Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Wantedly におけるマッチング体験を最大化させるための推薦システム

Shuhei Goda
January 17, 2024

Wantedly におけるマッチング体験を最大化させるための推薦システム

Recommendation Industry Talks #1 の登壇資料です
https://recommendation-industry-talks.connpass.com/event/305091/

概要:ウォンテッドリーでは、多様なユーザーと会社の理想的なマッチングを実現するために、会社訪問アプリ「Wantedly Visit」の推薦システム開発に力を入れています。サービスにおけるマッチング体験を向上させるための推薦システム開発に関する知見や具体的な取り組みについて紹介します。

Shuhei Goda

January 17, 2024
Tweet

More Decks by Shuhei Goda

Other Decks in Technology

Transcript

  1. © 2024 Wantedly, Inc. ໊લɿ 
 ߹ా पฏ(Shuhei Goda) ॴଐͱ໾ׂɿ

    
 ɾ΢ΥϯςουϦʔגࣜձࣾ(2019~) 
 ɾਪનνʔϜϦʔμʔ 
 ɾML Tech Lead & Product Manager 
 
 Kaggle Tierɿ 
 Kaggle Competitions Grandmaster @jy_msc ࣗݾ঺հ
  2. © 2024 Wantedly, Inc. iOS, Android and Web ؾܰʹձࣾ๚໰ ϛογϣϯ΍Ձ஋؍΁ͷڞײͰϚονϯά

    • څ༩΍෱རްੜͳͲͷ৚݅Ͱ͸ͳ͘ɺ૝͍͕͋Ε͹ձࣾͷ ن໛ʹͱΒΘΕͳ͍ ·ͣʮ࿩Λฉ͖ʹߦ͘ʯͱ͍͏৽͍͠ମݧ • ݸਓͱاۀ͕ϑϥοτͳ໨ઢͰग़ձ͑Δ͜ͱͰɺΑΓັྗ తͳ৔ॴΛݟ͚ͭΔ͜ͱ͕Մೳʹ ձࣾ๚໰ΞϓϦʮWantedly Visitʯ
  3. © 2024 Wantedly, Inc. ϚονϯάମݧͷϞσϧԽ Ϣʔβʔ اۀ 1. ίϯςϯπͷӾཡ 2.

    Ԡื 3. ϝοηʔδฦ৴ p(match = 1|u, c) = p(exam = 1|u, c) × p(apply = 1|u, c, exam = 1) × p(reply = 1|u, c, apply = 1) Ϣʔβʔ͸ͲͷاۀΛݟΔ͔ Ϣʔβʔʹͱͬͯ 
 اۀͷίϯςϯπ͸ັྗత͔ اۀʹͱͬͯ 
 ϢʔβʔͷϓϩϑΟʔϧ͸ັྗత͔ ※ຊൃද༻ʹ؆қԽ
  4. © 2024 Wantedly, Inc. αʔϏεͷ֦େʹ൐ͬͯັྗతͳاۀΛݟ͚ͭΔ͜ͱ͕ࠔ೉ʹ • Ϣʔβʔ΍اۀͷ਺͕େ͖͘૿Ճ • Ϣʔβʔ΍اۀͷଐੑ΍ᅂ޷ੑ͕ଟ༷Խ αʔϏεͷϑΣʔζ͕ਐΉதͰݦࡏԽ͢Δ՝୊

    p(exam = 1|u, c) × p(apply = 1|u, c, exam = 1) × p(reply = 1|u, c, apply = 1) Ϣʔβʔ͸ͲͷاۀΛݟΔ͔ Ϣʔβʔʹͱͬͯ 
 اۀͷίϯςϯπ͸ັྗత͔ اۀʹͱͬͯ 
 Ϣʔβʔͷίϯςϯπ͸ັྗత͔ ͕͜͜େ͖ͳϘτϧωοΫʹ αʔϏε಺ͷෳࡶ౓͕૿͍ͯ͘͠ 
 ͜ͱʹΑΔϛεϚονͷ૿Ճ
  5. © 2024 Wantedly, Inc. ਪનٕज़Λ׆༻͢Δ͜ͱͰϚονϯάΛੜΈग़͢ • ͲͷϢʔβʔʹͲ͏͍͏اۀΛදࣔ͢Δͷ͔ • ͲͷاۀʹͲ͏͍͏ϢʔβʔΛදࣔ͢Δͷ͔ දࣔΛ੍ޚ͢ΔͨΊͷιϦϡʔγϣϯ

    p(exam = 1|u, c, rec) × p(apply = 1|u, c, exam = 1) × p(reply = 1|u, c, apply = 1) Ϣʔβʔ͸ͲͷاۀΛݟΔ͔ Ϣʔβʔʹͱͬͯ 
 اۀͷίϯςϯπ͸ັྗత͔ اۀʹͱͬͯ 
 Ϣʔβʔͷίϯςϯπ͸ັྗత͔ ͜͜Λ੍ޚ͢Δ p(match = 1|u, c, rec) =
  6. © 2024 Wantedly, Inc. Ϛονϯάͷਪનͷ੒ޭɾࣦഊ Ϣʔβʔ اۀ 1. Ԡื 2.

