![Search icon](https://hackernoon.imgix.net/search-new.png?w=19&h=19)
Ukuqala iprojekthi entsha yokufunda koomatshini kuzisa ungxamiseko yomdla, kwaye kunokubahenda ukutsibela ngqo esiphelweni esinzulu. Zininzi iimodeli zamva nje zokusika okanye ii-algorithms ezinzima onokuthi ufunde ngazo. Bathembisa iziphumo ezisisiseko, kwaye ukuphepha isihendo sokuzama ngabo kanye nje elulwaneni ngumsebenzi onzima.
Nawuphi na usomashishini wale mihla uzimisele ukuvavanya iindlela zobugcisa zale mihla kwaye abonise iiprojekthi eziphucukileyo (kunye neziyimpumelelo) kuluntu. Nangona kunjalo, lo mdla, ngelixa ulungile, ngamanye amaxesha unokuthatha ixesha elibalulekileyo njengoko ulungisa kakuhle i-hyperparameters kwaye udibana nobunzima bokuphumeza iimodeli ezinzima.
Kule nkqubo, kukho umbuzo omnye ophambili ekufuneka ubuzwe: Ngaba ngokwenene silinganisa njani ukusebenza kwemodeli yethu?
Ukufumanisa ukuba ubunzima bemodeli yethu bufanelekile okanye ukuba ukusebenza kuphezulu ngokwenene kunokuba ngumngeni. Oku kwenzeka xa kungekho ndawo ilula yokubhekisela kuyo. Apha, ukuba nemodeli yesiseko kubaluleke kakhulu. Isiseko sinika loo ngongoma ibalulekileyo yereferensi - ithe ngqo, iyakha ngokukhawuleza, kwaye iyacaciswa ngokwemvelo. Okumangalisayo kukuba, kaninzi imodeli esisiseko, enokuthi ithathe kuphela i-10% yemizamo yophuhliso iyonke, inokufezekisa ukuya kuthi ga kwi-90% yentsebenzo efunekayo, ivelise indlela efanelekileyo kakhulu kwiziphumo ezifanelekileyo.
Uluvo lokuqalisa olulula ayisiyondlela nje elula yabaqalayo - luqheliselo olusisiseko oluhlala lufanelekile kuwo onke amanqanaba omsebenzi wesayensi yedatha. Sisixhobo esisisiseko kunye nesikhumbuzo esikhulu sokulungelelanisa umnqweno wethu wobunzima kunye nokusebenza kwezisombululo ezicacileyo, ekulula ukuziqonda, kunye nezilawulekayo.
Imodeli esisiseko yeyona nguqulelo isisiseko esetyenziselwa ukujongana nengxaki. Ngokuqhelekileyo, le mifuziselo ibandakanya ukuhlehliswa komgca kwiziphumo eziqhubekayo okanye ukuhlehliswa kwenkqubo yeziphumo zecandelo. Ngokomzekelo, ukuhlehliswa komgca kunokuqikelela imbuyekezo yesitokhwe ngokusekelwe kwidatha yexabiso lembali, ngelixa ukuhlehliswa kwezinto eziphathekayo kunokuhlela abafaki zicelo zekhredithi njengomngcipheko ophezulu okanye ophantsi.
Le ndlela yokusebenzela yohlukile kwiimodeli ezintsonkothileyo ezinje ngothungelwano lwe-neural okanye iindlela ezidityanisiweyo, ezithi, ngelixa zinamandla, zinokwenza ukubamba ingxaki kube nzima ngakumbi kwaye konyuse ixesha elifunekayo kuphuhliso ngenxa yokuntsokotha kwazo kunye nezixhobo ezibalulekileyo zokubala.
