Bin Ƙwararrun Ilimi Mai Adalci a Koyon Harshe Na Biyu: Nazarin Karkatarwar Algorithm

1. Gabatarwa

Tsarin hasashe a fannin ilimi, musamman Bin Ƙwararrun Ilimi (KT), yana nufin ƙirƙirar tsarin yanayin ilimin ɗalibi don keɓance koyo. Hanyoyin gargajiya sun dogara da hukuncin ɗan adam, wanda ke da saukin karkata saboda iyakokin ƙwaƙwalwar ajiya, gajiya, da kuma son abubuwa masu kyau. Bin Ƙwararrun Ilimi na kwamfuta, wanda Corbett da Anderson (1994) suka gabatar, yana amfani da bayanan hulɗar ɗalibi (maki, ra'ayi, shiga ciki) don hasashen ayyukan gaba da daidaita koyarwa.

Duk da cewa daidaito shine abin da aka fi mayar da hankali, wannan binciken ya nuna wata gibi mai mahimmanci: adalcin algorithm. Binciken ya bincika ko tsarin hasashe a cikin koyon harshe na biyu (ta amfani da bayanan Duolingo) yana nuna karkata ba da gangan ba ga wasu ƙungiyoyi bisa dandamali (iOS, Android, Yanar Gizo) ko matsayin ci gaban ƙasa (masu ci gaba da masu tasowa).

2. Hanyoyin Bincike & Tsarin Gwaji

Binciken ya yi amfani da tsarin nazarin kwatance don kimanta adalci tare da daidaito.

2.1 Bayanai & Hanyoyin Koyo

An yi amfani da hanyoyin koyo guda uku daga cikin bayanan aikin haɗin gwiwa na Duolingo na 2018:

en_es: Masu magana da Ingilishi suna koyon Sifen.
es_en: Masu magana da Sifen suna koyon Ingilishi.
fr_en: Masu magana da Faransanci suna koyon Ingilishi.

Bayanai sun haɗa da jerin ayyukan ɗalibi, daidaito, da metadata (dandamali na abokin ciniki, ƙasa). An rarraba ƙasashe a matsayin "Masu Ci Gaba" ko "Masu Tasowa" bisa ma'auni na tattalin arziki (misali rarrabuwar IMF).

2.2 Tsarin Hasashe

An kimanta nau'ikan tsare-tsare guda biyu:

Koyon Injini (ML): Tsare-tsaren gargajiya kamar Regression na Logistic, Dazuzzukan Bazu.
Koyo Mai Zurfi (DL): Tsare-tsare na tushen hanyar sadarwar jijiyoyi, mai yuwuwa sun haɗa da bambance-bambancen Bin Ƙwararrun Ilimi Mai Zurfi (DKT) ko gine-ginen Transformer.

Aikin farko shine hasashe na binary: shin ɗalibin zai amsa aikin gaba daidai?

2.3 Ma'aunin Adalci

An kimanta adalci ta amfani da ma'auni na adalcin ƙungiya, tare da kwatanta aikin tsari a tsakanin ƙungiyoyin da aka kare:

Adalcin Dandamali: Kwatanta daidaito, makin F1, ko AUC tsakanin masu amfani akan abokan ciniki na iOS, Android, da Yanar Gizo.
Adalcin Ƙasa: Kwatanta ma'auni na aiki tsakanin masu amfani daga ƙasashe masu ci gaba da masu tasowa.

Bambance-bambance a cikin waɗannan ma'auni suna nuna karkatarwar algorithm. Tsarin da ya cika adalci zai kasance da aiki daidai a duk ƙungiyoyi.

3. Sakamako & Bincike

Binciken ya samar da mahimman bincike guda huɗu, wanda ya bayyana manyan yin zabi da karkata.

3.1 Daidaito da Adalci: Yin Zabi

Tsare-tsaren Koyo Mai Zurfi (DL) gabaɗaya sun fi Tsare-tsaren Koyon Injini (ML) aiki a cikin daidaito da adalci. Ikon DL na ɗaukar rikitattun alamu, waɗanda ba su da layi a cikin bayanan koyo na jeri yana haifar da hasashe masu ƙarfi waɗanda ba su dogara da alaƙar ƙarya da ke da alaƙa da halaye masu mahimmanci.

