会社訪問アプリ「Wantedly Visit」における人・仕事・会社の関係性を考慮した相互推薦システム

Slide 1

Slide 1 text

Slide 2

Slide 2 text

©2023 Wantedly, Inc. 1. اۀ঺հ • ձࣾͱϓϩμΫτͷ঺հ • ϓϩμΫτʹ͓͚ΔσʔλαΠΤϯεͷ׆༻ࣄྫ • ΞΧσϛΞʹ͓͚Δ׆ಈ 2. ૬ޓਪનγεςϜͱ͸ • ૬ޓਪનγεςϜͷ֓ཁ • HRྖҬʹ͓͚Δઌߦݚڀ 3. ਓɾ࢓ࣄɾձࣾͷؔ܎ੑΛߟྀͨ͠૬ޓਪનγεςϜ • ໰୊ఏى • ఏҊख๏ • ࣮ݧɾߟ࡯ ձࣾ๚໰ΞϓϦʮWantedly Visitʯʹ͓͚Δਓɾ࢓ࣄɾձࣾͷؔ܎ੑΛߟྀͨ͠૬ޓਪનγεςϜ

Slide 3

Slide 3 text

Slide 4

Slide 4 text

©2023 Wantedly, Inc. ࣗݾ঺հ ໊લ   ߹ా पฏ(Shuhei Goda) ܦྺ   2016೥3݄ ๺ւಓେֶେֶӃཧֶӃӉ஦ཧֶઐ߈म࢜՝ఔमྃ ॴଐɾ໾ׂ   ΢ΥϯςουϦʔגࣜձࣾɾData Science Tech Lead ձࣾ๚໰ΞϓϦ Wantedly Visit ͷϚονϯά෦෼Λ୲͏   ਪનγεςϜͷ։ൃΛਪਐ͍ͯ͠ΔɻKaggle Masterɻ @hakubishin3 @jy_msc

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

©2023 Wantedly, Inc. ࠃ಺ֶձ΁ͷڠࢍͱൃද https://www.ai-gakkai.or.jp/jsai2022/industrial-session/ https://www.ai-gakkai.or.jp/jsai2022/sponsor/ JSAI DEIM https://deim-management-system.github.io/deim2023_program/index.html#3b-5 https://event.dbsj.org/deim2023/

Slide 10

Slide 10 text

Slide 11

Slide 11 text

©2023 Wantedly, Inc. ૬ޓਪનγεςϜ(Reciprocal Recommeder Systems)ͱ͸ʁ User Item ैདྷͷҰൠతͳਪનγεςϜ ૬ޓਪનγεςϜ e.g. Amazon, Netflix ਪન͞ΕΔ ʮαʔϏε಺ͷϢʔβ͕ޓ͍ʹਪન͞Ε߹͏γεςϜʯ User(Job Seeker) User(Recruiter/Company) e.g. Wantedly, LinkedIn ਪન͞Ε߹͏ User(Female) User(Male) e.g. Tinder, Pairs ਪન͞Ε߹͏

Slide 12

Slide 12 text

Slide 13

Slide 13 text

©2023 Wantedly, Inc. ૬ޓਪનγεςϜͷ࿮૊Έ 1. γεςϜ಺ͷϢʔβ͸ A ͱ B ͷ̎ͭͷάϧʔϓʹ෼͔Ε͓ͯΓɺҟͳΔάϧʔϓͷϢʔβ͕ޓ͍ਪન͞ΕΔ   ΋ͷͱ͢Δʢe.g. σʔςΟϯάαʔϏεʹ͓͚ΔஉঁɺٻਓαʔϏεʹ͓͚Δٻ৬ऀͱاۀʣ 2. ୯ํ޲ͷᅂ޷ͷେ͖͞Λද͢ Preference Score ΛɺA ͔Β B ΁ͷϢʔβٴͼ B ͔Β A ͷϢʔβͷͦΕͧΕ   ʹ͍ͭͯܭࢉ 3. Aggregation Function Λར༻ͯ͠ɺ̎ͭͷ Preference Score Λ૊Έ߹Θͤͯ૒ํ޲ͷᅂ޷ͷେ͖͞Λද͢   Reciprocal Preference Score Λܭࢉ Predict User A to User B Preference Score Predict User B to User A Preference Score Aggregate   Preference Scores Reciprocal   Preference Score

