【論文紹介】ImRec: Learning Reciprocal Preferences Using Images / recsys2020-imrec

2020年10月17日 RecSys2020論文読み会（オンライン） (https://connpass.com/event/189192/) における発表資料です。

以下の論文について概要を紹介しました。

- J Neve, R McConville.2020. Proceedings of the 14th ACM Conference on Recommender Systems, 170-179.
- https://dl.acm.org/doi/10.1145/3383313.3411476

最近盛り上がってきている相互推薦システム（Reciprocal Recommender Systems）についての論文で、オンラインデーティングサービスにおけるユーザのプロフィール画像を用いた内容ベースのアルゴリズム ImRec の提案がなされています。

Transcript

1. Yuya Matsumura @yu-ya4
   ʲ࿦จ঺հʳ
   ImRec: Learning Reciprocal Preferences Using Images
   RecSys2020 ࿦จಡΈձʢΦϯϥΠϯʣ
   17 Oct. 2020
   J Neve, R McConville.2020. Proceedings of the 14th ACM Conference on Recommender Systems, 170-179.
   Photo by Agustín Diaz on Unsplash
   

   View Slide

2. ✓ দଜ ༏໵ʢYuya Matsumuraʣ
   ✓ ژ౎େֶେֶӃ৘ใֶݚڀՊ म࢜՝ఔमྃʢాதݚڀࣨ -> ٢઒ݚڀࣨʣ
   ✓ Wantedly, Inc. Matching Squad / Visit Data Chapter
   ✓ νʔϜϦʔυɺσʔλαΠΤϯεɺϓϩμΫτϚωδϝϯτ
   ✓ Wantedly Visit ʹ͓͚ΔσʔλΛ׆༻ͨ͠ϓϩμΫτʢਪનγεςϜʣͷ։ൃ
   @yu-ya4
   @yu__ya4
   ࣗݾ঺հ
   

   View Slide

3. ૬ޓਪનγεςϜʢReciprocal Recommender Systemsʣͱ͸ʁ
   

   View Slide

4. ʮαʔϏε಺ͷϢʔβΛޓ͍ʹਪન͢ΔγεςϜʯ
   User Item
   ैདྷͷҰൠతͳਪનγεςϜ ૬ޓਪનγεςϜ
   User(Female)
   User(Male)
   User(Job Seeker) User(Recruiter/Company)
   ex. Amazon, Netflix
   ex. Tinder, Pairs
   ex. Wantedly, LinkedIn
   ૬ޓਪનγεςϜʢReciprocal Recommender Systemsʣͱ͸ʁ
   

   View Slide

5. 1. ਪનͷ"੒ޭ"
   • ਪનΛड͚ΔϢʔβ (subject) ͱਪન͞ΕΔϢʔβ (object) ͷ྆ํͷᅂ޷͕Ұகͯ͠ॳΊͯਪન͕"੒ޭ"
   2. ਪનγεςϜʹର͢ΔϢʔβͷೝࣝ
   • subject Ϣʔβࣗ਎͕ "૬ޓਪનγεςϜ"Ͱ͋Δ͜ͱΛೝ্ࣝͨ͠Ͱɺobject Ϣʔβͷᅂ޷Λߟ্ྀͨ͠Ͱߦಈ
   3. Ϣʔβ͔Β໌ࣔతʹఏڙ͞ΕΔϓϩϑΟʔϧ΍ᅂ޷ʹ͍ͭͯͷ৘ใ
   • Ϣʔβ͕ࣗ਎Ͱఏڙ͢Δ໌ࣔతͳϓϩϑΟʔϧ΍ᅂ޷৘ใ͕๛෋͕ͩɺͦΕΒ͕ඞͣ͠΋ਖ਼͘͠ͳ͍
   4. ϢʔβͷϥΠϑαΠΫϧ
   • ϢʔβͷϥΠϑαΠΫϧ͕୹͘ɺϢʔβͷߦಈʹجͮ͘҉໧తͳϢʔβͷᅂ޷৘ใ͕ू·Γʹ͍͘
   5. ϢʔβͷγεςϜ಺ʹ͓͚Δߦಈͷ࢟੎
   • ೳಈతͳϢʔβͱडಈతͳϢʔβ͕ଘࡏ͠ɺडಈతͳϢʔβ͸ਪન͞Εͳ͍ͱ੒ޭମݧΛಘ೉͍
   ૬ޓਪનγεςϜͷಛ௃΍՝୊
   

   View Slide

6. ૬ޓਪનγεςϜͷʢయܕతͳʣ࢓૊Έ
   Aggregation
   Function
   User A to User B
   Preference Score
   User B
   User A
   User B to User A
   Preference Score
   User B
   User A
   Reciprocal
   Preference Score
   User B
   User A
   Arithmetic Mean, Harmonic Mean, or ...
   