3.2 Karkata Ta Kan Dandamali (iOS/Android/Yanar Gizo)

Dukansu algorithms na ML da DL sun nuna wani karkata mai mahimmanci ga masu amfani da wayoyin hannu (iOS/Android) fiye da waɗanda ba na wayar hannu ba (Yanar Gizo). Wannan na iya fitowa daga bambance-bambancen ingancin bayanai (misali, tsarin hulɗa, tsawon zaman), ƙirar fuska, ko bayanan al'ummar da aka saba danganta su da kowane dandamali. Wannan karkatarwar tana haifar da haɗari ga masu koyo waɗanda ke samun damar kayan aikin ilimi ta kwamfutoci na tebur.

3.3 Karkata Ta Kan Ƙasa (Masu Ci Gaba da Masu Tasowa)

Algorithms na ML sun nuna karkata mai zurfi ga masu amfani daga ƙasashe masu tasowa idan aka kwatanta da algorithms na DL. Wannan wani bincike ne mai mahimmanci, domin tsare-tsaren ML na iya koyo da haɓaka rashin daidaiton tarihi da ke cikin bayanan horo (misali, bambance-bambance a cikin damar ilimi ta baya, amincin intanet). Tsare-tsaren DL, duk da cewa ba su da kariya, sun nuna juriya mafi girma ga wannan karkatarwar ta ƙasa.

Zaɓin Mafi Kyawun Tsari: Binciken ya ba da shawarar hanya mai zurfi:

Yi amfani da Koyo Mai Zurfi don hanyoyin en_es da es_en don mafi kyawun daidaiton adalci da daidaito.
Yi la'akari da Koyon Injini don hanyar fr_en, inda bayanansa na adalci-daidaito aka ga ya fi dacewa da wannan yanayin na musamman.

4. Nazarin Fasaha & Tsarin Aiki

4.1 Tsarin Bin Ƙwararrun Ilimi

A ainihinsa, Bin Ƙwararrun Ilimi yana ƙirƙirar tsarin yanayin ilimin ɗalibi. Idan aka ba da jerin hulɗa $X_t = \{(q_1, a_1), (q_2, a_2), ..., (q_t, a_t)\}$, inda $q_i$ aiki/tambaya ne kuma $a_i \in \{0,1\}$ shine daidaito, manufar ita ce hasashen yuwuwar daidaito akan aikin gaba: $P(a_{t+1}=1 | X_t)$.

Bin Ƙwararrun Ilimi Mai Zurfi (Piech et al., 2015) yana amfani da Hanyar Sadarwar Jijiyoyi ta Maimaituwa (RNN) don ƙirƙirar wannan:

$h_t = \text{RNN}(h_{t-1}, x_t)$

$P(a_{t+1}=1) = \sigma(W \cdot h_t + b)$

inda $h_t$ shine yanayin ɓoyayye wanda ke wakiltar yanayin ilimi a lokacin $t$, $x_t$ shine haɗakar shigarwar $(q_t, a_t)$, kuma $\sigma$ shine aikin sigmoid.

4.2 Tsarin Kimanta Adalci

Binciken a ɓoyayye yana amfani da tsarin adalcin ƙungiya. Don mai hasashe na binary $\hat{Y}$ da sifa mai mahimmanci $A$ (misali, dandamali ko ƙungiyar ƙasa), ma'auni na gama gari sun haɗa da:

Bambancin Daidaiton Ƙididdiga: $|P(\hat{Y}=1|A=0) - P(\hat{Y}=1|A=1)|$
Bambancin Daidaiton Damar: $|P(\hat{Y}=1|A=0, Y=1) - P(\hat{Y}=1|A=1, Y=1)|$ (Ana amfani da shi lokacin da aka san ainihin lakabin Y).
Bambance-bambancen Ma'aunin Aiki: Bambanci a cikin daidaito, AUC, ko makin F1 tsakanin ƙungiyoyi.

Ƙaramin bambanci yana nuna adalci mafi girma. Binciken ya nuna cewa tsare-tsaren DL suna rage waɗannan bambance-bambancen cikin inganci fiye da tsare-tsaren ML a duk ƙungiyoyin da aka ayyana.

5. Nazarin Lamari: Aiwatar da Tsarin

Yanayi: Wani kamfani na EdTech yana amfani da tsarin KT don ba da shawarar ayyukan bita a cikin app ɗin sa na koyon harshe. An horar da tsarin akan bayanan masu amfani na duniya.

