論文紹介 / metapath2vec: Scalable Representation Learning for Heterogeneous Networks

KDD 2017に投稿された「metapath2vec: Scalable Representation Learning for Heterogeneous Networks」の紹介スライドです。（https://ericdongyx.github.io/papers/KDD17-dong-chawla-swami-metapath2vec.pdf）

metapathを用いてノードのシーケンスを生成し、Skip-gramによってノードのEmbeddingを獲得する手法です。

研究室のゼミでの紹介用につき、口頭での説明がないぶん分かりにくいところがあるかもしれませんが、ご容赦くださいませ。

Transcript

1. NFUBQBUIWFD4DBMBCMF
   3FQSFTFOUBUJPO-FBSOJOHGPS
   )FUFSPHFOFPVT/FUXPSLT
   ,%%`
   
   :VYJBP %POH /JUFTI 7$IBXMB "OBOUISBN 4XBNJ
   ࿦จ঺հ!LVSJJWF
   

   View Slide

2. ໨࣍
   wݚڀഎܠ
   w࿦จ֓ཁ
   wఏҊख๏
   w࣮ݧ
   w݁࿦
   

   View Slide

3. ໨࣍
   wݚڀഎܠ
   w࿦จ֓ཁ
   wఏҊख๏
   w࣮ݧ
   w݁࿦
   

   View Slide

4. ݚڀഎܠʛ)FUFSPHFOFPVT/FUXPSL>
   
   ϊʔυɾΤοδͷλΠϓ͕
   ෳ਺ଘࡏ͢ΔωοτϫʔΫ
   >:J[IPV 4VOBOE+JBXFJ)BO.JOJOH)FUFSPHFOFPVT*OGPSNBUJPO/FUXPSLT1SJODJQMFTBOE.FUIPEPMPHJFT.PSHBO$MBZQPPM1VCMJTIFST
   

   View Slide

5. ݚڀഎܠʛXPSEWFD>Ԡ༻ख๏ͷڵོ
   
   • %FFQ8BML>
   ϥϯμϜ΢ΥʔΫͰੜ੒ͨ͠
   ϊʔυܥྻ͔Β෼ࢄදݱΛੜ੒
   • -*/&>
   ྡ઀͢Δϊʔυͱͷؔ܎͔Β෼ࢄදݱΛੜ੒
   • OPEFWFD>
   ্هख๏Λ૊Έ߹Θͤͯ෼ࢄදݱΛੜ੒
   >#SZBO1FSP[[J 3BNJ"M3GPV BOE4UFWFO4LJFOB%FFQ8BML0OMJOF-FBSOJOHPG4PDJBM3FQSFTFOUBUJPOT*O,%%`"$. r
   >+JBO5BOH .FOH 2V .JOH[IF 8BOH .JOH;IBOH +VO:BO BOE2JBP[IV .FJ-*/&-BSHFTDBMF*OGPSNBUJPO/FUXPSL&NCFEEJOH*O888`"$.
   >"EJUZB(SPWFSBOE+VSF-FTLPWFD/PEF7FD4DBMBCMF'FBUVSF-FBSOJOHGPS/FUXPSLT*O,%%`"$. r
   > 5PNBT.JLPMPW *MZB4VUTLFWFS ,BJ$IFO (SFH4$PSSBEP BOE+F%FBO%JTUSJCVUFESFQSFTFOUBUJPOTPGXPSETBOEQISBTFTBOEUIFJSDPNQPTJUJPOBMJUZ*O/*14`r
   ϊʔυͷ
   ෼ࢄදݱΛ
   ֫ಘ͢Δख๏
   

   View Slide

6. ݚڀഎܠʛطଘख๏ͷ՝୊
   
   )PNPHFOFPVT/FUXPSL
   ʹ͸༗ޮ
   )FUFSPHFOFPVT/FUXPSL
   Ҏ߱)/ͱදه
   ʹ͸ෆे෼
   >:J[IPV 4VOBOE+JBXFJ)BO.JOJOH)FUFSPHFOFPVT*OGPSNBUJPO/FUXPSLT1SJODJQMFTBOE.FUIPEPMPHJFT.PSHBO$MBZQPPM1VCMJTIFST
   ͔͠͠
   

