Koyon Kwatancen Niyya don Shawarwarin Tsari (ICLRec)

1. Gabatarwa

Shawarwarin Tsari (SR) yana nufin hasashen mu'amalar mai amfani na gaba bisa jerin halayensa na tarihi. Duk da cewa samfurori na zurfin koyo sun sami mafi kyawun aiki, sau da yawa suna yin watsi da ainihin niyyoyi masu boye da ke motsa halayen mai amfani (misali, "siyan kayan kamun kifi," "shirye-shiryen hutu"). Waɗannan niyyoyin ba a gani ba amma suna da mahimmanci don fahimtar dalilin mai amfani da inganta daidaiton shawara da ƙarfi, musamman a cikin yanayin bayanai marasa yawa ko masu hayaniya.

Wannan takarda ta gabatar da Koyon Kwatancen Niyya (ICL), wani sabon tsari wanda ke shigar da mai canjin niyya mai boye cikin samfuran SR. Babban ra'ayi shine koyon rarraba niyyar mai amfani daga jerin da ba a yiwa lakabi ba da kuma inganta samfurin SR ta amfani da koyon kai-kai na kwatancen, daidaita ra'ayoyin jerin tare da niyyoyinsu masu dacewa.

2. Bayan Fage & Ayyukan Da Suka Danganta

2.1 Shawarwarin Tsari

Samfura kamar GRU4Rec, SASRec, da BERT4Rec suna ɗaukar yanayin lokaci amma yawanci suna ƙirƙirar hali azaman jerin abubuwa kai tsaye, suna rasa siginonin niyya mafi girma.

2.2 Tsarin Niyya

Samfuran da suka sani da niyya a baya sau da yawa sun dogara da bayanan gefe na zahiri (misali, tambayoyi, rukuni). ICL ta ƙirƙira ta hanyar koyon niyyoyi kai tsaye daga jerin halaye na ɓoye.

2.3 Koyon Kwatancen

An yi wahayi ta hanyar nasarori a cikin hangen nesa na kwamfuta (misali, SimCLR, MoCo) da NLP, koyon kwatancen yana haɓaka yarjejeniya tsakanin ra'ayoyin da aka haɓaka daban-daban na bayanai iri ɗaya. ICL ta daidaita wannan don daidaita jerin halaye tare da niyyoyinsu masu boye.

3. Hanyar Aiki: Koyon Kwatancen Niyya (ICL)

3.1 Tsarin Matsala

Idan aka ba mai amfani $u$ tare da jerin mu'amala $S^u = [v_1^u, v_2^u, ..., v_t^u]$, manufar ita ce hasashen abu na gaba $v_{t+1}^u$. ICL ta gabatar da mai canjin niyya mai boye $z$ don bayyana jerin.

3.2 Mai Canjin Niyya Mai Boye

An ƙirƙira niyyar $z$ azaman mai canjin rukuni wanda ke wakiltar ainihin dalilin jerin. Samfurin yana koyon rarraba $p(z | S^u)$.

3.3 Koyon Rarraba Niyya Ta Hanyar Rarrabawa

Wakilcin jerin mai amfani ana rarraba su (misali, ta amfani da K-means) don gano samfuran niyya masu boye $K$. Kowane tsakiyar rukunin yana wakiltar wata niyya.

3.4 Koyon Kai-Kai na Kwatancen

Babban siginonin koyo ya fito ne daga asarar kwatancen. Don jerin $S$, an ƙirƙiri ra'ayoyi biyu da aka haɓaka ($S_i$, $S_j$). An horar da samfurin don jawo wakilcin jerin da wakilcin rukunin niyyar da aka sanya masa kusa da juna, yayin da ake tura shi daga sauran niyyoyi. Asarar kwatancen don ma'aurata masu kyau (jerin, niyyarsa) ya dogara ne akan asarar InfoNCE:

$\mathcal{L}_{cont} = -\log \frac{\exp(\text{sim}(f(S), g(z)) / \tau)}{\sum_{z' \in \mathcal{Z}} \exp(\text{sim}(f(S), g(z')) / \tau)}$

inda $f$ shine mai ɓoye jerin, $g$ shine aikin haɗa niyya, $\text{sim}$ shine aikin kamanceceniya (misali, cosine), kuma $\tau$ shine sigar zafin jiki.

3.5 Horarwa Ta Hanyar Tsarin EM Gabaɗaya

Horarwa yana musanya tsakanin matakai biyu a cikin tsarin Gabaɗayan Tsammanin-Matsakaici (EM):

1. Mataki-E (Ƙididdigar Niyya): Ƙididdige rarraba bayan baya na ainihin niyya mai boye $z$ don kowane jerin da aka ba da sigogin samfurin na yanzu.
2. Mataki-M (Sabunta Samfurin): Sabunta sigogin samfurin SR ta hanyar haɓaka tsammanin log-likelihood, wanda ya haɗa da asarar hasashen abu na gaba na yau da kullun da asarar kwatancen $\mathcal{L}_{cont}$.

Wannan tsari na maimaitawa yana inganta duka fahimtar niyya da ingancin shawara.

