Hei Habr! Isu tinokoka maInjiniya eData uye nyanzvi dzeKudzidza kwemuchina kune yemahara Demo chidzidzo . Isu tinoburitsawo chinyorwa Luca Monno - Musoro weFinancial Analytics kuCDP SpA.
Imwe yeanonyanya kubatsira uye ari nyore muchina kudzidza nzira ndeye Ensemble Kudzidza. Ensemble Kudzidza ndiyo nzira iri kuseri kweXGBoost, Bagging, Random Sango uye mamwe akawanda algorithms.
Kune zvakawanda zvezvinyorwa zvakanaka paKuenda kuData Science, asi ndakasarudza nyaya mbiri ( ΠΈ ) chandakanyanya kufarira. Saka sei kunyora imwe nyaya nezve EL? Nokuti ndinoda kukuratidza kuti inoshanda sei nemuenzaniso wakajeka, izvo zvakaita kuti ndinzwisise kuti hapana mashiripiti apa.
Pandakatanga kuona EL achiita (kushanda nemamwe maitiro akareruka ekuderedza) handina kukwanisa kutenda maziso angu, uye ndichiri kuyeuka purofesa akandidzidzisa nzira iyi.
Ini ndaive nemhando mbiri dzakasiyana (mbiri dzisina kusimba dzekudzidzira algorithms) nemametrics kunze-kwe-sample RΒ² yakaenzana ne0,90 uye 0,93, zvichiteerana. Ndisati ndatarisa mhedzisiro, ndaifunga kuti ndaizowana RΒ² kumwe kunhu pakati peaya maviri ekutanga maitiro. Mune mamwe mazwi, ndaitenda kuti EL inogona kushandiswa kuita kuti modhi isaite zvisina kunaka seyakaipisisa modhi, asi kwete pamwe neiyo yakanakisa modhi inogona kuita.
Kukushamisika kwangu kukuru, kungoita avhareji yekufungidzira kwakaburitsa RΒ² ye0,95.
Pakutanga ndakatanga kutsvaga kukanganisa, asi ndakazofunga kuti panogona kunge paine mashiripiti akavanda pano!
Chii chinonzi Ensemble Kudzidza
NeEL, unogona kusanganisa fungidziro yemhando mbiri kana kupfuura kuti ibudise yakasimba uye inoita modhi. Kune dzakawanda nzira dzekushanda nemuenzaniso ensembles. Pano ini ndichabata pane maviri anonyanya kubatsira kuti ndipe muchidimbu.
Nekubatsirwa kwe kuderera zvinokwanisika kuenzana nekuita kwemhando dziripo.
Nekubatsirwa kwe classification Iwe unogona kupa mienzaniso mukana wekusarudza mavara. Chiratidzo chakasarudzwa kazhinji ndicho chichasarudzwa nemuenzaniso mutsva.
Nei EL inoshanda zvirinani
Chikonzero chikuru chinoita kuti EL iite zvirinani ndechekuti kufanotaura kwega kwega kune chikanganiso (tinoziva izvi kubva pane zvingangoita dzidziso), kubatanidza kufanotaura kuviri kunogona kubatsira kuderedza kukanganisa, uye nekudaro kunatsiridza maitiro ekuita (RMSE, RΒ², nezvimwewo). d.).
Dhiagiramu inotevera inoratidza kuti maviri asina kusimba algorithms anoshanda sei pane data set. Yekutanga algorithm ine yakakura kutsetseka kupfuura inodiwa, nepo yechipiri iine inenge zero (zvichida nekuda kwekuwandisa-regularization). Asi Ensemble inoratidza zvibereko zviri nani zvikuru.
Kana iwe ukatarisa iyo RΒ² chiratidzo, saka yekutanga neyechipiri yekudzidzira algorithm ichave yakaenzana ne -0.01ΒΉ, 0.22, zvichiteerana, nepo ensemble ichave yakaenzana ne0.73.