Slide 14

Slide 14 text

©2023 Wantedly, Inc. HRྖҬʹ͓͚Δ૬ޓਪનγεςϜͷઌߦݚڀ ͜Ε·Ͱʮਓ(ީิऀ)ʯͱʮ࢓ࣄ(δϣϒɾσΟεΫϦϓγϣϯ)ʯͷཧ૝తͳϚονϯάΛ͍͔ʹ࣮ݱ͢Δ͔ɺ   ʹয఺Λ౰ͯͯɺHRྖҬʹ͓͚Δ૬ޓਪનγεςϜͷݚڀ͕ਐΜͰ͍Δ Tsunenori 2013 Yang 2022 Job User ʁ ɾ৬छ ɾ࢓ࣄ಺༰ ɾۈ຿஍ ɾۈ຿࣌ؒ ɾ࠾༻ܗଶ ɾ଴۰ ɹ….

Slide 15

Slide 15 text

Slide 16

Slide 16 text

©2023 Wantedly, Inc. ਓ͸ԿΛߟྀ͠ͳ͕Β࢓ࣄΛ୳͢ͷ͔ Ұൠతʹਓ͕࢓ࣄΛ୳࣌͢ʹߟྀ͢Δͷ͸ʮδϣϒɾσΟεΫϦϓγϣϯʯ͚ͩͰͳ͍͸ͣ • ྫ͑͹ɺձࣾͷϛογϣϯ΍ձࣾͰಇ͍͍ͯΔϝϯόʔͷงғؾ • ձࣾͷن໛΍੡඼ɺۀքɺઓུͳͲ΋ॏཁͩΖ͏ Job User ɾ৬छ ɾ࢓ࣄ಺༰ ɾۈ຿஍ ɾۈ຿࣌ؒ ɾ࠾༻ܗଶ ɾ଴۰ ɹ…. ʮ࢓ࣄʯʹؔ͢Δᅂ޷ Company ɾϛογϣϯ ɾҰॹʹಇ͘ਓͨͪ ɾձࣾن໛ ɾۀք ɾ੡඼ ɾઓུ ɹ… ʮձࣾʯʹؔ͢Δᅂ޷ Company಺

Slide 17

Slide 17 text

©2023 Wantedly, Inc. طଘݚڀͰߟྀͰ͖͍ͯͳ͍͜ͱ͸Կ͔ ͜Ε·Ͱʮਓʯͱʮ࢓ࣄʯͷؒͷΠϯλϥΫγϣϯ͔͠ѻΘΕͯ͜ͳ͔ͬͨͷͰɺ   ʮਓʯͱʮձࣾʯͷؔ܎Λද͢৘ใΛ໌ࣔతʹऔΓೖΕΔ͜ͱ͸Ͱ͖͍ͯͳ͍ • ࢓ࣄʹؔ͢Δᅂ޷͸நग़Ͱ͖Δ͕ɺձࣾʹؔ͢Δᅂ޷͸ʁ Job User ɾ৬छ ɾ࢓ࣄ಺༰ ɾۈ຿஍ ɾۈ຿࣌ؒ ɾ࠾༻ܗଶ ɾ଴۰ ɹ…. ʮਓʯͱʮ࢓ࣄʯͷؔ܎ੑ Company ɾϛογϣϯ ɾҰॹʹಇ͘ਓͨͪ ɾձࣾن໛ ɾۀք ɾ੡඼ ɾઓུ ɹ… ʮਓʯͱʮձࣾʯͷؔ܎ੑ Company಺

Slide 18

Slide 18 text

©2023 Wantedly, Inc. ձࣾ๚໰ΞϓϦ Wantedly Visit ͱ͸ ձࣾ΁ͷڞײΛ࣠ͱͨ͠࢓ࣄ୳͠ͷମݧΛఏڙ͢ΔαʔϏε • ࢓ࣄ΋ॏཁͳཁૉ͕ͩɺͦΕҎ্ʹձࣾʹڞײͰ͖Δ͔Λॏཁࢹ • ձࣾͷϛογϣϯ΍ͦ͜Ͱಇ͘ਓͷ૝͍͕఻ΘΓ΍͍͢ϓϩμΫτઃܭ Ϣʔβ ձࣾͷίϯςϯπ ɾձࣾϖʔδ ɾձࣾϒϩά ࢓ࣄͷίϯςϯπ ɾืू ճ༡ ืू΁ͷ   Ԡื ձࣾʹ࿩Λฉ͖ʹߦ͘

