kaggler-ja driven learning | Kaggle Tokyo Meetup #5

B1cc148711c6a37a5c922b6e72a4ad52?s=47 u++
November 28, 2018

kaggler-ja driven learning | Kaggle Tokyo Meetup #5

Kaggle Tokyo Meetup #5 (https://connpass.com/event/105298/) での発表資料。kaggler-ja slack参加はこちらから (https://kaggler-ja.herokuapp.com/)。

B1cc148711c6a37a5c922b6e72a4ad52?s=128

u++

November 28, 2018
Tweet

Transcript

  1. LBHHMFSKBESJWFOMFBSOJOH ͳ࣭ͥ໰ʹ౴͑ɺ Ͳ͏ֶΜͰ͍Δ͔ ,BHHMF5PLZP.FFUVQ -JHIUOJOH5BML %FDTU  V  !VQVSB

  2. V !VQVSB l ๭اۀͰӦۀɾϚʔέςΟϯά޲͚ ͷσʔλ෼ੳˍ։ൃ l ,BHHMF&YQFSU l ໺ٿϋοΧιϯʢΤϯδχΞϦϯά෦໳ʣҐ l

    ݚڀ಺༰ɿ৘ใٕज़ͷࣾձ࣮૷ ü %FTJHO.FUIPEPG3FNPUF.POJUPSJOH4FSWJDFGPS &MEFSMZ$POTJEFSJOH$PNNVOJUZ$IBSBDUFSJTUJDT  IUUQFCPPLTJPTQSFTTOMWPMVNFBSUJDMF
  3. ຊ೔ͳͥൃදʁ l ͸·ͩແ͍ɾɾɾ͚ΕͲ l LBHHMFSKBTMBDL UXJUUFS ͸ͯͳϒϩάͳͲͰ ੵۃతʹΞ΢τϓοτ͍ͯ͠Δ l ڧ͘ͳ͍ͱൃݴݖ͕ͳ͍

    ˡؒҧ͍ l ॳ৺ऀ΋ؾܰʹࢀՃͰ͖Δʮ೔ຊͷσʔλ෼ੳ ίϛϡχςΟʯͰࡏΓଓ͚ͯ΄͍͠ 
  4. CFHJOOFSTIFMQͰͷ౤ߘ਺  l աڈϩάΛ#FBVUJGVM4PVQͰॲཧ l ूܭظؒɿ ೥݄೔ʙ݄೔ ʢॳΊͯ౤ߘ͔ͯ͠Βूܭʣ l ͪͳΈʹҐ͸ULN͞Μ

  5. 5-%3 l LBHHMFSKB TMBDL XJLJ NFFUVQͳͲ ʹ͸ɺ ॳ৺ऀ΋ؾܰʹࢀՃͰ͖Δʮ೔ຊͷσʔλ෼ੳ ίϛϡχςΟʯͰࡏΓଓ͚ͯ΄͍͠ ü

    4UFQ ࢀՃͯ͠ΈΑ͏ ü 4UFQ ࣭໰ͯ͠ΈΑ͏ ü 4UFQճ౴ͯ͠ΈΑ͏ -FU`TLBHHMFSKBESJWFOMFBSOJOH  ಛʹTMBDLʹ͍ͭͯ
  6. ݴ͍͍ͨ͜ͱ͸Ҏ্ ͋ͱ͸֤4UFQʹ͍ͭͯ গ͠ࡉ͔͘Έ͍͖ͯ·͢

  7. গ͚ͩ͠༨ஊ ͳͥࢲ͕ʮ೔ຊͷσʔλ෼ੳ ίϛϡχςΟʯʹظ଴͢Δ͔

  8. σʔλ͸ʮਓͷຊ࣭ʯΛ๫͘ͱࢥ͏ l ড়౔फͷ͓ࣉͷ௕உͱͯ͠ੜ·ΕΔ l কདྷ͸ܧ͙ʜʜ͔΋ʁʁʁ l σʔλ෼ੳ͸ʮਓؒͷຊ࣭ʯʹܨ͕Δ ü ݸʑͷ෺ࣄͷʮ࠷దԽʯ͕ਐΜͰ͍͘ ü

    ෺ࣄ͕࠷దԽ͞Ε͍ͯͬͨઌʹɺԿ͕࢒Δͷ͔ ü ͋ΒΏΔ෼໺Ͱσʔλ׆༻͕ਐΜͰ΄͍͠ ü ͲΜͲΜ৽نࢀೖͯ͠΄͍͠ ü ਓจܥͷ஌ݟ͸ࠓޙҰ૚ॏཁʹͳ͍ͬͯ͘ 
  9. 4UFQࢀՃͯ͠ΈΑ͏

