxa yalokho, ukufaka ngokweqile kuhlotshaniswa ngokuphambene nevolumu yedatha. Indlela yokuqinisekisa ehlukahlukene ivumela ukufaka ngokweqile ukuthi kugwenywe kumadathasethi amancane. Idathasethi ihlukaniswa yaba izingxenye ezimbili kule ndlela.
Idathasethi yokuhlolwa nokuqeqeshwa izoqukatha lezi zingxenye ezimbili. Idathasethi yokuqeqeshwa isetshenziselwa ukudala imodeli, kuyilapho idathasethi yokuhlola isetshenziselwa ukuhlola imodeli kusetshenziswa okokufaka okuhlukile.
Lena indlela yokuvimbela ukugcwala ngokweqile.
15. Bayini ngempela abafundi bezigaba baseNaive Bayes?
Izindlela ezahlukahlukene zokuhlukanisa zakha abahlukanisi beNaive Bayes. Isethi yama-algorithms aziwa njengalezi zihlukanisi zonke zisebenza embonweni ofanayo oyisisekelo.
Umcabango owenziwe abahlukanisi bezigaba be-Bayes abangenangqondo ukuthi ukuba khona noma ukungabikho kwesici esisodwa akuthinti ukuba khona noma ukungabikho kwesinye isici.
Ngamanye amazwi, yilokhu esikubiza ngokuthi “ukunganaki” njengoba kwenza kucatshangwe ukuthi isibaluli sedathasethi ngayinye sibaluleke ngokulinganayo futhi sizimele.
Ukuhlukaniswa kwenziwa kusetshenziswa izihlukanisi ze-Bayes ezingenangqondo. Zilula ukuzisebenzisa futhi zikhiqize imiphumela engcono kunezibikezelo eziyinkimbinkimbi lapho isisekelo sokuzimela siyiqiniso.
Ekuhlaziyeni umbhalo, ukuhlunga ogaxekile, nezinhlelo zokuncoma, ziyasebenza.
16. Kusho ukuthini Imisebenzi Yezindleko kanye Nemisebenzi Yokulahlekelwa?
Inkulumo ethi "umsebenzi wokulahlekelwa" ibhekisela kunqubo yokulahlekelwa kwekhompuyutha lapho ucezu olulodwa lwedatha lucatshangelwa.
Ngokuphambene, sisebenzisa umsebenzi wezindleko ukuze sinqume inani eliphelele lamaphutha kudatha eminingi. Akukho mehluko obalulekile okhona.
Ngamanye amazwi, kuyilapho imisebenzi yezindleko ihlanganisa umehluko wayo yonke idathasethi yokuqeqeshwa, imisebenzi yokulahlekelwa iklanyelwe ukuthwebula umehluko phakathi kwamanani angempela nabikezelwe erekhodi elilodwa.
17. Yini ehlukanisa imodeli ekhiqizayo kunemodeli ebandlululayo?
Imodeli ebandlululayo ifunda umehluko phakathi kwezigaba ezimbalwa zedatha. Imodeli ekhiqizayo ithatha izinhlobo ezahlukene zedatha.
Ezinkingeni zokuhlukanisa, amamodeli abandlululayo avame ukudlula amanye amamodeli.
18. Chaza ukuhluka phakathi kwamaphutha oHlobo I kanye noHlobo II.
Okuhle okungelona iqiniso kuwela ngaphansi kwesigaba samaphutha ohlobo I, kuyilapho ukuphika okungamanga kungena ngaphansi kwamaphutha ohlobo lwe-II (ukuthi akwenzeki lutho ngenkathi kwenzeka ngempela).
19. Ekufundeni komshini, iyiphi indlela yokufunda ye-Ensemble?
Indlela ebizwa nge-ensemble learning ihlanganisa amamodeli amaningi okufunda omshini ukuze kukhiqizwe amamodeli anamandla kakhulu.
