YAYIN DA MANHAJAR GIDAJEN ABINCI TA FI DAIDAI, HAKA TAKE ƘARA RABA MASU KUƊI DA MATALAUTA

Taswirori (maps) suna jera gidajen abinci da ke kusa, manhajojin abinci suna ba da shawarar inda za a ci abinci na gaba, kuma manyan samfuran harshe (language models) yanzu suna amsa tambayoyi game da inda za a je. Ba kamar littafin jagora da aka buga ba, tsarin da ke bayan waɗannan shawarwari suna koyo daga ziyarorin da suke jagoranta. Ana sake horar da su (retrain) bisa inda mutane suka je — har da wuraren da su kansu suka ba da shawara. Wannan shi ne zagayen martani (feedback loop) tsakanin mutane da algorithms.

Dandamali (platforms) yawanci suna tantance waɗannan tsare-tsare ta daidaitonsu na offline (offline accuracy): yadda suke hasashen ziyarorin da mutane suka yi kafin kowace shawara. Wannan gwajin ba ya barin shawarwarin su sauya inda mutane ke zuwa, don haka ba zai iya ganin abin da zagayen ke yi a tsawon lokaci ba. Wata ƙungiya daga Northeastern University, Bruno Kessler Foundation a Trento, ISTI-CNR a Pisa da New York University ta tashi don auna shi.

Me ya sa birni ba jerin waƙoƙi (playlist) ba ne

An fi nazarin zagayen martani ne a kan waƙoƙi, fina-finai da kayayyaki, waɗanda kowane adadin mutane za su iya amfani da su ba tare da sun shafi juna ba. Gidajen abinci sun bambanta ta hanyoyi uku. Mutane suna zaɓa daga wuraren da ke kusa da inda suke. Gidan abinci zai iya hidimta wa adadin abokan ciniki iyakantacce kawai. Kuma shawarar da aka ba mutum ɗaya tana kuma sauya da wa sauran abokan ciniki ke haɗuwa.

Don haka ƙungiyar ta gina kwaikwayo bisa ainihin motsin mutane. Bayanan GPS da aka cire sunayensu (anonymised) daga masu amfani da suka amince, da aka naɗa tsakanin Oktoba 2016 da Maris 2017, sun ba da jadawalin yau da kullum na mazauna 4,000 a Boston, San Francisco da Detroit da 8,000 a Chicago, da aka zaɓa daidai-wa-daida daga rukunan samun kuɗi huɗu (quartiles) na unguwannin gidajensu. A kowace tafiya don cin abinci, wani tsarin shawarwari (recommender) ya ba da shawarar wurare 20 a cikin nisa mai ma’ana waɗanda har yanzu suke da sauran wuri; an saita ƙarfin karɓar baƙi na kowane wuri a mako a sau 1.5 na ainihin ziyarorinsa. An sake horar da tsarin kowane mako bisa watanni huɗu da suka gabata, na kusan shekaru uku. An riƙe ainihin ziyarorin kowa da ke cikin bayanan — har zuwa mutane 290,977 a kowane birni — don haka mazaunan kwaikwayon sun raba wurare da su.

An kwatanta algorithms huɗu: ɗaya da ke ba da shawarar wuraren da mutum ya fi ziyarta (UR-POI), biyu da ke koyo daga wurare masu kama da juna ko mutane masu kama da juna, da ɗaya da ke bin motsin dukkan jama’a.

Auna wa ke haɗuwa da wa

Ƙungiyar ta auna rabewa bisa samun kuɗi da ake fuskanta (experienced income segregation): ga kowane mutum, yadda lokacin da ake ɓatarwa a wurare ke rabe ba daidai ba tsakanin mutanen rukunan samun kuɗi huɗu. Maki 0 na nufin cakuɗa daidai; maki 1 na nufin haɗuwa da rukuni ɗaya kawai. Wurare masu cakuɗa suna da muhimmanci saboda alaƙa tsakanin masu matakan samun kuɗi daban-daban tana hasashen hawan matsayin tattalin arziki (upward mobility). Gidajen abinci wuri ne mai kyau na gwaji: rabewarsu ta kai kusan rabin ta unguwannin da mutane ke zaune.

