Ukuhlelwa kovavanyo lweealgorithms zoMatshini: Ukulungiswa komgca, ukuHlela kunye nokuHlanganiswa

UkuFunda ngomatshini kunezixhobo ezifanayo kunye nokusebenza kakuhle kwezibalo, ezibonelela ngeendlela, iithiyori kunye neendawo zokusetyenziswa. 

Isifundo somatshini senziwa "njengeengxaki zokunciphisa" zomsebenzi wokulahleka ngokuchasene neseti enikezelweyo yemizekelo (iseti yoqeqesho). Eli nqaku libonisa ukungangqinelani phakathi kwamaxabiso axelwe kwangaphambili kwimodeli yoqeqesho kunye namaxabiso alindelweyo kumzekelo ngamnye. 

Eyona njongo iphambili kukufundisa imodeli isakhono sokuqagela ngokuchanekileyo kwiseti yamaxesha ongekho kwiseti yoqeqesho.

Indlela ngokwendlela enokwenzeka ngayo ukwahlula iindidi ezahlukeneyo ze-algorithm luhlobo lwemveliso elindelekileyo kwinkqubo ethile yokufunda umatshini. 

Phakathi kwamacandelo aphambili esiwafumanayo:

• La ukwahlulahlula: igalelo lahlulahlulwe laba zimbini okanye ngaphezulu kwaye inkqubo yokufunda kufuneka ivelise imodeli enokwabela iklasi enye nangaphezulu kwezo zifumaneka kwigalelo.Ezi ntlobo zemisebenzi zihlala ziqwalaselwa kusetyenziswa iindlela zokufunda ezigadiweyo. 

  Umzekelo wokwahlulahlula-hlula kukwabiwa kweelebheli enye okanye ezingaphezulu kumfanekiso osekwe kwizinto okanye izifundo eziqulathwe kuyo;

• La regression: conceptually iyafana nokuhlelwa ngumahluko okhupha iziphumo zine-domain eqhubekayo nengeyo-discrete.Ihlala ilawulwa ngokufunda esweni. 

  Umzekelo wokuhlengahlengiswa kukuqikelelwa kobunzulu bendawo evela kumelo lwayo ngendlela yomfanekiso wombala. 

  Ngapha koko, i-domain yokukhutshwa kumbuzo phantse ayinasiphelo, kwaye ayikhawulelwanga kuseti oluthile lwezinto ezinokubakho;

• Il ukuqoqa: iphi iseti yedatha yahlulahlulwe yangamaqela, nangona kunjalo, ngokungafaniyo nokubekwa kodidi, awaziwa eyongori.Ubunjani beengxaki zeli nqanaba zidla ngokubenza babe yimisebenzi yokufunda engaqwalaselwanga.

Imodeli elula yokulandelanisa

Ukulungelelaniswa komgca ngu-amImodeli esetyenzisiweyo ngokubanzi esetyenziselwa ukuqikelela amaxabiso okwenyani anje ngala:

• iindleko zezindlu,
• Inani leminxeba,
• Intengiso iyonke

kwaye alandele isilinganiselo sokuguquguquka okuqhubekayo:

• iimitha ezisikwere,
• ubhaliso kwiakhawunti yangoku,
• imfundo yomntu

Kumgca wokuhlengahlengiswa komgca, ubudlelwane phakathi kokuhluka okuzimeleyo kunye nokuxhomekeka kokuxhomekeka kulandelwa ngomgca ovame ukubonisa ubudlelwane phakathi kwezi zinto zimbini eziguqukayo.

Umgca ofanelekileyo waziwa ngokuba ngumgca we-regression kwaye umelwe ngumlinganiso wolungelelaniso lohlobo lwe-Y = a * X + b.

Ifomula isekwe kwidatha yokudibanisa ukudibanisa iimpawu ezimbini okanye ngaphezulu kunye. Xa unikezela i-algorithm ngesimo sokungenela, ukuphindisela kubuyisela olunye uphawu.

