The first step of the Top-k Selection algorithm in search

Transcript

1. The first step of the Top-k Selection algorithm in search

   M3 tecktalk #164 hiromu-nakamura @po3rin

2. த
   ଜ
   ߂
   ෢ ΤϜεϦʔͷ"*ɾػցֶशνʔϜͰݕࡧج൫पΓ Λ୲౰͍ͯ͠Δɻ(P͕޷͖ɻ
   ֤छΞΧ΢ϯτ
   QPSJO
   Ͱ΍ͬͯ·͢ɻ )
   *
   S
   P
   N
   V
   
   /
   B
   L
   B
   N
   V
   S
   B

3. None

4. ݕࡧΤϯδϯ͸ΫΤϦΛड͚औΓɺυΩϡϝϯτ͝ͱʹ
   είΞܭࢉΛείΞ্Ґ
   L
   ݅Λฦ͢ɻ 4FBSDI
   &OHJOF ΫΤϦ L
   ݅ͷείΞॱυ Ωϡϝϯτ

5. &MBTUJDTFBSDIͳͲͷݕࡧΤϯδϯ͸ɺ୯ޠͷҐஔ৘ใΛ ֨ೲ͢ΔͨΊͷࡧҾσʔλߏ଄Ͱ͋ΔసஔΠϯσοΫε Λ࠾༻͍ͯ͠Δɻ 4FBSDI
   &OHJOF %
   ࢲ͸ম೑Λ৯΂Δ
   %
   ࢲ͸αϥμ΋৯΂Δ ʮম೑ʯ % సஔΠϯσοΫε

6. ݕࡧʹ͓͚ΔυΩϡϝϯτͷείΞϦϯάʹ͸
   0LBQJ
   #.
   ͳͲ୅දతͳ΋ͷ͕͋ΔLɻ͜ΕΛݕࡧʹώοτ ͢ΔυΩϡϝϯτʹରͯ͠ܭࢉ͍ͯ͘͠ɻ 4DPSJOH

7. υΩϡϝϯτશͯͷείΞϦϯά͕ऴ͔ྃͯ͠ΒͦΕΒΛฒ ͼସ͑Δͷ͸ϥϯΩϯά্ҐL൪໨·Ͱͷ݁Ռʹڵຯ͕͋Δ৔ ߹ඇޮ཰ɻͦ͜Ͱݕࡧʹ͓͚ΔL
   ݸͷ࠷େཁૉʢ·ͨ͸࠷খ ཁૉʣΛબ୒͢Δ໰୊Λߟ͑Δɻ 5PQL
   TFMFDUJPO

8. 5PQL
   TFMFDUJPO ೋ෼ώʔϓʹυΩϡϝϯτͱείΞΛอ͍࣋ͯ͘͜͠ͱͰ࠷ ऴతͳιʔτͷίετΛεϧʔͰ͖Δɻώʔϓʹ֨ೲ͢Δ͔ ͸ࠜͷείΞ͚ͩݟΕ͹ྑ͍ɻ ೋ෼ώʔϓ    

   

9. 5PQL
   TFMFDUJPO ࠓ·Ͱݕࡧʹώοτͨ͠શͯͷυΩϡϝϯτͷείΞΛܭࢉ ͍͕ͯͨ͠ɺ࣮͸είΞΛܭࢉ͠ͳͯ͘΋ɺ͋ΔυΩϡϝϯ τ͕UPQLʹೖΒͳ͍͜ͱΛ֬ఆͰ͖Δํ๏͕͋ΔɻͦΕ͕ NBY
   TDPSFͱݺ͹ΕΔख๏Ͱ͢ɻ .BY
   TDPSF

10. ͷ໾ׂ͸ΫΤϦλʔϜͦΕͧΕͷείΞ΁ͷߩݙ౓ͷ্ݶ Λܾఆ͢Δ͜ͱɻ ͷܦݧతॳظ஋Λ࠾༻͢Δͱɺ5'஋ ্ࣜͷୈҼࢠ ͸
    
    ͭ·Γ
    Λ௒͑ͳ͍ɻ*%'΋ࣄલ ʹܭࢉͰ͖Δɻ͜ΕʹΑΓ֤λʔϜ͕د༩͢Δ࠷େείΞ͸ ࣄલʹܾఆͰ͖ΔɻͦͷͨΊϥϯΩϯά্ҐʹݱΕͳ͍υ Ωϡϝϯτ͸͋Β͔͡ΊΘ͔Δɻ k1