Imodeli ingahlukahluka ngenxa yezizathu ezihlukahlukene. Ziningana izimbangela yilezi:
- Abantu Abahlukahlukene
- Ama-hypotheses ahlukahlukene
- Izindlela ezahlukahlukene zokumodela
Sizobhekana nenkinga ngenkathi sisebenzisa idatha yokuqeqeshwa nokuhlola yemodeli. Ukuchema, ukuhluka, kanye nephutha elingenakunqanyulwa yizinhlobo ezingaba khona zaleli phutha.
Manje, lokhu kubhalansi sikubiza phakathi kokuchema nokuhluka kwemodeli ngokuthi ukuhwebelana nokuhlukahluka, futhi kufanele kuhlale kukhona. Lokhu kuhwebelana kufezwa ngokusebenzisa izifundo ezihlangene.
Nakuba kunezindlela ezihlukahlukene zokuhlanganisa ezitholakalayo, kunezindlela ezimbili ezivamile zokuhlanganisa amamodeli amaningi:
- Indlela yomdabu ebizwa ngokuthi i-bagging isebenzisa isethi yokuqeqesha ukukhiqiza amasethi okuqeqesha engeziwe.
- I-Boosting, indlela eyinkimbinkimbi kakhulu: Njengokufaka izikhwama, i-boosting isetshenziselwa ukuthola ifomula yokukala efanelekile yesethi yokuqeqeshwa.
20. Ayini ngempela amamodeli we-parametric? Nikeza isibonelo.
Kunenani elilinganiselwe lamapharamitha kumamodeli wepharamitha. Ukuze ubikezele idatha, okudingeka ukwazi amapharamitha wemodeli.
Okulandelayo yizibonelo ezijwayelekile: ukuhlehla kwezinto, ukuhlehla komugqa, nama-SVM anomugqa. Amamodeli angewona amapharamitha ayavumelana nezimo njengoba angaqukatha inani elingenamkhawulo lamapharamitha.
Amapharamitha emodeli kanye nesimo sedatha ebhekiwe kuyadingeka ukuze kuqagelwe idatha. Nazi izibonelo ezijwayelekile: amamodeli wesihloko, izihlahla zesinqumo, nomakhelwane abaseduze kuka-k.
21. Chaza ukuhlunga ngokubambisana. Kanye nokuhlunga okusekelwe kokuqukethwe?
Indlela ezanyiwe neyiqiniso yokudala iziphakamiso zokuqukethwe ezenzelwe wena ukuhlunga ngokubambisana.
Uhlobo lwesistimu yokuncoma olubizwa ngokuthi ukuhlunga ngokubambisana lubikezela izinto ezintsha ngokulinganisa okuthandwayo komsebenzisi nezithakazelo ezabiwe.
Okuthandwayo komsebenzisi ukuphela kwento ecatshangwa yizinhlelo zesincomo esekwe kokuqukethwe. Ngokuvumelana nezinketho zangaphambilini zomsebenzisi, izincomo ezintsha zinikezwa ngezinto ezihlobene.
22. Usho ukuthini ngempela ngochungechunge Lwesikhathi?
Uchungechunge lwesikhathi luyiqoqo lezinombolo ngokulandelana okunyukayo. Esikhathini esinqunyiwe kusengaphambili, iqapha ukunyakaza kwamaphoyinti edatha akhethiwe futhi ngezikhathi ezithile ithwebula amaphoyinti edatha.
Akukho okokufaka kwesikhathi esincane noma esiphezulu sochungechunge lwesikhathi.
Uchungechunge lwesikhathi luvame ukusetshenziswa abahlaziyi ukuze bahlaziye idatha ngokuvumelana nezidingo zabo ezihlukile.
23. Chaza ukuhluka phakathi kwe-Gradient Boosting kanye ne-Random Forest algorithms.
Ihlathi Elingahleliwe:
- Inani elikhulu lezihlahla zesinqumo zihlanganiswe ndawonye ekugcineni futhi zaziwa njengamahlathi angahleliwe.
- Nakuba ukukhuliswa kwe-gradient kukhiqiza isihlahla ngasinye ngaphandle kwesinye, ihlathi elingahleliwe lakha isihlahla ngasinye ngesikhathi.