Imodeli ephindaphindiweyo yokulungelelanisa imodeli

Xa sinezinto ezahlukeneyo ezizimeleyo ezihlukileyo, emva koko sithetha ngohlobo oluhlengahlengiso lomgca, sithatha imodeli efana nale ilandelayo:

y=b₀ + b₁x₁ + b₂x₂ +… + B_nx_n

• y Impendulo kumaxabiso, oko kukuthi, imele isiphumo esixelwe kwangaphambili yimodeli;
• b₀ luphawu lokungena, lixabiso lika y xa x_i Zonke zilingana no-0;
• uphawu lokuqala b₁ kukwenzeka kwe-x₁;
• elinye uphawu b_n kukwenzeka kwe-x_n;
• x₁,x₂,…, X_n ziguquguqukayo ezizimeleyo zemodeli.

Ngokusisiseko i-equation icacisa unxibelelwano phakathi kokuguquguquka okuxhomekekileyo okuqhubekayo (y) kunye nezinto ezimbini okanye ngaphezulu ezizimeleyo (x1, x2, x3…). 

Umzekelo, ukuba sifuna ukuqikelela ukuphuma kwe-CO2 yemoto (exhomekeke ngokuguquguqukayo y) kuqwalaselwa amandla enjini, inani leesilinda kunye nokusebenzisa ipetroli. Ezi zinto zokugqibela ziinguqu ezizimeleyo x1, x2 kunye x3. I-constants bi ngamanani okwenyani kwaye abizwa ngokuba yimodeli yoqikelelo ye-coression yokuqikelelwa .. Y sisimo esiguquguqukayo esixhomekekileyo, okt sisixa esisisi-b0, b1 x1, b2 x2, njl. y iyakuba linani lokwenyani.

Uhlalutyo oluninzi lokubuyiswa kwakhona yindlela esetyenzisiweyo yokuchonga isiphumo esinokuxhomekeka esizimeleyo kwinguqu exhomekekileyo.

Ukuqonda indlela iinguqu ezixhomekekileyo eziguqukayo ngayo njengotshintsho oluzimeleyo eziguquguqukayo zisivumela ukuba siqikelele iziphumo okanye iimpembelelo zotshintsho kwiimeko zokwenyani.

Ngokusebenzisa ulungelelwaniso oluninzi lomgca kunokwenzeka ukuba uqonde indlela olutshintsha ngayo uxinzelelo lwegazi njengoko isindisi somzimba siguqula ukuqaphela izinto ezifana nobudala, isini, njl.

Ngokuhlengahlengiswa okuphindaphindiweyo sinokufumana uqikelelo kwimilinganiselo yamaxabiso, ezinje ngemeko yexesha elizayo yeoyile okanye igolide.

Okokugqibela, ukuhlengahlengisa okumgama ngokulandelelana ukufumana umdla omkhulu kwicandelo lokufunda kumatshini kunye nobukrelekrele bokuvumisa njengoko kuvumela ukufumana iimodeli zokufunda zisebenza nkqu nakwimeko yenani elikhulu leerekhodi eziya kuhlaziywa.

Imodeli yokuLawulwa koLwazi

Ukubuyiswa kwakhona kwezinto ezinesixhobo sisixhobo seenkcukacha-manani esijolise kwimodeli yesiphumo se-binomial ngolunye okanye ngaphezulu izinto ezichazayo.

Ihlala isetyenziselwa iingxaki ze-binary, apho kukho kuphela iiklasi ezimbini, umzekelo u-Ewe okanye u-Hayi, u-0 okanye u-1, oyindoda okanye umntu obhinqileyo njl ...

Ngale ndlela kunokwenzeka ukuba uchaze idatha kwaye uchaze ubudlelwane phakathi kokuxhomekeka okungaxhomekekanga okukhethekileyo kokunye kunye nangaphezulu okanye okungamagama azimeleyo.

Isiphumo sizimisele ngokubulela ukusetyenziswa komsebenzi welogistic, oqikelela ukuba kunokwenzeka kwaye emva koko defiiphelisa olona didi lukufutshane (elungileyo okanye embi) kwixabiso elinokwenzeka elifunyenweyo.

Singayijonga into yokuhlengahlengiswa kwezinto njengendlela yokuhlela usapho ukubeka iliso kwii-algorithms zokufunda.

Sebenzisa iindlela zobalo, ukulawulwa kokubonelelwa kwempahla kuvumela ukuvelisa iziphumo, enyanisweni, zimele ukuba kungenzeka ukuba ixabiso elinikiweyo lokufaka linelodidi olunikiweyo.

