Ukuhlaziywa kwe-Spotify hit: idatha, ama-algorithms kanye nesayensi yempumelelo yomculo

Informatec Digital » Izinsiza » Ukuhlaziywa kwe-Spotify hit: idatha, ama-algorithms kanye nesayensi yempumelelo yomculo

I-Spotify isibe yindawo ephelele yokufunda ukuthi yini eyenza ingoma ibe yimpumeleloSinedatha yokudlala ngesikhathi sangempela, izibalo ezithuthukisiwe zabaculi, kanye nezigidi zabalaleli ezenza izinqumo njalo ngomzuzwana. Kunokuba nje ipulatifomu yokulalela, iSpotify isishintshe yaba yindawo enkulu yokudlala. database lapho amaphethini, imizwa, izinhlobo, kanye namasu ohlu lwadlalwayo kungahlaziywa khona ukuze kuqondwe ukuthi kungani ezinye izingoma ziqala ukuduma kanti ezinye zishona zingasekho.

Eminyakeni yamuva nje, amathuluzi, ucwaningo lwezemfundo, kanye amamodeli wokufunda womshini Kunikezelwe Ukuhlaziywa kwe-Spotify hitKusukela ekufundeni amashadi okuhlela njenge-Today's Top Hits noma i-Rap Caviar, kuya ekwakheni ama-algorithms akwazi ukubikezela ukuthi ithrekhi izoba ngeyeqembu elidumile kakhulu kukhathalogi. Ngesikhathi esifanayo, i-Spotify for Artists isebenzise ukufinyelela kulezi zilinganiso ngendlela yentando yeningi, ivumela noma yimuphi umculi ukuthi abone ukuthi ubani omlalele, ukuthi utholakala kuphi futhi kanjani, futhi alungise isu lakhe lokudala nelokumaketha ngokufanele.

Kuyini ngempela ukuhlaziywa kwe-Spotify hit futhi kungani kubalulekile?

Uma sikhuluma ngakho Ukuhlaziywa kwe-Spotify hit Asikhulumi nje ngokubheka ukuthi ingoma idlalwe kangaki. Kuyindlela ebanzi ehlanganisa umsindo, umongo, kanye nokuziphatha komsebenzisi ukuphendula umbuzo ongakhululekile kodwa obalulekile: ungakwazi yini ukubikezela impumelelo yengoma ngaphambi kokuba ikhishwe? Besebenzisa idatha evela ku-API ye-Spotify, abacwaningi basebenza ngezinto eziguquguqukayo ezifana nokuthandwa, amandla, i-valence, ubude, ijubane, kanye nokudansa ukuze bazame ukusondela kuleyo mpendulo.

Le ndlela isekelwe kulokho okubizwa ngokuthi "isayensi yengoma eshayayo"Lo mbono, owathandwa kakhulu nguMike McCready ekuqaleni kwawo-2000, uhilela ukusebenzisa ama-algorithms namamodeli ezibalo ukubikezela ukuthi yiziphi izingoma ezizosebenza kahle kumashadi nasemsakazweni. Nakuba izifundo zokuqala ziphethe ngokuthi lokhu kwakunzima kakhulu (nokuthi amamodeli ayenganembi ngokwanele), isimo sishintshe ngokuphelele ngokufika kokusakaza, ukwanda okukhulu kwevolumu yedatha, kanye nokuthuthukiswa kwamasu okufunda komshini.

Namuhla, iSpotify inikeza i-API enokufinyelela kumathrekhi ayizinkulungwane ezingamashumi nge- izimfanelo zomculo kanye nomongoKusukela kukhiye kanye nemodi, kuya emathubeni okuba ithrekhi ibe yi-acoustic noma i-instrumental, kufaka phakathi izibalo eziklanyelwe ngqo ukuchaza indlela ingoma ebonwa ngayo (amandla, i-valence, ukudansa, njll.). Konke lokhu kusenza sikwazi ukusuka emibonweni yomuntu siqu siye ekuhlaziyeni okunembe kakhulu kwenani.

Ngesikhathi esifanayo, ithuluzi Spotify Abaculi Kusiza abadali ukuthi bagweme ukulahleka ezilinganisweni zokuzikhukhumeza futhi bagxile kulokho okubaluleke ngempela: ukuthuthukiswa kwezilaleli zesikhathi eside, ukuzibandakanya, ukugcinwa, kanye nendlela imizamo yokumaketha ethinta ngayo ukudalwa kwabalandeli bangempela. Ngamanye amazwi, izinombolo, yebo, kodwa ngomongo kanye nenhloso yamasu.

