user-behaviour-vol2

ユーザー行動予測に関する調査-vol2

Transcript

1. Ϣʔβʔߦಈ༧ଌʹؔ͢Δௐࠪ-vol2
   ∁໦ࢤ࿠
   

   View Slide

2. Ϣʔβʔߦಈ෼ੳ
   Web্ͰϢʔβʔ͸༷ʑͳߦಈΛߦ͏
   ͦͷߦಈͷϩάͷσʔλ͔ΒϢʔβʔ͕࣍ʹͲͷΑ͏ͳߦಈΛ͢Δͷ͔Λ༧ଌ͢Δ͜ͱ͸ɼϏδωε
   ্େ͖ͳՁ஋Λ࣋ͭ
   https://medium.com/@yahavofir/paid-information-is-everywhere-and-it-changes-the-way-we-make-decisions-ca82679ed2bd
   

   View Slide

3. SessionͱϢʔβʔߦಈ
   Ϣʔβʔߦಈ͸Ұൠʹsessionͱ͍͏·ͱ·ΓͰѻΘΕΔ
   Sessionͱ͸ಛఆظؒ಺ʹى͖ͨҰ࿈ͷϢʔβʔߦಈͷ·ͱ·Γͷ͜ͱ
   [https://support.google.com/analytics/answer/2731565?
   hl=ja#zippy=%2C%E3%81%93%E3%81%AE%E8%A8%98%E4%BA%8B%E3%81%AE%E5%86%85%E5%AE%B9]
   

   View Slide

4. Sequence-aware vs session-aware vs session-based
   session session session session session session
   2021/11/2 2021/11/6
   session-based
   sequence-aware
   ੲͷηογϣϯͷ৘ใ΋༻͍ͯߦಈ༧ଌΛߦ͏৔߹Λɼsession-awareɼಉҰsession಺ͷ৘ใͷΈΛ༻͍ͯߦಈ༧ଌΛߦ͏৔
   ߹Λɼsession-basedͱݺͿɽ·ͨɼҰൠʹsessionΛ੾Βͣաڈͷߦಈ͔Β༧ଌΛߦ͏৔߹Λɼsequence-awareͱݺͿɽ
   session-aware
   [Ludwelg 2020]
   user 1
   

   View Slide

5. ຊษڧձͰ΍Δ͜ͱ
   Session-awareͳϢʔβߦಈ༧ଌʹؔ͢Δݚڀͷ֓ཁͷ঺հ
   աڈͷϢʔβߦಈΛͲ͏࢖͏͔
   ֘౰Ϣʔβͷ௚ۙͷηογϣϯ৘ใ
   ֘౰Ϣʔβͷ͜Ε·ͰͷશͯͷཤྺΛूܭͨ͠৘ใ
   ಛ௃ྔΤϯδχΞϦϯά
   SIGIR eCom 21 Data Challengeʹ͓͚Δσʔλͷѻ͍ͷ঺հ
   Session-basedͳϢʔβʔߦಈ༧ଌʹؔ͢Δίϯϖ
   ͦͷଞsession-basedͳ༧ଌʹ͍ͭͯͷݚڀྫͷ঺հ
   ಛʹσʔλͷલॲཧ΍࢖༻͍ͯ͠Δಛ௃ྔʹ஫໨͠ͳ͕Βɼ͜ΕΒʹ͍ͭͯઆ໌
   

   View Slide

6. ಛ௃ྔͷ෼ྨ
   ߦಈಛ௃ྔ Ϣʔβʔಛ௃ྔ
   ಛ௃ྔ
   ߦಈಛ௃ྔ
   ʢintra-sessionʣ
   ߦಈಛ௃ྔ
   ʢinter-sessionʣ
   Ϣʔβʔߦಈ༧ଌͷಛ௃ྔ͸ʮߦಈಛ௃ྔʯʮϢʔβʔಛ௃ྔʯʮΞΠςϜಛ௃ྔʯʹେผ͞ΕΔ
   ຊௐࠪͰ͸ಛʹͲͷΑ͏ͳߦಈಛ௃ྔ͕༻͍ΒΕ͍ͯΔͷ͔ʹ஫໨ͯ͠ௐࠪ
   ΞΠςϜಛ௃ྔ
   ੑผɼ೥ྸɼɾɾ ঎඼໊ɼ஋ஈɼɾɾ
   ΫϦοΫɼɾɾ աڈͷ๚໰਺ɼɾɾ
   

   View Slide

7. ໨࣍
   1. Session-aware recommendationͷ֓ཁ
   2. Session-based/aware recommendationͷ༧ଌͰॏཁͳಛ௃ྔʹ͍ͭͯͷௐࠪ
   • SIGIR eCom 21 Data Challenge
   • جຊߦಈಛ௃ྔͱಛ௃ྔબ୒
   • ͍͔ͭ͘ͷྫ
   • Struggling search
   • Information Seeking Conversation
   • Churn Prediction
   • ͦͷଞ
   3. ·ͱΊ
   

   View Slide

8. ໨࣍
   1. Session-aware recommendationͷ֓ཁ
   2. Session-based/aware recommendationͷ༧ଌͰॏཁͳಛ௃ྔʹ͍ͭͯͷௐࠪ
   • SIGIR eCom 21 Data Challenge
   • جຊߦಈಛ௃ྔͱಛ௃ྔબ୒
   • ͍͔ͭ͘ͷྫ
   • Struggling search
   • Information Seeking Conversation
   • Churn Prediction
   • ͦͷଞ
   3. ·ͱΊ
   

   View Slide

9. Session-awareͳਪનϞσϧͷ෼ྨͱಈ޲ɹ
   [Wang+ 2021]
   [Fang+ 2020]
   Session-based/aware recommendation Sequence-aware neural recommendation
   [Wang+ 2021]
   ఻౷తʹ͸kۙ๣๏΍Ϛϧίϑ࿈࠯ͳͲ͕༻͍ΒΕ͖͕ͯͨɼݚڀʹ͓͍ͯ͸ۙ೥͸χϡʔϥϧωο
   τϫʔΫΛ༻͍ͨ΋ͷ͕૿Ճ
   Session-based/aware recommendation
   

   View Slide

10. Session-based/awareΞϧΰϦζϜͷpros/cons
    [Wang+ 2021]
    ݚڀͷத৺͸χϡʔϥϧωοτϫʔΫ
    ͔͠͠ɼݱঢ়ͦͷଞͷख๏͕χϡʔϥϧωοτʹྼޙ͍ͯ͠ΔΘ͚Ͱ͸ͳ͍ [Dacrema+ 2019] [Latifi+ 2021]
    Pros/consΛ͓͑ͯ͞͏·͘׆༻Ͱ͖Δͱྑ͍ؾ͕͢Δ
    

