Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
[読み会] Individually Fair Gradient Boosting
Search
mei28
April 13, 2021
0
32
[読み会] Individually Fair Gradient Boosting
読み会資料
Individually Fair Gradient Boosting (ICLR 2021)
mei28
April 13, 2021
Tweet
Share
More Decks by mei28
See All by mei28
[読み会] “Are You Really Sure?” Understanding the Effects of Human Self-Confidence Calibration in AI-Assisted Decision Making
mei28
0
91
[JSAI'24] 人間の判断根拠は文脈によって異なるのか?〜信頼されるXAIに向けた人間の判断根拠理解〜
mei28
1
530
[CHI'24] Fair Machine Guidance to Enhance Fair Decision Making in Biased People
mei28
0
59
[DEIM2024] 卓球の得点予測における重要要素の分析
mei28
0
45
[Human-AI Decision Making勉強会] 意思決定 with AIは個人vsグループで変わるの?
mei28
0
210
[読み会] Words are All You Need? Language as an Approximation for Human Similality Judgements
mei28
0
41
[参加報告] AAAI'23
mei28
0
91
[計算機構論] Learning Models of Individual Behavior in Chess
mei28
0
72
[計算機構論] Why do tree-based models still outperform deep learning on tabular data?
mei28
0
60
Featured
See All Featured
Bootstrapping a Software Product
garrettdimon
PRO
307
110k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
28
1.6k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
233
17k
VelocityConf: Rendering Performance Case Studies
addyosmani
328
24k
Designing Dashboards & Data Visualisations in Web Apps
