[読み会]Fair Adversarial Gradient Tree Boosting

読み会資料.
Fair Adversarial Gradient Tree Boosting(ICDM2019)

Transcript

1. 'BJS
   "EWFSTBSJBM
   
   (SBEJFOU
   5SFF
   #PPTUJOH ಡΈձ!
   ༶໌఩
2.  wஶऀ
   w7JODFOU
   (SBSJ
   4PSCPOOF
   6OJWFSTJUZ
   w#PSJT
   3VG
   "9"
   w4ZMWBJO
   -BNQSJFS
   4PSCPOOF
   6OJWFSTJUZ
   w.BSDJO
   %FUZOJFDLJ
   "9"
   wग़య
   *$%.
    ࿦จ৘ใ
3.  wܾఆ໦ͷϞσϧͰެฏͳϞσϧΛ֫ಘͰ͖ΔΑ͏ʹͨ͠ɽ
   w405"ͱಉ͘͡Β͍ͷੑೳΛୡ੒Ͱ͖ͨɽ
   wબΜͩཧ༝
   wެฏͳϞσϧ֫ಘΛ໨తͱͯ͠ɼܾఆ໦ϕʔεͷख๏ΛఏҊ͍ͯͨͨ͠Ίɽ
   wޯ഑ϒʔεςΟϯάܾఆ໦͕ςʔϒϧσʔλʹରͯ͠ڧྗͰ͋Δ͔Βɽ ֓ཁ ͲΜͳ࿦จʁ
4.  wόΠΞεͷܰݮʹର͢ΔλεΫͰ͸Α͘χϡʔϥϧωοτϫʔΫ͕༻͍ΒΕ͍ͯ Δ͠ɼੑೳͱͯ͠΋͍͍ɽ
   wςʔϒϧσʔλʹରͯ͠ɼ9(#PPTU΍-JHIU(#.ͳͲޯ഑ϒʔεςΟϯά໦͕ඇ ৗʹڧྗͰΑ͘༻͍ΒΕ͍ͯΔɽ
   wͦ͜Ͱɼޯ഑ϒʔεςΟϯά໦ͰόΠΞεΛܰݮ͢Δ ެฏੑΛୡ੒͢Δ Α͏ͳख ๏ΛఏҊ͢Δɽ ֓ཁ ݚڀഎܠ
5.  wطଘݚڀͰ͸΄ͱΜͲ͕χϡʔϥϧωοτϫʔΫΛ༻͍ͯɼόΠΞεΛܰݮͯ͠ Δɽ
   wχϡʔϥϧωοτϫʔΫͰ͸ੑೳ͸ग़͍ͤͯΔ͕ɼͳΜͰ͜͏ͳ͔͕ͬͨ෼͔Γ ʹ͍͘ɽ
   wςʔϒϧσʔλͰ͸χϡʔϥϧωοτϫʔΫΑΓ΋ޯ഑ϒʔεςΟϯά໦͕޷·Ε ͯ࢖ΘΕ͍ͯΔɽ ֓ཁ طଘݚڀʹ͍ͭͯ
6.  wҰൠతͳ෼ྨϞσϧʹରͯ͠ɼܾఆ໦ΛؚΉఢରతֶशͰެฏΛ֫ಘ͢Δํ๏Λ ఏҊ͢Δɽ
   wҟͳΔެฏੑͷن४Ͱ405"ͳϞσϧͱಉ౳ͷੑೳΛୡ੒ͨ͠ɽ ֓ཁ ߩݙ
7.  w
   