Uhlu lwadlalwayo olusemthethweni kanye nesu: isisindo sezinhlu zokuhlela

Enye yezingcezu ezibaluleke kakhulu kunoma yikuphi ukuhlaziywa okujulile kwezingoma ezidumile ku-Spotify yindima ye- uhlu lwadlalwayo lokuhlela olusemthethweniAmashadi afana ne-Discover Weekly, i-Release Radar, izingoma ezidumile zanamuhla, i-New Music Friday, i-Rap Caviar, i-Mint noma i-¡Viva Latino! asebenza njenge-amplifiers ezinkulu: ukungena kwenye yazo kungashintsha ngokuphelele indlela yengoma noma ngisho nomsebenzi.

Amathuluzi eqembu lesithathu afana ne-Soundcharts akuvumela ukuthi ubone indlela umculi aziphatha ngayo kulezi zinhlu zokudlalayoInani lokubonakala, ubude bokuhlala, izimakethe lapho zinethonya elikhulu khona, njll. Lolu hlobo lokuhlaziya lukwenza kucace ukuthi ukuba khona ohlwini lwabahleli akuyona into ehlotshisiwe, kodwa kuyisici esibalulekile sesu lokukhula epulatifomu.

Kunezindlela ezithile eziphindaphindayo ekwakhiweni kwezinhlu zokudlalayo zekhwalithi. Isibonelo, kuyaziswa ukuthi kukhona abaculi abahlukahlukene esikhundleni sokuphindaphinda amagama afanayo ngokuphindaphindiweUhlu lapho umculi ofanayo evela khona izikhathi eziningi luvame ukuthola izilinganiso ezimbi kakhulu, ngoba lunciphisa umuzwa wokuthola izinto ezintsha kumlaleli.

Futhi ngifuna i- ibhalansi yobulili ehambisanayoUma uhlu lwadlalwayo luhlanganisa izitayela eziningi kakhulu ngaphandle kwesiqondiso esicacile, isipiliyoni Iyahlukana futhi isilinganiso siyehla. Uhlu olubonwa kahle luvame ukugxila kolunye noma ezimbalwa izinhlobo ezichazwe kahle, okusiza umsebenzisi ukuthi aqonde ngokushesha ukuthi yini okufanele ayilindele lapho ecindezela ukudlala.

Esinye isici esibalulekile yi- ingxube yezindikimba ezaziwayo nezivelayoIzinhla zokudlalayo eziqukethe izingoma ezinkulu kuphela zilungile, kodwa azinikezi lutho olutheni ngendlela yokuthola umculo. Ngokuphambene nalokho, izinhla zokudlalayo ezihlanganisa izingoma ezidumile namagugu angaziwa kakhulu zivame ukukhiqiza ukuzibandakanya okukhulu futhi zisize ekudaleni izingoma ezintsha.

Okokugcina, kukhona ukuvumelana okuthile kokuthi uhlu lokudlala oluhle kufanele okungenani lube nalo amathrekhi angaba ngu-50 ukuze anikeze ulwazi olupheleleAmashadi anezingoma ezingaphansi kwe-10 avame ukubonwa njengabi futhi athola amaphuzu aphansi, okuthinta nokusebenza kwamathrekhi afakiwe.

Umlando omfushane weSpotify kanye nomsuka wokuhlaziya kwayo

Ngaphambi kokuba iSpotify ikwazi ukwenza ukuhlaziywa okuthuthukisiwe kwezingoma ezidumile, ipulatifomu ngokwayo kwadingeka ithole indawo yayo embonini yomculo. Lo mbono, owasungulwa nguDaniel Ek, wavela ngemva kokuwa kukaNapster ngo-2002. Wayefuna ukudala isevisi "engcono kunobugebengu bokushushumbisa, kodwa futhi eyayikhokhela imboni."Ngaleso sikhathi, u-Ek wayesebenzisa i-uTorrent, enye yamakhasimende amakhulu okulanda nokwabelana nge-P2P, ngakho wayenolwazi oluqondile ngomhlaba wokusabalalisa okungagunyaziwe.