Daidaito da cakuɗa

Sakamakon shi ne musaya (trade-off). Idan aka jera su bisa daidaiton offline da kuma bisa rabewar da suke haifarwa, algorithms huɗun suna fitowa a tsari ɗaya. Zaɓa bisa daidaito yana ɗaukar algorithm da ya fi haifar da rabewa a kowane birni. Mafi daidaito, UR-POI, ya bar mutane a rabe da 15–21% fiye da a ainihin ziyarorinsu. Yana cin nasara a daidaito ne daidai saboda mutane suna komawa wuraren da suka saba — kuma ta hanyar mayar da su can, yana ajiye su a cikin irin nasu.

Jerin (ranking) ne abin zargi, ba taswira ba. Zaɓa ba da tsari ba daga wurare iri ɗaya da ake iya kaiwa ya ba da rabewa mafi ƙanƙanta duka a Boston. Kuma nauyin ya fi faɗawa kan waɗanda tuni suka fi keɓewa: mazauna mafi ƙarancin samun kuɗi su ne rukunin da ya fi rabewa a ko’ina, kuma a Boston makinsu ya tashi daga 0.360 a zahiri zuwa 0.413 a ƙarƙashin UR-POI.

Gidajen abincin da ke ɓacewa

Zagayen yana kuma kwashe wurare. A ƙarƙashin kowane algorithm da a kowane birni, aƙalla kashi ɗaya cikin biyar na gidajen abinci sun rasa dukkan baƙinsu na kwaikwayo — tsakanin 21% da 71% daga cikinsu, idan aka kwatanta da 5% zuwa 30% ba tare da sake horarwa kowane mako ba. Sake horarwar kanta ta haifar da 55 zuwa 75% na asarar: wurin da ba ya samun ziyara yana ficewa daga bayanan kuma ba a sake ba da shawararsa. Asarar ta taru a unguwanni marasa yawan hada-hada, masu ƙarancin samun kuɗi.

Cakuɗa da tsira na iya rabuwa. Algorithms biyu da suka fi ƙarancin haifar da rabewa su ne suka fi koyo daga ziyarorin wasu mutane, suna tara kowa a wurare iri ɗaya — haɗuwar da ta kai ninki 4.4 zuwa 6.2 na ta ainihin ziyarori — don haka suna barin ƙarin wurare fanko, UserKNN fiye da kowa, a kowane birni.

Ɗan turawa a cikin jerin

Ƙungiyar ta gwada wani gyara mai sauƙi da ake kira tail-boost, da ake ƙarawa a kan kowane algorithm: ana haɗa makin kowane wuri da kari da ya fi girma ga wuraren da ba a fi ziyarta ba. A Boston, ba tare da sadaukar da fiye da 15% na daidaiton offline ba, ya rage ɓacewar wurare da 54 zuwa 75% a ƙarƙashin kowane algorithm kuma ya rage rabewa a ƙarƙashin uku daga cikin huɗun; a ƙarƙashin UR-POI, ƙarin rabewar ya faɗi daga 21% zuwa 13%. Ya yi hakan galibi ba tare da tura mutane zuwa wuraren da rukunin samun kuɗinsu ɗan tsiraru ne ba. Da daidaiton offline kaɗai, da an ƙi shi.

Kwaikwayon yana sauƙaƙa halayen mutane — a manyan gwaje-gwajen, kowa yana bin shawarwarin — kuma yana auna samun kuɗi bisa unguwa, ba bisa mutum ba. Marubutan sun saki lambar shirinsu (code) kuma suna jayayya cewa dandamali da masu tantancewa (auditors) ya kamata su kwaikwayi zagayen kafin su ƙaddamar da algorithm, a cikin ruhin Dokar Ayyukan Dijital ta Turai (European Digital Services Act). Binciken tsarin shawarwari ya daɗe yana ɗaukar kowace shawara a matsayin al’amari tsakanin mai amfani ɗaya da abu ɗaya; a birni, kamar yadda takardar ta faɗa, kowace shawarar da aka bi “tana kuma sauya da wa sauran mutane ke haɗuwa da kuma wuraren da ake ci gaba da ziyarta.”