- Multiclass ukutholwa kwento isebenza kahle namahlathi angahleliwe.
Ukukhulisa i-Gradient:
- Ngenkathi amahlathi angahleliwe ejoyina izihlahla zesinqumo ekupheleni kwenqubo, Imishini Yokukhuphula I-Gradient iyayihlanganisa kusukela ekuqaleni.
- Uma amapharamitha alungiswa ngokufanelekile, ukukhuphula i-gradient kusebenza ngaphezu kwamahlathi angahleliwe ngokwemiphumela, kodwa akuyona inketho ehlakaniphile uma isethi yedatha inokuningi kokuphumayo, okudidayo, noma umsindo njengoba kungabangela imodeli ukuthi ilingane ngokweqile.
- Uma kunedatha engalingani, njengoba kukhona ekuhlolweni kwengozi kwesikhathi sangempela, ukukhuliswa kwegradient kusebenza kahle.
24. Kungani udinga i-matrix yokudideka? Kwenzenjani?
Ithebula elaziwa ngokuthi i-matrix yokudideka, ngezinye izikhathi elaziwa ngokuthi i-matrix yephutha, lisetshenziswa kakhulu ukubonisa ukuthi imodeli yokuhlukanisa, noma isigaba, isebenza kahle kangakanani kusethi yedatha yokuhlola amanani angempela aziwa ngayo.
Kusenza sikwazi ukubona ukuthi imodeli noma i-algorithm isebenza kanjani. Kwenza kube lula ngathi ukubona ukungezwani phakathi kwezifundo ezahlukahlukene.
Isebenza njengendlela yokuhlola ukuthi imodeli noma i-algorithm yenziwa kahle kangakanani.
Izibikezelo zemodeli yokuhlukanisa zihlanganiswa zibe i-matrix yokudideka. Amanani wokubala ilebula yekilasi ngalinye asetshenziswe ukuze kuhlukaniswe inani lezibikezelo ezilungile nezingalungile.
Ihlinzeka ngemininingwane ngamaphutha enziwe ohlukanisa ngezigaba kanye nezinhlobo ezahlukene zamaphutha abangelwa abahlukanisa ngezigaba.
25. Kuyini ngempela ukuhlaziywa kwengxenye yesimiso?
Ngokunciphisa inani lezinto eziguquguqukayo ezihlotshaniswa nenye, umgomo uwukunciphisa ubukhulu bokuqoqwa kwedatha. Kodwa kubalulekile ukugcina ukuhlukahluka ngangokunokwenzeka.
Okuguquguqukayo kushintshwa kube isethi entsha ngokuphelele yokuguquguquka ebizwa ngokuthi izingxenye eziyinhloko.
Lawa ma-PC angama-orthogonal ngoba angama-eigenvector we-covariance matrix.
26. Kungani ukuzungezisa kwengxenye kubaluleke kakhulu ku-PCA (ukuhlaziywa kwengxenye eyinhloko)?
Ukuzungezisa kubalulekile ku-PCA ngoba kuthuthukisa ukuhlukana phakathi kokuhluka okutholwe ingxenye ngayinye, okwenza ukuhumusha kwengxenye kube lula.
Sidinga izingxenye ezinwetshiwe ukuveza ukuhluka kwengxenye uma izingxenye zingazungeziswa.
27. Ukujwayela nokujwayelekile kwehluka kanjani komunye nomunye?
Ukujwayela:
Idatha iyashintshwa ngesikhathi sokujwayelekile. Kufanele wenze idatha ibe ngokwejwayelekile uma inezikali ezihluke kakhulu, ikakhulukazi ukusuka kokuphansi kuye phezulu. Lungisa ikholomu ngayinye ukuze izibalo eziyisisekelo zihambisane zonke.
Ukuqinisekisa ukuthi akukho ukulahleka kokunemba, lokhu kungaba usizo. Ukuthola isignali ngenkathi ungaziba umsindo kungenye yezinhloso zokuqeqeshwa kwemodeli.