    View Slide

11. ਪનΞϧΰϦζϜͷϦϙδτϦ/ϥΠϒϥϦ
    ۙ೥ͷਪનΞϧΰϦζϜ͕͙͢ʹ࢖͑ΔϥΠϒϥϦ͕͋ΔͷͰɼ͜ͷลͰ৭ʑࢼͯ͠ΈΔͱָ͍͔͠΋͠Εͳ͍
    RecBole
    Sequential recommendation΋ؚΉ࣮ʹ70༨ΓͷϞσϧͱ28ͷσʔληοτ͕࢖༻ՄೳͳϥΠϒϥϦ
    SASRec΍GRU4Recͱ͍ͬͨϞσϧ͔ΒBERT4Rec·Ͱ
    https://github.com/RUCAIBox/RecBole
    Recommenders
    Microsoft͕ఏڙ͢ΔύοέʔδͰ30લޙͷϞσϧ͕࢖༻Մೳɽݹయతͳख๏͕ଟΊ
    https://github.com/microsoft/recommenders
    Session-rec
    Session-awareͳख๏΋ؚΊͨ25ͷϞσϧ͕࢖༻ՄೳͳϦϙδτϦ
    https://github.com/rn5l/session-rec
    Tutorial on Sequence-Aware Recommender Systems
    RecSys 2018ͰߦͳΘΕͨsequence-awareͳhands-on͕·ͱΊΒΕͨϦϙδτϦ
    https://github.com/mquad/sars_tutorial
    

    View Slide

12. Session-based/awareͳ༧ଌʹ࢖͑Δσʔληοτ
    [Wang+ 2021]
    σʔληοτͰ͸click stream͔Βҙຯͷ͋Δجૅతͳߦಈσʔλʹม׵͞Ε͍ͯΔ
    ॳظతʹ͸͜ΕΒͷσʔληοτͰͲͷΑ͏ͳಛ௃ྔ͕࡞੒͞Ε͍ͯΔͷ͔ݟͯΈΔͱࢀߟʹͳΔ͔΋͠Εͳ͍
    [Wang+ 2021]
    

    View Slide

13. Session-aware RecommendationͰॏཁͳཁૉʢ1/3ʣ
    [Fang+ 2020]
    [Fang+ 2020]
    Fang+ 2020͸ɼෳ਺ͷsequential recommendationͷݚڀΛൺֱͨ݁͠Ռɼӈͷ
    දͷཁૉ͕ॏཁͰ͋ΔͱΘ͔ͬͨͱओு
    ͜Ε͸ͦΕͧΕԼਤͷsequential recommendationͷֶशͷϑϩʔʹରԠ
    

    View Slide

14. Input
    Side Information
    ঎඼ಛ௃ྔɿ঎඼ͷΧςΰϦʔɼࣸਅɼઆ໌ͳͲ
    σʔλͷগͳ͞΍cold-startʹΑΔੑೳͷྼԽΛ؇࿨ [Wang+ 2016] [Zhang+ 2019] etc
    ߦಈಛ௃ྔɿdwell timeͳͲ
    Dwell timeͷ௥Ճ͸GRU4RecͷੑೳΛେ෯ʹվળ [Hidasi+ 2016]
    Behavior Type
    click, purchase, collectͳͲ
    Ϣʔβʔͷ௕ظͷᅂ޷ੑΛΑ͘ଊ͑Δͷ͸ߪങߦಈ͕ͩͦͷଞͷߦಈಛ௃΋୹ظͷᅂ޷ੑΛଊ͑Δͷʹ໾ཱͭ [Gao+ 2019] etc
    Repeated Consumption
    Ϣʔβʔͷ܁Γฦ͠ͷߪങߦಈΛߟྀͨ͠΄͏͕༧ଌʹΑ͍ [Ren+ 2019]
    recencyʢ࠷ऴߪങ͔Βͷ࣌ؒʣ͕repeated consumptionͷ༧ଌʹͱͬͯॏཁ [Anderson+ 2014]
    loyaltyΛද͢ಛ௃ྔΛ௥Ճ͢Δ͜ͱͰ༧ଌੑೳ͕޲্ [Wan+ 2018]
    RNN͕kNNʹਪનλεΫͰෛ͚Δͷ͸ɼrepeated consumptionΛ͏·͘ϞσϦϯάͰ͖ͳ͍͔Βʁ[Hu+ 2020]
    Session-aware RecommendationͰॏཁͳཁૉʢ2/3ʣ
    [Fang+ 2020]
    

    View Slide

15. Data processing
    Embedding Design
    sessionͷຒΊࠐΈΛࣄલֶश͢ΔͳͲʢsession information embeddingʣ[Wu+ 2017]
    Data Augmentation
    NݸͷΠϕϯτ͔ΒͳΔηογϣϯ͔Βk(1~N)൪໨·ͰͷΠϕϯτ͔ΒͳΔNݸͷηογϣϯΛ࡞੒ [Tan+ 2016]
    Model Structure
    Incorporating Attention Mechanisms
    AttentionϞσϧ͸طଘϞσϧͷੑೳΛେ෯ʹ྇կ [Li+ 2017] [Kang+ 2018] [Zhang+ 2019]
    Adding Explicit User Representation
    User embedded models
    Ϣʔβʔ৘ใͷຒΊࠐΈΛՃ͑Δ͜ͱͰੑೳ͕޲্ [Tang+ 2018]
    ͨͩ͠୯ͳΔϢʔβʔ৘ใͷ׆༻͚ͩͰ͸cold-startͷ໰୊΍ɼϢʔβʔͷؔ৺ͷมԽ΁ͷରԠͷ໰୊͕ൃੜ
    User recurrent models
    Ϣʔβʔ৘ใΛ࣌ܥྻతʹϞσϧԽ͢Δ͜ͱͰGRU4RecΛ3.5%վળ [Quadrana+ 2018]
    Session-aware RecommendationͰॏཁͳཁૉʢ3/3ʣ
    [Fang+ 2020]
    

    View Slide

16. Session Length
    Long sessionsʢ10 ~ʣɿҰ൪৘ใྔ͕ଟ͍͕ͦͷ෼ϊΠζ΋ଟ͘ɼΠϕϯτؒͷؔ܎΋ෳࡶ
    Medium sessionsʢ4 ~ 9ʣɿ࠷΋ଟ͍έʔεͰɼ͍͔ʹͯ͠ΑΓ๛͔ͳจ຺Λ׆༻Ͱ͖Δ͔͕େࣄ
    User Information
    Non-anonymous sessionsɿϢʔβʔͷաڈͷ৘ใΛར׆༻Ͱ͖ΔͷͰɼ͍͔ʹ௕ظͷᅂ޷ੑΛଊ͑Δ͔͕՝୊
    Anonymous sessionsɿݱࡏηογϣϯͷ৘ใ͚͔ͩΒɼ͍͔ʹᅂ޷ੑΛଊ͑Δ͔͕՝୊
    Session-data Structure
    Single-level session dataɿิॿ৘ใͷͳ͍anonymous sessionsͷ͜ͱͰɼcold-start΍data sparsityͷ໰୊΁ͷରॲ͕՝୊
    Multi-level session dataɿ
    Adding Explicit User Representation
    User embedded models
    Ϣʔβʔ৘ใͷຒΊࠐΈΛՃ͑Δ͜ͱͰੑೳ͕޲্ [Tang+ 2018]
    ͨͩ͠୯ͳΔϢʔβʔ৘ใͷ׆༻͚ͩͰ͸cold-startͷ໰୊΍ɼϢʔβʔͷؔ৺ͷมԽ΁ͷରԠͷ໰୊͕ൃੜ
    User recurrent models
    Ϣʔβʔ৘ใΛ࣌ܥྻతʹϞσϧԽ͢Δ͜ͱͰGRU4RecΛ3.5%վળ [Quadrana+ 2018]
    Session-aware/based Recommendationʹ͓͚Δ՝୊
    [Fang+ 2020]
    