Ngemva kokuthengisa i-uTorrent ku-BitTorrent ngasekupheleni kuka-2006, u-Ek wagxila ngokuphelele ekwakheni i-Spotify. Lolu hlelo lokusebenza lwethulwa ngokusemthethweni ngo-2008 eSweden, ngemuva kokufinyelela ku-inthanethi. izivumelwano zokunikeza amalayisense kanye nokubamba iqhaza ngokulingana nezinkampani ezinkulu zomculoI-Sony Music Entertainment, i-Universal Music Group, kanye ne-Warner Music Group. Ngemva konyaka, yanwebeka yaya e-United Kingdom, futhi ngo-2011, yafika e-United States.

Ngaleso sikhathi sokuqala, isisekelo sababhalisile abakhokhelwayo sakhula sisuka cishe esigidini esisodwa eYurophu saya cishe Izigidi ezi-4 emhlabeni jikelele ngo-2012Ngo-2016, iSpotify yayisivele imemezele abasebenzisi abakhokhayo abayizigidi ezingu-40 kanye nabasebenzisi abacishe babe yizigidi eziyi-100, okuhlanganisa ukusakaza njengohlobo oludumile lokusetshenziswa komculo emhlabeni jikelele.

Ngesikhathi esifanayo, amathuluzi edatha abaculi aqala ukuvela. Okokuqala kwaba Ukuqonda KwabalandeliLesi sixazululo sanikeza amanye amaqembu ukufinyelela okulinganiselwe kudatha yokusakaza: izibalo zabantu, indawo, kanye nezitayela eziyisisekelo. Ngo-2017, lesi sixazululo sashintsha saba yi-Spotify for Artists, savula umnyango kubo bonke abaculi ukuze babone izilinganiso zabo ezibalulekile futhi basebenzise idatha ukwenza izinqumo mayelana nokuvakasha, ukukhishwa, kanye nokukhushulwa.

Ngo-2018, inkampani yaqala ukusebenza esidlangalaleni ngenani lemakethe elilinganiselwa ku- Ama-dollar ayi-30.000Kusukela ngaleso sikhathi, inani lezimakethe lapho isebenza khona liqhubekile nokukhula, futhi kanye nalo, ukubaluleka kokuhlaziya ngaphakathi kwepulatifomu. Lokho okwaqala njengomdlali wezindaba onelayisensi sekuyisingqalasizinda yomhlaba wonke lapho idatha ibusa khona kakhulu.

Izilinganiso ze-Spotify: indlela yokulinganisa ingoma ngale kwemifudlana

Ukuze wakhe amamodeli okubikezela okuphumelelayo, kubalulekile kuqala ukuqonda ukuthi hlobo luni lwe- idatha enikezwe yi-Spotify APIKwesinye isifundo esigxile emakethe yaseNdiya, isibonelo, amarekhodi ezingoma angaphezu kuka-46.000 akhishwe ohlwini lwadlalwayo olukhiqizwe yiSpotify, oluhlanganisa izinhlobo ezahlukene zomculo kanye nezinhlobo ezincane.

Ulwazi luhlelwe ngezigaba eziningana. Ezingeni lokulandelela, sithola idatha efana nalokhu isihlonzi, isihloko, umculi, ukuthandwa, usuku lokukhishwa, kanye nobude besikhathi ngama-millisecondEzingeni le-albhamu, i-ID kanye negama kuyagcinwa. Ezingeni lohlu lokudlalayo, kuvela igama, i-ID, uhlobo, kanye nohlobo olungaphansi oluhambisanayo.

Kodwa into ethakazelisa kakhulu ekuhlaziyweni kwe-hit yi- izici zomsindoihlukaniswe ngezigaba ezahlukene: izici "zesimo sengqondo" (ukudansa, amandla, i-valence, ijubane), izakhiwo zomzimba (umsindo, ukukhuluma, izinsimbi zomculo), umongo (umsindo, impilo), kanye nezingxenye zomculo (ukhiye, imodi). Ngayinye yalezi ziguquguquko ichazwa ngokucophelela futhi ijwayelekile.

