Kwamfuta da AIKafin bugawaNazarin bayanaiKaratun minti 5

KUN FI ALGORITHM SANIN ABIN DA ZA KU SO

Tsarin shawarwari (recommender systems) ne ke yanke shawarar yawancin abin da mutane ke kallo, karantawa da saya; a Netflix, a cewar wata takarda ta 2015 da marubutan suka ambata, tsarin shawarwarin “yana tasiri a kan zaɓi na kusan 80% na sa’o’in da ake kallo”. Don ya yi aiki, dandali dole ne ya yi hasashen abin da kowane mutum zai so. Zai iya koyon hakan daga ɗabi’a — makin da aka bayar, dannawa (clicks), shirye-shiryen da ku da miliyoyin wasu kuka kalla har ƙarshe — ko kuma zai iya kawai ya tambaya.

Tambaya tana da tsada. A MovieLens, wani shafin ba da maki ga fina-finai, ba a amsa 98.8% na tambayoyi miliyan 3.0 ba. Kuma ka’ida (theory) tana ba da shawarwari masu cin karo da juna. Wata al’ada, wadda marubutan ke kira ra’ayin taron jama’a (crowd view), ta ce ƙididdigar da aka gina daga mutane da yawa ta fi hukuncin mutum ɗaya, kuma mutane ba su da ƙwarewa wajen hasashen jin daɗinsu. Wata kuma, ra’ayin ilimin da ya warwatsu (dispersed knowledge), ta ce kowannenmu yana riƙe da bayanan da babu wani rumbun bayanai (database) da ke rubuta su: kun ƙaunaci ayyukan baya na darektan amma ba ku taɓa ba su maki ba, tallan fim (trailer) bai burge ku ba, wani abokinku da kuke amincewa da ɗanɗanonsa ya ƙi daina magana a kansa.

Tambaya a shafin farko

Zihan Chen da Di Zhu (Stevens Institute of Technology) da Lei Nico Zheng (University of Massachusetts Boston) sun yi amfani da wani tarin bayanai (dataset) da ke ba da damar wannan gwaji. MovieLens, wani tsarin shawarwari wanda ba na kasuwanci ba da ɗakin bincike na GroupLens a University of Minnesota ke tafiyarwa, ya sanya wata tambaya a shafinsa na farko daga Maris 2023 zuwa Mayu 2024. Tana nuna sunan wani fim kuma tana tambayar memba ko ya kalla. Idan bai kalla ba: taurari nawa kuke ganin za ku ba shi, daga rabin tauraro zuwa biyar?

Daga cikin irin waɗannan hasashe 26,020, an biyo bayan 551 daga cikinsu — daga membobi 277 — da ainihin maki aƙalla kwana ɗaya daga baya. A kan waɗannan ma’aurata (pairs) kaɗai, ƙungiyar ta kwatanta hasashen da injuna huɗu: hasashen tsarin shawarwari na MovieLens kansa, algorithms na shawarwari na gargajiya guda biyu da aka horar a kan maki miliyan 33 na baya kuma aka sabunta su da duk makin da memban ya bayar har zuwa lokacin hasashen, da kuma “tari” (stack) da ke haɗa kowane hasashen injuna da abin da dandalin ya sani game da fim ɗin da memban.

Mutane sun yi nasara

Hasashen memba yana da alaƙa ta 0.69 da makin da aka bayar daga baya. Algorithms da aka horar sun kai 0.54 zuwa 0.59; hasashen MovieLens kansa, 0.34. A matsakaici, hasashen ya kauce wa makin na baya da taurari 0.61, idan aka kwatanta da aƙalla 0.72 ga kowane inji. Kuma kusan ba shi da son kai (bias), yayin da MovieLens ya ƙiyasta sama da makin na baya da rabin tauraro.

