ジョブマッチングサービスにおける相互推薦システムの応用事例と課題

Slide 1

Slide 1 text

Slide 2

Slide 2 text

© 2024 Wantedly, Inc. ໊લɿ   ߹ా पฏ Shuhei Goda ॴଐͱ໾ׂɿ   ΢ΥϯςουϦʔגࣜձࣾ   ɾData Team Manager   ɾMachine Learning Tech Lead   ɾProduct Manager     Kaggle Tierɿ   Kaggle Competitions Grandmaster @jy_msc ࣗݾ঺հ https://www.kaggle.com/shuheigoda

Slide 3

Slide 3 text

Slide 4

Slide 4 text

Slide 5

Slide 5 text

Slide 6

Slide 6 text

Slide 7

Slide 7 text

Slide 8

Slide 8 text

Slide 9

Slide 9 text

Slide 10

Slide 10 text

Slide 11

Slide 11 text

© 2024 Wantedly, Inc. Ϛονϯάޮ཰͕ߴ͘ҡ࣋͞ΕΔঢ়ଶΛͲ͏࣮ݱ͢Δ͔ p(match = 1|c, j) = p(exam = 1|c, j) × p(scout = 1|c, j, exam = 1) × p(reply = 1|c, j, scout = 1) اۀ͸ͲͷϢʔβʔΛ   ݟΔ͔ اۀʹͱͬͯͦͷϢʔβʔ͸   ັྗతʹײ͡ΒΕΔ͔ Ϣʔβʔʹͱͬͯͦͷืू͸   ັྗతʹײ͡ΒΕΔ͔ ਪનγεςϜͰϢʔβʔɾاۀ૒ํͷίϯςϯπͷදࣔΛ੍ޚ͢Δ ● To اۀɿاۀʹͱͬͯັྗతͰɺͦͷاۀͱϚονϯά͠΍͍͢ϢʔβʔΛදࣔ ● To ϢʔβʔɿϢʔβʔʹͱͬͯັྗతͰɺͦͷϢʔβʔͱϚονϯά͠΍͍͢ืूΛදࣔ j … Job seeker c … Company ͜͜ʹհೖ͢Δ p(match = 1| j, c) = p(exam = 1| j, c) × p(apply = 1| j, c, exam = 1) × p(reply = 1| j, c, apply = 1) Ϣʔβʔ͸ͲͷاۀΛ   ݟΔ͔ Ϣʔβʔʹͱͬͯͦͷืू͸   ັྗతʹײ͡ΒΕΔ͔ اۀʹͱͬͯͦͷϢʔβʔ͸   ັྗతʹײ͡ΒΕΔ͔ εΧ΢τ༝དྷ   ͷϚονϯά֬཰ Ԡื༝དྷ   ͷϚονϯά֬཰

Slide 12

Slide 12 text

© 2024 Wantedly, Inc. ղ͘΂͖ػցֶशλεΫ Ϛονϯά਺͕࠷େͱͳΔΑ͏ͳਪનϦετΛ֤ඃਪનऀʹରͯ͠࡞੒͢ΔλεΫ σ* c := argsortj∈ 𝒥 m(c, j) σ* j := argsortc∈ 𝒞 m(c, j) →֤اۀ c ʹͱͬͯཧ૝తͳϥϯΩϯάɺ m(c, j)͸ਅͷϚονϯά֬཰ →֤Ϣʔβʔ j ʹͱͬͯཧ૝తͳϥϯΩϯά • ཧ૝తͳϥϯΩϯά͸ਅͷϚονϯά֬཰ʹґଘ͍ͯ͠Δ͕ɺͦͷ֬཰ͷ஋͸෼͔Βͳ͍ • ֶशσʔλʹج͍ͮͯਪఆͨ͠Ϛονϯά֬཰Λ࢖ͬͯϥϯΩϯάΛ࡞੒͠ɺ࠷దԽΛਤΔ

