Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
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
Search
Shuhei Goda
November 07, 2024
Technology
3
680
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
第27回情報論的学習理論ワークショップ (IBIS2024)
企画セッション3:ビジネスと機械学習
https://ibisml.org/ibis2024/os/
Shuhei Goda
November 07, 2024
Tweet
Share
More Decks by Shuhei Goda
See All by Shuhei Goda
とある事業会社にとっての Kaggler の魅力
hakubishin3
8
1.9k
課題の解像度が荒かったことで意図した改善ができなかった話
hakubishin3
3
910
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
hakubishin3
4
990
Recommendation Industry Talks #1 Opening
hakubishin3
1
310
会社訪問アプリ「Wantedly Visit」での シゴトに関する興味選択機能と推薦改善
hakubishin3
0
560
論文紹介: Improving Implicit Feedback-Based Recommendation through Multi-Behavior Alignment(Xin Xin et al., 2023)
hakubishin3
0
530
Feedback Prize - English Language Learning における擬似ラベルの品質向上の取り組み
hakubishin3
0
880
ウォンテッドリーにおける推薦システムのオフライン評価の仕組み
hakubishin3
7
6.5k
機械学習でプロダクト成長させる技術と組織を3社のMLエン ジニアが語る - note/コネヒト/Wantedly
hakubishin3
0
240
Other Decks in Technology
See All in Technology
隣接領域をBeyondするFinatextのエンジニア組織設計 / beyond-engineering-areas
stajima
1
270
ISUCONに強くなるかもしれない日々の過ごしかた/Findy ISUCON 2024-11-14
fujiwara3
8
870
OTelCol_TailSampling_and_SpanMetrics
gumamon
1
120
Amplify Gen2 Deep Dive / バックエンドの型をいかにしてフロントエンドへ伝えるか #TSKaigi #TSKaigiKansai #AWSAmplifyJP
tacck
PRO
0
380
AGIについてChatGPTに聞いてみた
blueb
0
130
個人でもIAM Identity Centerを使おう!(アクセス管理編)
ryder472
3
200
社内で最大の技術的負債のリファクタリングに取り組んだお話し
kidooonn
1
550
Amazon CloudWatch Network Monitor のススメ
yuki_ink
1
200
OCI Network Firewall 概要
oracle4engineer
PRO
0
4.1k
Exadata Database Service on Dedicated Infrastructure(ExaDB-D) UI スクリーン・キャプチャ集
oracle4engineer
PRO
2
3.2k
第1回 国土交通省 データコンペ参加者向け勉強会③- Snowflake x estie編 -
estie
0
130
サイバーセキュリティと認知バイアス:対策の隙を埋める心理学的アプローチ
shumei_ito
0
380
Featured
See All Featured
Agile that works and the tools we love
rasmusluckow
327
21k
Done Done
chrislema
181
16k
Performance Is Good for Brains [We Love Speed 2024]
tammyeverts
6
410
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
31
2.7k
4 Signs Your Business is Dying
shpigford
180
21k
Documentation Writing (for coders)
carmenintech
65
4.4k
The Myth of the Modular Monolith - Day 2 Keynote - Rails World 2024
eileencodes
16
2.1k
How STYLIGHT went responsive
nonsquared
95
5.2k
Why Our Code Smells
bkeepers
PRO
334
57k
Designing on Purpose - Digital PM Summit 2013
jponch
115
7k
Typedesign – Prime Four
hannesfritz
40
2.4k
Adopting Sorbet at Scale
ufuk
73
9.1k
Transcript
© 2024 Wantedly, Inc. δϣϒϚονϯάαʔϏεʹ͓͚Δ ૬ޓਪનγεςϜͷԠ༻ࣄྫͱ՝ Nov. 7 2024
- Shuhei Goda ୈ27ճใతֶशཧϫʔΫγϣοϓ (IBIS2024) اըηογϣϯ3ɿϏδωεͱػցֶश
© 2024 Wantedly, Inc. ໊લɿ ߹ా पฏ Shuhei Goda
ॴଐͱׂɿ ΥϯςουϦʔגࣜձࣾ ɾData Team Manager ɾMachine Learning Tech Lead ɾProduct Manager Kaggle Tierɿ Kaggle Competitions Grandmaster @jy_msc ࣗݾհ https://www.kaggle.com/shuheigoda
©2024 Wantedly, Inc. ڀۃͷదࡐదॴʹΑΓɺ γΰτͰίίϩΦυϧͻͱΛ;͢ ࢲͨͪͷϛογϣϯ ©2024 Wantedly, Inc.