    View Slide

17. Session-aware recommendationͷݱঢ়
    session-awareͷݚڀࣗମ͕·ͩᴈ໌ظ
    session-awareͷݚڀ͸·ͩগͳ͘ɼͦͷେ൒͕2014೥Ҏ߱ʹ
    ग़൛͞Εͨ΋ͷʢACM Digital Library಺ʣ[Quadrana+ 2018]
    աڈͷཤྺ͕௕͍΄Ͳੑೳ޲্ʢw/๛෋ͳσʔλʣ [Pi+ 2019]
    AUC 0.02ͷվળ͕අ༻ʹݟ߹͏͔͸ཁݕ౼
    ~ 1000ͷܥྻ௕͕༻͍ΒΕΔ
    latencyͱstorage cost͕Ϣʔβʔߦಈܥྻͷ௕͞ʹґଘͯ͠ઢ
    ܗʹ૿Ճ [Pi+ 2019]
    աڈ50ͷϢʔβʔߦಈΛอ࣋ [Zhout+ 2018]
    աڈ100/200/1000ͷϢʔβʔߦಈΛอ࣋ [Pi+ 2019]
    Session-based΍ݹయతख๏ʹྼޙ͢Δ͜ͱ͕͋Δ [Dacrema+
    2019] [Latifi+ 2020]
    Session-aware͸ΑΓ๛͔ͳ৘ใΛ༻͍ΔͷͰ࠷దԽ͕೉͍͠
    ༧ଌର৅ʹ࣌ؒతʹۙ઀͢Δߦಈ͕Ұ൪ॏཁ [Chen+ 2019]
    [Quadrana+ 2018]
    [Pi+ 2019]
    [Zhout+ 2018] [Pi+ 2019]
    

    View Slide

18. session-aware vs session-based
    session session session session session session
    2021/11/2 2021/11/6
    session-based
    ΄ͱΜͲͷ৔߹session-aware΋ಛ௃ྔͷҙຯͰ͸session-basedͱมΘΒͳ͍
    ͲͪΒ΋session͕ೖྗͳͷ͸มΘΒͣɼͦͷ਺΍ͦΕΛѻ͏ΞϧΰϦζϜ͕ҧ͏͚ͩͳͨΊ
    →ɹsession-based/awareʹؔΘΒͣ༧ଌʹॏཁͦ͏ͳಛ௃ྔΛௐࠪ͢Δ
    session-aware
    [Ludwelg 2020]
    user 1
    

    View Slide

19. ໨࣍
    1. Session-aware recommendationͷ֓ཁ
    2. Session-based/aware recommendationͷ༧ଌͰॏཁͳಛ௃ྔʹ͍ͭͯͷௐࠪ
    • SIGIR eCom 21 Data Challenge
    • جຊߦಈಛ௃ྔͱಛ௃ྔબ୒
    • ͍͔ͭ͘ͷྫ
    • Struggling search
    • Information Seeking Conversation
    • Churn Prediction
    • ͦͷଞ
    3. ·ͱΊ
    

    View Slide

20. SIGIR eCom 21 Data Challenge
    SIGIR eCom 21 ͰCoveoʹΑͬͯ։࠵͞Εͨίϯϖ
    1ͭͷηογϣϯͷ৘ใ͔ΒϢʔβʔߦಈΛ༧ଌ͢Δ
    ࢀՃऀ͸ҎԼͷ2ͭͷλεΫʹऔΓ૊Ή
    ਪનλεΫʢsession-based recommendationʣ
    ॳΊͷkݸͷΠϕϯτ͔Βಉηογϣϯ಺Ͱͷ঎඼ΛΊ͙ΔߦಈΛ༧ଌ
    ҙਤ༧ଌλεΫʢcart-abandonment predictionʣ
    Χʔτʹ঎඼͕ೖΕΒΕͨޙɼಉηογϣϯ಺Ͱ঎඼͕ߪೖ͞Ε͔ͨΛ༧ଌ
    

    View Slide

21. σʔληοτ
    36MΠϕϯτΛؚΉσʔλͰҎԼͷ3ͭͷcsvϑΝΠϧ͔ΒͳΔ
    browsing_train.csv
    ঎඼ϖʔδͰͷϢʔβʔΠϕϯτͷϩάͱඇ঎඼ϖʔδʢFAQͱ͔ʣͰͷӾཡߦಈ
    Ӿཡ͔Πϕϯτ͔ɼΧʔτ΁௥Ճ͔ߪೖ͔঎඼ͷৄࡉΛݟΔ͔ͳͲ
    search_train.csv
    ηογϣϯʹؔΘΔݕࡧΠϕϯτ
    ݕࡧ୯ޠͷΫΤϦϕΫτϧɼݕࡧ݁Ռͷ঎඼ͷϦετɼΫϦοΫ͞Εͨ঎඼ͳͲ
    sku_to_content.csv
    ঎඼ͷϝλσʔλ৘ใ
    Ձ֨ɼ঎඼ΧςΰϦʔɼ঎඼ը૾ͱ঎඼ͷઆ໌จΛදݱ͢ΔϕΫτϧͳͲ
    ৄࡉ͸ϦϙδτϦΛࢀর
    https://github.com/coveooss/SIGIR-ecom-data-challenge
    

    View Slide

22. Accepted papers
    1 . Transformers with Multi-modal Features and Post-fusion Context for E-commerce Session- based Recommendation
    [Gabriel Moreira, Sara Rabhi, Ronay Ak, Md Yasin Kabir and Even Oldridge]
    Transformer-XLͱXLNetͷΞϯαϯϒϧ
    2 . Comparison of Transformer-Based Sequential Product Recommendation Models for the Coveo Data Challenge
    [Elisabeth Fischer, Daniel Zoller and Andreas Hotho]
    SASRecɼBERT4RecɼKeBERT4Recͷ૊Έ߹Θͤ
    3 . Utilizing Graph Neural Network to Predict Next Items in Large-sized Session-based Recommendation Industry Data
    [Tianqi Wang, Zhongfen Deng, Houwei Chou, Lei Chen and Wei-Te Chen]
    KNN, s-KNN, SASRec, DynamicRec, SRGNNΛ࢖༻
    4 . Session-based Recommender System Using an Ensemble of Multiple NN Models with LSTM and Matrix Factorization
    [Yoshihiro Sakatani]
    LSTMͱߦྻ෼ղ
    5 . Adversarial Validation to Select Validation Data for Evaluating Performance in E-commerce Purchase Intent Prediction
    [Shotaro Ishihara, Shuhei Goda and Hidehisa Arai]
    GBDT, NN, ϧʔϧϕʔεͷΞϯαϯϒϧ
    6 . A Session-aware DeepWalk Model for Session-based Recommendation
    [Kaiyuan Li, Pengfei Wang and Long Xia]
    page-viewͱproduct interactionͦΕͧΕͰάϥϑΛ࡞ͬͯDeepWalkΛ࢖༻ͯ͠ϊʔυຒΊࠐΈΛͯ͠ɼGRU
    