   View Slide

7. ؔ࿈ݚڀ/طଘख๏ 1/2
   • [RECON] (Pizzato 2010) RECON: A Reciprocal Recommender for Online Dating
   • աڈʹ޷ҙΛࣔͨ͠૬खϢʔβͷ໌ࣔతͳϓϩϑΟʔϧʹڞ௨͢Δཁૉ͔ΒɼϢʔβ͝ͱͷ҉໧తͳᅂ޷৘ใʢetc. ମܕɼੑ
   ֨ɼֶྺ…ʣΛநग़
   • ᅂ޷৘ใͱ૬खϢʔβͷ໌ࣔతͳϓϩϑΟʔϧͷҰக౓Λܭࢉ͢ΔίϯςϯπϕʔεͳΞϓϩʔνʹΑΓ୯ํ޲ͷద߹౓Λද
   ͢ Preference Score Λࢉग़
   • ̎ͭͷ Preference Score ͔ΒAggregation Functionʢௐ࿨ฏۉʣͰ૒ํ޲ͷద߹౓Λද͢ Reciprocal Preference Score Λࢉग़
   • [RCF] (Xia 2015) Reciprocal Recommendation System for Online Dating
   • ߦಈཤྺʢLike or Nopeʣʹجͮ͘ϢʔβϕʔεͷڠௐϑΟϧλϦϯάʢk-ۙ๣ʣͰ Preference Score Λࢉग़
   • Aggregation Function ͸ௐ࿨ฏۉΛར༻
   • RECON (Pizzato 2010) ͷੑೳΛେ෯ʹ௒͑Δ
   

   View Slide

8. • [LFRR](Neve 2019) Latent Factor Models and Aggregation Operator for Collaborative Filtering
   in Reciprocal Recommender Systems
   • ߦಈཤྺʹج͖ͮ࡞੒ͨ͠ User-User ߦྻʹ Matrix Factorization Λద༻ͯ͠ Preference Score Λࢉग़
   • Aggregation Function ʹ͍ͭͯɼௐ࿨ฏۉҎ֎ͷؔ਺ʹ͍ͭͯ΋ൺֱ࣮ݧʢCross-Ratio Uninormʣ
   • ੑೳ͸ RCF (Xia 2015) ͱಉఔ౓͕ͩɼࣄલʹϞσϧΛֶश͓͚ͯ͠ΔͷͰߴ଎
   ؔ࿈ݚڀ/طଘख๏ 2/2
   https://speakerdeck.com/yuya4/latent-factor-models-and-aggregation-operators-for-collaborative-
   filtering-in-reciprocal-recommender-systems
   

   View Slide

9. ʲ࿦จ঺հʳ*N3FD-FBSOJOH3FDJQSPDBM1SFGFSFODFT6TJOH*NBHFT
   J Neve, R McConville.2020. Proceedings of the 14th ACM Conference on Recommender Systems, 170-179.
   

   View Slide

10. Backgrounds / Motivation
    • ૬ޓਪનγεςϜ͸·ͩ·ͩݚڀ͕ਐΜͰͳ͍ྖҬ
    • ڠௐϑΟϧλϦϯάΞϓϩʔνͷݚڀ͸͍͔ͭ͘ݟΒΕΔΑ͏ʹͳ͕ͬͨɺ಺༰ϕʔεϑΟϧλϦϯ
    άΞϓϩʔνͷݚڀ͸ RECON (Pizzato 2010) Ҏ߱͋·ΓਐΜͰ͍ͳ͍
    • RECON ͸ϓϩϑΟʔϧͷΧςΰϦσʔλͷΈΛར༻ͨ͠΋ͷ
    • ҰํͰɺσʔςΟϯάαʔϏεʹ͓͍ͯը૾͸େมॏཁͳཁૉɻϓϩϑΟʔϧͷςΩετ৘ใΛݟͣ
    ʹը૾͚ͩͰ޷͖ݏ͍Λ൑அ͢Δ͜ͱ͕ଟ͍ͱ͍͏σʔλ΋ɻʢTinder ͷྫΛग़͍ͯͨ͠ʣ
    ը૾σʔλΛ༻͍ͨ಺༰ϕʔεϑΟϧλϦϯάͷΞϧΰϦζϜͷఏҊ
    

    View Slide

11. Contributions
    1. ը૾σʔλͱᅂ޷σʔλʹجͮ͘ personal attractiveness (not general
    attractivenessʣͷ༧ଌϞσϧͷ࡞੒
    2. personal attractiveness ϕʔεͷ૬ޓਪનγεςϜʢImRecʣͷߏங
    3. ࣮ࡏ͢ΔσʔςΟϯάαʔϏεʢpairsʣͷσʔλΛ༻͍ͯੑೳͷݕূΛߦ͍ɺ
    RECON Λ௒͑ΔੑೳΛ֬ೝ
    personal attractiveness: how attractive is 's image to user
    general attractiveness: how attractive is to the average potential partner
    y x
    y
    

