Ukuhlukaniswa kwama-Machine wokufunda ama-algorithms: Ukuqothuka komugqa, ukuhlukanisa kanye nokuhlanganiswa

Ukufundwa ngomshini kunokufana okuhle ngokulungela ukusebenza kwezibalo, okunikeza izindlela, imibono kanye nezizinda zezicelo. 

Ukufunda ngomshini kwakhiwa njenge "zinkinga zokunciphisa" zomsebenzi wokulahleka okuphikisana nesethi enikezwe yezibonelo (isethi yokuqeqesha). Lesi sici siveza ukungafani phakathi kwamanani abikezelwe imodeli eqeqeshiwe namanani alindelekile wesibonelo ngasinye. 

Umgomo wokugcina ukufundisa imodeli amandla okubikezela ngokunembile kusethi yezimo ezingekho kuqoqo lokuqeqeshwa.

Indlela okungenzeka ngayo ukuhlukanisa izigaba ezahlukene ze-algorithm wuhlobo lokukhiphayo olulindelekile ohlelweni oluthile lwe ukufunda imishini. 

Phakathi kwemikhakha ebalulekile esiyitholayo:

• La ukuhlukaniswa: okokufaka kuhlukaniswe ngezigaba ezimbili noma ngaphezulu kanti uhlelo lokufunda kumele lukhiqize imodeli ekwazi ukwaba isigaba esisodwa noma ngaphezulu kulokho okutholakala kokufaka.Lezi zinhlobo zemisebenzi ngokuvamile zibhekiswa kusetshenziswa amasu wokufunda ogadiwe. 

  Isibonelo sokwahlukaniswa kokwabiwa kwamalebula owodwa noma angaphezulu emfanekisweni osuselwa kwizinto noma izifundo eziqukethwe kuwo;

• La ukuzithoba: ngokomqondo kufana nokuhlukaniswa nomehluko wokuthi okuphumayo kunesizinda esiqhubekayo futhi esingahlukani.Imvamisa ilawulwa ngokufunda okugadiwe. 

  Isibonelo sokulawulwa kabusha ukulinganiselwa kokujula kwesimo kusuka kokumelwa ngendlela yesithombe sombala. 

  Eqinisweni, isizinda sokukhishwa okukhulunywa ngaso cishe asinamkhawulo, futhi asikhawulwanga kusethi ethile yamatshe;

• Il ukuhlanganisa: kuphi iqoqo ledatha lihlukaniswe ngamaqembu, kodwa, ngokungafani nesigaba, awaziwa nge-priori.Ubunjalo bezinkinga ezikulesi sigaba ngokujwayelekile bubenza babe yimisebenzi yokufunda engagadiwe.

Imodeli elula yokuhlehla yokulawula

Ukuqedwa komugqa yi-amkusetshenziswe imodeli esetshenziselwa ukulinganisa amanani wangempela anjengalawa:

• izindleko zezindlu,
• inombolo yezingcingo,
• ukuthengiswa okuphelele kumuntu ngamunye,

futhi ilandela umbandela wokuguquguquka okuqhubekayo:

• amamitha skwele,
• ukubhaliswa kwi-akhawunti yamanje,
• imfundo yomuntu

Ngokuhleleka okuqondile, ubudlelwano phakathi kokuhlukahluka okuzimele kanye nokuhlukahluka okuthembekile kulandelwa ngomugqa ovame ukumelela ubuhlobo phakathi kokuhluka okubili.

Umugqa ofanele ubizwa njengolayini we-regression futhi umelelwa yi-equation eqondile yohlobo Y = a * X + b.

Ifomula isuselwa ekuhlanganiseni idatha ukuhlanganisa izici ezimbili noma ngaphezulu nomunye. Lapho unikeza i-algorithm isimilo sokufaka, ukubekelwa phansi kubuyisela esinye isimilo.