    View Slide

23. ৘ใ͕গͳ͍σʔλͷѻ͍
    ग़ݱස౓/interaction͕গͳ͍঎඼ͷѻ͍
    [1] Interaction͕5ճҎԼͷ঎඼Λಉ͡঎඼IDͱͯ͠·ͱΊΔ
    [2] Ұ౓͔͠interaction͕ͳ͍ΞΠςϜΛআڈ
    [4] ग़ݱස౓͕8ճҎԼͷ঎඼͸single eventͱͯ͠·ͱΊΔ
    page-viewͷΈͷURLͷѻ͍
    [1] page-viewͷURLΛԾ૝తͳ঎඼ͱͯ͠ѻ͏
    ɹʢඇ঎඼ϖʔδӾཡʢ70%ʣ͸঎඼Ӿཡʢ28%ʣ΍ݕࡧΫϦοΫ( 2%)ΑΓଟ͍ʣ
    [3] ঎඼ͱͷinteractionͷͳ͍page-viewΠϕϯτͷআڈ
    ৘ใ͕গͳ͍sessionͷѻ͍
    [2] Ұ౓͔͠interaction͕ͳ͍sessionΛআڈ
    [3] Ұͭͷ঎඼ؚ͔͠ΜͰ͍ͳ͍sessionΛςετηοτ͔Βআڈ
    ॏෳ͢ΔΠϕϯτͷѻ͍
    [1] ঎඼ͱͷ࠷ॳͷinteractionͷΠϕϯτ͚ͩΛ࢒͢ʢશମͷ13%͕ॏෳʣ
    [4] ಉҰηογϣϯ಺ͰಉURLʹରͯ͠ى͖Δpage-viewΛ࡟আ
    [6] ॏෳͨ͠Ϣʔβʔinteraction͸࠷ޙͷ͚ͩ࢒͢
    

    View Slide

24. ঎඼ಛ௃ྔ
    [1] όέοτ঎඼Ձ֨Λ঎඼ʢαϒʣΧςΰϦͷฏۉͰׂͬͨྔ
    ಛఆ঎඼΁ͷinteractionͷ਺ɼಛఆͷ঎඼ͷৄࡉΛݟ͔ͨɼಛఆͷ঎඼ΛΧʔτʹೖΕ͔ͨΛಛ௃ྔͱͯ͠௥Ճ
    [4] hashed product IDͷ਺ɼhashed categoriesɼ֤ηογϣϯͷ࠷ॳͷ঎඼ɼ࠷ॳͷ঎඼ͷΧςΰϦɼϝλσʔλͷ༗ແ
    Χ΢ϯτಛ௃ྔ
    [4] Πϕϯτͷྦྷੵ࿨ɼ֤঎඼Πϕϯτؒͱ঎඼ΠϕϯτޙͷϖʔδϏϡʔͱݕࡧΠϕϯτͷճ਺΋ಛ௃ྔͱͯ͠࢖༻
    ࣌ؒಛ௃ྔ
    [1] ঎඼͕ॳΊͯݟΒΕ͔ͯΒͷ࣌ؒɽϩάʢ೔ʣɼ࣌ؒɼ༵೔
    [4] िɼ೔ɼ࣌ؒͱɼि຤͔൱͔Λಛ௃ྔͱͯ͠࢖༻
    ͦͷଞಛ௃ྔ
    [1] ݕࡧςʔϒϧͰར༻ՄೳͳݕࡧΫϦοΫΠϕϯτ΋Πϕϯτʹ௥Ճ
    ΠϕϯτλΠϓɼ঎඼ʢαϒʣΧςΰϦɼόέοτՁ֨ɼ঎඼IDΛྡ઀IDͷͱͯ͠·ͱΊΔ
    browsing interaction, search query information, categorical meta dataΛͻͱͭͷΠϕϯτͱͯ͠·ͱΊΔ
    [4] Πϕϯτͷ௕͞΋ಛ௃ྔͱͯ͠࢖༻
    ৽نಛ௃ྔͷ࡞੒
    

    View Slide

25. GBDTͷ༧ଌʹޮ͍͍ͯͨಛ௃ྔ
    [5] ঎඼ΛΧʔτʹೖΕ͔ͨ൱͔ɼ͍࣍Ͱ঎඼ͷৄࡉΛݟ͔͕ͨ༧ଌʹޮ͍͍ͯͨ
    ॏཁͳಛ௃ྔ
    

    View Slide

26. ଟ͘ͷάϧʔϓ͕ɼॏෳ͢Δಛ௃ྔͷू໿΍ɼ਺ͷগͳ͍ಛ௃ͷ࡟আɾม׵Λ͓͜ͳ͍ͬͯͨ
    ʮ๲Ε্͕Δσʔλʹ೗Կʹରॲ͢Δ͔ʁʯ͸Ϗδωε্Ұͭͷ՝୊Ͱ͋Δͱೝࣝ
    Ծʹ۩ମతͳ༧ଌର৅΍Ϣʔεέʔε͕ఆ·Γ͖͍ͬͯͳ͍৔߹ɼ༧ଌೳྗͷߴ͔ͭ͘େ͖͘ॖ໿͞Εͨಛ௃Λ
    ଈ࠲ʹ࡞Δͷ͸೉͍͠ͱࢥΘΕΔ
    େ఍ͷ৔߹શମͷ1~2ׂͷಛ௃Λอ͍࣋ͯ͠Ε͹શମͷ8~9ׂΛઆ໌Ͱ͖Δʢ͜ͱ͕ظ଴͞ΕΔʣ
    ·ͣ͸ɼجຊతͳಛ௃ྔͷ͏ͪ৘ใͷ৑௕ੑ͕͋Δ΋ͷΛ࡟ݮ͢Δ͜ͱͰɼσʔλอ࣋ͷίετΛݮΒ͢͜ͱ͕
    Մೳ͔΋͠Εͳ͍
    SIGIR e-commerce Data challengeʹ͍ͭͯͷௐ͔ࠪΒͷࣔࠦ
    