Kukhona ithuba lokugcwalisa ngokweqile uma imodeli inikezwa ukulawula okuphelele ukuze kuncishiswe iphutha.
Ukwenza njalo:
Ngokujwayelekile, umsebenzi wokubikezela uyashintshwa. Lokhu kungaphansi kokulawulwa okuthile ngokujwayela, okuvumela imisebenzi yokufaka elula kuneyinkimbinkimbi.
28. Ukujwayezwa kanye nokumiswa kuhluke kanjani kokunye?
Izindlela ezimbili ezisetshenziswa kakhulu zokulinganisa izici ukujwayela nokumiswa.
Ukujwayela:
- Ukukala kabusha idatha ukuze ivumelane nobubanzi be-[0,1] kwaziwa njengokujwayelekile.
- Uma wonke amapharamitha kufanele abe nesilinganiso esihle esifanayo, ukujwayela kuyasiza, kodwa okuphumayo kwesethi yedatha kuyalahleka.
Ukwenza njalo:
- Idatha ilinganiswa kabusha ukuze ibe nencazelo engu-0 kanye nokuchezuka okujwayelekile kuka-1 njengengxenye yenqubo yokumiswa (ukuhluka kweyunithi)
29. Kusho ukuthini ngempela ukuthi “variance inflation factor”?
Isilinganiso sokuhluka kwemodeli nokuhluka kwemodeli enokwahlukana okukodwa okuzimele kwaziwa ngokuthi i-variation inflation factor (VIF).
I-VIF ilinganisela inani le-multicollinearity elikhona kusethi yeziguquguquki ezimbalwa zokuhlehla.
Ukuhluka kwemodeli (VIF) Enokuhluka Okukodwa Okuhlukile Okuzimele
30. Ngokusekelwe kusayizi wesethi yokuqeqeshwa, ungamkhetha kanjani umuntu ohlukanisa isigaba?
Ukuchema okuphezulu, imodeli yokuhluka okuphansi isebenza kangcono kusethi yokuqeqeshwa emfushane njengoba ukufaka ngokweqile kuncane kakhulu. I-Naive Bayes iyisibonelo esisodwa.
Ukuze umelele ukusebenzisana okuyinkimbinkimbi yesethi enkulu yokuqeqesha, imodeli enokwenzelela okuphansi nokuhluka okuphezulu kuyakhethwa. Ukuhlehla kwezinto kuyisibonelo esihle.
31. Iyiphi i-algorithm ekufundeni komshini ebizwa ngokuthi "umfundi oyivila" futhi ngani?
Umfundi ovilaphayo, i-KNN iyi-algorithm yokufunda komshini. Ngenxa yokuthi i-K-NN ibala ngokuguqukayo ibanga isikhathi ngasinye lapho ifisa ukulihlukanisa esikhundleni sokufunda noma imaphi amanani afundiwe ngomshini noma okuhlukile kudatha yokuqeqeshwa, ibamba ngekhanda idathasethi yokuqeqeshwa.
Lokhu kwenza u-K-NN abe ngumfundi ovilaphayo.
32. Yini i-ROC Curve ne-AUC?
Ukusebenza kwemodeli yokuhlukanisa kuwo wonke ama-threshold kuvezwa ngesithombe ijika le-ROC. Inesilinganiso esihle sangempela kanye nemibandela yezinga lokuphozithivu okungamanga.
Kalula nje, indawo engaphansi kwejika le-ROC yaziwa ngokuthi i-AUC (Indawo Ngaphansi Kwejika Le-ROC). Indawo enezinhlangothi ezimbili zejika le-ROC ukusuka ku-(0,0) ukuya ku-AUC iyalinganiswa (1,1). Ukuze kuhlolwe amamodeli okuhlukaniswa okumbambili, kusetshenziswa njengezibalo zokusebenza.
33. Ayini ama-hyperparameter? Yini ebenza bahluke kumapharamitha wemodeli?
Okuhlukile kwangaphakathi kwemodeli kwaziwa njengepharamitha yemodeli. Kusetshenziswa idatha yokuqeqeshwa, inani lepharamitha liyalinganiselwa.