Ukudansa, isibonelo, kuvezwa njengenani eliphakathi kuka-0 no-1 elifingqa Kulula kangakanani ukudansela ingomaNgokusekelwe ezintweni ezifana nesigqi, ukuzinza kwejubane, kanye namandla okushaya, amandla asukela ku-0 kuya ku-1, ezama ukubamba ukuthi ithrekhi izwakala inamandla, ishesha, futhi inamandla, uma kuqhathaniswa nokuthile okuzolile noma okuthambile.

UValencia uchaza "umbono omuhle" obonwayo ngokomzwelo Kumsindo, amanani aphansi ahambisana nemizwa yokudabuka, yokuxineka, noma yokudabuka, kuyilapho amanani aphezulu ehambisana nomculo ojabulisayo, okhanyayo, noma ojabulisayo. I-acoustic ilinganisa amathuba okuba ithrekhi ibe yi-acoustic, impilo ukuba khona kwezilaleli ezibukhoma ekurekhodweni, kanti ukukhuluma kukhombisa isilinganiso samagama akhulunywayo engxubeni (isibonelo, kuma-podcast noma kumathrekhi akhulunywayo).

Eminye imingcele ebalulekile yile i-tempo ku-BPMubude besikhathi, ukhiye (ofakwe ikhodi njengenombolo ephelele), imodi (enkulu noma encane), kanye nokuthandwa kwethrekhi. Lokhu kokugcina kuyi-metric yangaphakathi ye-Spotify esukela ku-0 kuya ku-100. Kuncike kokubili kumthamo wemifudlana nokuthi isanda kuba khonaIngoma eyayinkulu eminyakeni edlule, kodwa engasadlalwa kakhulu, izobona izinga layo lehla ngokuhamba kwesikhathi.

Indlela yokulungisa nokuhlanza isethi yedatha ukuze ukwazi ukubikezela amagalelo

Isinyathelo esisodwa esivame ukunganakwa lapho kuxoxwa ngokuhlaziywa kwe-Spotify hit yilesi ukulungiswa kwesethi yedathaOcwaningweni olukhona, idatha yaqoqwa kusetshenziswa i-R kanye ne-RStudio, kubizwa i-Spotify API yezinhlanganisela ezahlukene zemakethe (e-India), izinhlobo kanye nezinhlobo ezingaphansi ezikhethiwe ngokwesisindo sazo somhlaba wonke kanye nesendawo.

Uma kuhlanganiswa izingoma ezivela ohlwini lwadlalwayo olunezihloko ezahlukahlukene, kuvamile ukuthi kuphindwe amathrekhi amaningi, ngoba ingoma efanayo ingavela ohlwini oluhlukene. Njengoba inhloso kwakungekona ukuhlaziya isu lokudlala ngokwalo, kodwa kunalokho izici zezingoma, kwakuwukuthi Basuse amathrekhi aphindaphindwayoukunciphisa isisekelo esiphelele kusuka ku-46.417 kuya cishe ezingomeni eziyingqayizivele ezingu-39.147.

Amakholomu ayengeke asetshenziswe njengeziguquguquko ezichazayo kumamodeli, njengomculi, i-ID ye-albhamu kanye negama, kanye nezinkambu ezithile zohlu lwadlalwayo, asusiwe. Ngesikhathi esifanayo, Balinganisela amagama ensimu futhi izinhlobo zedatha zalungiswa: isibonelo, ukuthandwa, imodi, ukhiye kanye nobude besikhathi kwaguqulwa kwaba amanani ezinombolo zohlobo lwe-float ukuze kube lula ukuphathwa kwezibalo.

Isinqumo esibalulekile kwakuwukuguqula ukuthandwa kube yinguquko elawuleka kalula. Esikhundleni sokusebenza ngenani eliqhubekayo kusukela ku-0 kuya ku-100, i- umkhawulo wokuhlukanisa izingoma ezidumile nezingathandwaUma sibheka iphesenti lama-85 lokusatshalaliswa (cishe ukuthandwa okungu-65), amathrekhi angaphezu kwalelo nani abhekwa "njengodumile" kanti amanye "angathandwa".

Ngale ndlela, isethi yedatha yahlukaniswa yaba cishe izingoma ezidumile ezingu-6.000 kanye nezingezona ezidumile ezingaba ngu-33.000Ekuhlaziyweni kokuhlola, idatha yahlukaniswa yaba yizigaba ezinhlanu zokuthandwa (eziphezulu kakhulu, eziphezulu, eziphakathi nendawo, eziphansi, neziphansi kakhulu), ngezikhawu zamaphuzu angu-20, okwavumela ukuqhathaniswa kwezitayela ezinhle phakathi kwamazinga empumelelo.