    View Slide

27. ͋Δ༗໊ͳECαΠτͷ2018೥ͷ2ϲ݄෼ͷϒϥ΢δϯάσʔλ͔Β৽͍͠σʔληοτΛ࡞੒ʢData Challengeͷσʔληοτʣ
    ϢʔβʔߦಈΛpage view, detail ,add, remove, purchase, clickͷ6ͭʹ෼ྨ
    ʮߪೖʯؚ͕·Ε͍ͯΔܥྻ͸ɼͦͷ௚લ·Ͱͷ෦෼ܥྻͷΈΛ࢖༻
    ͜ͷσʔληοτͰϢʔβʔ͕঎඼Λߪೖ͔ͨ͠൱͔Λ༧ଌ͠ɼ͔ͳΓجຊతͳಛ௃ྔ͚ͩͰ͋Δఔ౓༧ଌՄೳͰ͋Δ͜ͱΛओு
    લॲཧͳͲ
    5 clickҎ্155 clickҎԼͷσʔλͷΈ࢖༻
    ʮߪೖʯ͕͓͜Δܥྻͱى͜Βͳ͍ܥྻͷ਺͕ۉߧ͕औΕΔΑ͏ʹαϯϓϦϯά
    ಛ௃ྔΛ࡞੒ͯ͠XGBoostͰ༧ଌ͢Δ৔߹ʢhand-craft feature engineeringʣͱneural networkͰ༧ଌ͢Δ৔߹ʢneural
    networkʣͦΕͧΕݕ౼ʢຊൃදͰ͸ޙऀͷΈઆ໌ʣ
    Shopper Intent Prediction from Clickstream E-commerce Data with Minimal Browsing
    Information [Requena+ SciRepo 2020] (Coveo)
    

    View Slide

28. hand-craft feature engineering
    k-gram
    ϢʔβʔߦಈΛฒ΂ͨ΋ͷͷk-gram
    ྫɿ2-gramʮ34ʯʮ12ʯ
    K-gram୯ମ͚͔ͩΒ͸ϢʔβʔߦಈΛ༧ଌ͢Δͷ͸೉ͦ͠͏
    horizontal visibility graph motifs (HVGM)
    ϊʔυΛҰྻʹฒ΂ͨແ޲άϥϑͰɼk-gramΑΓߴ࣍ͷύλʔϯʢZ1 ~ Z6ʣ͕ܭࢉՄೳ
    ͜ΕΒͷ֬཰ʹ͍ͭͯͷΤϯτϩϐʔʢhzʣ΋ܭࢉ
    ͜ΕΒͷಛ௃ྔΛ࢖ͬͯXGboostͰߪೖ͔ͨ͠൱͔Λ༧ଌ
    Shopper Intent Prediction from Clickstream E-commerce Data with Minimal Browsing
    Information [Requena+ SciRepo 2020] (Coveo)
    

    View Slide

29. ॏཁͳಛ௃ྔ/GBDT
    add (3) ͚ͩͰͳ͘add-view (31) ͕େࣄ
    ͨͩͷӾཡߦಈͰ͋Δview-detail (12) ͕େࣄ
    ϖʔδؒભҠview-view (1)΍঎඼ؒભҠdetail-detail (2) ͸͋Δ͕ݕࡧ search (6) ͸গͳ͍
    ॏཁͳಛ௃ྔ/SHAP
    view-detail (12) ΍ detail (2)΍search (6)͕ߴ͍ͱߪങʹࢸΒͳ͍ (SHAP஋͕negative)
    ͜ͷͱ͖Z1͕ߴͯ͘hz͕௿͍ˠ͜ͷΑ͏ͳϢʔβʔߦಈ͸୯ௐ
    ྫɿview-view-view-detail-view-detail (111212)ͷ৔߹ʢϢʔβʔ͕঎඼Λݟ͚ͭΔ໨
    తͰϖʔδΛ୳ͯ͠໭ͬͨΓΛ͍ͯ͠Δʣ
    view-view (11) ΍detail (2) ͕ߴ͘ͳΔ
    Z1 ͕2ͭͰ Z4͕1ͭͳͷͰhz͸த͘Β͍ʹͳΔ
    ঎඼ϖʔδΛ๚Εͣଟ͘ͷϖʔδΛ๚ΕΔ৔߹Ͱ (high view P(1))ɼ༷ʑͳߦಈ͔ΒӾཡʹ
    ࢸΔ৔߹ (low view-view P(1, 1) and high detail-view P(2, 1), add-view P(3, 1))ɼ঎඼ߪೖ
    ଞͷ঎඼ϖʔδ͔Β௚઀ඈͿ৔߹ (high detail-detail P(2, 2))΋ߪങʹͭͳ͕Δ
    Τϯτϩϐʔhz͕ߴ͍ύλʔϯͱͯ࣍͠ͷΑ͏ͳ΋ͷ͕͋Δ
    view-add-detail-view-add-view (132131) ͕ͩ detail P(2) ͱ view-view P(1, 1)͕௿͘ɼ
    view P(1) ɼadd P(3), ɼadd-view P(3, 1) ɼdetail-view P(2, 1)͕ߴ͍
    ༧ΊԿ͕΄͍͔ܾ͠·͍ͬͯΔΑ͏ͳڍಈͰ͋Γɼ͜Ε͕ߪೖʹͭͳ͕Δͷ͸ଥ౰
    Shopper Intent Prediction from Clickstream E-commerce Data with Minimal Browsing
    Information [Requena+ SciRepo 2020] (Coveo)
    

    View Slide

30. ͋Δe-commerceͷσʔληοτΛ༻͍ͯsequence-aware recommendationʹ͓͍ͯԿ͕ॏཁ͔Λௐࠪ
    ໿95ͷಛ௃ྔΛ࡞੒͠ɼGain ratioɼChi-squaredɼInformation gain ratioɼGini indexʹΑͬͯॏཁ౓ΛධՁ
    શಛ௃ྔ͸HPࢀর https://rn5l.github.io/session-rec/umuai/
    Exploring Session-Awareness in E-Commerce [Ludewig 2020]
    [Jannach+ 2017]
    

    View Slide

31. SIGIR e-commerce Data challengeɼ[Requena+ 2020]ɼ [Ludewig 2020]ͰࢀߟʹͳΔͷ͸ɼૉ๿ͳಛ௃ྔ࡞੒
    → ಛ௃ྔબ୒ͱ͍͏ϓϩηε
    Ϣʔβʔߦಈ෼ੳʹ͓͍ͯ͸ɼ·ͣ͸͋Δಛఆͷ໰୊ʹରͯ͠ૉ๿ͳಛ௃ྔΛ࡞੒͠ɼ࣮ࡍʹ༧ଌ໰୊Λͱ͍ͯͦ
    ͷد༩ΛΈΔ͜ͱͰɼॏཁͳಛ௃ྔΛܦݧతʹධՁ͍ͯ͘͠ͷ͕ॏཁͩͱࢥΘΕΔ
    ॏཁͳಛ௃ྔͷԾઆ͕ੵΈ্͕Ε͹ɼΑΓཉ͍͠৘ใͷղ૾౓্͕͕ΔͷͰɼઌߦݚڀɾࣄྫͷαʔϕΠ΋ޮ཰
    Խ͠ɼԾઆ͕ΑΓڧԽ͞Εɽɽͱ͍͏ਖ਼ͷϑΟʔυόοΫϧʔϓ͕ճΔ
    اۀͰϢʔβʔߦಈ༧ଌʹܞΘ͍ͬͯΔํʑ΋݁ہͦΕ͕େࣄͩͱ͍͍ͬͯΔ
    [Challenges, Best Practices and Pitfalls in Evaluating Results of Online Controlled Experiments, KDD 2019]
    ͕ͨͬͯ͠ɼॏཁͳಛ௃ྔʹ͍ͭͯͷԾઆݕূΛճ͢࢓૊ΈࣗମΛͭ͘ΓͦΕΛճ͢͜ͱ͕ॏཁͰ͸ͳ͍͔ͱߟ͑Δ
    ͜͜·Ͱͷௐ͔ࠪΒͷࣔࠦ
    