Slide 19

Slide 19 text

©2023 Wantedly, Inc. Wantedly Visit ʹطଘͷ૬ޓਪનγεςϜΛద༻͢Δࡍͷ՝୊ • ҰൠతͳδϣϒϚονϯάαʔϏεͱൺ΂ͯɺϢʔβʔମݧʹ͓͍ͯ   ձࣾͷཁૉ͕ॏཁͳ໾ׂΛ୲͍ͬͯΔ • ͦͷͨΊɺʮਓʯͱʮ࢓ࣄʯͷΠϯλϥΫγϣϯ͚ͩΛߟྀͨ͠طଘͷ૬ޓਪનγεςϜ   ͷ࿮૊ΈΛͦͷ··ద༻͢Δͷ͸ෆे෼ͩͱߟ͑Δ - ⭕Ϣʔβͷ࢓ࣄʹؔ͢Δᅂ޷ɺ࢓ࣄͷީิऀʹؔ͢Δᅂ޷ʢ৬຿ΛՌͨͤΔظ଴͕͋Δ͔Ͳ͏͔ʣ - ❌Ϣʔβͷձࣾʹؔ͢Δᅂ޷ɺձࣾͷީิऀʹؔ͢Δᅂ޷ʢՁ஋؍ͷҰகɺWillͱMustͷҰகɺͳͲʣ

Slide 20

Slide 20 text

©2023 Wantedly, Inc. Wantedly Visit ʹదͨ͠૬ޓਪનγεςϜͱ͸ ʮਓʯͱʮ࢓ࣄʯɺʮਓʯͱʮձࣾʯͷؔ܎ੑΛ໌ࣔతʹऔΓೖΕͨ૬ޓਪનγεςϜͷఏҊ • ఏҊख๏Λ Wantedly Visit ͷ࣮σʔλΛར༻ͯ͠ݕূ͢Δ Job User ɾ৬छ ɾ࢓ࣄ಺༰ ɾۈ຿஍ ɾۈ຿࣌ؒ ɾ࠾༻ܗଶ ɾ଴۰ ɹ…. ࢓ࣄʹؔ͢Δᅂ޷ Company ɾϛογϣϯ ɾҰॹʹಇ͘ਓͨͪ ɾձࣾن໛ ɾۀք ɾ੡඼ ɾઓུ ɹ… ձࣾʹؔ͢Δᅂ޷ Company಺ Ϣʔβʹؔ͢Δᅂ޷ Ϣʔβʹؔ͢Δᅂ޷

Slide 21

Slide 21 text

©2023 Wantedly, Inc. ࣮ݧ֓ཁ Wantedly Visit ʹ͓͚ΔϢʔβͱืूͷ Matching Λ༧ଌ (ਪન୯Ґ͸ืूͰ͋ΔͨΊ) • Ϣʔβ͕Ԡื or ืू(اۀ)͕εΧ΢τૹ৴ͨ͠ࡍʹ૬ख͕ϝοηʔδΛฦ৴͢Ε͹ Match (positive) • Ԡื or εΧ΢τૹ৴Λ૬ख͕Ӿཡ্ͨ͠ͰϝοηʔδΛฦ৴͠ͳ͚Ε͹ Not Match (negative) ืू(اۀ) Ϣʔβ Ԡื ϝοηʔδฦ৴ ืू(اۀ) Ϣʔβ εΧ΢τૹ৴ ϝοηʔδฦ৴ ืू(اۀ) Ϣʔβ Ԡื ӾཡͷΈ ืू(اۀ) Ϣʔβ εΧ΢τૹ৴ ӾཡͷΈ Match Not Match