© 2024 Wantedly, Inc. iOS, Android and Web ؾܰʹձࣾ๚ ϛογϣϯՁ؍ͷڞײͰϚονϯά
• څ༩རްੜͳͲͷ݅Ͱͳ͘ɺ͍͕͋Εձࣾͷ نʹͱΒΘΕͳ͍ ·ͣʮΛฉ͖ʹߦ͘ʯͱ͍͏৽͍͠ମݧ • ݸਓͱاۀ͕ϑϥοτͳઢͰग़ձ͑Δ͜ͱͰɺΑΓັྗ తͳॴΛݟ͚ͭΔ͜ͱ͕Մೳʹ ձࣾ๚ΞϓϦʮWantedly Visitʯ
© 2024 Wantedly, Inc. తΛୡ͢ΔͨΊʹɺ֤ొਓҎԼͷΑ͏ʹߦಈ͢Δ αʔϏεΛར༻͢Δਓͷతͱߦಈ ొਓ αʔϏεΛར༻͢Δత Ϣʔβʔ
ઓతͰΓ͕͍ͷ͋Δࣄʹͭ͘ اۀ ࣗࣾͰ׆༂Ͱ͖ΔਓࡐΛ࠾༻͢Δ ࣗͷίϯςϯπ Λ࡞͢Δ ૬खͷίϯςϯπ ΛӾཡ͢Δ ૬खʹ໘ஊͷػձ Λਃ͠ࠐΉ
© 2024 Wantedly, Inc. ϓϥοτϑΥʔϜߏʢίϯςϯπʣ Ϣʔβʔ ձࣾһ ෭ۀ ϑϦʔϥϯε ֶੜ
اۀ ܦӦ ࣾһ ਓࣄ Wantedly Visit ϓϩϑΟʔϧ ࡞ Ӿཡ ձࣾϖʔδɾืू ࡞ Ӿཡ
© 2024 Wantedly, Inc. ϓϥοτϑΥʔϜߏʢϚονϯάʣ Ϣʔβʔ ձࣾһ ෭ۀ ϑϦʔϥϯε ֶੜ
اۀ ܦӦ ࣾһ ਓࣄ Wantedly Visit Ԡื Λฉ͍ͯΈ͍ͨͰ͢ ͥͻ͓͠·͠ΐ͏ʂ εΧτ ͓͠·ͤΜ͔ʁ ͓Λฉ͔͍ͤͯͩ͘͞
© 2024 Wantedly, Inc. ϓϥοτϑΥʔϜͷϞσϧʢུ֓ʣ ܧଓɾ෮ؼ ܧଓɾ෮ؼ Ԡื εΧτ ྲྀೖ
৽ن • 2-sided marketplaceͰ͋ΓɺاۀͱϢʔβʔͷ྆ํ͕υϥΠόʔ • ྑ࣭ͳίϯςϯπͷੵɺΓͳ͍Ϛονϯάʢޭମݧʣ͕ॏཁ
© 2024 Wantedly, Inc. ϓϥοτϑΥʔϜͷ՝ - ͷϘτϧωοΫԿ͔ 20248݄ظͷొͱਪҠ 403ສਓ 4.1ສࣾ
ෳࡶͷ૿ՃͱͦΕʹ͏ϚονϯάޮͷԼ Ϣʔβʔ૿Ճʹ͏՝ • ෳࡶͷ૿Ճ→ϚονϯάޮͷԼ →Ϛονϯά૿ՃͷಷԽ→ͷఀϦεΫ Ϛονϯάޮ ͷԼ Ϣʔβʔ Ϛονϯάޮ
© 2024 Wantedly, Inc. ϓϥοτϑΥʔϜͷʹඞཁͳऔΓΈ Ϣʔβʔ Ϛονϯάޮ Ϣʔβʔ Ϛονϯάޮ Ϛονϯάޮ
Ϛονϯάޮ Ϛονϯάޮ͕ߴ͘ҡ࣋͞ΕΔঢ়ଶΛ࡞Γ αʔϏεશମͷ࣋ଓతͳΛ࣮ݱ͢Δ Before After
© 2024 Wantedly, Inc. Ϛονϯάޮ͕ߴ͘ҡ࣋͞ΕΔঢ়ଶΛͲ͏࣮ݱ͢Δ͔ p(match = 1|c, j) =