Pane zvikonzero zvakawanda nei algorithm ichigona kuve yakashata modhi kunyangwe pamuenzaniso wekutanga wakadai seizvi: pamwe iwe wakafunga kushandisa nguva dzose kudzivirira kuwandisa, kana kuti wafunga kusatonga zvimwe zvinokanganisa, kana pamwe wakashandisa polynomial regression uye wakanganisa. dhigirii (somuenzaniso, takashandisa polynomial yedhigirii rechipiri, uye bvunzo data inoratidza asymmetry yakajeka iyo dhigirii rechitatu raizokodzera zviri nani).
Apo EL inoshanda zviri nani
Ngatitarisei maalgorithms maviri ekudzidza anoshanda nedata rimwechete.
Pano iwe unogona kuona kuti kusanganisa iwo maviri mamodheru hakuna kunatsiridza kuita zvakanyanya. Pakutanga, kune maviri ekudzidziswa algorithms, iyo RΒ² zviratidzo zvakaenzana ne -0,37 uye 0,22, zvichiteerana, uye kune ensemble yakazova -0,04. Ndiko kuti, iyo EL modhi yakagamuchira avhareji kukosha kwezviratidzo.
Zvisinei, pane musiyano mukuru pakati pemienzaniso miviri iyi: mumuenzaniso wekutanga, zvikanganiso zvemuenzaniso zvakanga zvisina kunaka, uye mune yechipiri, zvakabatanidzwa zvakanaka (makoefficients ematatu matatu haana kufungidzirwa, asi aingosarudzwa ne munyori semuenzaniso.)
Naizvozvo, Ensemble Kudzidza inogona kushandiswa kuvandudza bias / musiyano chiyero mune chero mamiriro, asi riini Zvikanganiso zvemuenzaniso hazvina kurongeka zvakanaka, kushandisa EL kunogona kutungamirira kukuvandudza kushanda.
Homogeneous uye heterogeneous modhi
Kazhinji EL inoshandiswa pane homogeneous modhi (semumuenzaniso uyu kana sango risingaverengeki), asi kutaura zvazviri unogona kusanganisa mhando dzakasiyana (mutsara regression + neural network + XGBoost) nemaseti akasiyana eanotsanangura akasiyana. Izvi zvingangoguma nekukanganisa kusingawirirani uye nekuvandudzika kwekuita.
Kuenzanisa neportfolio diversification
EL inoshanda zvakafanana kune kusiyanisa mune portfolio theory, asi zvakanyanya zvirinani kwatiri.
Paunenge uchisiyanisa, unoedza kudzikisa kusiyana kwekuita kwako nekudyara mumasheya asina kubatanidzwa. Iyo yakanaka-yakasiyana-siyana portfolio yemasheya ichaita zvirinani pane yakaipisisa yega stock, asi isingambove nani pane yakanyanya kunaka.
Kutaura nezveWarren Buffett:
"Kusiyana-siyana kuzvidzivirira pakusaziva; kune mumwe munhu asingazivi zvaari kuita, izvo [kusiyana-siyana] hazvina musoro."
Mukudzidza kwemichina, EL inobatsira kudzikisa musiyano wemodhi yako, asi zvinogona kukonzera modhi ine kuita kwese kurinani pane yakanakisa yekutanga modhi.
Ngationei mhinduro
Kubatanidza akawanda mamodheru kuita imwe inzira iri nyore iyo inogona kutungamira mukugadzirisa dambudziko rekusiyana kwekurerekera uye kugadzirisa mashandiro.
Kana uine maviri kana anopfuura mamodheru anoshanda zvakanaka, usasarudza pakati pawo: shandisa ese (asi nekuchenjerera)!
Unofarira kusimudzira munzira iyi? Saina kuti uwane yemahara Demo chidzidzo uye kutora chikamu mazviri - Machine Kudzidza Injiniya paMail.ru Boka.
Source: www.habr.com