Slide 13

Slide 13 text

Slide 14

Slide 14 text

© 2024 Wantedly, Inc. Ϛονϯά༧ଌͷయܕతΞϓϩʔν ҎԼͷೋछྨʹେผͨ͠৔߹ɺجຊతʹ࢖ΘΕΔͷ͸ Predict-then-Aggregate ͷܗࣜ ● Direct Match Prediction (DMP) : ؍ଌ͞ΕͨϚονϯάʹج͍ͮͯϚον֬཰Λ༧ଌ͢Δ ● Predict-then-Aggregate (PtA) : ૒ํͷબ޷ΛಠཱʹϞσϧԽ͠ɺͦΕΒͷ༧ଌΛू໿͢Δ Direct Match Prediction Predict-then-Aggregate 𝒟 = {(ci , ji , yc→j i , yj→c i )}n i=1 ֶशσʔληοτɿ 𝒟 = {(ci , ji , mi )}n i=1 ؍ଌ͞ΕͨϚονϯά ̂ m = argmin ̂ m′ n ∑ i=1 ℓ( ̂ m′ (ci , ji ), mi ) Ϛονϯά֬཰ͷ   ֶशɾ༧ଌɿ ֶशσʔληοτɿ اۀ͕ϢʔβʔʹΞΫγϣϯ͔ͨ͠ Ϣʔβʔ͔ΒΞΫγϣϯ͔͋ͬͨ ̂ pc→j = argmin ̂ p′ n ∑ i=1 ℓ( ̂ p′ (ci , ji ), yc→j i ) اۀ→Ϣʔβʔͷ   બ޷֬཰ɿ Ϣʔβʔ→اۀͷ   બ޷֬཰ɿ ̂ pj→c = argmin ̂ p′ n ∑ i=1 ℓ( ̂ p′ (ci , ji ), yj→c i ) ̂ m = M( ̂ pc→j, ̂ pj→c) Ϛονϯά֬཰ɿ ଛࣦؔ਺ ू໿ؔ਺ ● DMP͸௚઀తͳϚονϯάͷ༧ଌ͕ՄೳͰ͋Δ ● ҰํͰϚονͷϥϕϧ͸ඇৗʹεύʔεͰ͋Γɺֶश͕ࠔ೉ ● PtA͸ൺֱతີͳೋछྨͷϥϕϧΛͦΕͧΕ༧ଌ͢Δ   λεΫʹ෼ׂ͢Δ͜ͱͰɺεύʔεੑͷ໰୊ʹରॲ ※ ؆୯ԽͷͨΊɺҎ߱͸   اۀଆͷέʔεͷΈΛߟ͑Δ

Slide 15

Slide 15 text

© 2024 Wantedly, Inc. ૬ޓਪનγεςϜ(Reciprocal Recommender Systems) ૬ޓਪનγεςϜͱ͸ʮαʔϏε಺ͷϢʔβʔΛޓ͍ʹਪન͠߹͏γεςϜʯ ● ਪનΛड͚औΔϢʔβʔͱਪન͞ΕͨϢʔβʔͷ྆ํ͕ຬ଍ͯ͠ਪન੒ޭͱ͢Δ ● ૬ޓਪનγεςϜ͸ Predict-then-Aggregate(PtA) ΞϓϩʔνΛ࠾༻͢Δ ઌߦݚڀͷΞϓϩʔν ● ίϯςϯπϕʔε [Pizzato+, 2010] ● ڠௐϑΟϧλϦϯάϕʔε [Xia+, 2015] [Neve+, 2019] ● ϋΠϒϦοτϕʔε [Neve+, 2020] ● DLϕʔε [Yıldırım+, 2021] [Luo+, 2020] [Liu+, 2024] ू໿ؔ਺ ● 2ͭͷผʑͷ༧ଌΛ૊Έ߹ΘͤΔ໾ׂΛ࣋ͭ ● جຊతʹώϡʔϦεςΟοΫͳ΋ͷɻ୯७ੵɺௐ࿨ ฏۉɺزԿฏۉͳͲ [Pizzato+, 2010] [Neve+, 2019] جຊతͳߏ੒ ̂ pa→b Preference Score from a to b M( ̂ pa→b, ̂ pb→a) Aggregation ̂ pb→a Preference Score from b to a ߦಈϩά΍   ଐੑσʔλ   ͳͲ

Slide 16

Slide 16 text

© 2024 Wantedly, Inc. ࣮ݧ݁Ռͷ֓ཁ ϓϩμΫτͷ࣮ࡍͷσʔλΛ࢖࣮ͬͨݧͷ࣮ࢪ ● ΦϑϥΠϯɿDMP ͱෳ਺ͷू໿ؔ਺ͷύλʔϯͷ PtA Λൺֱɻ૒ํ޲ͷᅂ޷ͷू໿ͷ༗ޮੑΛ֬ೝ ● ΦϯϥΠϯɿPtA(Scout-Only)ͱൺֱͨ͠ PtA(Harmonic Mean) ͷੑೳΛݕূɺେ෯ͳKPIͷ޲্Λ֬ೝ ϕʔεϥΠϯ M( ̂ pc→j, ̂ pj→c) = ̂ pc→j • PtA (Scout-Only)ɿ M( ̂ pc→j, ̂ pj→c) = ̂ pj→c • PtA (Reply-Only)ɿ ݕ౼ख๏ M( ̂ pc→j, ̂ pj→c) = ̂ pc→j ⋅ ̂ pj→c • PtA (Multiplication)ɿ M( ̂ pc→j, ̂ pj→c) = 2 ̂ pc→j ⋅ ̂ pj→c ̂ pc→j + ̂ pj→c • PtA (Harmonic Mean)ɿ ΦϑϥΠϯධՁͷҰࣄྫ