Kwiingxaki zokubhalwa kwengxelo ye-binomial logical, amathuba okuba iziphumo zeklasi enye iya kuba ngu-P, ngelixa lelinye iklasi 1-P (apho u-P ephakathi kwenani le-0 ne-1 kuba echaza amathuba).

Ukulawulwa kokunikezelwa kwengxelo ye-binomial kusebenza kakuhle kuwo onke la matyala apho isiguquguqukayo esizama ukusiqikelela kuyinto engamabhanti, oko kukuthi, inokuthatha kuphela amaxabiso amabini: ixabiso 1 elimele udidi olungileyo, okanye ixabiso elibonisa udidi olubi.

Imizekelo yeengxaki ezinokusonjululwa ngohlobo lokurekhodwa kwezinto:

• i-imeyile ayixekile okanye hayi;
• ukuthenga kwi-Intanethi kunobuqhophololo okanye akunjalo, kuvavanywa iimeko zokuthenga;
• Isigulana sinokuqhekeka, kuvavanya radii yayo.

Ngokuhlengahlengiswa kwezinto, sinokwenza uhlalutyo oluqikelelweyo, ukulinganisa ubudlelwane phakathi kwento esifuna ukuyiqikelela (ukuxhomekeka ngokuxhomekekileyo) kunye nokuhluka okuzimeleyo okukodwa, i.e. Uqikelelo olunokwenzeka lwenziwa ngomsebenzi wokuqhutywa kwezinto.

Iimpawu ezinokuthi zitshintshwe kamva zibe ziinqobo zokugqibela, kwaye ukuze ubenakho ukwenza uqikelelo lube lolokwenyani, esi siphumo sabelwe iklasi esikuyo, ngokubhekisele ekubeni kufutshane okanye iklasi ngokwaso.

Umzekelo, ukuba usetyenziso lomsebenzi wokungena lubuyisa 0,85, oko kuthetha ukuba igalelo lenze iklasi elungileyo ngokuyinika iklasi 1. Ngokuchaseneyo ukuba ifumene ixabiso elinjenge-0,4 okanye ngaphezulu ngokubanzi <0,5 ..

Ukuhlengahlengiswa komzimba kusetyenziswa umsebenzi wokuvavanya ukuhlelwa kwexabiso lamagalelo.

Umsebenzi wokudibanisa, okwabizwa ngokuba yi-sigmoid, yjikajika ekwazi ukuthatha naliphi na inani lexabiso lokwenyani kwaye ulimise kwixabiso eliphakathi ko-0 no-1, ngaphandle kokugqithisa. Umsebenzi yile:

iphi:

• e: isiseko se-logarithms yendalo (inombolo ka-Euler, okanye i-Exel function exp ())
• b0 + b1 * x: lixabiso lelona nani lenani ofuna ukuliguqula.

Ummiselo osetyenziselwe ukuhlengahlengiswa kwezinto

Ukuhlengahlengiswa komzimba kusetyenziswa i-equation njengommeli, njengokucekeceka komgca

Amaxabiso okufaka (x) adityaniswa ngokulandelelana kusetyenziswa iintsimbi okanye amaxabiso afanelekileyo, ukuqikelela ixabiso lokukhupha (y). Umahluko ophambili ukusuka kulungelelwaniso lomgca kukuba ixabiso lokukhuphela limodareyitha (0 okanye 1) kunexabiso leenombolo.

Nanku umzekelo wokulinganiswa kokuhlelwa kwengxelo

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

Ijuba:

• y yinguqu exhomekekileyo, i.e. ixabiso eliqikelelweyo;
• b0 luphawu lokwenza izinto okanye inqanaba lokuthintela;
• b1 kukwenziwa kwexabiso lexabiso elinye lokufaka (x).

Ikholamu nganye kwidatha yokufaka ine-b ehambelana (ixabiso elihlala lihleli) ekufuneka ifundwe kwidatha yoqeqesho.

Eyona nto iboniswayo imodeli oya kuyigcina kwimemori okanye kwifayile yimiba yolungelelwaniso (beta okanye ixabiso le-b).