Matsala: Nazarin bayan aiwatarwa ya nuna cewa masu amfani a ƙasar X (ƙasa mai tasowa) suna da ƙimar kashi 15% mafi girma na ba da shawarar ayyukan da suka yi wahala da ba daidai ba, wanda ke haifar da takaici da daina aiki, idan aka kwatanta da masu amfani a ƙasar Y (ƙasa mai ci gaba).

Nazari ta amfani da tsarin wannan takarda:

Gano Ƙungiya Mai Muhimmanci: Masu amfani daga ƙasashe masu tasowa da masu ci gaba.
Bincika Tsari: Lissafa ma'auni na aiki (Daidaito, AUC) daban-daban ga kowace ƙungiya. Bambancin kashi 15% da aka lura a cikin "ƙimar shawarwarin wahalar da ta dace" cin zarafi ne na adalci.
Bincika Dalili: Shin tsarin ML ne ko DL? Bisa ga wannan binciken, tsarin ML yana da yuwuwar ya nuna wannan karkatarwar ta ƙasa. Bincika rarraba fasali—watakila tsarin ya dogara da yawa akan fasali masu alaƙa da ci gaban ƙasa (misali, matsakaicin saurin haɗi, nau'in na'ura).
Gyara: Yi la'akari da canzawa zuwa ginin KT na tushen DL, wanda binciken ya gano ya fi ƙarfin juriya ga wannan karkatarwar. Ko kuma, yi amfani da dabarun horo masu kula da adalci (misali, kawar da karkata ta hanyar adawa, sake nauyi) ga tsarin da ke akwai.
Sa ido: Ci gaba da bin diddigin ma'aunin adalci bayan shiga tsakani don tabbatar da an rage karkatarwar.

6. Aiwatar da Gaba & Hanyoyi

Tasirin wannan binciken ya wuce koyon harshe na biyu:

Keɓance Koyo a Girma: Tsare-tsaren KT masu adalci za su iya ba da damar tsarin koyo masu daidaitawa na gaske a cikin MOOCs (kamar Coursera, edX) da tsarin koyarwa mai hankali, tare da tabbatar da cewa shawarwari suna da tasiri ga dukkan al'umma.
Binciken Karkata don EdTech: Wannan tsari yana ba da tsarin aiki don bincika software na ilimi na kasuwanci don karkatarwar algorithm, wanda ke damun masu tsari da malamai.
Adalci a Fannoni Daban-daban: Aikin gaba ya kamata ya bincika adalci a kan wasu halaye masu mahimmanci: jinsi, shekaru, matsayin tattalin arziki da aka samo daga bayanai, da nakasar koyo.
Nazarin Adalci na Dalili: Ƙaura daga alaƙa don fahimtar dalilan karkata—shin bayanai ne, ginin tsari, ko yanayin koyo? Ana iya haɗa dabarun ƙididdiga na dalili.
Koyo Mai Adalci na Tarayya & Kare Sirri: Horar da tsare-tsare masu adalci akan bayanan masu amfani marasa tsari ba tare da lalata sirri ba, wata mahimmanciyar hanya don AI mai da'a a ilimi.

7. Nassoshi

Baker, R.S., Inventado, P.S. (2014). Educational Data Mining and Learning Analytics. In: Larusson, J., White, B. (eds) Learning Analytics. Springer, New York, NY.
Corbett, A. T., & Anderson, J. R. (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge. User modeling and user-adapted interaction, 4(4), 253-278.
Piech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L. J., & Sohl-Dickstein, J. (2015). Deep knowledge tracing. Advances in neural information processing systems, 28.
Barocas, S., Hardt, M., & Narayanan, A. (2019). Fairness and Machine Learning: Limitations and Opportunities. fairmlbook.org.
Duolingo. (2018). Second Language Acquisition Modeling (SLAM) Workshop Dataset. Retrieved from https://sharedtask.duolingo.com/
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54(6), 1-35.