   View Slide

7. ݚڀഎܠʛຊݚڀͷ໨త
   
   )/͕࣋ͭߏ଄ͱҙຯͷ৘ใΛ
   อ࣋ͨ͠෼ࢄදݱͷੜ੒
   

   View Slide

8. ໨࣍
   wݚڀഎܠ
   w࿦จ֓ཁ
   wఏҊख๏
   w࣮ݧ
   w݁࿦
   

   View Slide

9. ࿦จ֓ཁʛͲΜͳ΋ͷ͔ʁ
   
   ԿΛͨ͠ʁ
   )/ʹରͯ͠༗ޮͳNFUBQBUIWFDͱɺ
   ͦͷ೿ੜ൛NFUBQBUIWFDΛఏҊɻ
   ݁ՌͲ͏ͩͬͨʁ
   ༷ʑͳϚΠχϯάλεΫʹର͠
   ࠷ઌ୺ͷख๏Λ্ճΔ͚ͩͰͳ͘ɺ
   ߏ଄ͱҙຯΛอ࣋ͨ͠෼ࢄදݱ͕ੜ੒Ͱ͖ͨ
   

   View Slide

10. ໨࣍
    wݚڀഎܠ
    w࿦จ֓ཁ
    wఏҊख๏
    w࣮ݧ
    w݁࿦
    

    View Slide

11. ఏҊख๏ʛ֓ཁਤ
    
    *OQVUɿ)/(ɺNFUBQBUIεΩʔϜ1
    0VUQVUɿ෼ࢄදݱߦྻ9
    ࣍ݩ E
    ௖
    ఺
    ਺
    7
    ͋Δ௖఺ͷ෼ࢄදݱ
    ผͷ௖఺ͷ෼ࢄදݱ
    

    View Slide

12. ఏҊख๏ʛ)/
    
    )FUFSPHFOFPVT/FUXPSLͷఆٛ
    άϥϑ(ɺ௖఺7ɺล&ɺλΠϓ5ɺ
    Ϛοϓؔ਺Ї W
    
    7ˠ57
    П F
    
    &ˠ5&
    ͷ࣌
    |TV
    | + | TE
    | > 2
    Λຬͨ͢άϥϑ
    

    View Slide

13. ఏҊख๏ʛNFUBQBUI
    
    NFUBQBUITDIFNF JT
    ɿ7
    ͱ7M
    ͷϊʔυλΠϓͷؔ܎
    

    View Slide

14. ఏҊख๏ʛNFUBQBUI
    
    NFUBQBUITDIFNF
    

    View Slide

15. ఏҊख๏ʛNFUBQBUI
    
    NFUBQBUIͷԿ͕خ͍͔͠
    ҙຯͷ͋ΔNFUBQBUIΛઃఆ͢Ε͹ɺ
    ຒΊࠐΈʹ൓өͤ͞ΒΕΔ
    ྫʣ"1" ˠ ڞஶؔ܎
    "171" ˠ ಉֶ͡ձʹ࿦จΛ
    ग़ͨ͠ஶऀͷؔ܎
    

    View Slide

16. ఏҊख๏ʛNFUBQBUIWFD
    
    ৽نੑ͸Ͳ͜Ͱ͔͢ʁ
    .FUB1BUI#BTFE3BOEPN8BMLT
    ͰೖྗͷϊʔυܥྻΛੜ੒
    ˣ
    )FUFSPHFOFPVT4LJQ(SBN
    Ͱ෼ࢄදݱΛੜ੒
    

    View Slide

17. ఏҊख๏ʛ.FUB1BUI#BTFE3BOEPN8BMLT
    
    ྫ
    
    XBMLBQ"$-QBQ
    ,%%QB
    ୯७ͳϥϯμϜ΢ΥʔΫ
    Ͱ͸ͳ͘NFUBQBUIʹ
    Ԋ͍ͬͯΔ
    

    View Slide

18. ఏҊख๏ʛ)FUFSPHFOFPVT4LJQ(SBN
    XPSEWFDͰ͸
    ೖྗ୯ޠ͔Βपล୯ޠͷ༧ଌΛ
    ͢ΔχϡʔϥϧωοτϫʔΫͷɺ
    ೖྗ૚͔ΒӅΕ૚ͷॏΈ͕ೖྗ
    ୯ޠͷ෼ࢄදݱ
    *XBOUUPFBU BOBQQMFFWFSZEBZ
    *XBOUUPFBU BOBQQMFFWFSZEBZ
    