Slide 17

Slide 17 text

© 2024 Wantedly, Inc. ٕज़త՝୊ - ਪનػձͷภΓʹΑΔҰ෦Ϣʔβʔ΁ͷूத ਪનػձͷภΓ͕ੜ͡Δ͜ͱͰɺϓϥοτϑΥʔϜશମͷརӹ(Ϛον૯਺)͕େ͖͘ͳΒͳ͍ ● ֤Ϣʔβʔʹ͸ΩϟύγςΟ(Ϛονͷ্ݶ)͕ଘࡏɺͦΕΛ௒͑Δ޷ҙΛ΋Βͬͯ΋ରԠͰ͖ͳ͍ ● ඃਪનػձͷগͳ͍Ϣʔβʔ͸ɺ੒ޭମݧͱͳΔϚονϯάΛ࣮ݱ͢Δػձ͕ݶΒΕͯ͠·͏ ● طଘͷ૬ޓਪનγεςϜ͸ݸผͷϚον਺Λ࠷దԽ͠ϥϯΩϯά͝ͱʹಠཱͯ͠ܭࢉ͍ͯ͠ΔͨΊɺਪ નػձͷภΓΛੜͤͯ͡͞͠·͏ શ෦͸ରԠ Ͱ͖ͳ͍… εΧ΢τ͕ དྷͳ͍… ՝୊ʹର͢ΔΞϓϩʔν ● ٻ৬ऀ͕اۀ͔ΒͷεΧ΢τʹԠ౴͢Δ֬཰͕ɺ   ٻ৬ऀ͕ΑΓଟ͘ͷεΧ΢τΛड͚ΔʹͭΕͯ   ݮগ͢ΔՄೳੑΛߟྀ͠ɺϚον૯਺͕࠷େԽ   ͞ΕΔΑ͏ϥϯΩϯάΛ࠷దԽ [Su+, 2022] ● Ϛονϯάཧ࿦ʹج͖ͮɺ૒ํ޲ͷϢʔβͷᅂ޷ ͚ͩͰͳ͘ΩϟύγςΟΛߟྀͨ͠ू໿Λߦ͏ [Tomita+, 2022]

Slide 18

Slide 18 text

© 2024 Wantedly, Inc. ٕज़త՝୊ - ૒ํͷᅂ޷ͷू໿ํ๏ ᅂ޷ͷू໿ํ๏͸αʔϏε΍ͦΕΛར༻͢ΔϢʔβʔͷੑ࣭ʹԠͯ͡ઃܭ͢Δඞཁ͕͋Δ ● ૒ํͷᅂ޷༧ଌ݁ՌΛͲͷΑ͏ʹू໿͢Δ͔͸ࣗ໌Ͱͳ͍ ● Ұൠతʹɺௐ࿨ฏۉͱ͍ͬͨɺͲͪΒ͔ยํͷείΞ͕௿͍ͱू໿ͨ͠είΞ΋௿͘ͳΔͱ͍͏ੑ࣭ Λ࣋ͭ͜ͱ͕޷·͍͠ͱ͞Ε͍ͯΔ [Palomares+, 2021] [Neve+, 2019] ● ׬ᘳʹ֬཰ΛਪఆͰ͖ͨͷͰ͋Ε͹ɺू໿   ؔ਺͸͍Βͳ͍͸ͣ ɻ   ֤ଆͷ༧ଌ෼෍ͷζϨΛमਖ਼͢Δ੹຿Λ   ू໿ؔ਺͕୲͍ͬͯΔɺͱղऍͰ͖Δɻ ՝୊ʹର͢ΔΞϓϩʔν ● ૒ํͷᅂ޷ͷॏΈΛϢʔβʔ͝ͱʹ࠷దԽ͢Δ   ख๏ΛఏҊ [Kleinermann+, 2018] ू໿ํ๏ͷ·ͱΊ [Palomares+, 2021]