destraynor
231
53k
Practical Orchestrator
shlominoach
187
10k
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
331
21k
Intergalactic Javascript Robots from Outer Space
tanoku
270
27k
We Have a Design System, Now What?
morganepeng
51
7.5k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
44
7.1k
StorybookのUI Testing Handbookを読んだ
zakiyama
28
5.6k
The Straight Up "How To Draw Better" Workshop
denniskardys
232
140k
Transcript
Individual Fair Gradient Boosting 2021/04/13 @ಡΈձ ༶໌
•ஶऀใ •Alexander Vargo, Fan Zhang, Mikhail Yurochkin, Yuekai Sun
•ϛγΨϯେֶɼ্ւՊٕେֶɼMIT-IBM Watson AI Lab •ग़య: ICLR2021 •ͳΜͰબΜ͔ͩʁ •ݸผެฏੑ+ܾఆΛߟ͍͑ͯΔͷҙ֎ʹগͳ͍ɽˠ͜Ε͕ॳΊͯΒ͍͠ จใ ݸผެฏੑ + GBDTʹͨ͠ݚڀ
•ػցֶश(ML)͕ҙࢥܾఆͷͰ͘ΘΕ࢝Ί͍ͯΔ •ಛఆͷάϧʔϓ(ਓ)ʹରͯ͠ෆެฏͳධՁΛ͍͚ͯ͠ͳ͍ •Amazonͷཤྺॻ৹ࠪγεςϜͰࠩผ͕ߦΘΕ͍ͯͨ͜ͱ͕໌Β͔ʹͳͬ ͨɽ ΠϯτϩμΫγϣϯ ެฏੑΛߟྀ͍͔ͯ͠ͳ͍ͱ͍͚ͳ͍
•MLք۾Ͱେ͖͘ೋछྨͷެฏੑΛߟ͑Δ •ݸผެฏੑ: ࣅ͍ͯΔݸਓಉ͡ධՁΛड͚Δ͖ •ूஂެฏੑ: ूஂ͝ͱʹධՁͷࠩผ͕ͳ͍Α͏ʹ͢Δ͖ •ूஂެฏੑ͕Α͘औΓ্͛ΒΕ͍ͯΔ •ݸਓͷྨࣅΛ͖ͪΜͱఆٛ͢Δ͜ͱ͕ࠔ͔ͩͬͨΒ ΠϯτϩμΫγϣϯ ࠓճݸผެฏੑΛରͱ͍ͯ͘͠
•දσʔλʹGBDTΛ༻͍Δͷ͕ओྲྀʹͳ͖͍ͬͯͯΔɽ •ैདྷͷFair-awarness MLͰnon-smoothͳϞσϧ ϊϯύϥϝτϦοΫMLͰ͋·Γྑ͍ޮՌ͕ಘΒΕͯ ͍ͳ͔ͬͨɽ ΠϯτϩμΫγϣϯ ޯϒʔεςΟϯάܾఆ(GBDT)Λରͱ͢Δ
•ݸผެฏੑΛରʹͨ͠GBDTʹΑΔख๏ΛఏҊͨ͠ɽ •ϞσϧͷͦΕͧΕͷެฏੑΛূ໌͢Δ͜ͱ͕Մೳɽ •ݸผެฏੑ͚ͩͰͳ͘ूஂެฏੑΛ্ͤͭͭ͞ɼਫ਼Λҡ࣋ ͢Δख๏ʹͳ͍ͬͯΔ͜ͱΛ࣮ݧతʹࣔͨ͠ɽ ΠϯτϩμΫγϣϯ ߩݙ
•ೖྗ: , ग़ྗ: •อޢ͢Δଐੑ: ͍ΘΏΔηϯγςΟϒଐੑ •αϯϓϧ͝ͱެฏࢦඪ: ͜Εαϯϓϧ͕͍ۙ΄Ͳࣅ͍ͯΔ •ඪ:
ɹαϯϓϧ͝ͱʹެฏͳϞσϧ Λ֫ಘ͢Δ͜ͱ 𝒳 ∈ ℝd 𝒴 = {0,1} 𝒵 = 𝒳 × {0,1} dx f : 𝒳 → {0,1} ४උ ͏ه߸Λఆٛ͢Δ