Slide 22

Slide 22 text

Slide 23

Slide 23 text

Slide 24

Slide 24 text

©2023 Wantedly, Inc. ϕʔεϥΠϯख๏ Preference Scoreͷਪ࿦ʹߦྻ෼ղΛ༻͍ͯɺू໿ؔ਺͸ௐ࿨ฏۉΛ࠾༻(Neve 2019) • طଘݚڀʹ͓͚ΔHRͷ૬ޓਪનͰ͸ɺϢʔβͱ࢓ࣄͷΠϯλϥΫγϣϯΛߦྻ෼ղ΍GNNͰѻ͍ͬͯΔ • ݕূͷ୯७ԽͷͨΊʹɺࠓճ͸ߦྻ෼ղΛબ୒ Predict User to Project Preference Score Predict Project to User Preference Score Aggregate using   Harmonic Mean Ϣʔβ × ืू ͷ   ΠϯλϥΫγϣϯߦྻ

Slide 25

Slide 25 text

©2023 Wantedly, Inc. ఏҊख๏ 2छྨͷΠϯλϥΫγϣϯߦྻΛ༻͍ͯ૬ޓਪનείΞΛਪ࿦͢Δ • ϢʔβͱձࣾͷؒͷΠϯλϥΫγϣϯ͔Βɺձࣾʹؔ͢ΔϢʔβͷ܏޲Λநग़ • ձࣾʹؔ͢Δᅂ޷ͱ࢓ࣄʹؔ͢Δᅂ޷Λू໿(ௐ࿨ฏۉ or ࢉज़ฏۉ)͔ͯ͠Βɺ૬ޓਪનείΞΛࢉग़͢Δ Predict User to Company Preference Score Predict Company to User Preference Score Aggregate using   HM / AM Predict User to Project Preference Score Predict Project to User Preference Score Aggregate using   HM / AM Ϣʔβ × ืू ͷ   ΠϯλϥΫγϣϯߦྻ Ϣʔβ × ձࣾͷ   ΠϯλϥΫγϣϯߦྻ Aggregate using   Harmonic Mean

Slide 26

Slide 26 text

©2023 Wantedly, Inc. ࣮ݧ݁Ռᶃ ϕʔεϥΠϯʹൺ΂ͯɺఏҊख๏ʹΑΔ Reciprocal Preference Score ͷ༧ଌੑೳ͸޲্ • ௐ࿨ฏۉΑΓ΋ࢉज़ฏۉͷ΄͏͕ߴ͍ੑೳʹɺ͔͠͠ू໿ؔ਺ͷҧ͍ʹΑΔࠩ͸େ͖͘ͳ͍ • େ͖ͳ޲্ͷ༨஍͸͋Γͦ͏ɺ཈͑Δ΂͖ϙΠϯτ͕͋ΔͷͰ͸ʁ AUC Reciprocal Preference Score Rate of change (%) ϕʔεϥΠϯ 0.7122 - ఏҊख๏(ௐ࿨ฏۉ) 0.7269 +2.08% ఏҊख๏(ࢉज़ฏۉ) 0.7301 +2.52%

Slide 27

Slide 27 text

©2023 Wantedly, Inc. Ծઆग़͠ͱͦΕʹجͮ͘ఏҊख๏ͷमਖ਼ ϢʔβʹΑͬͯ࢓ࣄͷᅂ޷ͱձࣾͷᅂ޷ͷॏΈͷڧ͕͞ҟͳΔͷͰ͸ʁ • ྫ͑͹ɺస৬ҙཉͷߴ͍Ϣʔβ΄Ͳձࣾʹؔ͢ΔཁૉΛॏཁࢹ͢Δ͔΋͠Εͳ͍ • Ϣʔβͷᅂ޷ͷू໿ؔ਺ʹϩδεςΟοΫճؼΛ࠾༻͠ɺ࠷దͳॏΈͰू໿Ͱ͖ΔΑ͏ʹ͢Δ Predict User to Company Preference Score Predict Company to User Preference Score Aggregate using   HM / AM Predict User to Project Preference Score Predict Project to User Preference Score Aggregate using   Logistic Reg Ϣʔβ × ืू ͷ   ΠϯλϥΫγϣϯߦྻ Ϣʔβ × ձࣾͷ   ΠϯλϥΫγϣϯߦྻ Aggregate using   Harmonic Mean

