Upgrade to Pro — share decks privately, control downloads, hide ads and more …

説明可能な機械学習と数理最適化

Avatar for kelicht kelicht
October 17, 2025

 説明可能な機械学習と数理最適化

2025/10/17
第37回 RAMP 数理最適化シンポジウム (RAMP2025)
https://orsj.org/ramp/ramp2025

タイトル:
説明可能な機械学習と数理最適化

概要:
深層学習に代表される機械学習技術の発展にともない,機械学習モデルの予測に基づくデータ駆動型の意思決定が,医療や金融などといった実社会の様々な領域で注目されている.このような領域における重大な意思決定タスクでは,予測に基づく意思決定結果が人間の生活に大きな影響を与える恐れがある.したがって,機械学習モデルの透明性を担保して説明責任を果たすために,予測結果に関する何らかの追加情報を提示できる説明可能性(explainability)の実現が喫緊の課題となっている.しかし,ユーザからの信頼を得るために説明が満たすべき要件と,それを満たす説明を生成するための具体的な方法は,まだ十分に解明されていない.本稿では,機械学習モデルから肯定的な予測結果を得るための行動指針(アクション)を説明として提示する枠組みであるアルゴリズム的償還(algorithmic recourse)に着目する.具体的には,ユーザにとって望ましいアクションを提示するタスクをどのように定式化するかというモデリングの観点と,定式化した最適化問題をどのように解くかというアルゴリズムの観点から,筆者らがこれまでに取り組んできた研究について紹介する.

Avatar for kelicht

kelicht

October 17, 2025
Tweet

More Decks by kelicht

Other Decks in Research

Transcript

  1. 1 આ໌Մೳͳػցֶशͱ ਺ཧ࠷దԽ Explainable Machine Learning and Mathematical Optimization ۚ৿

    ݑଠ࿕ [email protected] | https://sites.google.com/view/kentarokanamori ෋࢜௨גࣜձࣾ ਓ޻஌ೳݚڀॴɹϓϦϯγύϧϦαʔνϟʔ ୈ37ճ RAMP ਺ཧ࠷దԽγϯϙδ΢Ϝ (RAMP2025) * ຊݚڀ͸ JST ACT-X JPMJAX23C6 ͷॿ੒Λ෦෼తʹड͚ͨ΋ͷͰ͢
  2. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲࣗݾ঺հʳۚ৿ ݑଠ࿕ (Kentaro Kanamori, Ph.D.)

    2 ػցֶश΍਺ཧ࠷దԽΛ༻͍ͨҙࢥܾఆࢧԉʹڵຯ͕͋Γ·͢ʂ ۚ৿ ݑଠ࿕ ෋࢜௨גࣜձࣾ ਓ޻஌ೳݚڀॴ ϓϦϯγύϧϦαʔνϟʔ ത࢜ʢ৘ใՊֶʣ ུྺ • April 2020 - March 2022 ๺ւಓେֶ େֶӃ৘ใՊֶӃ ത࢜ޙظ՝ఔ (୹ॖमྃ) ೔ຊֶज़ৼڵձ ಛผݚڀһ (DC1) • April 2022 - Present ෋࢜௨גࣜձࣾ ਓ޻஌ೳݚڀॴ ݚڀһ • October 2023 - Present JST ACT-X ʮ࣍ੈ୅AIΛங͘਺ཧɾ৘ใՊֶͷֵ৽ʯྖҬ Ұظੜ • April 2025 - Present ਓ޻஌ೳֶձ ਓ޻஌ೳجຊ໰୊ݚڀձ (SIG-FPAI) װࣄ ݚڀ෼໺ • આ໌Մೳͳػցֶश (͍ΘΏΔ “આ໌ՄೳAI”) ΞϧΰϦζϜతঈؐ / ൓࣮Ծ૝આ໌๏ / ղऍՄೳͳϧʔϧϞσϧֶश • ౷ܭతҼՌ୳ࡧ • ਺ཧ࠷దԽ (੔਺ܭը๏, ྼϞδϡϥ࠷దԽ, …) ݸਓϖʔδ
  3. 2025/10/17 K. Kanamori, Fujitsu Ltd. ػցֶशͷઆ໌Մೳੑ 4 આ໌Մೳੑͷ࣮ݱ͸, ػցֶशͷ৴པੑ޲্΁ͷୈҰา •