Ngaphezu kwalokho, kukhiqizwe i-variable entsha ezihlokweni zezingoma: a isibonakaliso semizwaKusetshenziswa umtapo wolwazi we-TextBlob ku-Python, i-polarity yesihloko ngasinye (inombolo ephakathi kuka--1 no-1) yabalwa futhi yahlukaniswa njenge-positive, negative, noma i-neutral. Leli nani lezinombolo lanezelwa ku-dataframe ukuze kufundwe ukuthi ithoni yesihloko ihlobene yini nokuthandwa kwaso.

Ukuhlola idatha: yini ehlukanisa izingoma ezidumile kakhulu

Ngaphambi kokuqeqesha noma yimuphi imodeli yokufunda komshini, kubalulekile ukunikela ngesikhathi ukuhlola idatha ngokubona nangokubalaKulesi sifundo, kwahlaziywa ama-curve okusabalalisa, izilinganiso ngeqembu ngalinye lokuthandwa, kanye nobudlelwano phakathi kwemizwa nempumelelo.

Enye yezinto ezitholakele eziphawulekayo ihlobene nokuthandwa kwezingoma eziphumelele kakhulu. Lapho kubhekwe amathrekhi anesilinganiso sokuthandwa esingaphezu kwama-90, kwaphawulwa ukuthi Inani elikhulu labo lalingaphansi kwenani lika-0,5 eValenciaNgamanye amazwi, babezwakala bedabukile, bedabukile, noma bedabukile kunokuba bejabule. Kwakungebona ukubusa okuphelele, kodwa kwakuwumkhuba ocacile.

Lapho ukusatshalaliswa kokuthandwa ngokobulili kuhlelwe, iningi lama-curve lasebenzisa i- isimo sensimbi esilinganiselwengamathrekhi amaningi acishe abe maphakathi futhi ambalwa ngokwedlulele. Kodwa-ke, izinhlobo zomculo ezifana ne-rock, i-R&B, i-EDM, umhlaba, kanye nomculo waseNdiya zibonise ukungalingani ngenxa yenani elikhulu lamathrekhi angadumile nhlobo. Ngokuphambene nalokho, izitayela ezifana ne-pop, i-rap, isiLatini, kanye ne-desi zibonakale zilinganisela kakhulu futhi zingenawo ukugxila okukhulu kwamathrekhi anesilinganiso se-zero.

Lokhu kusikisela ukuthi kulezo zinhlobo ezivame kakhulu, ngokuvamile, kulula ukufinyelela ezingeni elithile lokuthandwanoma izingoma zingenazo zonke izici ezifanelekile, kanti kwezinye izitayela ukusatshalaliswa kwempumelelo kuhlukaniswe kakhulu.

Lapho kuqhathaniswa isilinganiso sezici ezahlukene ngesigaba sokuthandwa, kwavela iphethini ecacile kakhulu: izihloko eziphumelela kakhulu zivame ukuba nazo amandla amakhulu kanye nomsindo omkhulu (ivolumu ebonwayo)kanye nokusetshenziswa kwezinsimbi zomculo eziningi kanye nokukhuluma okuncane. Kalula nje, izingoma ezidlala kahle kakhulu ekusakazweni zivame ukuba namandla kakhulu futhi zigxile kakhulu emculweni kunasemazwini akhulunywayo.

Kuphinde kwaphawulwa ukuthi amathrekhi adumile afaka ukuzwakala okuphansi kanye nokuphila okuphansiNgamanye amazwi, azwakala “njenge-acoustic” encane futhi “njenge-live” encane. Akhiqizwa kakhulu estudiyo, acwebezelisiwe kakhulu, anomsindo omncane noma umuzwa wekhonsathi.

Okunye okubalulekile ukuthi izingoma ezidumile zivame ukuba zimfushane kunezingathandekiEsimweni lapho imali engenayo incike enanini lemifudlana kanye nokuphindaphinda kwemivuzo yama-algorithms, kunengqondo ukuthi amathrekhi amafushane angaqoqana ukudlala ngokushesha futhi abe "nobungane" kakhudlwana ohlwini lwadlalwayo.