    View Slide

32. GRUΛ࣋ͭDNNʹΑΔϚϧνλεΫֶशͰuser intent prediction
    ๲େͳσʔλʹର͢Δ౷ܭతͳ෼ੳͷ݁ՌɼҎԼͷಛ௃ྔ͕Ϣʔβʔͷߪങ༧ଌʹॏཁͰ͋Δ͜ͱΛൃݟ
    Ϣʔβʔͷجຊతͳ৘ใ
    ੑผɼ೥ྸɼաڈҰ೥ͷߪೖϨϕϧɼॅΜͰ͍Δ஍ҬͳͲ
    Ϣʔβʔߦಈʢ೔୯Ґɼ݄୯Ґʣ
    աڈ਺ճͷߪങߦಈɼొ࿥͔Βͷ೔਺ɼ࠷େ࿈ଓΞΫςΟϒ೔਺ɼաڈ਺ճͷ׆ಈͷมԽɼηʔϧεϓϩϞʔγϣϯ࣌ͷߪങߦಈ
    ਺஋ಛ௃ྔΛ͢΂ͯΧςΰϦΧϧಛ௃ྔʹม׵
    One-hotͩͱεύʔεʹͳΔͷͰɼຒΊࠐΈ
    TPG-DNN: A Method for User Intent Prediction Based on Total Probability Formula and
    GRU Loss with Multi-task Learning [Jiang+ ACM 2017] (Alibaba)
    

    View Slide

33. ΤοδσόΠεʹ͓͚ΔDNNʹΑΔuser intent prediction
    swipe-interactive behaviorɼtap-interactive behaviorɼbrowse-interactive behaviorͷߦಈάϧʔϓΛ࡞੒
    swipe-interactive behavior
    open page, leave page, swipe, tap͕جຊߦಈʢtapҐஔ΍swipe࢝ऴҐஔ΍ܧଓ࣌ؒ΋ه࿥ʣ
    time duration of a swipe, time gap between two actions, positional coordinates of actions
    (tap position X/Y, swipe start position X/Y, swipe end position X/Y and swipe length on X/Y)
    page indices, action indices and swipe directions
    tap-interactive behavior
    timestampɼpage indexɼevent indexɼbutton index
    browse-interactive behavior
    browse, search, collect, add to cart, purchaseͷ̐ͭͰɼ֤ߦಈʹରͯ࣍͠ͷ6ͭͷಛ௃Λநग़
    type index, top category index, leaf category index, page index, page stay time and timestamp
    ༧ଌ݁Ռʹج͍ͮͯɼϙοϓΞοϓ޿ࠂɼϓογϡ௨஌ʹ͍ͭͯෳ਺ͷઓུΛߟ͑ABςετ
    ఏҊख๏͸Ϣʔβʔͷreal-timeͳᅂ޷ੑΛ͏·͘ଊ͍͑ͯΔ͜ͱΛ࣮ূ
    EdgeDIPN: a Unified Deep Intent Prediction Network Deployed at the Edge [Guo+
    VLDB 2020] (Alibaba, Peking University)
    

    View Slide

34. جຊతͳߦಈಛ௃ྔΛ༻͍ͨɼݕࡧʹ͓͚Δຬ଍౓ͷ༧ଌ
    Deep Sequential Models for Task Satisfaction Prediction / User Interaction Sequences
    for Search Satisfaction Prediction [Mehrotra+ CIKM/SIGIR 2017] (UCL, Microsoft)
    

    View Slide

35. ୳ࡧతݕࡧߦಈͱۤઓ͍ͯ͠Δ࣌ͷݕࡧߦಈͷफ़ผ
    ಛ௃ྔ
    Struggling or Exploring? Disambiguating Long Search Sessions [Hassan+ WSDM 2014]
    (Microsoft)
    ۤઓ͍ͯ͠Δ͍Δͱ͖ͷݕࡧߦಈ
    ۤઓ͍ͯ͠Δͱ͖͸࠷ॳͷΫΤϦͱͷྨࣅ౓͕ߴ͍͕୳ࡧͷ৔߹͸গͣͭ͠ྨࣅ౓͕Լ͕Δ
    ୳ࡧͷ৔߹͸ɼޙ൒ͷηογϣϯʹͳΔ΄ͲΫϦοΫ਺͕૿͑Δ
    

    View Slide

36. ۤઓ͍ͯ͠Δ͕࠷ऴతʹॴ๬ͷ݁ՌΛಘΒΕͨ࣌ͷݕࡧߦಈ [Odijk+ 2015]
    ΫϦοΫ͕গͳ͍ɼΫΤϦΛΑΓ۩ମతͳ΋ͷʹมߋ͢ΔͳͲ
    ݕࡧ݁Ռʹຬ଍͢Δ͔൱͔ͷ༧ଌʹ͸࣍ͷΫΤϦͱͷྨࣅ౓ͳͲͷಛ௃ྔ͕େࣄ [Diriye+ 2012]
    ΫΤϦͷมߋ΍ݕࡧ݁ՌͷΫϦοΫ͸ݕࡧͷࠔ೉͞΍ݕࡧʹର͢ΔϑϥετϨʔγϣϯΛද͢ [Aula+ 2010] [Allan+ 2010]
    ΫϦοΫΛڬ·ͣΫΤϦ͕Կ౓΋มߋ͞Ε͍ͯΔͱ͖͸ݕࡧ݁ՌʹෆຬΛ͍͍ͩͯΔ [White+ 2009]
    E-commerceͷݕࡧʹ͓͍ͯɼΫΤϦ͸มߋ͞ΕΔͨͼʹ௕͘ͳΓɼΫΤϦมߋ͸ΫϦοΫ਺ͷ૿Ճͱ࠷ऴతͳߪങ΁ͱͭͳ
    ͕ͬͨ ʢҎԼͷಛ௃ྔͰΫΤϦมߋͷ༧ଌΛߦͬͨʣ[Hirsch+ 2020]
    ݕࡧߦಈͱຬ଍/ෆຬ଍/ۤઓ
    

    View Slide

37. conversational assistantͱͷϚϧνλʔϯͷձ࿩ʹ͓͚Δuser intent prediction
    લॲཧ
    Ϣʔβʔͷ࣭໰ҙਤΛͦͷ಺༰͔Β12ͷΫϥεʹϥϕϧ෼͚
    Greetings/Gratitude, Junk, Others͸༧ଌʹߩݙ͠ͳ͍ͷͰআ֎
    ͜ͷϥϕϧͷ૊Έ߹Θͤʹରͯ͠Ϣʔβʔߦಈ͔Β༧ଌΛߦ͏
    ग़ݱස౓͕গͳ͍ϥϕϧͷ૊Έ߹Θͤ͸ͦΕΛߏ੒͢Δ͍ͣΕ͔ͷϥϕϧʹϥϯμϜʹஔ͖׵͑
    ϥϕϧͷ૊Έ߹ΘͤͰग़ݱස౓͕େ͖͘ҟͳΔʢτοϓ32ͷϥϕϧͷ૊Έ߹Θ͕ͤશମͷ90ˋΛ઎ΊΔʣ
    ྫɿCQ+FD+IR+RQ → CQ
    User Intent Prediction in Information-seeking Conversations [Qu+ CHIIR 2019] (U M
    Amherst, RMIT U, Alibaba)
    ࢀߟ
    