    ػցֶशϞσϧͷ༧ଌΛ༻͍ͨҙࢥܾఆ͕࣮ࣾձʹਁಁͭͭ͋͠Δ ‣ Ϟσϧͷ༧ଌ݁Ռ͕ਓؒͷੜ׆ʹେ͖ͳӨڹΛ༩͑Δ (ྫ: ҩྍ਍அ, ϩʔϯ৹ࠪ) ͋ͳͨ͸ ౶೘පϦεΫ͕ ߴ͍Ͱ͢ ʁ Ϟσϧ Ϣʔβ આ໌Մೳੑ ࣮ݱ ࠜڌ͸ BMIͷ஋͕ߴ͍ ͔ΒͰ͢ Ϟσϧ Ϣʔβ • ػցֶशϞσϧ͕ग़ྗ͢Δ༧ଌ݁Ռʹؔ͢Δ “આ໌” ΛఏࣔͰ͖Δ આ໌Մೳੑ (explainability) ͷ࣮ݱ͕٤ۓͷ՝୊ͱͳ͍ͬͯΔ
  4. 2025/10/17 K. Kanamori, Fujitsu Ltd. આ໌Մೳੑ΁ͷࣾձతͳཁ੥ 5 ࠃ಺֎ͷΨΠυϥΠϯ΍๏ྩͰઆ໌Մೳੑ͕ॏཁࢹ͞Ε͍ͯΔ AIࣄۀऀΨΠυϥΠϯ (ܦࡁ࢈ۀলʴ૯຿ল,

    2024೥4݄) ᶆ ؔ࿈͢ΔεςʔΫϗϧμʔ΁ͷઆ໌ՄೳੑɾղऍՄೳੑͷ޲্ ؔ࿈͢ΔεςʔΫϗϧμʔͷೲಘײٴͼ҆৺ײͷ֫ಘɺ·ͨɺͦͷͨ Ίͷ AI ͷಈ࡞ʹର͢Δূڌͷఏࣔ౳Λ໨తͱͯ͠ɺઆ໌͢Δओମ͕ ͲͷΑ͏ͳઆ໌͕ٻΊΒΕΔ͔Λ෼ੳɾ೺ѲͰ͖ΔΑ͏ɺઆ໌Λड͚ Δओମ͕ͲͷΑ͏ͳઆ໌͕ඞཁ͔Λڞ༗͠ɺඞཁͳରԠΛߨ͡Δ (ୈ̎෦ʮAI ʹΑΓ໨ࢦ͢΂͖ࣾձٴͼ֤ओମ͕औΓ૊Ήࣄ߲ʯB.ʮݪଇʯΑΓ) ࠃ֎Ͱ͸ EU AI Act (EU, 2024೥) ΍ Ethically Aligned Design (IEEE, 2019೥) ͳͲʹ໌ه
  5. 2025/10/17 K. Kanamori, Fujitsu Ltd. આ໌Մೳੑ΁ͷΞϓϩʔν 6 ղऍ͠΍͍͢ϞσϧΛ࢖͏ ʗ Ϟσϧ͔Βہॴతͳઆ໌Λநग़

    :FT /P :FT /P ݂౶஋ ≤ 127 #.* ≤ 29.5 ݈߁ ౶೘ප ݈߁ Ϟσϧࣗମ͕ ༧ଌ݁Ռͷઆ໌Λ ఏࣔͰ͖Δ ಛ௃ྔॏཁ౓ ݂౶஋ BMI ೥ྸ ੑผ ղऍՄೳͳϞσϧͷֶश ༧ଌͷաఔ΍ࠜڌΛਓ͕ؒཧղͰ͖Δ ղऍՄೳͳϞσϧΛֶश ✓ ྫ: εύʔεઢܗϞσϧ (Lasso), ܾఆ໦, … ہॴઆ໌ͷࣄޙతநग़ ݸʑͷ༧ଌ݁Ռʹؔ͢Δہॴઆ໌Λ ֶशࡁΈϞσϧ͔Βநग़ ✓ ྫ: LIME [Ribeiro+ 16], SHAP [Lundberg+ 17], … ݸʑͷ༧ଌʹ͓͚Δ ಛ௃ྔͷߩݙ౓Λఏࣔ
  6. 2025/10/17 K. Kanamori, Fujitsu Ltd. ΋ͬͱৄ͘͠஌Γ͍ͨਓ޲͚ 7 ػցֶशͷઆ໌ՄೳੑɾղऍՄೳੑʹؔ͢ΔղઆهࣄͱڭՊॻ Christoph Molnar:

    Interpretable Machine Learning https://christophm.github.io/interpretable-ml-book/ ݪ૱: ࢲͷϒοΫϚʔΫʮઆ໌ՄೳAIʯ ਓ޻஌ೳ, Vol. 34, No. 4 (2019) ݪ૱: ػցֶशϞσϧͷ൑அࠜڌͷઆ໌ʢVer.2ʣ https://www.slideshare.net/slideshow/ver2-225753735/225753735
  7. 2025/10/17 K. Kanamori, Fujitsu Ltd. 1. ͲͷΑ͏ͳػցֶशϞσϧɾઆ໌ܗࣜΛର৅ʹ, 2. ͲͷΑ͏ͳ࠷దԽ໰୊ͱͯ͠ఆࣜԽ͠, 3.

    ͲͷΑ͏ʹղ͘ʁ ‣ ͲͷΑ͏ͳϞσϧ͕ղऍ͠΍͍͢ʁ ͲͷΑ͏ͳ৘ใΛؚΉઆ໌͕໾ʹཱͭʁ ‣ ͲͷΑ͏ͳධՁࢦඪͰઆ໌ͷ࣭ΛධՁ͢Δʁ ͲͷΑ͏ͳ੍໿৚݅Λߟྀ͢Δඞཁ͕͋Δʁ ‣ DNNͷֶशͷΑ͏ʹޯ഑๏Ͱղ͚Δʁ ʲࢲݟʳݚڀͷϙΠϯτ 8 Ϣʔβʹͱͬͯ “ྑ͍આ໌” ΛͲͷΑ͏ʹ਺ཧϞσϧԽ͢Δ͔ʁ આ໌ՄೳAI ݱ৔ͷϢʔβ Ϊϟοϓ Ͳ͏ղ͔͘ (ΞϧΰϦζϜ) ͚ͩͰͳ͘, ԿΛղ͔͘ (ϞσϦϯά) ΋େࣄʂ ͲΜͳઆ໌͕ Մೳʁ ͲΜͳઆ໌͕ ඞཁʁ
  8. 2025/10/17 K. Kanamori, Fujitsu Ltd. ΞϧΰϦζϜతঈؐ (Algorithmic Recourse; AR) 10

    ൱ఆతͳ༧ଌ݁ՌΛ෴ͨ͢Ίͷ “ΞΫγϣϯ” Λઆ໌͢Δ࿮૊Έ ैདྷͷہॴઆ໌๏ (ྫ: LIME) ༧ଌͷࠜڌͱͳͬͨಛ௃ྔΛఏࣔ ‣ࠜڌͷఏ͚ࣔͩͰ͸ೲಘͯ͠΋Β͑ͳ͍ ౶೘පϦεΫ͕ߴ͍Ͱ͢ ࠜڌ͸݂౶஋ͱBMIͱ೥ྸͰ͢ ;ʔΜ……Ͱ, Ͳ͏͢Ε͹༧๷Ͱ͖Δͷʁ ‣ϢʔβʹͱͬͯΑΓݐઃతͳઆ໌͕Մೳ ΞϧΰϦζϜతঈؐ [Ustun+ 19] ༧ଌΛ෴͢ಛ௃ྔͷมߋํ๏Λఏࣔ μΠΤοτ͢Δ͔ʂ ΞΫγϣϯ ΋͠BMI͕27.3ͩͬͨΒ ϦεΫ͕௿͍ͱ༧ଌ͞Ε·ͨ͠
  9. 2025/10/17 K. Kanamori, Fujitsu Ltd. ΞϧΰϦζϜతঈؐ (Algorithmic Recourse; AR) 11