•ఢରֶशʹΑͬͯୡ͢Δํ๏ଘࡏ͍ͯ͠Δ • ֶश͕ೖྗʹରͯ͠Β͔Ͱ͋Δ͜ͱ͕લఏʹͳ͍ͬͯΔ •Β͔Ͱͳ͍Ϟσϧʢܾఆͱ͔ʣʹରͯ͠ఢରֶशΛߦ͑ ΔΑ͏ʹ͍ͨ͠ʂ • ੍ݶ͖ఢରతίετؔΛఆٛͨ͠Αʂ طଘख๏Ͳ͏ͩͬͨͷʁ Non-smoothͳϞσϧͰ͏·͍͔͘ͳ͔ͬͨɽ
•Transport cost function: ݸผͷαϯϓϧ͕͍ۙ΄Ͳখ͍͞ •Zͷ্֬ͷ࠷ద༌ૹڑ : ͷۙ͞Λߟ͍͑ͯΔ c ((x1
, y1), (x2 , y2)) ≜ d2 x (x1 , x2) + ∞ ⋅ 1 {y1 ≠y2} W W (P1 , P2) ≜ inf Π∈C(P1 , P2) ∫ 𝒵×𝒵 c (z1 , z2) dΠ (z1 , z2) ४උ αϯϓϧ͝ͱʹެฏͳϞσϧΛֶश͍ͨ͠
• σʔλੜɼ ͷඍখͳڐ༰ύϥϝʔλ •ඪຊ্ۭؒͰ1) σʔλੜ͕͍ۙ ɹɹɹɹɹ ɹ2) MLϞσϧͷଛࣦΛେ͖͘ͳΔͷ Λ୳͍ͨ͠
Lr (f) ≜ sup P:W(P, P* )≤ϵ 𝔼P [ℓ(f(X), Y)] P⋆ ϵ > 0 ४උ ఢରతϦεΫؔΛఆٛ͢Δɽ
•ྨࣅͨ͠αϯϓϧʹରͯ͠ϞσϧͷੑೳࠩΛݟ͚ͭΒΕΔ •ੑೳࠩΛ୳ࡧ͢Δ͜ͱͰʹରͯ͠ؤ݈ͳެฏੑͩͱଊ͑ ΒΕΔɽ •ݱঢ়ͩͱ·ͩsmoothͳϞσϧͷޯ͔͠ಘΒΕͳ͍ɽ ४උ ϩόετͰެฏͳΛಘ͍ͨʂ
•σʔληοτΛ֦ு͢Δ: •࠷ద༌ૹؔʹ੍ݶΛՃ͑Δ: ҧ্͍ͷσʔληοτ͔Ͳ͏͔ 𝒟0 ≜ {(xi , yi), (xi
,1 − yi)} n i=1 W𝒟 (P1 , P2) ≜ inf Π∈C0(P1 , P2) ∫ 𝒵×𝒵 c (z1 , z2) dΠ (z1 , z2) ఏҊख๏ ੍ݶΛՃ͑ͯnon-smoothͷͨΊʹ͢Δɽ
• σʔληοτΛՃ͑Δ͜ͱͰ্ք ʹࢦࣔ͞Εͨʹ੍ݶ ͞ΕΔ •͜ΕʹΑͬͯ༗ݶ࣍ݩઢܗܭը๏ʹΑͬͯղ͚ΔΑ͏ʹͳΔɽ •ଛࣦ ʹ͔͠ґଘͯ͠ͳ͍ ͔ΒඇฏͳϞσϧͰద༻Ͱ͖Δɽ D0
ℓ (f (xi), yi) and ℓ (f (xi) ,1 − yi) ఏҊख๏ ͬͱඇฏʹద༻Ͱ͖ΔΑ
ޯϒʔεςΟϯάͰ ΛٻΊΔඞཁ͕͋Δɽ μϯεΩϯͷఆཧΛ༻͍Δͱޯɼ ∂L ∂ ̂ y ∂L ∂
̂ yi = ∂ ∂f (xi) [ sup P:W𝒟(P, Pn)≤ϵ 𝔼P [ℓ (f (xi), yi)]] = ∑ y∈𝒴 ∂ ∂f (xi) [ℓ (f (xi), y)) P* (xi , y) ఏҊख๏ ޯϒʔεςΟϯάͰ͑ΔΑ͏ʹ͢Δ
•ઌड़ͷޯͰɼϞσϧΛඍ͢Δඞཁ͕ͳ͍͔ΒඇฏͳϞ σϧͰؔޯΛධՁ͢Δ͜ͱ͕Ͱ͖Δʂ •͋ͱ ΛٻΊΕྑ͍ɽ •ઢܗܭը๏ʹΑͬͯ ΛٻΊΔํ๏ΛఏҊ͢Δɽ P⋆ P⋆ ఏҊख๏
ؔޯΛߟ͑Δ
• ʹΑΔҙͷ ʹରͯ͠ɼ ͱ͢Δͱ ࣍ͷΑ͏ͳߦྻ ͰදͤΔɽ 1. 2. D0