Imodeli yokubuyiswa kwemigqa eminingi

Lapho sinokuhlukahluka okuzimele okukodwa, bese sikhuluma ngokubuyisa okuqondile ngokuqondile, sithatha imodeli efana nale elandelayo:

y=b₀ + b₁x₁ + b₂x₂ +… + B_nx_n

• y Impendulo yamanani, okusho ukuthi imelela umphumela obikezelwe imodeli;
• b₀ umuzwa, lelo inani lika-y lapho x_i bonke bayalingana no-0;
• isimilo sokuqala b₁ kungukukodwa kwe-x₁;
• esinye isici b_n kungukukodwa kwe-x_n;
• x₁,x₂,…, X_n kukhona okuhlukile okuzimele kwemodeli.

Ngokuyisisekelo i-equation ichaza ubudlelwano phakathi kokuhlukahluka okuqhubekayo okuthembekile (y) nokuhlukahluka okuzimele okubili (x1, x2, x3…). 

Isibonelo, uma sifuna ukulinganisa ukuphuma kwe-CO2 kwemoto (okuncike kokuguquguquka y) kucatshangelwa amandla enjini, inani lamasilinda kanye nokusebenzisa uphethiloli. Lezi zinto zokugcina zingukuhlukahluka okuzimele okungu-x1, x2 kanye no-x3. Ama-constants bi ziyizinombolo zangempela futhi abizwa ngokuthi ama-coefficients alinganiselwa emodeli .. Y ukuguquguquka okuqhubekayo okuthembekile, isb.ukuba isamba se-b0, b1 x1, b2 x2, njll. y kuzoba inombolo yangempela.

Ukuhlaziywa kokubuyiselwa okuningana okuyindlela esetshenziselwa ukukhomba umphumela okuhlukahluka okuzimele okunakho kokuhlukahluka okuthembekile.

Ukuqonda ukuthi ukuguquguquka okuthembekile okuguqukayo njengokuguquguquka kokuguquguquka okuzimele kusivumela kanjani ukubikezela imiphumela noma umthelela wezinguquko ezimeni zangempela.

Ngokusebenzisa ukubuyiselwa kwemigqa eminingana kungenzeka ukuthi uqonde ukuthi umfutho wegazi ushintsha kanjani njengoba inkomba yesisindo somzimba ishintsha ngokucabanga ngezinto ezinjengobudala, ucansi, njll.

Ngokubuyiselwa okuningana singathola izilinganiso zokuthambekela kwamanani, okufana nethambiso yakusasa yamafutha noma yegolide.

Okokugcina, ukubuyiswa kabusha kwemigqa eminingi kuthola isasasa elikhudlwana emkhakheni wokufunda ngomshini nobuhlakani bokufakelwa njengoba kuvumela ukuthola amamodeli wokufunda noma esimeni senani elikhulu lamarekhodi ahlaziywayo.

Isimodeli Sokucabanga Kwezinto Ezinhle

Ukuhlehliswa kokuhlelekile kuyithuluzi lezibalo elihlose ukulingisa umphumela we-binomial ngokuhluka okukodwa noma okuningi okuchazayo.

Ngokuvamile isetshenziselwa izinkinga kanambambili, lapho kunamakilasi amabili kuphela, isibonelo u-Yebo noma Cha, u-0 noma u-1, owesilisa noma owesimame njll.

Ngale ndlela kungenzeka ukuthi uchaze idatha futhi uchaze ubudlelwane phakathi kokuhlukahluka okuthembekile kanambambili nokuguquguquka okuhlukile noma okungokomqondo okuhlukile.

Umphumela unqunywa ngenxa yokusetshenziswa komsebenzi welogistic, olinganisela amathuba bese kamuva defiiphetha isigaba esiseduze kakhulu (esihle noma esibi) kunani lamathuba atholiwe.

Singabheka ukuhlelwa kabusha kwempahla njengendlela yokuhlukanisa umndeni we kugadwe ubuchule bokufunda.

Kusetshenziswa izindlela zezibalo, ukuhlehlisa kabusha amandla okuvumelekile ukukhipha umphumela, okuyiqiniso, okubonisa ukuthi kungenzeka ukuthi ivelu yenani elinikeziwe yesigaba esinikeziwe.

Ezinkinga ze-binomial logression regression, kungenzeka ukuthi okukhishwayo kungokwesigaba esisodwa kuzoba ngu-P, kuyilapho kungokwelinye ikilasi le-1-P (lapho u-P enombolo ephakathi kuka-0 no-1 ngoba kuveza amathuba).