    View Slide

19. ఏҊख๏ʛ)FUFSPHFOFPVT4LJQ(SBN
    
    NFUBQBUIWFDͰ͸
    ೖྗϊʔυ͔Βपลϊʔυͷ༧
    ଌΛ͢Δχϡʔϥϧωοτϫʔ
    Ϋͷɺೖྗ૚͔ΒӅΕ૚ͷॏΈ
    ͕ೖྗϊʔυͷ෼ࢄදݱ
    BQ"$-Q BQ,%%QB
    BQ"$-Q BQ,%%QB
    

    View Slide

20. ఏҊख๏ʛग़ྗ
    
    X ∈ R|V|×d ,
    d << |V|
    ͳΔE࣍ݩ෼ࢄදݱΛֶश
    ࣍ݩ E
    ௖
    ఺
    ਺
    7
    ͋Δ௖఺ͷ෼ࢄදݱ
    ผͷ௖఺ͷ෼ࢄදݱ
    

    View Slide

21. ఏҊख๏ʛNFUBQBUIWFD
    
    ͸Կʁ
    /FHBUJWF4BNQMJOH
    Λ֦ு
    

    View Slide

22. ఏҊख๏ʛNFUBQBUIWFD
    
    /FHBUJWF4BNQMJOHʢҎલͷख๏ʣ
    ˠपล୯ޠͷ֬཰͕ʹɺपลʹ͸ͳ͍୯ޠͷ֬཰͕
    ʹۙͮ͘Α͏ʹχϡʔϥϧωοτϫʔΫͷॏΈΛֶश
    BQ"$-Q BQ,%%QB
    )FUFSPHFOFPVT/FHBUJWF4BNQMJOHʢʣ
    ˠϊʔυͷλΠϓ͝ͱʹ/FHBUJWF4BNQMJOH
    BQ"$-Q BQ,%%QB
    

    View Slide

23. ໨࣍
    wݚڀഎܠ
    w࿦จ֓ཁ
    wఏҊख๏
    w࣮ݧ
    w݁࿦
    

    View Slide

24. ࣮ݧʛ࣮ݧ֓ཁ
    
    ̏ͭͷ࣮ݧʹΑΓఏҊख๏ͷ༗ޮੑΛௐࠪ
    • ଟΫϥε෼ྨ
    • ΫϥελϦϯά
    • ྨࣅ౓ݕࡧ
    Ճ͑ͯɺՄࢹԽͱεέʔϥϏϦςΟΛௐࠪ
    

    View Slide

25. ࣮ݧʛσʔληοτ
    
    • ".JOFS $PNQVUFS4DJFODF>
    ສҎ্ͷݚڀऀɺສҎ্ͷ࿦จɺ
    ۙ͘ͷֶձ
    • %BUBCBTFBOE*OGPSNBUJPO4ZTUFNT>
    ۙ͘ͷֶձɺͦ͜ʹ͓͚Δͷ
    τοϓஶऀɺສҎ্ͷ࿦จ
    > +JF 5BOH +JOH;IBOH -JNJO :BP +VBO[J -J -J;IBOH BOE;IPOH 4V"S OFU.JOFS&YUSBDUJPOBOE.JOJOHPG"DBEFNJD4PDJBM/FUXPSLT*O,%%`r
    > :J[IPV 4VO +JBXFJ)BO 9JGFOH :BO 1IJMJQ4:V BOE5JBOZJ 8V1BUITJN.FUBQBUICBTFEUPQLTJNJMBSJUZTFBSDIJOIFUFSPHFOFPVTJOGPSNBUJPOOFUXPSLT*O
    7-%#`r
    

    View Slide

26. ࣮ݧʛ࣮ݧઃఆ
    
    • ϊʔυ͋ͨΓͷࢼߦճ਺ɿ
    • ࢼߦ͋ͨΓͷา਺ɿ
    • ϕΫτϧͷ࣍ݩ਺ɿ
    • ۙྡαΠζɿ
    • ωΨςΟϒαϯϓϧ਺ɿ
    

    View Slide

27. ࣮ݧʛൺֱख๏
    
    • %FFQ8BML>OPEFWFD>
    ύϥϝʔλʔΛௐ੔ͯ͠ಉҰख๏ͱͯ͠ѻ͏
    • -*/&>
    ྡ઀͢Δϊʔυͱͷؔ܎͔Β෼ࢄදݱΛੜ੒
    • 15&>
    ୯ޠɾจॻɾϥϕϧͷؔ܎Λάϥϑʹද͠ɺ
    ͦͷάϥϑ͔ΒϕΫτϧදݱΛֶश
    