Ukulinganisa linyathelo lokuqala elibaluleke kakhulu ekuphuhlisweni kwayo nayiphi na imodeli yeML. Xa useta imodeli esisiseko, useka i-metric yokusebenza esisiseko ekufuneka yonke imifuziselo ezayo emva kwayo (ehlala intsonkothe ngakumbi) kufuneka igqithe ukuthethelela ukuntsonkotha kwazo kunye nokusetyenziswa kobutyebi. Le nkqubo ayisositsheki nje sokuziphatha kakuhle kodwa ikwamisela izinto ozilindeleyo kwaye ikunika umlinganiselo ocacileyo wenkqubela phambili.
Umzekelo, khawufane ucinge ukuphuhlisa imodeli yokuqikelela iintsingiselo zemarike yezemali usebenzisa i-avareji elula ehambayo (SMA) njengesiseko. Le SMA inokusebenzisa idatha yembali yexesha elifutshane ukuqikelela amaxabiso esitokhwe kwixesha elizayo, ukufezekisa ukuchaneka kokuqala kwe-60% ekubikezelweni kweentshukumo zemarike ngokuchanekileyo. Le modeli emva koko ibeka ibhenchmark yazo naziphi na iimodeli eziphambili ezilandelayo. Ukuba imodeli entsonkothileyo, efana nenethiwekhi yeMemori yeXesha eliFutshane (LSTM), iphuhliswa kamva kwaye izuze ukuchaneka kwama-65%, ukonyuswa komsebenzi kunokulinganiswa ngokuchanekileyo ngokuchasene nesiseko sokuqala sama-60%.
Olu thelekiso lubalulekile ekumiseleni ukuba ngaba ukuphucuka kwe-5% kokuchaneka kuthethelela ukuntsonkotha okongezelelweyo kunye neemfuno zokubala ze-LSTM. Ngaphandle kwesiseko esifana nesi, ukwenza izigqibo ezizizo malunga nobukhulu kunye nokusetyenziswa okusebenzayo kweemodeli ezinzima ngakumbi kuba ngumngeni.
Le ndlela yokulinganisa iqinisekisa ukuba ukuphuculwa kobunzima bemodeli bufanelekile kwaye kuya kubangela uphuculo lokwenyani, lonke eli xesha lwenza inkqubo yophuhliso ihambelane neziphumo ezisebenzayo.
Ukulandela indlela yokunciphisa iindleko kwi-ML kubalulekile. Ngokukodwa xa ubeka kwinjongo yokulungelelanisa iinkqubo zakho kunye nemigaqo ebeka ixabiso eliphezulu ngelixa unciphisa inkunkuma njengento ephambili. Xa uqala ngemodeli esisiseko, unciphisa izixhobo kunye nexesha elifunekayo kuphuhliso lwemodeli yokuqala kunye novavanyo. Oku kuthetha ukuchwetheza okukhawulezileyo - kwaye oko kubalulekile kwingxelo ekhawulezileyo kunye nophuculo oluphindaphindayo.
Ngolu sisiseko, nakuphi na ukuntsokotha owongezayo ngoku kungaphononongwa ngononophelo.
Umzekelo, ukuba ufuna ukwenza utshintsho kwi-algorithm entsonkothileyo njenge vector autoregression (VAR) kwaye ufumanise ukuba yonyusa kancinci ukuchaneka koqikelelo lwengqikelelo, kufuneka ucinge kwakhona ukuba olu phuculo luncinci luthethelela iimfuno ezongezelelweyo zokubala kunye nobunzima. Impendulo isenokuba nguhayi. Emva koko imodeli elula ihlala iyona ndlela ingabizi kakhulu.
Ngokugxila ekusebenzeni kakuhle kweendleko, uqinisekisa ukuba izixhobo zisetyenziswa ngokufanelekileyo kwaye uphumeza ngaphezu kophuculo lobugcisa. Kwakhona, inika izisombululo ezisebenzayo, ezongezelelekileyo ezithethelelayo malunga nokuphuculwa komsebenzi kunye nokwabiwa kwezibonelelo. Ngale ndlela, utyalo-mali ngalunye kwimodeli entsonkothileyo luqinisekisiwe, olunegalelo kwiinjongo zeprojekthi ngokubanzi ngaphandle kweendleko ezingaphandle komlinganiselo.