  10. LBHHMFSKBTMBDL IUUQTLBHHMFSKBIFSPLVBQQDPN 

  11. ༗༻ͳDIBOOFM͕ͨ͘͞Μ  l CFHJOOFSTIFMQʢॳ৺ऀ࣭໰෦԰ʣ l MFBEFSCPBSEMJWFʢ-#࣮گɺղ๏ڞ༗ʣ l LBHHMF@HFOFSBMʢίϯϖ৘ใʣ l SBOEPNʢࡶஊɺྫɿUFOTPSGMPXͷಡΈํʣ

    l CFHJOOFST@CPPUDBNQʢJODMBTTίϯϖ։࠵ʣ ͳͲ
  12. CFHJOOFSTIFMQ 

  13. ࢀՃਓ਺ͷਪҠ  'SPN!ULN `TUXFFU

  14. ࢀՃਓ਺ͷਪҠʢਪఆʣ  ग़యɿ૯຿ল౷ܭہ ਓޱਪܭ IUUQTXXXTUBUHPKQEBUBKJOTVJOFXIUNM ೥݄ ! = 0.012'( +

    1.75' + 232
  15. ݟ͍ͯΔ͚ͩͰֶͼʹͳΔ l ೔ຊޠ͔ͩΒαϥοͱಡΊΔ l LBHHMFNBTUFSΒ͕஌ݟΛམͱ͍ͯ͠Δ l ࣮͸ʮ30.ઐʯͰ΋े෼ʢਓʣ l ελϯϓΛԡͦ͏ 

  16. 4UFQ࣭໰ͯ͠ΈΑ͏

  17. ࣭໰͢Δ࣌͋Δ͋Δ l ॳาతͳ࣭໰͗ͯ͢അࣛʹ͞Εͳ͍͔ʜʜʁ l ͜Μͳ͜ͱฉ͍ͯྑ͍ͷͩΖ͏͔ʜʜʁ l ୭ʹ΋൓Ԡ͞Εͳ͔ͬͨΒݏͩͳʜʜɻ 

  18. ৺഑͝ແ༻ 

  19. ʮ൓Ԡ͕ͳ͔ͬͨʯ৔߹ l ߴ౓ͳଞਓ͕औΓ૊ΜͰ͍ͳ͍಺༰ʁ ü ઌۦऀʹͳΕΔྖҬ͔΋ʂʁ ü ͋Δछͷࣗ৴ʹͳΔ l ࣭໰͕ཁྖΛಘ͍ͯͳ͍ʁ ü

    ʮྑ͍࣭໰ʯʹͳ͍ͬͯΔ͔ʁ ü ͬͪ͜͸஫ҙ 
  20. ʮྑ͍࣭໰ʯͷͨΊʹ l l1MFBTFJODMVEFBNJOJNBMSFQSFYz ü SFQSPEVDJCMFFYBNQMFʢ࠶ݱՄೳͳ࣮ྫʣ ü ඞཁ࠷௿ݶͷ৘ใΛఏࣔ͢Δʢ༨෼Λ࡟Δʣ ü 3ίϛϡχςΟͰ࢖ΘΕ͍ͯΔ ü

    3ͷύοέʔδ໊ IUUQTHJUIVCDPNUJEZWFSTFSFQSFY 
  21. NJOJNBMSFQSFYͷྫ 2ϦετΛTVN ͯ͠΋ΤϥʔʹͳΓ·͢ɻ 2Ϧετͷ஋ͷ࿨Λܭࢉ͍ͨ͠ͷͰ͕͢ɺҎԼͷ ίʔυ͕ΤϥʔʹͳΓ·͢ɻ Y<  > TVN Y

    5SBDFCBDL NPTUSFDFOUDBMMMBTU  'JMFTUEJO MJOF JONPEVMF 5ZQF&SSPSVOTVQQPSUFEPQFSBOEUZQF T GPS JOUBOETUS` l ʮྑ͍࣭໰ʯΛ͠Α͏ͱ͢Δաఔ΋ษڧʹͳΔ  ը૾͡Όͳͯ͘จࣈͰ
  22. ࢀߟɿNJOJNBMSFQSFY l 3͓͡͞Μͷҟৗͳଈճ౴·ͨ͸ࢲ͸೗Կʹͯ͠৺഑͢ ΔͷΛࢭΊͯSXBLBMBOHͰ࣭໰͢ΔΑ͏ʹͳ͔ͬͨ IUUQTTQFBLFSEFDLDPNZVUBOOJIJMBUJPOIPXJMFBSOFEUPTUPQXPSSZJOH BOEBTL l ࣭໰͓଴͓ͪͯ͠Γ·͢ʂͰ͖Ε͹SFQSFY࢖ͬͯͶ IUUQTVSJCPIBUFOBCMPHDPNFOUSZ l

    SFQSFYύοέʔδͷ࢖͍ํɺ͋Δ͍͸ӋௗʹౖΒΕΔ લʹ஌Δ΂͖ͭͷ͜ͱ IUUQTOPUDIBJOFEIBUFOBCMPHDPNFOUSZ 
  23. ࢀߟɿSXBLBMBOHίϛϡχςΟ IUUQTUXJUUFSDPNIFBWZXBUBMTUBUVT T 