    View Slide

12. Network: predicting personal attractiveness
    • Ϣʔβ ʹରͯ͠ɺ ͕ Like ͨ͠Ϣʔβ (anchor user)ɺ ͱ͸ผͷ ͕ Like ͨ͠Ϣʔβ ɺ ͕
    Nope ͨ͠Ϣʔβ ͱ͍͏ triplet Λ࡞੒
    • → 1ɺ → 0 ͱͳΔΑ͏ʹɺͦΕͧΕͷϢʔβͷը૾Λೖྗͱͯ͠ Siamese Network Λֶश
    • ͕ Like ͨ͠ϢʔβͲ͏͕ۙ͘͠ɺLike ͨ͠Ϣʔβͱ Nope ͨ͠Ϣʔβ͕ԕ͘ͳΔΑ͏ʹֶश
    • 200,000 ϖΞͷإࣸਅΛར༻
    • 100,000 ͸ޓ͍ʹ Like ͯ͠ Match ͨ͠΋ͷɺ100,000 ͸Ұํ͸ Like ͯ͠΋͏Ұํ͕ Nope ͨ͠΋ͷ
    x x ya
    ya
    x yp
    x
    yn
    (ya
    , yp
    ) (ya
    , yn
    )
    x
    ڞ௨
    W
    

    View Slide

13. Recommender Systems Architecture
    1. Ϣʔβ ͱϢʔβ ͷϖΞʹରͯ͠ɺͦΕͧΕͷϢʔβ
    ͕ Like ͨ͠ϢʔβΛूΊΔʢ্ݶ30݅ʣ
    2. ( → Λߟ͑Δ) ͷը૾Λ ɺ ͕ Like ͨ͠ը૾Λ
    ͱͯ͠ɺ ʹର͢Δग़ྗʢෳ਺ʣΛ֫ಘ
    3. ෳ਺ͷ ʹର͢Δग़ྗΛ̍ͭͷ Preference Score
    ʹ aggregate ͢Δʢࣄલʹ܇࿅͓͍ͯͨ͠ RFʣ
    4. → ɺ ̎ͭͷ Preference Score Λ Aggregate
    ʢௐ࿨ฏۉʣͯ͠ɺReciprocal Preference Score Λ֫ಘ
    x y
    x y y ya
    x
    yp
    (ya
    , yp
    )
    (ya
    , yp
    )
    x y y → x
    

    View Slide

14. Experiments
    • ϓϩϑΟʔϧͷΧςΰϦσʔλʹج͍ͮͨ಺༰ϕʔεͷ RECON ͱͷൺֱ࣮ݧ
    • ڠௐϑΟϧλϦϯάϕʔεͷ LFRR ͱͷൺֱ࣮ݧ
    • ಛʹɺcold-start ͷঢ়گʹ͓͚ΔൺֱΛߦͬͨ
    • ʢCV on 20,000 pairs of usersʣ
    

    View Slide

15. Result ᶃ ImRec vs RECON
    • ImRec ͷํ͕େ͖͘ྑ͍݁Ռʹ
    • RECON ͸΄΅ϥϯμϜͳਪનͱಉ͘͡Β͍ͷਫ਼౓ʹ
    • RECON ͷΦϦδφϧͳ࣮ݧͰ༻͍ΒΕͨσʔληο
    τ͸ɺϓϩϑΟʔϧ৘ใ͕ॏཁͳ΋ͷͰ͋ͬͨʢը૾
    σʔλ͸ͳ͔ͬͨʁʣ
    

    View Slide

16. Result ᶄ ImRec vs LFRR (cold-start situation)
    • cold-start ͷঢ়گΛγϛϡϨʔγϣϯ͢ΔͨΊʹɺ
    Ϣʔβ͋ͨΓͷᅂ޷σʔλͷ਺ʹ੍ݶΛֶ͔͚ͯश
    ΛߦͬͨϞσϧͰൺֱ࣮ݧ
    • Indicator: ̍Ϣʔβ͋ͨΓͷ Like ͱ Nope ͷσʔλ
    ͕ ݸͣͭ
    • ֶशσʔλ͕গͳ͍ͱ͖ʢcold-start ͕ݦஶͳͱ
    ͖ʣɺLFRR ͸΄΅ϥϯμϜͳਪનͱಉ͘͡Β͍ͷਫ਼
    ౓͕ͩɺImRec ͸ͦ͜·Ͱਫ਼౓͕མͪͳ͍
    • ֶशσʔλ͕૿͑ͯ͘ΔͱɺLFRR ͷਫ਼౓ͷํ͕ྑ͘
    ͳ͍ͬͯ͘
    • ॳճମݧ͕ॏཁͳσʔςΟϯάΞϓϦͰ͸ cold-start
    ໰୊΁ͷରॲ͕େ੾ͳͷͰɺ͜ΕΒΛ͏·͘૊Έ߹
    Θ͍ͤͯ͘ඞཁ
    K
    K
    

    View Slide

17. Conclusions
    • ը૾ʹج͍ͮͨ personal attractiveness Λ༧ଌ͢ΔϞσϧΛ࡞੒ͨ͠ɻ
    • ImRec ͸ίϯςϯπϕʔεͷ RRSs ͷϕʔεϥΠϯͰ͋Δ RECON ͷੑೳΛ্ճͬͨɻ
    • ڠௐϑΟϧλϦϯάϕʔεͷ RRSs ʹରͯ͠ɺcold-start ͳঢ়گʹ͓͍ͯੑೳΛ্ճͬͨɻ
    

    View Slide