p(exam = 1|c, j) × p(scout = 1|c, j, exam = 1) × p(reply = 1|c, j, scout = 1) اۀͲͷϢʔβʔΛ ݟΔ͔ اۀʹͱͬͯͦͷϢʔβʔ ັྗతʹײ͡ΒΕΔ͔ Ϣʔβʔʹͱͬͯͦͷืू ັྗతʹײ͡ΒΕΔ͔ ਪનγεςϜͰϢʔβʔɾاۀํͷίϯςϯπͷදࣔΛ੍ޚ͢Δ • To اۀɿاۀʹͱͬͯັྗతͰɺͦͷاۀͱϚονϯά͍͢͠ϢʔβʔΛදࣔ • To ϢʔβʔɿϢʔβʔʹͱͬͯັྗతͰɺͦͷϢʔβʔͱϚονϯά͍͢͠ืूΛදࣔ 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) ϢʔβʔͲͷاۀΛ ݟΔ͔ Ϣʔβʔʹͱͬͯͦͷืू ັྗతʹײ͡ΒΕΔ͔ اۀʹͱͬͯͦͷϢʔβʔ ັྗతʹײ͡ΒΕΔ͔ εΧτ༝དྷ ͷϚονϯά֬ Ԡื༝དྷ ͷϚονϯά֬
© 2024 Wantedly, Inc. ղ͖͘ػցֶशλεΫ Ϛονϯά͕࠷େͱͳΔΑ͏ͳਪનϦετΛ֤ඃਪનऀʹରͯ͠࡞͢ΔλεΫ σ* c := argsortj∈
𝒥 m(c, j) σ* j := argsortc∈ 𝒞 m(c, j) →֤اۀ c ʹͱͬͯཧతͳϥϯΩϯάɺ m(c, j)ਅͷϚονϯά֬ →֤Ϣʔβʔ j ʹͱͬͯཧతͳϥϯΩϯά • ཧతͳϥϯΩϯάਅͷϚονϯά֬ʹґଘ͍ͯ͠Δ͕ɺͦͷ֬ͷ͔Βͳ͍ • ֶशσʔλʹج͍ͮͯਪఆͨ͠Ϛονϯά֬ΛͬͯϥϯΩϯάΛ࡞͠ɺ࠷దԽΛਤΔ
© 2024 Wantedly, Inc. ػցֶश͕αʔϏεͷίΞՁͰ͋ΔϚονϯάΛΓཱͨͤΔ ػցֶशϞσϧͷ༧ଌਫ਼্͕͢Δ΄ͲɺͦΕʹͬͯϓϩμΫτͷॏཁࢦ ඪ্͕͍ͯ͘͠ߏ → ػցֶश͕ϓϩμΫτͷՁΛఏڙ͢ΔͨΊͷ
“must have” ͳঢ়ଶ ػցֶशϞσϧͷੑೳ αʔϏεͷ ఏڙՁ
© 2024 Wantedly, Inc. Ϛονϯά༧ଌͷయܕతΞϓϩʔν ҎԼͷೋछྨʹେผͨ͠߹ɺجຊతʹΘΕΔͷ Predict-then-Aggregate ͷܗࣜ • Direct
Match Prediction (DMP) : ؍ଌ͞ΕͨϚονϯάʹج͍ͮͯϚον֬Λ༧ଌ͢Δ • Predict-then-Aggregate (PtA) : ํͷબΛಠཱʹϞσϧԽ͠ɺͦΕΒͷ༧ଌΛू͢Δ 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 ) Ϛονϯά֬ͷ ֶशɾ༧ଌɿ ֶशσʔληοτɿ اۀ͕ϢʔβʔʹΞΫγϣϯ͔ͨ͠ Ϣʔβʔ͔ΒΞΫγϣϯ͔͋ͬͨ ̂ pc→j = argmin ̂ p′  n ∑ i=1 ℓ( ̂ p′  (ci , ji ), yc→j i ) اۀ→Ϣʔβʔͷ બ֬ɿ Ϣʔβʔ→اۀͷ બ֬ɿ ̂ pj→c = argmin ̂ p′  n ∑ i=1 ℓ( ̂ p′  (ci , ji ), yj→c i ) ̂ m = M( ̂ pc→j, ̂ pj→c) Ϛονϯά֬ɿ ଛࣦؔ ूؔ • DMPతͳϚονϯάͷ༧ଌ͕ՄೳͰ͋Δ • ҰํͰϚονͷϥϕϧඇৗʹεύʔεͰ͋Γɺֶश͕ࠔ • PtAൺֱతີͳೋछྨͷϥϕϧΛͦΕͧΕ༧ଌ͢Δ λεΫʹׂ͢Δ͜ͱͰɺεύʔεੑͷʹରॲ ※ ؆୯ԽͷͨΊɺҎ߱ اۀଆͷέʔεͷΈΛߟ͑Δ