4. Gwaje-gwaje & Sakamako

4.1 Bayanan Gwaji & Ma'auni

An gudanar da gwaje-gwaje akan bayanan duniya na gaske guda huɗu: Beauty, Sports, Toys, da Yelp. Ma'auni sun haɗa da samfuran SR na zamani (SASRec, BERT4Rec) da hanyoyin kai-kai (CL4SRec).

4.2 Kwatancen Aiki

ICL ta sami ingantacciyar ci gaba a cikin ma'auni na Tunawa da NDCG (misali, +5.2% NDCG@10 akan Beauty akan mafi kyawun ma'auni), yana nuna tasirin ƙirar niyya mai boye.

4.3 Bincike na Ƙarfi

Babban gudunmawa shine ingantaccen ƙarfi. ICL ta nuna mafi kyawun aiki a ƙarƙashin ƙarancin bayanai (ta amfani da gajerun jerin) da kuma a gaban mu'amaloli masu hayaniya (abubuwan da ba su dace ba da aka saka a bazuwar). Koyon kwatancen matakin niyya yana ba da siginoni masu daidaitawa waɗanda ba su da ƙarfi ga abubuwan hayaniya na mutum ɗaya.

4.4 Nazarin Cire Sassa

Nazarin cire sassa ya tabbatar da wajibcin duka sassa biyu: (1) cire asarar kwatancen ya haifar da faɗuwa mai mahimmanci, kuma (2) amfani da niyyoyi masu ƙayyadaddun/ba da gangan ba maimakon waɗanda aka koya suma sun cutar da aikin, yana tabbatar da ƙirar haɗin koyon niyya da daidaitawar kwatancen.

5. Muhimman Bayanai & Bincike

Babban Bayani: Babban nasarar takardar ba wai kawai wata dabara ta kwatancen ba ce; ita ce sake gabatar da ƙirar mai canji mai boye cikin masu ba da shawara na zamani na zurfi. Duk da yake samfura kamar SASRec ƙwararrun masu koyon jerin ne, su ainihin "baƙar fata" ne na kai-da-kai. Hazakar ICL ta ta'allaka ne wajen tilasta samfurin ya bayyana jerin ta hanyar ainihin niyya mai boye mai fassara $z$, ƙirƙirar kwalbar kwalbar da ke tace hayaniya da kuma ɗaukar "dalilin" bayan "abin da". Wannan yana tunawa da canjin falsafa a cikin samfuran samarwa kamar VAE, amma an yi amfani da su don nuna bambanci don shawara.

Kwararar Ma'ana: Hanyar aiki tana da sauƙi mai kyau. 1) Rarraba jerin don samun samfuran niyya (wakilin Mataki-E). 2) Yi amfani da waɗannan samfuran a matsayin madaidaici don asarar kwatancen. 3) Asarar kwatancen tana horar mai ɓoye jerin don samar da wakilcin da suka dace da waɗannan madaidaitan ma'anoni. 4) Wannan daidaitawar, bi da bi, yana inganta rukunin da manufar shawara gabaɗaya. Zagaye ne mai kyau na koyon wakilci da rarrabawa, wanda tsarin EM ya daidaita—wani ra'ayi na gargajiya da aka yi da ƙarfi tare da koyon kwatancen na zamani.

Ƙarfi & Kurakurai: Babban ƙarfi shine ƙarfi da aka nuna ta hanyar gwaji. Ta hanyar koyo a matakin niyya, samfurin ya zama ƙasa da rauni ga ƙarancin bayanai da hayaniya—kuskure mai mahimmanci a yawancin masu ba da shawara masu zurfi. Tsarin kuma ba ya sani ga tushen gine-ginen SR. Duk da haka, babban aibi shine tsammanin niyya mai tsayi. Samfurin yana ɗaukar ainihin niyya guda ɗaya a kowane jerin, amma a zahiri, zaman mai amfani na iya zama mai fuskoki da yawa (misali, bincika don kyauta da kuma kansa). Matakin rarrabawa kuma yana gabatar da hyperparameters (adadin niyyoyi K) da yuwuwar hankali ga farawa, wanda takardar ta yi watsi da shi. Idan aka kwatanta da ƙarin hanyoyin raba niyya masu ƙarfi a cikin binciken RL ko bincike, wannan mafita ce mai ƙaramin ƙima.

Bayanai Masu Aiki: Ga masu aiki, abin da za a ɗauka a bayyane yake: Shigar da tsarin da za a iya fassara cikin samfuran ku na zurfin koyo. Kar ku jefa manyan masu canzawa kawai a jerin. Tsarin ICL za a iya daidaita shi fiye da shawara—kowane aiki tare da hanyoyin mai amfani (misali, kewayawa UI, hanyoyin ilimi) zai iya amfana daga koyon kwatancen niyya mai boye. Mataki na gaba kai tsaye ga masu bincike ya kamata su canza wannan daga niyyoyi guda ɗaya, masu tsayi zuwa niyyoyi masu matsayi ko na tsari. Shin za mu iya ƙirƙirar yadda niyyar mai amfani ke tasowa yayin zaman? Ƙari ga haka, haɗa wannan tare da tsarin ƙididdiga na dalili zai iya raba ayyukan da niyya ke motsa su daga waɗanda ba su da mahimmanci, yana tura zuwa ga samfuran tsari na gaske masu bayyanawa da ƙarfi. Sakin lambar babbar fa'ida ce don kwafi da faɗaɗawa.