Kumacandelo afana nezezimali apho izigqibo kufuneka zihambelane nemigangatho engqongqo yolawulo, ukungafihli kweemodeli ayisiyonzuzo nje yoshishino. Yindlela yobuchule enceda kakhulu kwinkqubo yokuhlangana kwemimiselo kwaye iququzelele unxibelelwano olulula kunye nabachaphazelekayo abangenayo imvelaphi yobugcisa (enzulu).
Masithathe imodeli yethu ye-SMA. Itolika ngokulula kuba iziphumo zayo zihambelana ngokuthe ngqo kwidatha yegalelo. Oku kwenza kube lula ukuchaza ukuba igalelo ngalinye linefuthe njani kwisiphumo esiqikelelweyo. Xa izigqibo ezisekelwe kuqikelelo lwemodeli kufuneka zichaneke kubalawuli bangaphandle okanye ngaphakathi kumalungu eqela elingengobuchwepheshe, oku kulula kungundoqo kwiinkqubo zakho.
Ukuba isigqibo esisekelwe kuqikelelo lwemodeli ye-SMA siyabuzwa, ukucaca komzekelo kuvumela inkcazo ekhawulezayo nelula yengqiqo emva komsebenzi wayo. Oku kunokunceda ngophononongo lolawulo kunye nophicotho kunye nokuphucula ukuthembana kunye nokwamkelwa phakathi kwabasebenzisi kunye nabenzi bezigqibo. Ngaphaya koko, njengoko imodeli entsokothileyo isanda, umzekelo ukuya kwii-algorithms ezintsonkothileyo njenge-ARIMA okanye imodeli ye-VAR yoqikelelo oluninzi, ukutolika kwesiseko se-SMA sokuqala kuba luphawu lokuba leliphi inqanaba lengcaciso ekufuneka uyiveze.
Ngokusebenzisa iiregressors ezifana neempawu zokubaluleka kwamanqaku okanye amaxabiso e-SHAP adityaniswe neemodeli ezintsonkothileyo, inkqubela phambili yayo nayiphi na imodeli eyongezelelweyo yokusebenza ihlala iselubala. Oku kunceda injongo yenkqubo yokhuseleko ukuba ingalahlwa kwiimodeli eziphambili. Ingongoma yomzekelo wesiseko esilula kukuhlala kuzalisekisa imeko yokuba isakhiwo esipheleleyo kunye nokubaluleka kuya kugcinwa njengoko inqanaba lobunzima linyuka. Oku kuqinisekisa izibonelelo zokuthotyelwa kunye nonxibelelwano oluya kusebenza ngempumelelo.
Ulawulo lomngcipheko yenye inkalo ebalulekileyo yokuphuhlisa iimodeli zokufunda koomatshini, ngakumbi kumacandelo afana nezezimali apho uqikelelo oluchanekileyo noluthembekileyo lunempembelelo ekwenzeni izigqibo. Ukuba nomzekelo olula wesiseko sisicwangciso esihle sokulawula le mingcipheko.
Isiseko esithe ngqo sibonelela ngesiqalo esiqondakalayo, esikuvumela ukuba ngokuthe ngcembe (kwaye ngokukhuselekileyo) wongeze izixhasi kwimodeli enzima.
Ngokomzekelo, imodeli ye-SMA (ngelixa isisiseko) yenza isiseko esiluqilima sokufumana iipateni ezisisiseko kunye nezinto ezinokuthi zibe khona kwiintshukumo zexabiso lesitokhwe. Ukuyisebenzisa kunceda ukuchonga iimpawu zokuqala zokuguquguquka okanye ukuziphatha okungaqhelekanga kwimarike. Ukwenza oko kubalulekile, ukuphepha imingcipheko ebalulekileyo yezemali ngaphambi kokufaka i-algorithms enzima yokuqikelela.