   
   ࣍ݩͷಛ௃ྔʢඇηϯγςΟϒͳಛ௃ʣ
   w
   
   ηϯγςΟϒͳଐੑɽࠓճ͸ೋ஋Ͱߟ͍͑ͯΔɽ
   w
   ໨తม਺ɽࠓճ͸෼ྨ໰୊Λѻ͏ͷͰೋ஋Ͱߟ͑Δɽ
   wಛʹ؍ଌ஋͸
   ༧ଌ஋͸ ͱ͢Δ
   wO͜ͷαϯϓϧͷσʔληοτ
   Λ༻͍ͯҰൠతͳڭࢣ͋ΓֶशΛߦ͏ɽ x ∈ ℝd d s y y ̂ y (xi , si , yi )n i=1 ެฏੑ഑ྀܕػցֶश දهͷఆٛ
8.  w%FNPHSBQIJD
   1BSJUZ
   ηϯγςΟϒଐੑʹ͔͔ΘΒͣ༧ଌ஋ͷൺ཰͸Ұக͢Δ
   w
   w%1Λ࢖ͬͨධՁ1SVMFΛ࣍ͷΑ͏ʹఆٛ͢Δ
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   %1
   P( ̂ Y = 1|S = 0) = P( ̂ Y = 1|S = 1) 1SVMF = min ( P( ̂ Y = 1|S = 0) P( ̂ Y = 1|S = 1) , P( ̂ Y = 1|S = 1) P( ̂ Y = 1|S = 0)) ެฏੑ഑ྀܕػցֶश ެฏੑͷఆٛ
9.  w΋͏Ұͭͷج४ͱͯ͠EJTQBSF
   JNQBDUʢ%*ʣΛఆٛ͢Δɽ
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   
   w%*͸ʹͳΔ΄ͲެฏͰ͋Δɽ %* : |P( ̂ Y = 1|S = 1) − P( ̂ Y = 1|S = 0)| ެฏੑ഑ྀܕػցֶश ެฏੑͷఆٛ
10.  w؍ଌ஋ ΋৚݅ʹೖΕΔن४
    &RVBMJ[FE
    0EET͕͋Δɽ
    w
    w͜ͷ&0ʹΑͬͯٻΊΒΕΔ'13ͱ'/3ͷ஋͕ʹ͍ۙ΄Ͳެฏੑ͕֫ಘͰ͖Δɽ Y &0 : P( ̂ Y = 1|S = 0,Y = y) = P( ̂ Y = 1|S = 1,Y = y), ∀y ∈ 0,1 ެฏੑ഑ྀܕػցֶश ެฏੑͷن४
    ͦͷ
11.  wຊݚڀͰ͸ɼҰൠతͳޯ഑ϒʔεςΟϯά໦ͷ࿮૊ΈʹɼఢରతֶशΛ༻͍ͯɼ ެฏੑΛ֫ಘ͍ͯ͘͠ɽ ఏҊख๏
12.  wऑֶश໦ΛϒʔεςΟϯάͯ͠༧ଌੑೳͷ޲্Λ໨ࢦ͍ͯ͠Δɽ
    wֶश࣌
     ؍ଌ஋ ͱ༧ଌ஋ ɹͱͷޡࠩΛݮΒ͢Α͏ʹܾఆ໦ͰֶशΛߦ͏ɽ
     ɽΛ܁Γฦͯ͠ɼޡࠩΛݮΒ͢ɽ
    w༧ଌ࣌
    wֶशʹΑͬͯٻΊͨ࢒ᕓͱ࢖ͬͯ༧ଌΛߦ͏ɽ y ̂ y ఏҊख๏ ޯ഑ϒʔεςΟϯά໦ (5# ʹ͍ͭͯ
13.  w'"(5#ͰެฏੑΛ֫ಘ͢Δํ਑ͱͯ͠ɼ໨తม਺ͱηϯγςΟϒଐੑͷ༧ଌਫ਼౓Λ ఢରతʹֶशΛߦ͍ɼ໨ࢦ͍ͯ͘͠ɽ
    w໨తม਺ͷ༧ଌثͱఢରత༧ଌث ηϯγςΟϒଐੑͷ༧ଌ ΛϛχϚοΫε๏ͱͯ͠ ࠷దԽ͍ͯ͘͠ɽ
    