    k1 k1 + 1 4DPSJOH 5PQL
    TFMFDUJPO

11. 4DPSJOH λʔϜ .BY4DPSF UIF  TFBSDI  FOHJOF  

        ͜ͷ࣌఺ͰzUIFzؚ͔͠·ͳ͍ϙεςΟϯάϦετͷείΞܭ ࢉ͸ແࢹͰ͖ΔɻͳͥͳΒ࠷େείΞ͸ݱࡏͷώʔϓͷࠜ ͷείΞΛ௒͑ͳ͍͔Βɻ 5PQL
    TFMFDUJPO

12. 4DPSJOH λʔϜ .BY4DPSF UIF  TFBSDI  FOHJOF  

        ͜ͷ࣌఺Ͱ͸zUIFzͱzTFBSDIzؚ͔͠·ͳ͍υΩϡϝϯτͷε ίΞܭࢉ͸ແࢹͰ͖ΔɻͳͥͳΒ࠷େείΞ ͸ݱࡏͷ ώʔϓͷࠜͷείΞΛ௒͑ͳ͍͔Βɻ 5PQL
    TFMFDUJPO

13. *OUFSWBMCBTFE
    QSVOJOH TFBSDI FOHJOF %PDVNFOU 
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    ◾ ◾
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    
    ◾ NBY NBY 
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    

    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
     TDPSF 
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
     ͜ͷ࣌఺ͰzFOHJOFz͚͔ͩ͠ग़ݱ͍ͯ͠ͳ͍υΩϡϝϯτ͸ແࢹͰ͖Δ ͜ͷ࣌఺ͰzTFBSDIz͚͔ͩ͠ग़ݱ͍ͯ͠ͳ͍υΩϡϝϯτ͸ແࢹͰ͖Δ XBML
    EPDT

14. *OUFSWBMCBTFE
    QSVOJOH ࣮ફͰ͸ϙεςΟϯάϦετ͸ѹॖ͞Ε͍ͯΔͷͰϒϩοΫ αΠζ͝ͱʹNBY
    TDPSFΛอ࣋͠ ֤ϒϩοΫͷ۠੾ΓͷΠϯ λʔόϧʹ6QQFS
    #PVOE
    4DPSF 6#4 Λܭࢉ͓ͯ͘͠ɻ͜Ε ʹΑΓϒϩοΫղౚΛεΩοϓͰ͖Δɻ TFBSDI FOHJOF

    %PDVNFOU 
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    ◾ ◾
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    
    ◾
    
    
    
    
    
    
    
    
    
    
    ◾ NBY NBY NBY NBY NBY  6#4    

15. $PODMVTJPO 5PQL
    TFMFDUJPO
    Ͱ͸ೋ෼ώʔϓͳͲͷσʔλߏ଄͕ར༻Ͱ͖Δɻ
    ݕࡧείΞʹ͓͚Δ֤λʔϜ͕ߩݙ͢ΔNBYTDPSF͸ࣄલʹܭࢉͰ͖ΔͷͰείΞܭࢉΛεΩοϓͰ͖Δɻ
    *OUFSWBMCBTFE
    QSVOJOH
    Λར༻͢Δ͜ͱͰѹॖΠϯσοΫε΋είΞܭࢉΛޮ཰తʹߦ͑Δɻ
    -VDFOFͰ͸ߋʹൃలܥͷ8"/%ͱ͍͏ΞϧΰϦζϜ͕ಋೖ͞Ε͍ͯΔɻ

16. ࢀߟ -VDFOF
    
    ͷ
    5PQL
    ΫΤϦϓϩηογϯά࠷దԽ 
    .BY4DPSFΞϧΰϦζϜฤ
    IUUQTMJOLNFEJVNDPN&I%"D$HRKDC
    
    0LBQJ
    #.
    IUUQTKBXJLJQFEJBPSHXJLJ0LBQJ@#.
    ৘ใݕࡧ
    ݕࡧΤϯδϯͷ࣮૷ͱධՁ
    UUQTXXXBNB[PODPKQEQSFGDN@TX@FN@S@NU@EQ@$H'C&;7('9
    

17. The first step of the Top-k algorithm in search M3

    tecktalk #164 hiromu-nakamura @po3rin