    View Slide

38. ಛ௃ྔ
    Ϣʔβʔͷൃ࿩ʹ͍ͭͯɼ಺༰ʹؔ͢Δ΋ͷʢContent/Conʣɼର࿩ͷߏ଄ʹؔ͢Δ΋ͷʢStructure/Strʣɼײ৘ʹؔ͢Δ΋ͷ
    ʢSentiment/senʣͦΕͧΕʹରͯ͠ಛ௃ྔΛ࡞੒
    ߏ଄ʹؔ͢Δಛ௃ྔ͕Ϣʔβʔҙਤͷ༧ଌʹॏཁ
    ಛʹɼൃݴ͕Ͳ͜Ͱൃੜ͔ͨ͠ʢAbsolute Position/AbsPosɼNormalized Position/NormPosʣͳͲ͕ॏཁͩͬͨ
    ಺༰ʹؔ͢Δ΋ͷͰ͸࠷ॳͷൃ࿩ͷྨࣅ౓ʢInitial Utterance Similarity/InitSimʣ΍ର࿩ͷྨࣅ౓ʢDialog Similarity/DigSimʣ͕ॏཁ
    15Ґ͔Β࠷ԼҐ·Ͱ͸͢΂ͯʮ͋ΔΩʔϫʔυΛؚΜͰ͍͔ͨʁʹؔ͢Δಛ௃ྔʯ
    User Intent Prediction in Information-seeking Conversations [Qu+ CHIIR 2019] (U M
    Amherst, RMIT U, Alibaba)
    ࢀߟ
    

    View Slide

39. Frequency
    Ϣʔβʔ͕ͲΕ͚ͩසൟʹαʔϏεΛར༻͢Δ͔
    Recency
    Ϣʔβʔ͕࠷ޙʹαʔϏεΛར༻͔ͯ͠ΒͲΕ͚ͩܦ͔ͭ
    Increasing/Decreasing
    Ϣʔβʔͷར༻ස౓͸૿͍͑ͯΔ͔ɾݮ͍ͬͯΔ͔
    Churn Prediction (1/3)
    https://www.iguazio.com/sessions/product-madness-an-aristocrat-co-on-predicting-1st-day-churn-in-real-time/
    ར༻ස౓
    ࣌ؒ
    

    View Slide

40. ήʔϜʹ͓͚ΔϢʔβʔ཭୤༧ଌ
    Churn Prediction (2/3)
    [Xi+ 2019 A Latent Feelings-aware RNN Model for User Churn Prediction with Behavioral Data]
    

    View Slide

41. Telco Customer ChurnΛ༻͍ͨɼ௨৴αʔϏεղ໿༧ଌ
    Churn Prediction (3/3)
    EDA ಛ௃ྔͷ෼ੳ
    [AI ʹΑΔୀձ཈ࢭ γϦʔζ ʮࣄۀܦӦ΁ͷ AI ׆༻ʯᶃ ࡾඛUFJϦαʔν&ίϯαϧςΟϯά]
    

    View Slide

42. Ϣʔβʔͱ঎඼ͷinteractionͷಛ௃ྔΛೖΕΔ͜ͱͰ༧ଌਫ਼౓͕͕͋ͬͨ [Zhang+ 2020]
    ಛఆͷ঎඼ΛΫϦοΫͨ͠ճ਺ɼಛఆͷ঎඼ͷuser collectionͷ਺ɼಛఆͷ঎඼ΛΧʔτʹೖΕͨ਺
    ͦͷଞ
    

    View Slide

43. ໨࣍
    1. Session-aware recommendationͷ֓ཁ
    2. Session-based/aware recommendationͷ༧ଌͰॏཁͳಛ௃ྔʹ͍ͭͯͷௐࠪ
    • SIGIR eCom 21 Data Challenge
    • جຊߦಈಛ௃ྔͱಛ௃ྔબ୒
    • ͍͔ͭ͘ͷྫ
    • Struggling search
    • Information Seeking Conversation
    • Churn Prediction
    • ͦͷଞ
    3. ·ͱΊ
    

    View Slide

44. ݁࿦
    ༧ଌʹͱͬͯԿ͕ॏཁͳಛ௃ྔ͕Կ͔͸༧ଌର৅/υϝΠϯʹґଘ͢Δ
    ߪങߦಈͷ༧ଌʹ͸ߪങʹؔ࿈͢Δߦಈಛ௃ྔ͕ޮ͍ͯ͘Δ͠ɼ཭୤཰ͷ༧ଌʹ͸recency͕ޮ͘
    ۜߦआೖΕͷ༧ଌͳΒआ͕ۚ͋Δ͔͕ޮ͍ͯ͘Δ͠ɼήʔϜͷ཭୤༧ଌʹ͸ͲΜͳήʔϜΛ΍͔͕ͬͨޮ͘
    Session-awareʹϢχʔΫͳಛ௃ྔͷٞ࿦͸͋·Γͳ͞Ε͍ͯͳ͍ҹ৅
    Session-awareͳख๏͸݁ہsession͕جຊ୯ҐͳͷͰɼsession-basedͱൺ΂ͯѻ͏sessionͷ਺͕૿͚͑ͨͩ
    Session-awareͷݱࡏͷख๏ͷओྲྀ͸neural network͕ͩneural networkͰ͸ಛ௃ྔΛ࡞Γࠐ·ͳ͍
    ίϯϖͰ͸৘ใ͕গͳ͍sessionͷ৘ใΛ͍͔ʹѻ͏͔͕ҰͭϙΠϯτͩͬͨҹ৅
    ొ৔ස౓͕গͳ͍΋ͷΛ࡟আͨ͠Γ৑௕ͳ৘ใΛ΋ͭ΋ͷΛ·ͱΊͨΓ
    ૉ๿ʹEDAΛ΍ͬͨ݁ՌΘ͔ͬͨ͜ͱͩͱ͍͏ͷ͕େࣄ
    

    View Slide

45. ܥྻ৘ใͷอ࣋ͱτϨʔυΦϑ
    ৘ใྔ
    ѹॖ཰
    ߦಈಛ௃ྔ
    ൚༻ੑ
    ੜσʔλ
    ྫɿclick
    جຊಛ௃ྔ
    ྫɿclick਺
    υϝΠϯಛԽಛ௃ྔ
    ྫɿsnap_view
    ߴ
    ௿
    গ
    ଟ
    ௿
    ߴ
    ɾɾɾ
    ҰൠʹͲΕ͘Β͍ಛ௃ྔΛੜσʔλʹ͍ۙܗͰอ࣋͢Δ͔͸ଞͷ༷ʑͳཁૉͱτϨʔυΦϑͷؔ܎ʹ͋Δ͸ͣ
    Ϗδωε্ಛʹॏཁͳ෼ੳ࣠ຖʹͲ͜·ͰଥڠͰ͖Δ͔Λߟ͑Δ͜ͱͰɼͲͷཻ౓ͰσʔλΛอ࣋͢Δͷ͔͕ܾ·Γͦ͏
    