  24. ࣭໰͸Ғ͍  ࣭໰͢Δ΍ͭ͸ΊͪΌͪ͘ΌҒ ͍ɻͲΕ͘Β͍Ғ͍͔Λղઆ͠ ͨͭͷ߲໨͕࿩୊ʹ ΫιϦ ϓେձ΋ຄൃʛUPHFUUFS IUUQTUPHFUUFSDPNMJ 

  25. 1SJWBUF4IBSJOHʹ͸஫ҙ l ։࠵தίϯϖʹؔͯ͠ɺνʔϜ֎Ͱͷίʔυ΍ σʔλͷڞ༗͸ېࢭ l ެࣜͷLFSOFMT EJTDVTTJPOTΛར༻͢Δ l ٞ࿦΋LBHHMF্͕๬·͍͠ 

    IUUQTXXXLBHHMFDPNDQMBOFUVOEFSTUBOEJOHUIFBNB[POGSPN TQBDFEJTDVTTJPO BENJO͝ͱʹҙݟ͸ҟͳΔ
  26. 4UFQճ౴ͯ͠ΈΑ͏

  27. ճ౴͢Δ࣌͋Δ͋Δ l ࣗ෼ΑΓݰਓ͕ͨ͘͞Μ͍Δ͔Βʜʜɻ l ࣗ෼͸߹͍ͬͯΔͱࢥ͏͚Ͳɺຊ౰͸ؒҧͬͯ ͍Δ͔΋ʜʜʁ 

  28. ࣗ෼ͳΓͷରԠࡦʢߟ͑ํʣ l ݰਓͩͱॳ৺ऀͷ࣌ͷهԱ͕ബΕ͕ͪ ü ʮͳͥ෼͔Βͳ͍͔ʯ͕෼͔Βͳ͍৔߹΋ l ෼͔Δͱ͜Ζ·Ͱճ౴ͯ͠ɺո͍͠෦෼͸ਖ਼௚ ʹ఻͑Δ ü ʮ͖ͨͨ୆ʯ͕͋Δͱޙʹଓ͖΍͍͢

    
  29. ճ౴͢ΔͨΊͷํ๏ l Πϯλʔωοτݕࡧ ü ެࣜυΩϡϝϯτ ü (JU)VCʢιʔείʔυɺJTTVFTʣ ü BS9JWͳͲͷ࿦จ ü

    ͦͷଞϒϩάͳͲ l σόοά͢Δ ü ࣭໰͕ίʔυ෇͖ͳΒɺΤϥʔ͕࠶ݱ͢Δ͔֬ೝ 
  30. ճ౴͢ΔͨΊͷ5JQT l (PPHMFݕࡧͷػೳΛ࢖͏ ü l zɺݴޠઃఆɺظؒࢦఆͳͲ l ӳޠͰݕࡧ͢Δ ü ΍ͬͺӳޠͷํ͕৘ใྔ͕ଟ͍ʢάάΕΔӳޠྗʣ

    l ೔ࠒ͔ΒΞϯςφΛ޿͘ு͓ͬͯ͘ ü ʮ͋ͷลΛௐ΂Ε͹ྑͦ͞͏ʯΛ஌͓ͬͯ͘ 
  31. ճ౴͢Δͷ͕Ұ൪ษڧʹͳΔ l ճ౴͢ΔͨΊʹ͸ɺਂ͍ཧղͷͨΊʹʮී௨ʯ ͷษڧΛ͢Δ l ࣗ෼͕ᐆດʹ͝·͔͍ͯͨ͠෦෼ʹؾ෇͘ l ࣗ෼ͷࢥ͍͔ͭͳ͔ͬͨख๏ɾ໰୊Λఏࣔͯ͠ ͘ΕΔ 

  32. ࠷ޙʹվΊͯ

  33. 5-%3 l LBHHMFSKB TMBDL XJLJ NFFUVQFUD ʹ͸ɺ ॳ৺ऀ΋ؾܰʹࢀՃͰ͖Δʮ೔ຊͷσʔλ෼ੳ ίϛϡχςΟʯͰࡏΓଓ͚ͯ΄͍͠ ü

    4UFQ ࢀՃͯ͠ΈΑ͏ ü 4UFQ ࣭໰ͯ͠ΈΑ͏ ü 4UFQճ౴ͯ͠ΈΑ͏ -FU`TLBHHMFSKBESJWFOMFBSOJOH  ಛʹTMBDLʹ͍ͭͯ
  34. 8JOPS-FBSO :PV/FWFS-PTF  #Z.BSJPT .JDIBJMJEJT ,B["OPWB ,%%BU-POEPO GSPN!WFSGJU`TUXFFU IUUQTUXJUUFSDPNWFSGJUTUBUVT T

  35. উ͚͕ͭͩLBHHMF͡Όͳ͍ ࣗ෼ͳΓͷֶͼ͕ಘΒΕΕ͹ ͦΕ͸ʮෛ͚ʯ͡Όͳ͍ TMBDLࢀՃ͸ͪ͜Β͔Β IUUQTLBHHMFSKBIFSPLVBQQDPN