Slide 28

Slide 28 text

©2023 Wantedly, Inc. ࣮ݧ݁Ռᶄ Ϣʔβ͕ઃఆͨ͠స৬ҙཉΛ༻͍ͯɺϢʔβͷᅂ޷ΛϩδεςΟοΫճؼͰू໿ͤ͞Δ • ͜ͷ΍Γํ͸ޮՌͳ͠ AUC Reciprocal Preference Score Rate of change (%) ϕʔεϥΠϯ 0.7122 - ఏҊख๏(ௐ࿨ฏۉ) 0.7269 +2.08% ఏҊख๏(ࢉज़ฏۉ) 0.7301 +2.52% ఏҊख๏(ϩδճؼ) 0.7255 +2.07%

Slide 29

Slide 29 text

Slide 30

Slide 30 text

©2023 Wantedly, Inc. ߟ࡯ᶃ ϕʔεϥΠϯͷ࣌఺Ͱ͢ͰʹɺձࣾͷཁૉΛ͋Δఔ౓நग़Ͱ͖͍ͯΔͷͰ͸ͳ͍͔ • Wantedly Visit ʹ͓͚Δืूϖʔδͷߏ੒͸ҎԼͷ௨Γɺձࣾͷཁૉͷํ͕༏ઌ౓ߴ͍ • Ϣʔβ͸͜ͷߏ੒ʹڧ͘Өڹ͞ΕɺืूͱϢʔβͷΠϯλϥΫγϣϯߦྻʹ΋ձࣾͷཁૉؚ͕·ΕΔͷͰ͸ Χόʔࣸਅ ձࣾͷϝϯόʔ঺հ ձࣾͷ8)"5 ࣄۀ಺༰ͷઆ໌ͳͲ ձࣾͷ8): ϏδϣϯɺϛογϣϯͳͲ ձࣾͷ)08 ࣄۀͷಛ௃ɺಇ͘஥ؒͳͲ ϙδγϣϯͷৄࡉɺۀ຿಺༰ͳͲ ืूͷλΠτϧɺ৬छɺ࠾༻ܗଶɺձ໊ࣾ ืूϖʔδͷߏ੒ ϖʔδ্෦ ϖʔδԼ෦ ձࣾͷཁૉ ࢓ࣄͷཁૉ ྆ํͷཁૉ

Slide 31

Slide 31 text

©2023 Wantedly, Inc. ߟ࡯ᶄ ఏҊख๏Ͱ͸ɺձࣾͷཁૉΛ͏·͘நग़Ͱ͖͍ͯͳ͍ͷͰ͸ͳ͍͔ • Wantedly Visit ʹ͓͍ͯϢʔβ͸࢓ࣄʹؔ͢Δ࠷௿৚݅ʢޏ༻ܗଶɺۈ຿஍ɺ৬छͳͲʣΛ   ઃఆͯ͠ϑΟϧλϦϯάΛ͔ͯ͠ΒձࣾΛ୳͢ • ྫ͑͹ɺձࣾͷཁૉ͕ຬͨ͞Ε͍ͯͯ΋ɺ΍Γ͍ͨ࢓ࣄ͕ͳ͍ͳΒԠื͠ͳ͍ͩΖ͏ • ձࣾ - ϢʔβͷΠϯλϥΫγϣϯʹ͸ɺ࢓ࣄͷཁૉ΋େؚ͖͘·Ε͍ͯΔͷͰ͸ͳ͍͔ Wantedly Visit ʹ͓͚Δืूͷݕࡧ৚݅ ޏ༻ܗଶɺۈ຿஍ɺ৬छͳͲͷ৚݅ઃఆ͕Ͱ͖Δ