    View Slide

46. େࣄͩͱࢥͬͨ͜ͱ
    ར༻໨తͷ໌֬Խ
    Կʹ༻͍Δ͔͕ఆ·Γ͖͍ͬͯͳ͍ͳΒੜσʔλɾجૅతͳߦಈಛ௃ྔΛ࣋ͭ֎ͳ͍
    ఆ·͍ͬͯͳ͍৔߹ɼͱΓ͋͑ͣجૅతͳಛ௃ྔΛ࡞੒্ͨ͠Ͱ৑௕ͳಛ௃ྔΛ࡟আ͢Δ͘Β͍͸Ͱ͖ͦ͏
    ग़ݱස౓ͷগͳ͍/ॏෳ͢Δaction/session/itemͷ࡟আ΍ू໿Ͱσʔλอ࣋ͷίετΛҰ࣍తʹԼ͛ΔͳͲ
    ಛ௃ྔબ୒ͷԾઆݕূαΠΫϧ
    Ͳͷಛ௃ྔ͕ޮ͔͘ɾ൚༻త͔͸݁ہ΍ͬͯΈͳ͍ͱΘ͔Βͳ͍ͷͰԾઆݕূΛճ͍ͯ͘͠࢓૊ΈࣗମΛ࡞Δͷ͕େࣄͦ͏
    ·ͣ͸ԾઃΛཱͯͯૉ๿ͳಛ௃ྔΛ୔ࢁ࡞ΓɼGBDT΍random forestͳͲͰɼෳ਺ͷҟͳΔυϝΠϯͰਪ࿦͢ΔͳͲ
    ͦͷޙɼ૬ؔ΍SHAP஋ͳͲΛௐ΂ͯ൚༻తʹ༧ଌʹޮ͍͍ͯͦ͏ͳಛ௃ྔΛ୳͍ͯ͘͠ͳͲ
    Session-basedͰ༗ྗͳಛ௃ྔͷར༻
    Session-awareͷಛ௃ྔͱsession-basedͷಛ௃ྔͱ͸େ͖͘มΘΒͳ͍ʢҧ͍͸ओʹΞϧΰϦζϜͷ෦෼ʣ
    ݁ہ༧ଌΠϕϯτʹ࣌ؒతʹ͍ۙߦಈಛ௃ྔ͕ॏཁͰ͋Δ͜ͱ͕ࢦఠ͞Ε͍ͯΔ
    ϢʔβʔID͕खʹೖΒͳ͍ͱsession-awareͳख๏͸࢖͑ͳ͍
    Session-awareʹϢχʔΫͳ࣌ؒํ޲ʹ௵ͨ͠ಛ௃ྔΛ࡞Δ৔߹΋ɼ·ͣ͸ૉ๿/Ұൠతͳ΋ͷΛ࡞੒͠ɼԾઆݕূΛճ͢
    RecencyɼϘοτͷ࢖༻ճ਺ɼొ࿥͔Βͷ೔਺ͳͲ
    

    View Slide

47. [Ludewig 2020] Advances in Session-based and Session-aware Recommendation
    [Fang+ 2020] Deep Learning for Sequential Recommendation: Algorithms, Influential Factors, and Evaluations
    [Wang+ 2021] A Survey on Session-based Recommender Systems
    [Wang+ 2016] Collaborative Recurrent Autoencoder: Recommend While Learning to Fill in the Blank
    [Zhang+ 2019] Deep Learning Based Recommender System: A Survey and New Perspectives
    [Hidashi+ 2016] Session-based Recommendations with Recurrent Neural Networks
    [Gao+ 2019] Learning to Recommend with Multiple Cascading Behaviors
    [Ren+ 2019] RepeatNet: A Repeat Aware Neural Recommendation Machine for Session-based Recommendation
    [Anderson+ 2014] The Dynamics of Repeat Consumption
    [Wan+ 2018] Representing and Recommending Shopping Baskets with Complementarity, Compatibility and Loyalty
    [Hu+ 2020] Modeling Personalized Item Frequency Information for Next-basket Recommendation
    [Wu+ 2017] Session-aware Information Embedding for e-commerce Product Recommendation
    [Tan+ 2016] Improved Recurrent Neural Networks for Session-based Recommendations
    [Li+ 2017] Neural Attentive Session-based Recommendation
    [Kang+ 2018] Translation-based Recommendation
    [Tang+ 2018] Personalized top-n Sequential Recommendation via Convolutional Sequence Embedding
    [Quadrana+ 2018] Sequence-aware Recommender Systems
    ࢀߟจݙ
    

    View Slide

48. [Zhou+ 2018] Deep Interest Evolution Network for Click-Through Rate Prediction
    [Pi+ 2019] Practice on Long Sequential User Behavior Modeling for Click-Through Rate Prediction
    [Dacrema+ 2019] Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Recommendation Approaches
    [Laitifi+ 2020] Session-aware Recommendation: A Surprising Quest for the State-of-the-art
    [Chen+ 2019] AntProphet: an Intention Mining System behind Alipay’s Intelligent Customer Service Bot
    [Jannach+ 2017] Determining Characteristics of Successful Recommendations from Log Data—A Case Study
    [Odijk+ 2015] Struggling and Success in Web Search
    [Diriye+ 2012] Leaving so soon? Understanding and predicting web search abandonment rationales
    [Aula+ 2010] How Does Search Behavior Change as Search Becomes More Difficult?
    [Feild+ 2010] Predicting Searcher Frustration
    [Hirsch+ 2020] Query Reformulation in E-Commerce Search
    [White+ 2009] Characterizing and Predicting Search Engine Switching Behavior
    [Zhang+ 2020] An Improved Deep Forest Model for Prediction of E-commerce Consumers’ Repurchase Behavior
    ࢀߟจݙ
    

    View Slide

49. ΧϥΫϦͰ͸αΠϨϯτΧελϚʔʢ*1ʣΛٹࡁ͢ΔͨΊͷαʔϏε
    ։ൃ΍σʔλ෼ੳʹऔΓ૊ΜͰ͓ΓɺػցֶशΤϯδχΞ΍WebΤϯ
    δχΞΛืू͓ͯ͠Γ·͢ɻ
    Ұॹʹಇ͘͜ͱʹڵຯͷ͋Δํ͸ੋඇ͝࿈བྷ͍ͩ͘͞ʂ
    Twitter: @not_oohikata
    Mail: k.obinata@karakuri.ai
    *1ɿαʔϏεӡӦऀଆʹٙ໰΍ෆ҆Λ໰͍߹ΘͤͣɺαΠτ͔Β཭୤
    ͯ͠͠·͏ސ٬ ( https://karakuri.ai/lp_silentcustomer )
    We’re hiring !
    

    View Slide