© 2024 Wantedly, Inc. ૬ޓਪનγεςϜ(Reciprocal Recommender Systems) ૬ޓਪનγεςϜͱʮαʔϏεͷϢʔβʔΛޓ͍ʹਪન͠߹͏γεςϜʯ • ਪનΛड͚औΔϢʔβʔͱਪન͞ΕͨϢʔβʔͷ྆ํ͕ຬͯ͠ਪનޭͱ͢Δ
• 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] ूؔ • 2ͭͷผʑͷ༧ଌΛΈ߹ΘͤΔׂΛ࣋ͭ • جຊతʹώϡʔϦεςΟοΫͳͷɻ୯७ੵɺௐ 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
© 2024 Wantedly, Inc. ࣮ݧ݁Ռͷ֓ཁ ϓϩμΫτͷ࣮ࡍͷσʔλΛ࣮ͬͨݧͷ࣮ࢪ • ΦϑϥΠϯɿDMP ͱෳͷूؔͷύλʔϯͷ PtA
Λൺֱɻํͷᅂͷूͷ༗ޮੑΛ֬ೝ • ΦϯϥΠϯɿPtA(Scout-Only)ͱൺֱͨ͠ PtA(Harmonic Mean) ͷੑೳΛݕূɺେ෯ͳKPIͷ্Λ֬ೝ 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
© 2024 Wantedly, Inc. ٕज़త՝ - ਪનػձͷภΓʹΑΔҰ෦Ϣʔβʔͷूத ਪનػձͷภΓ͕ੜ͡Δ͜ͱͰɺϓϥοτϑΥʔϜશମͷརӹ(Ϛον૯)͕େ͖͘ͳΒͳ͍ • ֤ϢʔβʔʹΩϟύγςΟ(Ϛονͷ্ݶ)͕ଘࡏɺͦΕΛ͑ΔҙΛΒͬͯରԠͰ͖ͳ͍
• ඃਪનػձͷগͳ͍ϢʔβʔɺޭମݧͱͳΔϚονϯάΛ࣮ݱ͢Δػձ͕ݶΒΕͯ͠·͏ • طଘͷ૬ޓਪનγεςϜݸผͷϚονΛ࠷దԽ͠ϥϯΩϯά͝ͱʹಠཱͯ͠ܭࢉ͍ͯ͠ΔͨΊɺਪ નػձͷภΓΛੜͤͯ͡͞͠·͏ શ෦ରԠ Ͱ͖ͳ͍… εΧτ͕ དྷͳ͍… ՝ʹର͢ΔΞϓϩʔν • ٻ৬ऀ͕اۀ͔ΒͷεΧτʹԠ͢Δ͕֬ɺ ٻ৬ऀ͕ΑΓଟ͘ͷεΧτΛड͚ΔʹͭΕͯ ݮগ͢ΔՄೳੑΛߟྀ͠ɺϚον૯͕࠷େԽ ͞ΕΔΑ͏ϥϯΩϯάΛ࠷దԽ [Su+, 2022] • Ϛονϯάཧʹج͖ͮɺํͷϢʔβͷᅂ ͚ͩͰͳ͘ΩϟύγςΟΛߟྀͨ͠ूΛߦ͏ [Tomita+, 2022]
© 2024 Wantedly, Inc. ٕज़త՝ - ํͷᅂͷूํ๏ ᅂͷूํ๏αʔϏεͦΕΛར༻͢ΔϢʔβʔͷੑ࣭ʹԠͯ͡ઃܭ͢Δඞཁ͕͋Δ • ํͷᅂ༧ଌ݁ՌΛͲͷΑ͏ʹू͢Δ͔ࣗ໌Ͱͳ͍