P Pi,k = P({(xi , k}), k ∈ {0,1} WD (P, Pn ) ≤ ϵ Π Π ∈ Γ with Γ = {Π ∣ Π ∈ ℝn×n + , ⟨C, Π⟩ ≤ ϵ, ΠT ⋅ 1n = 1 n 1n} Π ⋅ y1 = (P1,1 , …, Pn,1), and Π ⋅ y0 = (P1,0 , …, Pn,0) ఏҊख๏ Λઢܗܭը๏ͰٻΊΔ P⋆
•ߦྻ ɹˠ ϥϕϧjͰ͋Δαϯϓϧj͕αϯϓϧiʹ ͳͬͨͱ͖ͷଛࣦ •ٻΊ͍ͨߦྻ ࣍ͷΑ͏ʹͳΔ Ri,j = l(f(xi
), yj ) Π⋆ Π⋆ ∈ arg max Π∈Γ ⟨R, Π⟩ ఏҊख๏ ͞Βʹఆ͍ٛͯ͘͠Α
•݁ہ࠷ޙͷ ΛٻΊΔ͜ͱ͕Ͱ͖Εྑ͍ɽ •ٻΊΔʹ͋ͨͬͯɼؔFʹԿԾఆΛஔ͍͍ͯͳ͍ͷͰɼඇ ฏͳؔʹద༻Ͱ͖Δɽ Π⋆ ఏҊख๏- ·ͱΊ ͜ΕͰඇฏͳؔʹద༻Ͱ͖Δʂ
•3ͭͷσʔληοτ(German Credit, Adult, COMPASS)Ͱݕূ •ఏҊख๏Ͱ༻͍ΔܾఆΞϧΰϦζϜɼXGBoostͱ͢Δɽ •ଛࣦؔϩδεςΟοΫଛࣦΛ༻͍Δɽ ࣮ݧ
•YurochikinΒͷΛར༻͢Δ: •QηϯγςΟϒ෦ۭؒͱߦ͢ΔࣹӨߦྻ •อޢ͞ΕΔηϯγςΟϒଐੑҎ֎ͷใ͕ಉ͡ͳΒಉʹѻΘ ΕΔ͖Ͱ͋Δͱ͍͏ߟ͔͑Β࡞ΒΕͨɽ d2 x = (x1 −
x2 , Q(x1 − x2 )) ࣮ݧ ެฏੑࢦඪʹ͍ͭͯ(ݸผͷαϯϓϧʹؔͯ͠)
•ܾఆख๏ʹؔͯ͠ɼର߅͕ͳ͍ͨΊόχϥΛ༻͍Δɽ •σʔλͷલॲཧΛ༻͍Δख๏ͱൺֱ͢Δ •อޢଐੑΛͳ͘͠ɼ෦ۭؒʹӨ͢Δ(Yurochkin et al., 2020) •ݸਓʹҟͳΔॏΈΛద༻ͯ͠όϥϯεΛͱΔ(Kamiran & Calders,
2011) ࣮ݧ ର߅ख๏ʹ͍ͭͯ
•อޢ͞ΕΔ͍ͯΔଐੑͱ૬͕ؔ͋Δଐੑ(e.g. ͔?࠺͔?)ΛͣΒ ͢͜ͱͰࣄ࣮ͷਓΛ࡞ɽ •→΄΅ಉ͡ਓ͔ͩΒಉ͡ධՁΛ͞ΕΔ͖ •อޢଐੑ͝ͱͷTPR,TNRͷࠩ(GAPMax)→Ϟσϧͷެฏੑࢦඪ •อޢଐੑ͝ͱͷRMSEͷࠩ(GAPRMSE) →Ϟσϧͷ༧ଌੑೳ ࣮ݧ ධՁʹ͍ͭͯ(طଘख๏ʹର͠༏ྼ͕ͳ͍Α͏ʹՃΛ͢Δ)
•ྸΛηϯγςΟϒଐੑʹઃఆ → ถࠃͰྸΛ͚ͭͯ༩৴ அ͢Δͷҧݑ •ࣹӨʹΑΔલॲཧఏҊ΄ͲݸਓͷެฏੑΛ্ͤ͞ͳ͔ͬͨɽ ࣮ݧ݁Ռ ᶃ German Credit
•ఏҊख๏GBDTͷੑೳͷྑ͞ΛҾ͖ܧ͗ͭͭɼެฏͳϞσϧʹ ͳ͍ͬͯͨʂ ࣮ݧ݁Ռ ᶄ Adult
•NNϞσϧͷํ͕ਫ਼جຊతʹྑ͔ͬͨɽ •͔͠͠ެฏੑʹ͍ͭͯɼఏҊͷํ͕ྑ͔ͬͨɽ ࣮ݧ݁Ռ ᶅCOMPASS
•ݸผެฏੑΛୡ͢Δ՝ΛMLϞσϧͷੑೳࠩΛ୳ࡧͰ͖ͳ͍ ͜ͱ → ୳ࡧۭؒΛ༗ݶ۠ؒʹ੍ݶ͢Δ͜ͱͰࠀͨ͠ɽ •ࠓճઃఆ੍ͨ͠ݶ͖ఢରଛࣦؔଞͷnon-smoothख๏(ϥϯ μϜϑΥϨετ)ͳͲʹద༻Ͱ͖Δ͔͠Εͳ͍ɽ •࣮ײͱͯ͠ɼNNϞσϧΑΓܾఆϕʔεͷ΄͏͕ਫ਼ʴެฏ ੑΛୡͰ͖ͦ͏. ·ͱΊ
ݸผެฏੑʴܾఆͷख๏ΛఏҊͨͧ͠
•࡞ऀ͕͍ࣔͯ͠ΔཧΛͪΌΜͱཧղͰ͖ͳ͍ͯ͘͘͠ɽ •ݸผެฏੑΛߟ͍͑ͯΔจΛಡΊͯྑ͔ͬͨɽ ײ ͖ͪΜͱཧΛ͑Δֶྗ͕ཉ͍͠