Fa’idar ba wai sanin waɗanne fina-finai ne masu kyau ba ce kawai, ko yadda mutum ke da karimci wajen ba da maki. Idan aka cire sha’awar fim ɗin a matsakaici da karimcin memban na yau da kullum, hasashen har yanzu yana hasashen abin da ya rage — dacewar da ke tsakanin wannan mutum da wannan fim — fiye da injuna nesa ba kusa ba (0.63 idan aka kwatanta da 0.47 a mafi yawa).

Idan aka ƙara shi a kan duk abin da injunan suka sani, hasashen ya rage kuskuren hasashe da 26%. Akasin haka ba gaskiya ba ne: ƙara dukan bayanan injunan a kan hasashen bai kawo wata riba da za a iya ganowa ba.

Inda injuna ke cim musu

Hoton ya fi daidaito a fannin jera abubuwa (ordering). Idan aka ba da fina-finai biyu da memba ya ba maki daban-daban daga baya, hasashen ya sa su a tsarin da ya dace 72% na lokaci. MovieLens ya samu 56% — jefa kwabo (coin flip) yana samun 50% — amma algorithms da aka horar sun kai 64 zuwa 67%, wanda marubutan suka ga cewa ba za a iya bambance shi da na mutumin ta fuskar ƙididdiga (statistically) ba.

Kuma a kan abubuwan da aka saba da su, injunan suna yin daidai ko fiye da haka. Ga fina-finan da membobi suka riga suka ba maki a zahiri shekaru da suka gabata — tsofaffin fina-finai, masu maki da yawa, waɗanda masu amfani da shafin da suka fi ba da maki suka ba maki — algorithms sun yi daidai da mutumin ko sun fi shi. Fifikon mutumin ya fi girma inda bayanai suka yi ƙaranci: ga fina-finan da aka ba su maki kaɗan (0.65 idan aka kwatanta da 0.41 ga wani algorithm da 0.13 ga MovieLens) da kuma nau’ikan fina-finai (genres) da memban ba kasafai yake ba maki ba.

Abin da ba ya taimakawa

Gajerun hanyoyi biyu sun gaza. Hasashen baya na memba game da wasu fina-finai, ko da an auna su gwargwadon kamancen waɗannan fina-finai, ba sa maye gurbin sabon hasashe game da fim ɗin da ake magana a kai. Kuma tambayar yadda mutane ke da tabbaci kusan ba ta ƙara komai: tabbaci galibi yana nuna ra’ayoyi masu tsauri, kuma membobin da suka fi tabbaci sun zama masu ɗan yawan kyakkyawan fata (overoptimistic).

Marubutan sun kuma duba cewa mutane ba sa kwafin hasashen da shafin ke nunawa kawai: hasashensu ya yi daidai da shi kusan gwargwadon yadda daidaituwa ta bazata (chance) za ta bayar.

Ƙa’idoji huɗu ga dandaloli

Daga waɗannan sakamako ƙungiyar ta fitar da ƙa’idojin tsari (design principles) guda huɗu: a tambayi makin da ake tsammani ga takamaiman abu; a tambaya game da abubuwan da ake yanke shawara a kansu yanzu; a auna tsarin shawarwari a kan manyan zaɓuɓɓukan da suke sake jerawa a zahiri, inda MovieLens ya jera 56% kawai na ma’aurata daidai; kuma a ba kowane algorithm duk abin da dandalin ya sani kafin a kwatanta shi da mutane.

Iyakokin a fili suke a cikin takardar. Ma’aurata 551 ɗin sune 2.1% na dukan hasashe, daga masu sha’awa a shafi guda, kuma mutane suna ba da maki ga abin da suka zaɓa su kalla. Wani gwaji na bazuwar zaɓi (randomised experiment), a cewar marubutan, zai nuna ko tambaya a zahiri tana inganta shawarwarin da mutane ke bi. Har zuwa lokacin, ƙarshen ya tsaya kamar yadda taken takardarsu ya faɗa: kun san wani abu game da fim ɗin da ke gabanku wanda babu wani tarihin maki da ya rubuta shi.