Akwaziwa kumodeli, i-hyperparameter iwukuguquguquka. Inani alikwazi ukunqunywa kusuka kudatha, ngakho-ke asetshenziswa njalo ukubala amapharamitha angamamodeli.
34. Kusho ukuthini i-F1 Score, khumbula, kanye nokunemba?
I-confusion Measure iyimethrikhi esetshenziswa ukukala ukusebenza kahle kwemodeli yokuhlukanisa. Imishwana elandelayo ingasetshenziswa ukuchaza kangcono imethrikhi yokudideka:
I-TP: Okuhle Kweqiniso - Lawa amanani amahle abelindelwe ngokufanelekile. Iphakamisa ukuthi amanani ekilasi elihlongozwayo kanye nekilasi langempela kokubili kuhle.
TN: I-True Negatives- Lawa amanani angalungile ayebikezelwe ngokunembile. Iphakamisa ukuthi kokubili inani lekilasi langempela kanye nesigaba esilindelwe kunegethivu.
Lawa manani—okuphozithiyo okungamanga kanye nokuphikisa okungamanga—kwenzeka uma isigaba sakho sangempela sihluka ekilasini okulindelekile.
Manje,
Isilinganiso senani leqiniso elihle (i-TP) kukho konke ukubonwa okwenziwe ekilasini langempela kubizwa ngokuthi ukukhumbula, okubuye kwaziwe njengokuzwela.
Ukukhumbula kabusha yi-TP/(TP+FN).
Ukunemba isilinganiso senani elihle lokubikezela, eliqhathanisa inani lezinto ezinhle imodeli ezibikezela ngempela ukuthi mangaki amaphothizithi alungile ewabikezela ngokunembile.
Ukunemba kungu-TP/(TP + FP)
Imethrikhi yokusebenza okulula ukuyiqonda iwukunemba, okuyingxenye nje yokuqashelwa okubikezelwe ngokufanelekile kukho konke ukubuka.
Ukunemba kulingana ne-(TP+TN)/(TP+FP+FN+TN).
Ukunemba kanye Nokukhumbula kukalwe futhi kukalwe ukuhlinzeka nge-F1 Score. Njengomphumela, leli phuzu licubungula kokubili okuhle nokubi okungamanga.
I-F1 ivamise ukuba yigugu kakhulu kunokunemba, ikakhulukazi uma unokusatshalaliswa kwekilasi okungalingani, ngisho noma ngokunembile akulula ukukuqonda njengokunemba.
Ukunemba okungcono kakhulu kufinyelelwa lapho izindleko zemibono engamanga kanye nemiphumela engemihle iqhathaniswa. Kungcono ukufaka kokubili i-Precision kanye ne-Recall uma izindleko ezihlotshaniswa nemibono engamanga kanye ne-negative engamanga zihluka kakhulu.
35. Kuyini ngempela ukuqinisekiswa okuphambene?
Indlela yokuphinda yesampula yezibalo ebizwa ngokuthi i-cross-validation ekufundeni komshini isebenzisa amasethi edathasethi ambalwa ukuze iqeqeshe futhi ihlole i-algorithm yokufunda yomshini kuyo yonke imizuliswano eminingana.
Inqwaba yedatha engazange isetshenziselwe ukuqeqesha imodeli ihlolwa kusetshenziswa ukuqinisekiswa okuphambene ukuze kubonakale ukuthi imodeli iyibikezela kahle kangakanani. Ukufakwa ngokweqile kwedatha kuvinjelwa ngokuqinisekisa okuphambene.
I-K-Fold Indlela yokuphinda isampula esetshenziswa kakhulu ihlukanisa yonke idathasethi ibe amasethi angu-K anosayizi abalinganayo. Kubizwa ngokuthi i-cross-validation.
36. Ake sithi uthole ukuthi imodeli yakho inokwehluka okukhulu. Iyiphi i-algorithm, ngokubona kwakho, efaneleke kakhulu ukusingatha lesi simo?