8. Nazarin Kwararru & Sharhi

Fahimta ta Asali: Wannan takarda ta kawo wata gaskiya mai mahimmanci, wacce ake yawan yin watsi da ita a cikin EdTech: daidaito mai girma baya daidaita da ilimi mai daidaito. Marubutan sun nuna cikin gaskiya cewa daidaitattun tsare-tsaren Bin Ƙwararrun Ilimi, idan aka tura su cikin sauki, suna cimma rashi ga gaba ɗaya ƙungiyoyin masu koyo—musamman, waɗanda ke amfani da dandamalin yanar gizo da waɗanda ke cikin ƙasashe masu tasowa. Binciken da ya fi ban mamaki shi ne cewa sauƙaƙan tsare-tsaren Koyon Injini ba kawai sun fi ƙarancin daidaito ba; suna da adalci ƙasa da yawa, suna aiki azaman masu haɓaka rarrabuwar kawuna na zamantakewa da na dijital da ke akwai. Wannan yana sanya adalcin algorithm ba a matsayin damuwa na ɗabi'a na musamman ba, amma a matsayin babban ɓangaren aikin tsari da ingancin koyarwa.

Tsarin Ma'ana: Hujja tana da tsari. Ta fara ne da kafa manyan matakai (ilimi na keɓance) da makafin tarihi (adalci). Sannan ya kafa gwaji mai tsafta, na kwatance na binary (ML vs. DL) a cikin yanayin koyon harshe guda uku daban-daban. Zaɓin gatari na adalci—dandamali da yanayin ƙasa—yana da hikima, yana nuna ma'auni na turawa na ainihi waɗanda ke tasiri kai tsaye ga ƙwarewar mai amfani. Sakamakon yana gudana cikin ma'ana: ƙarfin wakilci mafi girma na DL yana haifar da ba kawai hasashe mafi kyau ba, amma mafi adalci. Shawarar da ba ta dace ba (DL don en_es/es_en, ML don fr_en) tana da sabuntawa, tana guje wa dokar da ta dace da kowa kuma tana yarda da dogaro da yanayi, alamar nazari mai zurfi.

Ƙarfi & Kurakurai: Babban ƙarfinsa shine mayar da hankali kan aiki, na zahiri. Ya wuce tattaunawar adalci na ka'ida don ba da shaidar aunawa na karkata a cikin bayanan da aka yi amfani da su sosai (Duolingo). Wannan tsari ne mai ƙarfi don binciken tsari na ciki. Duk da haka, binciken yana da iyakoki. Yana ɗaukar "masu ci gaba" da "masu tasowa" a matsayin tubalan guda ɗaya, yana ɓoye babban bambance-bambance a cikin waɗannan nau'ikan (misali, masu amfani na birane da na karkara). Binciken kuma bai shiga cikin dalilin da ya sa karkatarwar ta wanzu ba. Shin wakilcin fasali ne, ƙarar bayanai a kowace ƙungiya, ko bambance-bambancen al'adu a cikin tsarin koyo? Kamar yadda aka lura a cikin cikakken binciken na Mehrabi et al. (2021), gano tushen karkata yana da mahimmanci don haɓaka ingantattun hanyoyin ragewa. Bugu da ƙari, duk da cewa DL ya bayyana mafi adalci a nan, yanayinsa na "akwatin baƙi" zai iya ɓoye ƙarin karkata masu sauƙi, waɗanda ba a iya gano su cikin sauƙi ba, ƙalubale da aka nuna a cikin wallafe-wallafen adalci.

Fahimta Mai Aiki: Ga shugabannin EdTech da manajoji samfur, wannan binciken umarni ne na canji. Na farko, dole ne a haɗa ma'auni na adalci cikin daftarin aiki na kimanta tsari na yau da kullun, tare da daidaito da AUC. Kafin tura kowane fasalin koyo mai daidaitawa, gudanar da bincike irin na wannan binciken. Na biyu, ba da fifiko ga gine-ginen Koyo Mai Zurfi don ayyukan ƙirƙira ɗalibi na asali, kamar yadda suke ba da kariya mafi kyau ga karkata, suna tabbatar da yanayin da aka gani a wasu fannoni inda cibiyoyin sadarwa masu zurfi suka koyi fasali masu ƙarfi. Na uku, raba bayanan ku. Kar ku kalli aikin "duniya" kawai. Rarraba ma'auni ta dandamali, yanki, da sauran bayanan al'umma masu dacewa a matsayin aiki na yau da kullun. A ƙarshe, saka hannun jari a cikin nazarin dalili don matsawa daga lura da karkata zuwa fahimta da kuma fitar da shi ta hanyar injiniya. Makomar EdTech mai daidaito ya dogara ne da kula da adalci da ƙwazo iri ɗaya da daidaiton hasashe.