Ukulungiswa kokubhalwa kokucabanga kwe-binomial kusebenza kahle kuzo zonke lezo zimo lapho ukuguquguquka esizama ukukubikezela kungokwamanani, okungukuthi, kungathatha amanani amabili kuphela: inani 1 elimele isigaba esihle, noma inani 0 elimele isigaba esibi.

Izibonelo zezinkinga ezingasombululwa ngokubhalwa phansi kwezinto:

• i-imeyili ingogaxekile noma cha;
• ukuthenga oku-inthanethi kuyinkohliso noma cha, kuhlolwe imibandela yokuthenga;
• isiguli sinokulimala, sihlaziya i-radii yaso.

Ngokucabanga okubonakalayo singenza ukuhlaziywa kokuqagela, ukulinganisa ubudlelwano phakathi kwalokho esifuna ukukuqagela (ukuxhomekeka okuthembekile) kanye noshintsho olulodwa noma ngaphezulu oluzimele, okungukuthi, izici. Ukulinganisa okungenzeka kwenziwa ngomsebenzi wokuhleleka.

Amaphuzu aguqulwa kamuva abe ngamanani kanambambili, futhi ukuze kwenziwe isibikezelo sibe ngokoqobo, lo mphumela unikezwa ikilasi okulo, kususelwa ekuthini siseduze nesigaba uqobo.

Isibonelo, uma ukusetshenziswa komsebenzi we-logistic kubuyisa u-0,85, kusho ukuthi okokufaka kukhiqize isigaba esihle ngokusinikeza ikilasi 1. Ngakolunye uhlangothi uma belitholile inani elifana no-0,4 noma ngaphezulu ngokuvamile <0,5 ..

Ukuhlehliswa kokuhleleka kwezinto kusebenza umsebenzi wokuhlola ukuhlola ukuhlukaniswa kwamanani wokufaka.

Umsebenzi wokuphawula, obizwa nangokuthi i-sigmoid, ijika elikwazi ukuthatha noma yiliphi inani lenani langempela futhi ulimise ngenani eliphakathi kuka-0 no-1, ngaphandle kokuqhela. Umsebenzi:

kuphi:

• e: isisekelo sama-logarithms wemvelo (Inombolo ka-Euler, noma i-Exel function exp ())
• b0 + b1 * x: inani langempela lezinombolo ofuna ukuliguqula.

Isethulo esisetshenziselwa ukucindezelwa kwezinto

Ukuhlehliswa kokucabanga kwezinto ezibonakalayo kusetshenziswa i-equation njengesethulo, kufana nokuqedwa komugqa

Amanani wokufaka (x) ahlanganiswe ngokuqondile kusetshenziswa izinsimbi noma amanani afanayo, ukubikezela inani lokuphumayo (y). Umehluko obalulekile kusuka ekuqedweni komugqa ukuthi inani lokulinganisa okukhiphile ivelu kanambambili (0 noma 1) kunani lezinombolo.

Nasi isibonelo se-equation regression equation:

y = e^(b0 + b1 * x) / (1 + e^(b0 + b1 * x))

Ijuba:

• y kungukuhluka okuthembekile, isib. inani elibikezelwe;
• i-b0 yigama lokugoqa;
• i-b1 iyisibambiso senani elilodwa lokufaka (x).

Ikholomu ngayinye kwidatha yokufaka ine-b ehambelana (inani elijwayelekile langempela) okufanele lifundwe kusuka kudatha yokuqeqesha.

Isifanekiselo sangempela semodeli ongayigcina kwimemori noma ifayela yizihlanganisi ezilinganisweni (inani le-beta noma le-b).

Ukuhlehlisa kokucabanga kubikezela amathuba (ububanzi bobuchwepheshe)

Izinhlobo zokucabanga ezinengqondo zinokwenzeka ngamathuba wesigaba esizenzakalelayo.