Ngokuphathelene nenjabulo ebonwayo (i-valence), lo mkhuba uyamangaza: amazinga akhuphuka kusukela ezigabeni eziphansi zokuthandwa kuya esigabeni "esiphezulu", kodwa esigabeni esibi kakhulu, "esiphezulu kakhulu," kukhona ukwehla okuphawulekayo. Ngamanye amazwi, Izingoma ezidumile kakhulu zivame ukuzwakala zidabukisa kakhulu kunezingoma "eziphumelelayo" nje.Lokhu kuvula umnyango wezincazelo eziningi mayelana nokunambitheka komphakathi kanye nomongo wezenhlalo-ngokomzwelo.

Ukuhlaziywa kwezibalo: umthelela wezinhlobo kanye nezimfanelo zomsindo

Uma idatha isihloliwe, isinyathelo esilandelayo kwakuwukusebenzisa ukuhlolwa kwezibalo okusemthethweni ukubona ukuthi yiziphi iziguquguquko ezingabhekwa njengezifanele ekuchazeni ukuthandwa. Ukuze kufundwe ukuthi impumelelo ithonywe ubulili, kwasetshenziswa ukuhlaziywa kokwehluka (i-ANOVA) ngezinhlobo eziyisishiyagalolunye eziyinhloko.

Umphumela we-ANOVA uveze inani eliphezulu kakhulu le-F kanye nokubaluleka okucishe kube yi-zero, okusho ukuthi Kukhona umehluko obalulekile ngokwezibalo ekuthandweni phakathi kwezinhlobo zezinhloboKodwa-ke, lokho akwanele: kubalulekile futhi ukwazi ukuthi lo mehluko uvela phakathi kwaziphi izibili zobulili nokuthi ingabe lokhu okuguquguqukayo kuzoba wusizo yini kumodeli yokubikezela.

Njengoba ukuhlukahluka kwakungafani futhi inani lezingoma ngohlobo ngalunye lahlukahluka, kusetshenziswe okulandelayo: Ukuhlolwa kwe-Games-Howell ngemuva kwe-hocLokhu kuhlaziywa kwenze kwaba nokwenzeka ukuqinisekisa ukuthi yiziphi izinhlanganisela zobulili ezingabonisi umehluko omkhulu ekuthandweni, okwenza, ngokujwayelekile, kungabi ngcono ukusebenzisa ubulili njengesibikezelo esinamandla kumamodeli okufunda komshini.

Ngesikhathi esifanayo, okulandelayo kwenziwe ukuhlolwa kwe-t okuzimele Kwisici ngasinye somsindo, izindlela zaqhathaniswa phakathi kweqembu lezingoma ezidumile neqembu lezingoma ezingezona ezidumile. Ngezinga lokubaluleka elingu-95%, kwatholakala ukuthi cishe zonke izinto eziguquguqukayo (ukudansa, amandla, umsindo, ukuzwakala, impilo, ubude, ijubane, i-valence, kanye nomculo wezinsimbi) zibonise umehluko omkhulu phakathi kwamaqembu amabili.

Okuwukuphela kwento esobala engavamile kwakuyi- ukukhulumaEsivivinyweni sokuqala, umehluko wezindlela wawungabalulekile, kodwa ukuhlaziywa okwengeziwe kwembula ukuthi ukukwazi ukukhuluma kuhluka ezihlokweni ezidumile kakhulu (ezinamanani aphakathi kuka-0,024 no-0,685) kuwele ngaphakathi kobubanzi besigaba esingavamile (phakathi kuka-0 no-0,964). Lokhu kugqagqana akuzange kuvimbele ukukwazi ukukhuluma, lapho kusetshenziswa ngempumelelo, ekunikezeni ulwazi kumodeli, ngakho-ke kwanqunywa ukuthi kugcinwe njengesibikezelo.

Ngamafuphi, ibhlogo lezici zomsindo libonisiwe amandla acacile achazayo mayelana nokuthandwa, kuyilapho ubulili buphathwe ngokucophelela okukhulu futhi kwakhishwa njengento eyinhloko kwezinye izindlela zokubikezela.