Ukuphatha ukuhlukahluka okuphezulu
Kufanele sisebenzise inqubo yokufaka izikhwama ngezinkinga ezinokuhlukahluka okukhulu.
Ukusampula okuphindaphindiwe kwedatha okungahleliwe kuzosetshenziswa i-algorithm yesikhwama ukuze kuhlukaniselwe idatha ngamaqembu amancane. Uma idatha isihlukanisiwe, singasebenzisa idatha engahleliwe kanye nenqubo ethile yokuqeqesha ukuze sikhiqize imithetho.
Ngemuva kwalokho, ukuvota kungasetshenziswa ukuhlanganisa ukuqagela kwemodeli.
37. Yini ehlukanisa ukuhlehla kwe-Ridge nokuhlehla kwe-Lasso?
Izindlela ezimbili ezisetshenziswa kabanzi zokujwayela iLasso (ebuye ibizwe ngokuthi i-L1) kanye ne-Ridge (ngezinye izikhathi ebizwa ngokuthi i-L2). Zisetshenziselwa ukuvimbela ukugcwala ngokweqile kwedatha.
Ukuze kutholwe isixazululo esingcono kakhulu futhi kuncishiswe ubunkimbinkimbi, lawa masu asetshenziswa ukujezisa ama-coefficients. Ngokujezisa inani lamanani aphelele ama-coefficients, ukuhlehla kwe-Lasso kuyasebenza.
Umsebenzi wenhlawulo ku-Ridge noma ukuhlehla kwe-L2 uthathwa enanini lezikwele zama-coefficients.
38. Yikuphi okubaluleke kakhulu: ukusebenza kwemodeli noma ukunemba kwemodeli? Iyiphi futhi kungani uzoyithanda?
Lona ngumbuzo okhohlisayo, ngakho-ke umuntu kufanele aqale aqonde ukuthi iyini iModel Performance. Uma ukusebenza kuchazwa njengejubane, khona-ke kuncike ohlotsheni lohlelo lokusebenza; noma yiluphi uhlelo lokusebenza olubandakanya isimo sesikhathi sangempela lungadinga isivinini esikhulu njengengxenye ebalulekile.
Isibonelo, imiphumela yosesho ehamba phambili izoba sengozini uma imiphumela yombuzo ithatha isikhathi eside ukufika.
Uma Ukusebenza kusetshenziswa njengesizathu sokuthi kungani ukunemba nokukhumbula kufanele kubekwe kuqala ngaphezu kokunemba, umphumela we-F1 uzoba usizo kakhulu kunokunemba ekuboniseni icala lebhizinisi lanoma iyiphi isethi yedatha engalinganiselwe.
39. Ungayiphatha kanjani idathasethi enokungalingani?
Idathasethi engalingani ingazuza kumasu amasampula. Ukusampula kungenziwa ngendlela engaphansi noma eyeqile.
Ngaphansi kweSampling kusivumela ukuthi sinciphise usayizi wesigaba esiningi ukuze sifane nesigaba esincane, esiza ekwenyuseni isivinini ngokuphathelene nokugcinwa nokusebenza kwesikhathi sokusebenza kodwa futhi okungaholela ekulahlekelweni kwedatha ebalulekile.
Ukuze kulungiswe inkinga yokulahleka kolwazi okubangelwa ukusampula ngokweqile, senza isampula yesigaba Sabancane; nokho, lokhu kusenza sibhekane nezinkinga eziwumthwalo ngokweqile.
Amasu engeziwe afaka:
- I-Cluster-Based Over Sampling- Izimo zesigaba esincane neningi zingaphansi kwenqubo yokuhlanganisa ye-K kulesi simo. Lokhu kwenzelwa ukuthola amaqoqo edathasethi. Bese, iqoqo ngalinye lenziwa isampula ngokweqile ukuze wonke amakilasi abe nosayizi ofanayo futhi wonke amaqoqo ngaphakathi kwekilasi abe nenani elilinganayo lezenzakalo.