Slide 19

Slide 19 text

© 2024 Wantedly, Inc. ٕज़త՝୊ - Ϛονϯάͷεύʔεੑͷରॲ ϚονϯάϓϥοτϑΥʔϜͰ͸Ϛονϯάͱ͍͏৘ใ͕ಘʹ͍͘ಛੑ͕͋Δ ● δϣϒϚονϯάͷ৔߹ʮస৬ʯ͕੒ཱ͢Δͱɺ࣍ͷߦಈΛى͜͢·Ͱʹ௕͍͕͔͔࣌ؒΔ ● ਪનଆͱඃਪનଆͷ૒ํͷ޷ҙͱߦಈ͕߹கͯ͠ॳΊͯϚονϯά͕੒ཱ͢Δ ● Ϛονϯάͷ༧ଌਫ਼౓Λ্͛ΔͨΊʹɺͲͷΑ͏ͳ৘ใΛͲ͏ѻ͏΂͖͔͕՝୊ͱͳΔ ՝୊ʹର͢ΔΞϓϩʔν ● ஌ࣝάϥϑ͔ΒϝλύεΛநग़ͯ͠Ϛονϯάͷ ϞσϦϯάʹऔΓೖΕΔ͜ͱͰɺΠϯλϥΫγϣ ϯ͚ͩͰͳ͘ίϯςϯπ৘ใΛ༗ޮతʹ׆༻͢Δ [Lai+, 2024]

Slide 20

Slide 20 text

© 2024 Wantedly, Inc. ݚڀ։ൃ - Ϛον༧ଌਫ਼౓ͷ޲্ ● ϚονϥϕϧΛֶश͢Δ௚઀తͳΞϓϩʔν͕ͩɺ   Ϛονϥϕϧͷۃ౓ͳεύʔεੑ͕໰୊ͱͳΓɺ   ੑೳͷߴ͍ϞσϧΛ࡞Εͳ͍ →ΞΠσΞɿҟͳΔੑ࣭Λ࣋ͭ2छྨͷ৘ใΛޮՌతʹ૊Έ߹Θͤͯɺີ౓ͱਫ਼౓ͷʮ͍͍ͱ ͜औΓʯΛ࣮ݱ͢Δ Predict-then-Aggregate(PtA)ͷ՝୊ Direct Match Prediction(DMP)ͷ՝୊ ● Ϛονϯάͱ͍͏ϞσϧԽΛɺಠཱͨ͠2छྨͷϞσ ϧʹ෼ׂ͢Δ͜ͱʹΑͬͯ໰୊͕ੜ͡Δ ● σʔλ΍λεΫͷੑ࣭ʹ߹Θͤͨू໿ؔ਺Λద੾ʹઃ ܭ͠ͳ͍ͱύϑΥʔϚϯε͕ෆे෼ ● ֤Ϟσϧͷ༧ଌޡࠩͷ܏޲͕࠷ऴతͳϥϯΩϯάύ ϑΥʔϚϯεʹӨڹ͢Δ छྨ ਫ਼౓ ີ౓ ਅͷϚονϥϕϧ ਖ਼֬ ↑ ૄ ↓ Ϛον༧ଌ஋ ൺֱతෆਖ਼֬ ↓ ີ ↑ S. Goda, Y. Hayashi, Y. Saito, A Best-of-Both Approach to Improve Match Predictions and Reciprocal Recommendations for Job Search. arXiv preprint arXiv:2409.10992 (2024).

Slide 21

Slide 21 text

© 2024 Wantedly, Inc. ݚڀ։ൃ - Ϛον༧ଌਫ਼౓ͷ޲্ ఏҊख๏ spseudo (c, j; αc,j ) = αc,j ⋅ m(c, j) + (1 − αc,j ) ⋅ ̂ pc→j ⋅ ̂ pj→c ਅͷϚονϥϕϧͱϚον༧ଌ஋Λ૊Έ߹Θͤͨ   Pseudo Match Scores Λੜ੒͠ɺϝλϞσϧΛֶश ਅͷϚονϥϕϧ Ϛον༧ଌ஋ ̂ f = argminf′ ∑ (c,j) ℓ( f′ (c, j), spseudo (c, j; α)) ΦϑϥΠϯධՁ݁Ռ ● ϝλϞσϧΛ࢖༻͢Δ͜ͱͰɺैདྷͷPtAͷू໿ϑΣʔζͰ   ൃੜ͢ΔΤϥʔ఻೻ͷӨڹΛ௿ݮ͍ͯ͠ΔՄೳੑ͕͋Δ ● ҟͳΔείΞ৘ใΛ૊Έ߹ΘͤΔ͜ͱʹΑͬͯɺΞϯαϯϒϧ   తͳޮՌ͕ಘΒΕ͍ͯΔՄೳੑ͕͋Δ ղऍ S. Goda, Y. Hayashi, Y. Saito, A Best-of-Both Approach to Improve Match Predictions and Reciprocal Recommendations for Job Search. arXiv preprint arXiv:2409.10992 (2024).