Ukwakha amamodeli okufunda komshini ukubikezela amagalelo

Njengoba kukhethwe isethi yedatha ehlanzekile kanye neziguquguquko ezifanele, kwase kuyisikhathi sokudala amamodeli okuhlukaniswa kwe-binary ikwazi ukubikezela ukuthi ingoma ingeyeqembu elidumile kakhulu elingu-15%. Ukuze kwenziwe lokhu, ukhiye kanye nemodi yezinguquko zesigaba kwaqala kwaguqulwa kwaba yizinguquko ze-dummy kusetshenziswa umsebenzi we-get_dummies kaPandas.

Ngemva kokufaka la ma-dummies, amakholomu okuqala e-key kanye ne-mode, kanye nokuthandwa okuqhubekayo, kwasuswa, futhi i-target variable yagcinwa. insimu yodumo enama-binary (okuthandwayo vs. okungathandwa). Ngaphambi kokuqeqesha amamodeli, zonke izici zezinombolo zahlelwa nge-StandardScaler, njengoba zazisebenza ezikalini ezihlukene kakhulu futhi kwakubalulekile ukuthi zibe nesisindo esifanayo.

Isethi yedatha yahlukaniswa yaba ukuqeqeshwa nokuhlolwa kusetshenziswa umsebenzi we-train_test_split, okugcina Ama-20% okubhalisela ukuhlolwaNgale ndlela, ukuvuza kolwazi kwagwenywa futhi kwaqinisekiswa ukuthi izilinganiso zokusebenza zibonisa ukuguquguquka kwangempela kwamamodeli, hhayi nje ikhono lawo lokukhumbula idatha yokuqeqeshwa.

Imodeli yokuqala esetshenzisiwe kwakuyi- Ukuhlehla kwezintoLe ndlela isetshenziswa kabanzi ekuhlukaniseni okubili. Inguqulo ephindaphindayo enezinga lokufunda elingu-0,01 kanye nokuphindaphinda okungu-200 yasetshenziswa futhi yahlolwa kusetshenziswa ukuqinisekiswa okuphambeneyo. Ukunemba okumaphakathi kufinyelele cishe ku-84,7%, umphumela onamandla ngokumangalisayo wenkinga eyinkimbinkimbi njengokubikezela impumelelo yomculo.

Ngemuva kwalokho i-algorithm yahlolwa Omakhelwane Abaseduze Kakhulu (KNN)okuhlukanisa ingoma ngayinye ngokwezindawo eziseduze ze-k esikhaleni sesici. Ukulungisa inani le-ky kusetshenziswa i-Grid Search ukuthola ukucushwa okuhle kakhulu kuveze ukunemba okufanayo, cishe ngo-84,8%, kanye nokuthuthuka okuncane kumaphuzu okuqinisekisa okuphambene uma kuqhathaniswa nokubuyela emuva kwe-logistic.

Imodeli elandelayo kwakuyi- Umshini Wokusekela Amavektha (i-SVM)Kwasetshenziswa enye indlela eqondisiwe ekwazi ukusingatha ubudlelwano obuqondile nobungebona obuqondile. Futhi, ngokufakwa kwe-hyperparameter, ukunemba okucishe kube ngu-84,6% kwafinyelelwa, ngamanani alinganayo ekuqinisekisweni okuphambene, okubonisa ukusebenza okuzinzile.

Okulandelayo kuhlolwe futhi: Isihlungi se-Naive BayesNgokusekelwe emcabangweni wokuzimela phakathi kwezici. Naphezu kokuba yimodeli elula kakhulu ngokwemibono, imiphumela yayo yokunemba (cishe u-84,6%) yayilingana ne-SVM futhi isondele kakhulu ekubuyiselweni kwe-logistic kanye ne-KNN, okuqinisa umqondo wokuthi isakhiwo senkinga sizivumela kahle kulolu hlobo lwendlela yokucabanga okungenzeka.

Ekugcineni, amamodeli asekelwe ezihlahleni zezinqumo ahlolwe. Isihlukanisi Sesihlahla Sezinqumo Ifinyelele ukunemba okuphansi kakhulu, cishe ngo-75,4%, futhi ukuqinisekiswa okuphambene kuqinisekisile lokhu kwehla kokusebenza, okungenzeka ukuthi kungenxa yezinkinga zokufaka ngokweqile. I-Random Forest Classifier, ehlanganisa izihlahla eziningi ukuze ilungise le miphumela, ithuthuke kakhulu, ifinyelela cishe ukunemba okungu-84,1% kanye nokuqinisekiswa okungaphezu kuka-84% ekuhlanganisweni.