Ukuhlengahlengiswa kwengqondo kuqikelela okunokwenzeka (Uluhlu lobugcisa)

Imodeli yokuhlengahlengiswa kwezinto ezinokwenzeka

Njengomzekelo, masicinge ukuba sibonisa ukuba isini sabantu njengamadoda okanye amabhinqa ukusuka ekuphakameni kwabo, iklasi yokuqala inokuba yeyamadoda, kwaye imodeli yokubuyela emxholweni inokubhalwa njengonokwenzeka wokuba indoda inikwe ubude bomntu, okanye nangaphezulu. ngokusesikweni:

P (ngokwesondo = yindoda | ukuphakama)

Ibhalwe ngenye indlela, senza umzekelo wokuba igalelo (X) leleklasi yangaphambilidefinite (Y = 1), sinokuyibhala ngolu hlobo:

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

Uqikelelo olunokwenzeka kufuneka luguqulwe lube ngamaxabiso asisiseko (0 okanye 1) ukwenzela ukwenza uqikelelo olunokwenzeka.

Ukuhlengahlengiswa kwezinto ezinokusetyenziswa njengendlela ehambelanayo, kodwa uqikelelo luguqulwa kusetyenziswa umsebenzi wokwenza izinto. Impembelelo yoku kukuba asisakwazi ukuqonda uqikelelo njengenxalenye yomgca wokudibanisa kwamagalelo njengoko sinakho ngesiphelo somgca, umzekelo, ukuqhubeka ukusuka phezulu, imodeli inokubonakaliswa njenge:

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

Ngoku singakwazi ukuyibuyisela kwakhona le nxaki ngolu hlobo lulandelayo. Ukubuyisa umva sinokuqhubeka ngokususa i-e kwelinye icala ngokongeza i-logarithm yendalo kwelinye icala.

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

Ngale ndlela sifumana isibakala sokuba udibaniso lwezinto ezikhutshelwe ekunene lungumtya kwakhona (njengodweliso lomgca), kwaye igalelo elisekhohlo liyincwadi yegama elinokwenzeka lodidi olungagqibekanga.

Amathuba abalwa njengomlinganiso wamandla wesiganeko ohlulelwe kukungabikho komsitho, o.k.t. 0,8 / (1-0,8) isiphumo sayo ngu-4. Singabhala ngoku:

ln (iingxaki) = b0 + b1 * X

Ukusukela ukuba utshintsho luguqulwe, sibiza le-od-esecaleni yelogo-engathandekiyo.

Singawubuyisela umthengisi ekunene kwaye usibhale ngolu hlobo:

amathuba = e ^ (b0 + b1 * X)

Konke oku kusinceda ukuba siqonde ukuba ngenene imodeli iseyindibaniselwano yomgama wamagalelo, kodwa ukuba olu didi lwelayini lubhekisa kwilog enokwenzeka yodidi lwangaphambili.definita.

Kufundwa imodeli yokuhlengahlengiswa kwempahla

I-coefficients (i-beta okanye i-b yexabiso) yelogression regression algorithm iqikelelwa kwinqanaba lokufunda. Ukwenza oku, sisebenzisa uqikelelo olunokwenzeka.

Ubuninzi boqikelelo lokunokwenzeka yi-algorithm yokufunda esetyenziswa ngoomatshini abaninzi be-algorithms yokufunda. Ii-coefficients eziphuma kwimodeli ziqikelela ixabiso elikufutshane kakhulu ku-1 (umzekelo, indoda) kwiklasi yaphambi kokuya esikolweni.definite kunye nexabiso elikufutshane kakhulu ku-0 (umz. female) kwenye iklasi. Obona buninzi bamathuba obuyiselo lolungiselelo yinkqubo yokufumana amaxabiso ee-coefficients (Beta okanye amaxabiso e-ob) acutha impazamo kwizinto ezinokwenzeka eziqikelelweyo ngumfuziselo ngokunxulumene nezo zikwidatha (umzekelo, amathuba okuba 1 ukuba idatha yiklasi yokuqala) .

Siza kusebenzisa i-algorithm yokunciphisa ukunciphisa amanqanaba afanelekileyo edatha yoqeqesho. Oku kuhlala kuphunyezwa kuqheliselo kusetyenziswa i-algorithm efanelekileyo yokusebenza kwamanani.

Ercole Palmeri

---