- I-SMOTE: I-Smotetic Minority Over-sampling Technique- Ucezu lwedatha oluvela ekilasini labancane lusetshenziswa njengesibonelo, ngemva kwalokho izimo zokwenziwa ezengeziwe eziqathaniswa nayo zikhiqizwa futhi zengezwe kudathasethi yoqobo. Le ndlela isebenza kahle ngamaphoyinti edatha yezinombolo.
40. Ungahlukanisa kanjani phakathi kwe-boosting ne-bagging?
I-Ensemble Techniques inezinguqulo ezaziwa ngokuthi i-bagging and boosting.
Isikhwama-
Kuma-algorithms anokwehluka okuphezulu, i-bagging iyindlela esetshenziselwa ukwehlisa ukuhluka. Omunye umndeni onjalo wabahlukanisi othanda ukuchema umndeni wesihlahla sesinqumo.
Uhlobo lwedatha izihlahla zesinqumo eziqeqeshelwa kulo lunomthelela omkhulu ekusebenzeni kwazo. Ngenxa yalokhu, ngisho nokucutshungulwa okuphezulu kakhulu, ukuhlanganiswa kwemiphumela kwesinye isikhathi kunzima kakhulu ukukuthola kuyo.
Uma idatha yokuqeqeshwa kwezihlahla zesinqumo ishintshiwe, imiphumela iyahluka kakhulu.
Ngenxa yalokho, kusetshenziswa izikhwama, lapho kwakhiwa khona izihlahla eziningi zesinqumo, ngasinye sazo esiqeqeshwa kusetshenziswa isampula yedatha yoqobo, futhi umphumela wokugcina uyisilinganiso sawo wonke lawa mamodeli ahlukene.
Ukukhuthaza:
I-Boosting indlela yokwenza izibikezelo ngohlelo lwe-n-weak classifier lapho isigaba ngasinye esibuthaka senza khona ukushiyeka kwabahlukanisi baso abanamandla. Sibhekisela kusihlukanisi esisebenza kabi kudatha enikeziwe "njengesihlukanisa esibuthakathaka."
Ukukhuphula ngokusobala kuyinqubo kune-algorithm. Ukuhlehla kwezinto kanye nezihlahla zesinqumo ezingajulile ziyizibonelo ezivamile zabahlukanisi ababuthakathaka.
I-Adaboost, i-Gradient Boosting, ne-XGBoost zingama-algorithms amabili okuthuthukisa aziwa kakhulu, nokho, maningi amanye.
41. Chaza umehluko phakathi kokufunda ngokufunda nokudonsela phansi.
Lapho ifunda ngesibonelo eqoqweni lezibonelo ezibhekiwe, imodeli isebenzisa ukufunda kokufunda ukuze ifinyelele esiphethweni esijwayelekile. Ngakolunye uhlangothi, ngokufunda okudonsela phansi, imodeli isebenzisa umphumela ngaphambi kokwenza owayo.
Ukufunda nge-inductive kuyinqubo yokuthola iziphetho ngokubhekwayo.
Ukufunda ngokudonsela phansi kuyinqubo yokudala ukuqaphela okusekelwe kulokho okucatshangwayo.
Isiphetho
Halala! Lena imibuzo ephezulu engama-40 nangaphezulu yenhlolokhono yokufunda ngomshini manje osuzazi izimpendulo zayo. Isayensi yedatha kanye ukuhlakanipha okungekhona okwangempela imisebenzi izoqhubeka nokudingeka njengoba ubuchwepheshe buthuthuka.
Abafundi ababuyekeza ulwazi lwabo lwalobu buchwepheshe obuphambili futhi bathuthukise isethi yabo yamakhono bangathola inhlobonhlobo yamathuba omsebenzi ngenkokhelo encintisanayo.
Ungaqhubeka nokuphendula izingxoxo manje njengoba usunokuqonda okuqinile kokuthi ungayiphendula kanjani eminye yemibuzo ebuzwa kabanzi yokufunda ngomshini.
Kuye ngezinjongo zakho, thatha isinyathelo esilandelayo. Lungiselela izingxoxo ngokuvakashela i-Hashdork's Uchungechunge Lwezingxoxo.
shiya impendulo