    ൱ఆతͳ༧ଌ݁ՌΛ෴ͨ͢Ίͷ “ΞΫγϣϯ” Λઆ໌͢Δ࿮૊Έ ैདྷͷہॴઆ໌๏ (ྫ: LIME) ༧ଌͷࠜڌͱͳͬͨಛ௃ྔΛఏࣔ ‣ࠜڌͷఏ͚ࣔͩͰ͸ೲಘͯ͠΋Β͑ͳ͍ ౶೘පϦεΫ͕ߴ͍Ͱ͢ ࠜڌ͸݂౶஋ͱBMIͱ೥ྸͰ͢ ;ʔΜ……Ͱ, Ͳ͏͢Ε͹༧๷Ͱ͖Δͷʁ ‣ϢʔβʹͱͬͯΑΓݐઃతͳઆ໌͕Մೳ ΞϧΰϦζϜతঈؐ [Ustun+ 19] ༧ଌΛ෴͢ಛ௃ྔͷมߋํ๏Λఏࣔ μΠΤοτ͢Δ͔ʂ ΞΫγϣϯ ΋͠BMI͕27.3ͩͬͨΒ ϦεΫ͕௿͍ͱ༧ଌ͞Ε·ͨ͠ ൓࣮Ծ૝આ໌๏ [Wachter+ 18] ݱࡏͷ༧ଌ݁ՌΛ෴͢ઁಈϕΫτϧΛ ఏࣔ͢Δہॴઆ໌๏ ‣ ΍͍ͬͯΔ͜ͱ͸΄΅ಉ͡ʂ
  10. 2025/10/17 K. Kanamori, Fujitsu Ltd. ΞΫγϣϯੜ੒໰୊ͷఆࣜԽ (1/2) 12 Ϣʔβʹͱͬͯ “࣮ݱՄೳ”

    ͳΞΫγϣϯͰͳ͍ͱҙຯ͕ͳ͍ ؍఺2. ࣮ߦίετ ࣮ߦʹ͔͔Δ࿑ྗ͸ͳΔ΂͘গͳ͘ ؍఺1. ࣮ߦՄೳੑ ม͑ΒΕͳ͍ಛ௃ྔ͸มߋ͠ͳ͍ ମॏΛ20kgݮΒ͠·͠ΐ͏ʂ ࣮ߦίετ͕ݱ࣮తͰͳ͍ ೥ྸΛ5ࡀݮΒ͠·͠ΐ͏ʂ อޢଐੑ͸มߋෆՄೳ աڈͷϩʔϯฦࡁ஗Ԇ݅਺Λ 2݅ݮΒ͠·͠ΐ͏ʂ ཤྺ΍࣮੷͸ม͑ΒΕͳ͍ ݄ऩΛ3ສԁ૿΍ͯ͠ ࢿ֨Λ1݅औಘͯ͠స৬ͯ͠… มߋ͢Δಛ௃ྔ͕ଟ͍
  11. 2025/10/17 K. Kanamori, Fujitsu Ltd. ΞΫγϣϯੜ੒໰୊ͷఆࣜԽ (2/2) 13 ༧ଌ݁ՌΛ෴࣮͢ߦՄೳͳΞΫγϣϯͷதͰ࣮ߦίετΛ࠷খԽ ΞΫγϣϯੜ੒໰୊

    [Wachter+ 18; Ustun+ 19] Πϯελϯε ͱ෼ྨث ʹରͯ͠, ࣍ͷ໰୊ͷ࠷దղͱͳΔΞΫγϣϯ ΛٻΊΔ: ͜͜Ͱ, ͸࣮ߦՄೳΞΫγϣϯू߹, ͸࣮ߦίετΛධՁ͢Δίετؔ਺ (ྫ: -ϊϧϜ ). x = (x1 , …, xD ) ∈ ℝD h: ℝD → {−1, + 1} a* ∈ ℝD mina∈ 𝒜 (x) c(a ∣ x) subject to h(x + a) = + 1 𝒜 (x) ⊆ ℝD c ℓ1 ∥a∥1 x1 x2 x x + a •: •: h(x) = + 1 h(x) = − 1
  12. 2025/10/17 K. Kanamori, Fujitsu Ltd. ΞΫγϣϯੜ੒໰୊ͷղ๏ 14 ࿈ଓ࠷దԽ΍཭ࢄ࠷దԽͷٕज़ͰΞΫγϣϯੜ੒໰୊Λղ͘ mina∈ 𝒜