    w ͕ఢରֶशͷׂ߹Λௐ੔͢ΔϋΠύʔύϥϝʔλ
    w-͸ͦΕͧΕଛࣦؔ਺Ͱ͋Δɽࠓճ͸෼ྨ໰୊͔ͩΒɼෛͷର਺໬౓Λ࠾༻ɽ arg min F max θA n ∑ i=1 LFi (F (xi)) − λ n ∑ i=1 LAi (F (xi); θA) λ ఏҊख๏ 'BJS
    "EWFSTBSJBM
    (SBEJFOU
    5SFF
    #PPTUJOH '"(5#
14.  ఏҊख๏ '"(5#ͷΞϧΰϦζϜ
15.  ఏҊख๏ '"(5#ͷશମਤ
16.  w࣮ݧͰ͸ɼਓ޻σʔλͱ࣮σʔλͷ྆ํΛ࢖ͬͯݕূ͢Δɽ
    w·ͨ405"ͳख๏ͱൺ΂ͯɼ༗༻ੑΛ͍ࣔͯ͘͠ɽ
    wఢରతֶशͷϋΠύʔύϥϝʔλ ͕GBJSOFTTͱBDDVSBDZͰͲ͏͍͏ӨڹΛ༩͑Δ ͔ௐ΂Δ λ ࣮ݧ ࣮ݧ֓ཁ
17.  wঢ়گઃఆͱͯ͠ɼࣗಈंอݥͰΫϨʔϜΛड͚Δ͔Ͳ͏͔Λߟ͑Δɽ
    wಛ௃ྔͱͯ͠ɼंͷ৭ɼ೥ྸɼੑผɼڟ๫ੑɼෆ஫ҙ͞Λ࣋ͭɽ
    w໨తม਺ͱͯ͠ɼΫϨʔϜΛ͔ͨ͠Ͳ͏͔Λೋ஋Ͱද͢ɽ
    wηϯγςΟϒଐੑͱͯ͠ɼੑผΛ༻͍Δɽ
    wͦΕͧΕͷੜ੒ํ๏͸ӈͷ௨Γ ࣮ݧ ਓ޻σʔλΛ༻͍Δ
18.  ࣮ݧ݁Ռ ਓ޻σʔλ
19.  wطଘͷ405"ͳख๏ͱൺ΂ΔͨΊɼ࣮σʔλͭΛ࢖͍ൺֱΛߦ͏ɽ
    w࢖༻͢Δσʔληοτ͸"EVMU
    $0.1"4
    %FGBVMU
    #BOLͷ̐छྨ
    w ࣮ݧ ࣮σʔλ
20.  w%1Λߟ͍͑ͯΔͱ͖͸ɼ Λຬ͍ͨͯ͠Δঢ়گͷ΋ͱͰɼBDDVSBDZ Λ༻͍ͯൺֱΛߦ͏ɽ
    w·ͨ&0Λߟ͍͑ͯΔͱ͖͸ɼ\'13 '/3^͕ҎԼΛຬ͍ͨͯ͠Δ΋ͱͰͷ BDDVBSDZΛൺֱ͢Δɽ 1SVMF = 90 % ࣮ݧ ໰୊ઃఆ
21.  ࣮ݧ݁Ռ %FNPHSBQIJD
    1BSJUZ
22.  ࣮ݧ݁Ռ &RVBMJ[FE
    0EET
23.  ࣮ݧ݁Ռ ϋΠύʔύϥϝʔλ
    ʹΑΔӨڹ λ
24.  wܾఆ໦Λ༻͍ͯɼެฏੑ഑ྀܕػցֶशΛߦ͏ख๏ΛఏҊͨ͠ɽ
    wຊख๏Ͱ͸طଘͷχϡʔϥϧωοτϫʔΫͷϞσϧͱಉ౳ͷੑೳ͕͋Δ͜ͱΛࣔ ͨ͠ɽ
    wܾఆ໦Λ༻͍Δ͜ͱͰ//ϞσϧΑΓ΋આ໌ੑ͕૿͍ͯ͠Δɽ ·ͱΊ
25.  wޯ഑ϒʔεςΟϯά໦ΛษڧͰ͖ͯΑ͔ͬͨ
    wϒʔεςΟϯάͩͱઆ໌ੑ͕͋Δͷ͔ٙ໰ʹࢥͬͨɽ
    wਓ޻σʔλͷ࡞Γํ͸͍͍ͳͬͯࢥͬͨɽ "9"Β͍͠
    wਂ૚χϡʔϥϧωοτܾఆ໦
    
    ఢରֶश
    ͰެฏੑΛ֫ಘͰ͖ͨΒࣗ෼ͷ΍Γͨ ͍͜ͱ ެฏੑͱઆ໌ੑ ΛຬͨͤΔ
    w%FFQ
    /FVSBM
    %FDJTJPO
    5SFFT
    IUUQTBSYJWPSHBCT ײ૝
26.  wޯ഑ܾఆϒʔεςΟϯά໦Ͱͱͯ΋෼͔Γ΍͔ͬͨ͢:PVUVCFS
    w4UBU2VFTU
    XJUI
    +PTI
    4UBSNFS
    ͞Μ
    wIUUQTXXXZPVUVCFDPNXBUDI W$$/[(+D ͓·͚ ษڧʹͳͬͨಈը