Slide 32

Slide 32 text

©2023 Wantedly, Inc. Future work Wantedly Visit ʹదͨ͠ɺਓɾ࢓ࣄɾձࣾͷؔ܎ੑΛߟྀͨ͠૬ޓਪનͷߏ੒͸ʁ • ࢓ࣄ - ਓͷΠϯλϥΫγϣϯཤྺΛϕʔεʹɺձࣾͷ৘ใΛ෇༩͢ΔʢFactorization MachineͳͲʣ • ͜ͷߏ੒ͷ৔߹ɺձࣾͷಛ௃நग़Λ໌ࣔతʹߦ͏ඞཁ͕͋Δʢϛογϣϯ΍ձࣾͷงғؾɺ   ঎඼΍ۀքɺձࣾن໛΍ઓུͳͲʣ - Wantedly Visit ʹ৘ใ͕ଘࡏ͠ͳ͍΋ͷ͸ɺ৘ใΛऔಘͰ͖ΔΑ͏ʹ͢Δମ੍ͷߏங - Wantedly Visit ʹ৘ใ͸ଘࡏ͍ͯ͠Δ͕ඇߏ଄ԽσʔλʢςΩετɺը૾ʣͱͯ͋͠Δ΋ͷ͸ɺ໌ࣔతͳಛ௃   ͱͯ͠ѻ͑ΔΑ͏ʹ৘ใநग़͕ඞཁ - ಛʹϛογϣϯ΍ձࣾͷงғؾͳͲͷந৅తͳཁૉ͸ɺը૾৘ใ΍ςΩετ͔Β͏·͘நग़͢Δ࿮૊Έ͕ඞཁ Wantedly Visit ʹ͓͚Δืूը૾ͷऔΓ૊Έ͸   DEIM2022ͷٕज़ใࠂͰൃද

Slide 33

Slide 33 text

©2023 Wantedly, Inc. ݁࿦ ໨త: Wantedly Visit ʹదͨ͠ɺਓɾ࢓ࣄɾձࣾͷؔ܎ੑΛߟྀͨ͠૬ޓਪનͷߏ੒ݕ౼ Ϣʔβͷ࢓ࣄͱձࣾͷᅂ޷ਪఆΛ੾Γ෼͚ͯɺ࢓ࣄͷᅂ޷ͱձࣾͷᅂ޷ΛϢʔβ͝ͱʹՃॏฏۉ͢Δ   ͜ͱΛఏҊɺେ͖ͳվળ͸ݟࠐΊͳ͔ͬͨ ߟ࡯: ձࣾͱ࢓ࣄͷᅂ޷Λ͏·͘੾Γ෼͚ΒΕ͍ͯͳ͍Մೳੑ͕͋Δ 1. αʔϏεʹ͓͚Δืूίϯςϯπͷಛ௃: ձࣾͷཁૉ͕ଟؚ͘·Ε͍ͯΔ 2. αʔϏεʹ͓͚ΔϢʔβߦಈͷಛ௃: ࢓ࣄʹؔ͢Δ࠷௿৚݅Λઃఆ͔ͯ͠ΒձࣾΛ୳͢ ࠓޙ: ਓɾ࢓ࣄɾձࣾͷؔ܎ੑΛ׆͔͠੾ΔͨΊʹɺͲ͏͍͏ߏ੒͕ద੾ͳͷ͔ • ࢓ࣄ - ਓͷΠϯλϥΫγϣϯཤྺΛϕʔεʹɺձࣾͷ৘ใΛ෇͚Ճ͑Δ • ը૾΍ςΩετͳͲͷඇߏ଄Խσʔλ͔Βɺཉ͍͠ձࣾͷಛ௃Λ͏·͘৘ใநग़͢Δඞཁ͕͋Δ

Slide 34

Slide 34 text

©2023 Wantedly, Inc. 3FGFSFODFT • (Pizzato 2010) Luiz Pizzato, Tomek Rej, Thomas Chung, Irena Koprinska, and Judy Kay. 2010. RECON: a reciprocal recommender for online dating. Proceedings of the fourth ACM conference on Recommender systems, 207-214. • (Hongtao 2011)Hongtao Yu, C. Liu and Fuzhi Zhang. 2011. Reciprocal Recommendation Algorithm for the Field of Recruitment. Journal of Information & Computational Science, 4061-4068. • (Tsunenori 2013)Tsunenori Mine, Tomoyuki Kakuta, Akira Ono. 2013. Reciprocal Recommendation for Job Matching with Bidirectional Feedback. 2013 Second IIAI International Conference on Advanced Applied Informatics, 39-44. • (Neve 2019)J Neve, I Palomares.2019. Latent Factor Models and Aggregation Operators for Collaborative Filtering in Reciprocal Recommender Systems Proceedings of the 13th ACM Conference on Recommender Systems, 219-227. • (Yang 2022)Yang, Chen and Hou, Yupeng and Song, Yang and Zhang, Tao and Wen, Ji-Rong and Zhao and Wayne Xin. 2022. Modeling Two-Way Selection Preference for Person-Job Fit. Proceedings of the 16th ACM Conference on Recommender Systems,102-112.