    (x) c(a ∣ x) subject to h(x + a) = + 1 x x + a ΞΫγϣϯੜ੒໰୊ (࠶ܝ) ղ๏1. ϥάϥϯδϡ؇࿨&ޯ഑๏ ๏ ࣮૷͕༰қ ๏ GPUΛ༻͍ͨߴ଎Խ͕Մೳ ✓ Ϟσϧ΍ίετͷඍ෼ՄೳੑΛԾఆ ✓ ࠷దղͱ͸ݶΒͳ͍ ղ๏2. ࠞ߹੔਺ઢܗ࠷దԽ (MILO) ๏ ඍ෼ෆՄೳͳϞσϧ΍ίετʹରԠ ๏ ιϧόʔΛ༻͍ͯ࠷దղ͕ಘΒΕΔ ✓ େن໛໰୊ʹ͸εέʔϧ͠ͳ͍ ͦͷଞͷղ๏: SAT, GA, ہॴ୳ࡧ, etc.
  13. 2025/10/17 K. Kanamori, Fujitsu Ltd. զʑͷݚڀͷग़ൃ఺ 15 طଘͷΞΫγϣϯੜ੒໰୊Λޮ཰ྑ͘ղ͚ͩ͘Ͱे෼ͳͷ͔ʁ mina∈ 𝒜

    (x) c(a ∣ x) subject to h(x + a) = + 1 x x + a ΞΫγϣϯੜ੒໰୊ (࠶ܝ) ղ๏1. ϥάϥϯδϡ؇࿨&ޯ഑๏ ๏ ࣮૷͕༰қ ๏ GPUΛ༻͍ͨߴ଎Խ͕Մೳ ✓ Ϟσϧ΍ίετͷඍ෼ՄೳੑΛԾఆ ✓ ࠷దղͱ͸ݶΒͳ͍ ղ๏2. ࠞ߹੔਺ઢܗ࠷దԽ (MILO) ๏ ඍ෼ෆՄೳͳϞσϧ΍ίετʹରԠ ๏ ιϧόʔΛ༻͍ͯ࠷దղ͕ಘΒΕΔ ✓ େن໛໰୊ʹ͸εέʔϧ͠ͳ͍ ͦͷଞͷղ๏: SAT, GA, ہॴ୳ࡧ, etc. ՝୊ 1. ΞΫγϣϯͷ࣮ݱՄೳੑΛ ద੾ʹධՁ͢Δʹ͸ ͲΜͳίετؔ਺͕ඞཁʁ ՝୊ 2. ॴ༩ͷֶशࡁΈϞσϧʹ͓͍ͯ ࣮ݱՄೳͳΞΫγϣϯ͕ ଘࡏ͢Δอূ͸͋Δʁ
  14. 17 DACE: Distribution-Aware Counterfactual Explanation by Mixed-Integer Linear Optimization Kentaro

    Kanamori1, Takuya Takagi2, Ken Kobayashi2,3, Hiroki Arimura1 1Hokkaido University, 2Fujitsu Laboratories Ltd., 3Tokyo Institute of Technology 29th International Joint Conference on Artificial Intelligence (IJCAI2020)
  15. 2025/10/17 K. Kanamori, Fujitsu Ltd. ՝୊ 18 طଘͷίετؔ਺Ͱ͸σʔλ෼෍ͷಛੑΛे෼ʹߟྀͰ͖ͳ͍ ے೑ྔ ମॏ