    View Slide

28. ࣮ݧʛଟΫϥε෼ྨʢਫ਼౓ʣ
    
    ʙͷ޲্
    ʙͷ޲্
    

    View Slide

29. ࣮ݧʛଟΫϥε෼ྨʢύϥϝʔλʔʣ
    
    ɾBVUIPSDMVTUFSJOH ਫ
    
    XͱMͱE͸େ͖͘ͳΔʹͭΕਫ਼౓΋޲্
    L͕େ͖͘ͳΔʹͭΕਫ਼౓͸ѱԽɺ
    ͜Ε͸IPNPHFOFPVTFOWJSPONFOU> Ͱͷ݁Ռͱਅٯ
    ɾWFOVFDMVTUFSJOH ੨
    
    ͋·Γਫ਼౓ʹӨڹ͠ͳ͍
    

    View Slide

30. ࣮ݧʛϊʔυΫϥελϦϯάʢਫ਼౓ʣ
    
    15&ΑΓˋ޲্ 15&ΑΓˋ޲্
    

    View Slide

31. ࣮ݧʛϊʔυΫϥελϦϯάʢύϥϝʔλʔʣ
    
    ɾBVUIPSDMVTUFSJOH ੨
    
    XͱM͸େ͖͘ͳΔʹͭΕਫ਼౓΋޲্
    EͱL͸େ͖͘ͳΔʹͭΕਫ਼౓͸ѱԽ
    ɾWFOVFDMVTUFSJOH ੺
    
    L͸େ͖͘ͳΔʹͭΕҰ౓ਫ਼౓͕޲্͠ɺͦͷޙѱԽ͢Δ
    

    View Slide

32. ࣮ݧʛྨࣅݕࡧ
    
    ଟ͘ͷ෼໺ʹ͓͍ͯɺ
    ݕࡧΫΤϦͱಉ͡෼໺ͷτοϓֶձ͕
    τοϓͷதʹೖ͍ͬͯΔ
    

    View Slide

33. ࣮ݧʛՄࢹԽ
    
    ఏҊख๏Ͱ͸XPSEWFDʹ͓͚Δ
    ࠃͱट౎ͷରԠؔ܎ͷΑ͏ʹɺ
    ֶձͱͦͷ෼໺ͷஶ໊ͳݚڀऀͷ
    ରԠؔ܎ֶ͕शͰ͖͍ͯΔ
    

    View Slide

34. ࣮ݧʛՄࢹԽ
    
    ࣍ݩʹ౤Ө͢Δͱ
    ͍ۙ෼໺ͷֶձ͕ۙ͘ʹଘࡏ
    

    View Slide

35. ࣮ݧʛεέʔϥϏϦςΟ
    
    ཧ૝తͳઢܗͰ͸ͳ͍͕
    ४ઢܗͷύϑΥʔϚϯε
    

    View Slide

36. ໨࣍
    wݚڀഎܠ
    w࿦จ֓ཁ
    wఏҊख๏
    w࣮ݧ
    w݁࿦
    

    View Slide

37. ݁࿦ʛ'VUVSF8PSL
    
    NFUBQBUIੜ੒࣌ͷܭࢉྔ࡟ݮ
    NFUBQBUIͷࣗಈֶश
    ωοτϫʔΫͷਐԽ΁ͷରԠ
    ϞσϧͷҰൠԽ
    

    View Slide

38. ݁࿦ʛ·ͱΊ
    
    ԿΛͨ͠ʁ
    )/ʹରͯ͠༗ޮͳNFUBQBUIWFDͱɺ
    ͦͷ೿ੜ൛NFUBQBUIWFDΛఏҊɻ
    ݁ՌͲ͏ͩͬͨʁ
    ༷ʑͳϚΠχϯάλεΫʹର͠
    ࠷ઌ୺ͷख๏Λ্ճΔ͚ͩͰͳ͘ɺ
    ߏ଄ͱҙຯΛอ࣋ͨ͠෼ࢄදݱ͕ੜ੒Ͱ͖ͨ
    

    View Slide