Ukuze kuqedwe ukuhlaziywa, i-matrix yokudideka kanye nombiko wokuhlukanisa ye-Random Forest. Imodeli ibonise ikhono elihle kakhulu lokuhlonza izingoma ezingezona ezidumile (isigaba seningi), ngokunemba okungu-0,85 kanye nokukhunjulwa okuseduze no-0,99, kuyilapho ukusebenza kwayo ekilasini lezingoma ezidumile (iqembu elincane) kwakuphansi kakhulu, ngokunemba okungu-0,40 kanye nokukhunjulwa okungu-0,05. Lokhu kugqamisa inselele yakudala yokungalingani kwezigaba kulolu hlobo lwenkinga.

Ukubaluleka kweziguquguquko: yini esinda kakhulu lapho kudalwa i-hit

Ngaphandle kokwazi ukuthi iyiphi imodeli enembile kakhulu, kubalulekile ukuqonda yiziphi izici ezinikeza ulwazi oluwusizo ngempela lapho kubikezelwa ukuthi ingoma izothandwa yini. Ngale njongo, i-XGBoost (XGBClassifier) ​​​​yasetshenziswa ukuthola amaphuzu okubaluleka okuguquguqukayo, ngokusekelwe kumnikelo wesici ngasinye ekunciphiseni amaphutha ezihlahleni zezinqumo.

Imiphumela ikhombisile ukuthi Ubude bengoma bucacile ukuthi buhlukile kwabanyengesilinganiso sika-F esingaphezu kuka-400. Lokhu okutholakele kuhambisana nomqondo wokuthi amathrekhi amafushane akhuthaza ukulalela okuphindaphindiwe futhi ngaleyo ndlela aqoqe ukudlala okwengeziwe ngesikhathi esincane, into evame ukuvuzwa yi-algorithm yezincomo ze-Spotify.

Ngakolunye uhlangothi, iziguquguquko ezifana nokhiye, imodi, noma umuzwa othathwe esihlokweni Bathole amaphuzu okubaluleka angaphansi kwama-50, okuphakamisa ukuthi umnikelo wabo emandleni okubikezela jikelele emodeli wawumncane. Lokhu akusho ukuthi abanawo nhlobo umphumela, kodwa kunalokho, uma kuqhathaniswa nezinye izici, isisindo sabo sincane kakhulu.

Izimfanelo zomsindo ezisele (njengamandla, ukudansa, i-valency, umsindo, i-acoustic, impilo, ukukhuluma, kanye ne-instrumental) bezisezingeni eliphakathi nendawo lapho amaphuzu e-F ephakathi kuka-200 no-300, okubonisa ukuthi Bakha ibhulokhi yolwazi elinganiselayoAkukho neyodwa kuzo ebusa ngokuphelele, kodwa konke ndawonye kusiza ekwakheni isithombe esinembile samathuba engoma okuphumelela.

Sekukonke, amamodeli asekelwe ezicini zomsindo akwazile ukubikezela cishe izinga lempumelelo elingu-84-85% uma inkomba ingaba yingxenye ye-15% ethandwa kakhuluLokhu kunikeza ukwesekwa okuthile okubonakalayo emqondweni wakudala we-"hit song science": ngemininingwane emihle namasu afanele, impumelelo yomculo ayiyona nje into engahleliwe, yize kusekhona ingxenye ebalulekile engabikezeleki.

Isithombe esidwetshwe yilolu hlaziyo sibonisa ukuthi ukuhlanganisa idatha ye-Spotify, izibalo zendabuko, kanye namamodeli okufunda komshini kusenza sikwazi ukwenza okungaphezu kokubala nje imifudlana. Ukuqonda umthelela wobude besikhathi, amandla, i-valence, kanye nezinsimbi zomculo, kanye nendima yezinhlu zokudlalayo zokuhlela kanye nezincomo zesikhulumi, kunikeza abaculi, amalebula, kanye nabahlaziyi imephu eqondile yokuhumusha ukuthi kungani izingoma ezithile ziba yizingoma ezidumile nokuthi bangawandisa kanjani amathuba abo okujoyina lelo qembu elikhethiwe ngaphandle kokulahlekelwa ngokuphelele yisici sobuntu nesokudala, ngenhlanhla, esihlala singenakuncishiswa kunoma iyiphi ifomula.