Ngaphezu koko, ukusebenzisa imodeli yesiseko kunciphisa umngcipheko wokugqithisa. Ngumgibe oqhelekileyo kumzekelo wemali. Ukufakwa ngokugqithisileyo kwenzeka xa imodeli ilungelelaniswe kakuhle kakhulu kwidatha yembali kwaye ibamba ingxolo kunendlela esisiseko. Ngenxa yoku, unokufumana iingqikelelo ezilahlekisayo kwaye ufumane izicwangciso zokurhweba ezingathembekanga ngenxa yoko. Imodeli elula kunye neeparamitha ezimbalwa ayifuni kakhulu kulo mba, iqinisekisa ukuba uqikelelo olunikezelayo lusebenza ngokubanzi kwidatha engabonakaliyo.
Ukwandisa ubunzima njengoko i-SMA iqhubela phambili kwimodeli encinci ehambayo efana ne-ARIMA kunye ne-VAR ibe nzima ngakumbi, ulwakhiwo olulula lwe-SMA lunokusinceda siqwalasele ngokucwangcisekileyo ukusebenza kobunzima obudityanisiweyo. Olu phuculo oluhamba ngenyathelo lokuntsonkotha lunceda ukugcina ulawulo ekusebenzeni komzekelo, uqinisekisa ukuba umaleko ngamnye owongezelelweyo onzima ubonelela ngenzuzo ecacileyo kwaye awuzisi umngcipheko ongafunekiyo.
Le ndlela ecwangcisiweyo yokunyuka kobunzima bemodeli inceda ekuqondeni ukuba utshintsho kwimodeli luchaphazela njani indlela yokuziphatha kunye nokuthembeka kwayo. Ikwaqinisekisa ukuba imingcipheko isoloko ilawulwa kakuhle. Xa uqala ngesiseko esilula kwaye ulawule ngokucophelela isigaba ngasinye sophuhliso, uqinisekisa ukuba iimodeli zokubikezela zihlala zinamandla kwaye zikhuselekile, zixhasa ukwenza izigqibo zemali.
Ukukhetha imodeli yesiseko efanelekileyo kakhulu, kufuneka uqonde ingxaki yeshishini kunye neempawu zedatha. Umzekelo, uqikelelo lwexesha leemarike zezemali lunokuqala ngemodeli ye-ARIMA njengesiseko sokubamba amandla exeshana ngendlela elula. Umgangatho wedatha kunye nokulungiswa kwangaphambili kukwadlala indima ebalulekileyo; neyona modeli ilula inokuqhuba kakubi ukuba yondliwe ngokunganelanga okanye idata elungiswe kakubi kwangaphambili.
Kwaye okokugqibela, ukwazi ixesha lokutshintsha ukusuka kwisiseko ukuya kwimodeli enzima ngakumbi kubalulekile. Esi sigqibo kufuneka sikhokelwe ngokunyuka kovavanyo kunye nokuqinisekiswa, ngokuhambelana nendlela ye-Agile yokuphindaphinda.
Ukuqala iiprojekthi zakho zokufunda ngomatshini ngokwazisa imodeli yesiseko elula ayilonyathelo lokuqala. Licebo. Isicwangciso esihambelana neendlela ze-Agile ezikhuthaza ukusebenza kakuhle, ukusebenza, kunye nokulungelelanisa. Ukusondela kwiprojekthi yakho ngale ndlela kunokuphucula kakhulu iziphumo zeprojekthi ngokuqinisekisa ukuba konke ukwanda kobunzima kuyathetheleleka kwaye kongeza ixabiso elibambekayo. Ukwamkela ubulula yinto enamandla. Sisicwangciso esikhulu ngakumbi kwiinkalo ezifana nezemali apho izigqibo kufuneka zikhawuleze.