6. Cikakkun Bayanai na Fasaha & Tsarin Lissafi

Gabaɗayan aikin manufa ya haɗa da asarar hasashen abu na gaba na yau da kullun (misali, giciye-entropy) tare da asarar niyya ta kwatancen:

$\mathcal{L} = \mathcal{L}_{pred} + \lambda \mathcal{L}_{cont}$

Inda $\lambda$ ke sarrafa nauyin kalmar kwatancen. Asarar hasashe $\mathcal{L}_{pred}$ ita ce:

$\mathcal{L}_{pred} = -\sum_{u} \log P(v_{t+1}^u | S_{1:t}^u, z^u)$

An haɗa mai canjin niyya $z$ cikin mai ɓoye jerin. Misali, a cikin mai ɓoye na tushen canzawa, haɗa niyya $g(z)$ za a iya fara shi azaman alamar musamman `[NIYYA]` zuwa jerin abu, yana ba da damar samfurin ya halarci mahallin niyya lokacin samar da hasashe.

7. Tsarin Bincike: Misalin Lamari

Yanayi: Bincika zaman mai amfani akan dandalin kasuwanci na e-commerce.

Ba tare da ICL ba: Samfurin yana ganin jerin User A: ["takalmin tafiya", "kwalbar ruwa", "sanduna na kuzari"]. Yana hasashen "jakar baya" bisa tsarin haɗuwa.

Tare da ICL:

1. Rarrabawar Niyya: Samfurin ya koyi rukunin niyya don "Shirye-shiryen Waje." An sanya wakilcin jerin User A zuwa wannan rukunin.
2. Koyon Kwatancen: Yayin horo, wakilcin ["takalmin tafiya", "kwalbar ruwa", "sanduna na kuzari"] an ja shi kusa da haɗa niyya na "Shirye-shiryen Waje."
3. Ingantaccen Hasashe: A ƙididdiga, samfurin, sanin niyyar "Shirye-shiryen Waje," yanzu yana iya kuma ba da shawarar "magungunan sauro" ko "alkibla"—abubuwan da ke da alaƙa da niyya amma ba lallai ba ne tare da ainihin jerin tarihi—yana nuna mafi kyawun haɗawa da ƙarfi ga bayanai marasa yawa.

8. Aikace-aikacen Gaba & Hanyoyi

• Shawara Mai Yawan Yankuna & Tsallaken Dandamali: Niyyoyi masu boye (misali, "na motsa jiki") za a iya raba su a cikin yankuna (kayan wasanni, aikace-aikacen abinci mai gina jiki, abun ciki na bidiyo), yana ba da damar canja wurin koyo.
• AI Mai Bayyanawa (XAI): Bayar da shawarwari tare da alamun niyya ("An ba da shawara saboda da alama kuna shirin tafiyar kamun kifi") zai iya ƙara amincewar mai amfani da gamsuwa sosai.
• Tsarin Masu Ba da Shawara na Tattaunawa: Niyyoyi za su iya zama gada tsakanin tattaunawar harshe na halitta da shawarar abu, inganta daidaiton wakilan tattaunawa.
• Ƙirar Niyya Mai Ƙarfi: Faɗaɗa ICL don ƙirƙirar canjin niyya a cikin zaman ɗaya (misali, daga "bincike" zuwa "siye") ta amfani da hanyoyin lokaci-lokaci ko samfuran sararin samaniya.
• Haɗawa tare da Manyan Samfuran Harshe (LLMs): Amfani da LLMs don samar da cikakkun bayanai, na rubutu na rukunin niyyoyin da aka koya don mafi kyawun fassara, ko amfani da haɗa LLM don farawa da samfuran niyya.

9. Nassoshi

1. Chen, Y., Liu, Z., Li, J., McAuley, J., & Xiong, C. (2022). Intent Contrastive Learning for Sequential Recommendation. Proceedings of the ACM Web Conference 2022 (WWW '22).
2. Kang, W. C., & McAuley, J. (2018). Self-attentive sequential recommendation. 2018 IEEE International Conference on Data Mining (ICDM).
3. Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W., & Jiang, P. (2019). BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. Proceedings of the 28th ACM International Conference on Information and Knowledge Management.
4. Xie, X., Sun, F., Liu, Z., Wu, S., Gao, J., Zhang, J., ... & Cui, B. (2022). Contrastive learning for sequential recommendation. 2022 IEEE 38th International Conference on Data Engineering (ICDE).
5. Oord, A. v. d., Li, Y., & Vinyals, O. (2018). Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748.
6. Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.
7. Jannach, D., & Jugovac, M. (2019). Measuring the business value of recommender systems. ACM Transactions on Management Information Systems (TMIS), 10(4), 1-23.