    ϝοηʔδฦ৴ ੒ޭ Ϣʔβʔ اۀ 1. Ԡื ʢϝοηʔδฦ৴ͳ͠ʣ ࣦഊ
  7. © 2024 Wantedly, Inc. ૬ޓਪનγεςϜʢReciprocal Recommender Systemsʣ αʔϏε಺ͷϢʔβʔΛޓ͍ʹਪન͠߹͏γεςϜ • ਪનΛड͚औΔϢʔβʔͱਪન͞ΕͨϢʔβʔͷ྆ํ͕ຬ଍ͯ͠੒ޭ

    • ਪનΛड͚औΔϢʔβʔ͕ຬ଍͢Ε͹੒ޭ͢Δ item-to-user ਪનͱ͸ 
 ੒ޭͷج४͕ҟͳΔ item-to-user ਪનͷ੒ޭ Ϣʔβʔ ΞΠςϜ ૬ޓਪનͷ੒ޭ ߪೖ Ԡื ฦ৴ Ϣʔβʔ اۀ ฦ৴ εΧ΢τ Ϣʔβʔ اۀ
  8. © 2024 Wantedly, Inc. 1. ϢʔβʔA→ϢʔβʔBͱɺϢʔβʔB→ϢʔβʔAͷ Preference Score Λਪ࿦ 2.

    ૬ޓͷ Preference Score Λू໿͢Δ ૬ޓਪનͷجຊతͳΞʔΩςΫνϟ User A to User B Preference Score User B to User A Preference Score Reciprocal Preference Score Aggregation Function ू໿ؔ਺͸ྫ͑͹ࢉज़ฏۉɺزԿฏۉɺௐ࿨ฏۉͳͲ 
 ྆ํͱ΋είΞ͕ߴ͚Ε͹ߴ͘ɺͲͪΒ͔௿͍ͱ௿͘ͳΔ 
 ੑ࣭Λ࣋ͭؔ਺͕ద੾ͱݴΘΕ͍ͯΔ ௐ࿨ฏۉ ࢉज़ฏۉ
  9. © 2024 Wantedly, Inc. Item-to-user ਪનΛ࠾༻͍ͯͨ࣌͠ظ͕͋ͬͨ • ඃਪનϢʔβʔͷຬ଍౓Λఆٛͯ͠ద߹౓Λઃܭ͠ɺͦΕΛ࠷దԽ ͔ͭͯͷԠืͱεΧ΢τͷਪનʹ͍ͭͯ Ϣʔβʔ

    اۀ ԠืλεΫ εΧ΢τλεΫ اۀ Ϣʔβʔ ྫ͑͹ɺ 
 ɾϚονϯά=2 
 ɾϚονϯά੒ཱ͠ͳ͔ͬͨԠื=1 
 ɾԠื͠ͳ͍=0 ྫ͑͹ɺ 
 ɾϚονϯά=2 
 ɾϚονϯά੒ཱ͠ͳ͔ͬͨεΧ΢τ=1 
 ɾεΧ΢τ͠ͳ͍=0
  10. © 2024 Wantedly, Inc. • ద߹౓͸ਓखͰઃఆ͍ͯ͠Δ͜ͱͰϚονϯά࠷దԽ͕ࠔ೉ʹ • Ϛονϯά੒ཱͨ͠ΞΠςϜ͸ɺϚονϯά੒ཱ͠ͳ͍͚ͲԠืͨ͠ 
 ΞΠςϜͷ2ഒຬ଍ͱ͢ΔɻͳΜͰ4ഒ͸ବ໨ʁ

    • Ϟσϧ಺Ͱಛ௃ྔ͕ҙਤ௨Γʹ༗ޮతʹ࢖ΘΕͳ͍ • ϚονϯάʹࢸΔ·Ͱͷεςοϓຖʹಉ͡ಛ௃ྔͰ΋ར༻ҙਤ͸ҟͳΔ • ϚονϯάϞσϧΛ1ͭͷϞσϧʹԡ͠ࠐΊ͍ͯΔ Item-to-user ਪનͷ՝୊ ※ϓϩμΫτ՝୊͸ࠓճ৮Εͳ͍ Ϣʔβʔ اۀ 1. ίϯςϯπͷӾཡ 2. Ԡื 3. ϝοηʔδฦ৴ User to Company Preference Model ద߹౓
  11. © 2024 Wantedly, Inc. Item-to-user ਪન͔Β૬ޓਪન΁ • ԠืͱεΧ΢τͱ΋ʹओཁKPI͕େ͖͘৳ͼͨ 🎉🎉🎉 ԠืͱεΧ΢τͷਪનվળΛ࣮ࢪ

    User to Company Preference Model User to Company Preference Model Company to User Preference Model Aggregation Function