• Ұൠతʹɺௐฏۉͱ͍ͬͨɺͲͪΒ͔ยํͷείΞ͕͍ͱूͨ͠είΞ͘ͳΔͱ͍͏ੑ࣭ Λ࣋ͭ͜ͱ͕·͍͠ͱ͞Ε͍ͯΔ [Palomares+, 2021] [Neve+, 2019] • ᘳʹ֬ΛਪఆͰ͖ͨͷͰ͋Εɺू ͍ؔΒͳ͍ͣ ɻ ֤ଆͷ༧ଌͷζϨΛमਖ਼͢ΔΛ ू͕ؔ୲͍ͬͯΔɺͱղऍͰ͖Δɻ ՝ʹର͢ΔΞϓϩʔν • ํͷᅂͷॏΈΛϢʔβʔ͝ͱʹ࠷దԽ͢Δ ख๏ΛఏҊ [Kleinermann+, 2018] ूํ๏ͷ·ͱΊ [Palomares+, 2021]
© 2024 Wantedly, Inc. ٕज़త՝ - Ϛονϯάͷεύʔεੑͷରॲ ϚονϯάϓϥοτϑΥʔϜͰϚονϯάͱ͍͏ใ͕ಘʹ͍͘ಛੑ͕͋Δ • δϣϒϚονϯάͷ߹ʮస৬ʯཱ͕͢Δͱɺ࣍ͷߦಈΛى͜͢·Ͱʹ͍͕͔͔࣌ؒΔ
• ਪનଆͱඃਪનଆͷํͷҙͱߦಈ͕߹கͯ͠ॳΊͯϚονϯάཱ͕͢Δ • Ϛονϯάͷ༧ଌਫ਼Λ্͛ΔͨΊʹɺͲͷΑ͏ͳใΛͲ͏ѻ͏͖͔͕՝ͱͳΔ ՝ʹର͢ΔΞϓϩʔν • ࣝάϥϑ͔ΒϝλύεΛநग़ͯ͠Ϛονϯάͷ ϞσϦϯάʹऔΓೖΕΔ͜ͱͰɺΠϯλϥΫγϣ ϯ͚ͩͰͳ͘ίϯςϯπใΛ༗ޮతʹ׆༻͢Δ [Lai+, 2024]
© 2024 Wantedly, Inc. ݚڀ։ൃ - Ϛον༧ଌਫ਼ͷ্ • ϚονϥϕϧΛֶश͢ΔతͳΞϓϩʔν͕ͩɺ
Ϛονϥϕϧͷۃͳεύʔεੑ͕ͱͳΓɺ ੑೳͷߴ͍ϞσϧΛ࡞Εͳ͍ →ΞΠσΞɿҟͳΔੑ࣭Λ࣋ͭ2छྨͷใΛޮՌతʹΈ߹Θͤͯɺີͱਫ਼ͷʮ͍͍ͱ ͜औΓʯΛ࣮ݱ͢Δ Predict-then-Aggregate(PtA)ͷ՝ Direct Match Prediction(DMP)ͷ՝ • Ϛονϯάͱ͍͏ϞσϧԽΛɺಠཱͨ͠2छྨͷϞσ ϧʹׂ͢Δ͜ͱʹΑ͕ͬͯੜ͡Δ • σʔλλεΫͷੑ࣭ʹ߹ΘͤͨूؔΛదʹઃ ܭ͠ͳ͍ͱύϑΥʔϚϯε͕ෆे • ֤Ϟσϧͷ༧ଌޡࠩͷ͕࠷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).
© 2024 Wantedly, Inc. ݚڀ։ൃ - Ϛον༧ଌਫ਼ͷ্ ఏҊख๏ spseudo (c,
j; αc,j ) = αc,j ⋅ m(c, j) + (1 − αc,j ) ⋅ ̂ pc→j ⋅ ̂ pj→c ਅͷϚονϥϕϧͱϚον༧ଌΛΈ߹Θͤͨ Pseudo Match Scores Λੜ͠ɺϝλϞσϧΛֶश ਅͷϚονϥϕϧ Ϛον༧ଌ ̂ f = argminf′  ∑ (c,j) ℓ( f′  (c, j), spseudo (c, j; α)) ΦϑϥΠϯධՁ݁Ռ • ϝλϞσϧΛ༻͢Δ͜ͱͰɺैདྷͷPtAͷूϑΣʔζͰ ൃੜ͢ΔΤϥʔͷӨڹΛݮ͍ͯ͠ΔՄೳੑ͕͋Δ • ҟͳΔείΞใΛΈ߹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).