Njengesibonelo, ake sicabange ukuthi silinganisa ubulili babantu njengowesilisa noma ngowesifazane ukusuka ekuphakameni kwabo, isigaba sokuqala singaba ngowesilisa, futhi imodeli yokucabanga okuqondakalayo ingabhalwa njengokuthi kungenzeka ukuthi owesilisa anikezwe ukuphakama komuntu, noma ngaphezulu. ngokusemthethweni:

P (ucansi = owesilisa | ukuphakama)

Okubhalwe ngenye indlela, senza imodeli yamathuba okuthi okokufaka (X) kungokwangaphambili kwekilasidefinite (Y = 1), singabhala kanje:

P(X) = P(Y = 1 | X)

Ukuqagela okungenzeka kufanele kuguqulwe kumanani wenketho kanambambili (0 noma 1) ukuze empeleni enze ukubikezela okungenzeka.

Ukuhlehliswa kokuhleleka kuyindlela eqondile, kepha ukubikezela kuguqulwa kusetshenziswa umsebenzi wokulungiselela. Umthelela walokhu ukuthi asisakwazi ukuqonda ukubikezela njengokuhlanganiswa okuqondile kokufakwayo njengoba nje sikwazi ngokuhlehlisa okuqondile, ngokwesibonelo, ukuqhubeka kusuka phezulu, imodeli ingavezwa njenge:

p(X) = e ^ (b0 + b1 * X) / (1 + e ^ (b0 + b1 * X))

Manje sesingaphenduka i-equation kanjena. Ukuyihlehlisa singakwazi ukuqhubeka ngokususa i-e ngakolunye uhlangothi ngokungeza i-logarithm yemvelo ngakolunye uhlangothi.

ln (p (X) / 1 - p (X)) = b0 + b1 * X

Ngale ndlela sithola iqiniso lokuthi ukuhlanganiswa kokukhishwa ngakwesokudla kuqondile futhi (njenge-regarra regression), futhi okokufaka ngakwesobunxele kuyinkomba yokuthi kungenzeka kwesigaba esizenzakalelayo.

Amathuba abalwa njengesilinganiso sokuvama komcimbi esihlukaniswe ngamathuba okuthi kungabikho umcimbi, isb. 0,8 / (1-0,8) onomphumela ongu-4. Ngakho-ke singabhala:

ln (izingqinamba) = b0 + b1 * X

Njengoba amathuba eshintshwa nge-log, sibiza le-log-od engakwesokunxele noma i-probit.

Singakwazi ukubuyisa okhipha impahla ngakwesokunene futhi sikubhale njengoku:

amathuba = e ^ (b0 + b1 * X)

Konke lokhu kusisiza ukuba siqonde ukuthi ngempela imodeli iseyinhlanganisela yomugqa yokokufaka, kodwa ukuthi le nhlanganisela yomugqa ibhekisela emathubeni elogi ekilasi langaphambili.definita.

Kufundwa imodeli yokuhlehla yokulawula

Ama-coefficients (ama-beta noma ama-b amanani) we-algorithm yokulawulwa kokuqondisa acatshangelwa esigabeni sokufunda. Ukuze wenze lokhu, sisebenzisa isilinganiso esilinganiselayo sokulinganisa.

Isilinganiso esiphezulu sokungenzeka i-algorithm yokufunda esetshenziswa ama-algorithms wokufunda womshini ambalwa. Ama-coefficients aphuma kumodeli abikezela inani elisondele kakhulu ku-1 (isb. Owesilisa) wesigaba sangaphambilidefinite kanye nenani elisondele kakhulu ku-0 (isb. owesifazane) kwelinye ikilasi. Amathuba amaningi okuhlehla kwezinto inqubo yokuthola amanani ama-coefficients (i-Beta noma amanani we-ob) anciphisa iphutha emathubeni abikezelwe imodeli ngokuhlobene nalawo akudatha (isb., amathuba okuthi 1 uma idatha iyisigaba esiyinhloko) .

Sizosebenzisa i-algorithm yokunciphisa ukwengeza amanani amahle aphelele wemininingwane yokuqeqesha. Lokhu kuvame ukwenziwa ekusebenzeni kusetshenziswa i-algorithm esebenza kahle yokubala.

Ercole Palmeri

---