    ࣮ࡍͷ ίετେ ֎Ε஋ طଘͷίετؔ਺ͷଟ͕͘ ಛ௃ྔؒͷ૬ؔؔ܎ΛߟྀͰ͖ͳ͍ ୯ͳΔίετؔ਺࠷খԽͰ͸ ΞΫγϣϯͷ࣮ߦ݁Ռ͕֎Ε஋ʹͳΔ ‣ σʔλ෼෍ͷಛੑΛߟྀ͠ͳ͍ͱ࣮ݱՄೳͳΞΫγϣϯ͸ಘΒΕͳ͍ ಛ௃ྔ̍ ಛ௃ྔ̎
  16. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ໨ඪʳσʔλ෼෍Λߟྀͨ͠ΞΫγϣϯઆ໌ 19 σʔλ෼෍ͷಛੑΛߟྀ࣮ͯ͠ݱՄೳͳΞΫγϣϯΛఏ͍ࣔͨ͠ ے೑ྔ ମॏ

    ࣮ࡍͷ ίετେ ࣮ࡍͷ ίετখ ಛ௃ྔ̍ ಛ௃ྔ̎ ֎Ε஋ ਖ਼ৗ஋ ໨ඪ ಛ௃ྔؒͷ૬ؔͱ֎Ε஋ϦεΫΛߟྀͨ͠৽ͨͳίετؔ਺Λಋೖ͠࠷దԽ͢Δ ૬ؔؔ܎ʹԊͬͨΞΫγϣϯ ΞΫγϣϯͷ࣮ߦ݁Ռ͕ਖ਼ৗ஋
  17. 2025/10/17 K. Kanamori, Fujitsu Ltd. σʔλ෼෍Λߟྀͨ͠ίετؔ਺ 20 ಛ௃ྔؒͷ૬ؔͱ֎Ε஋ϦεΫΛߟྀͨ͠ίετؔ਺Λಋೖ Distribution-Aware Counterfactual

    Explanation (DACE) Πϯελϯεू߹ , ڞ෼ࢄߦྻ , ύϥϝʔλ ʹ ରͯ͠, ҎԼͷίετؔ਺ͷ΋ͱͰΞΫγϣϯੜ੒໰୊Λղ͘: ͜͜Ͱ, ͸ϚϋϥϊϏεڑ཭, ͸ Local Outlier Factor (LOF). X ⊆ ℝD Σ ∈ ℝD×D λ ≥ 0, k > 0 cDACE (a ∣ x) := d2 M (x, x + a ∣ Σ−1) + λ ⋅ qk (x + a ∣ X) dM qk ‣ ಛ௃ྔؒͷ૬ؔͱ֎Ε஋ϦεΫΛߟྀͯ͠ΞΫγϣϯΛ࠷దԽ͢Δʂ
  18. 2025/10/17 K. Kanamori, Fujitsu Ltd. ࠷దԽํ๏ͷ֓ཁ 21 ࠞ߹੔਺ઢܗ࠷దԽ໰୊ͱͯ͠ఆࣜԽ͠ޮ཰Αۙ͘ࣅతʹղ͘ • ಋೖͨ͠ίετؔ਺

    Λ໨తؔ਺ͱ͢ΔΞΫγϣϯੜ੒໰୊͸ ࠞ߹੔਺ઢܗ࠷దԽ໰୊ͱͯ͠ఆࣜԽՄೳ ‣ ม਺ͱ੍໿ࣜͷ૯਺͕ಛ௃ྔ਺ʹґଘͯ͠૿Ճ͠ ਺े࣍ݩͰ΋ݱ࣮తͳ࣌ؒͰٻղෆՄೳʹ cDACE mina∈ 𝒜 (x) cDACE (a ∣ x) ΞΫγϣϯੜ੒໰୊ subject to h(x + a) = + 1 • ޮ཰Α͘ఆࣜԽՄೳͳ୅ཧؔ਺ Λಋೖ 1. ڞ෼ࢄߦྻ ͷ෼ղʹΑΔۙࣅϚϋϥϊϏεڑ཭ 2. LOF ͷ -ۙ๣఺ू߹Λ Ͱݻఆ ̂ cDACE Σ k k = 1 mina∈ 𝒜 (x) ̂ cDACE (a ∣ x) ୅ཧ໰୊ subject to h(x + a) = + 1 ୅ཧؔ਺ͷಋೖ
  19. 2025/10/17 K. Kanamori, Fujitsu Ltd. ϚϋϥϊϏεڑ཭ͷ MILO ఆࣜԽ (1/2) 22

    ϚϋϥϊϏεڑ཭͸ಛ௃ྔؒͷ૬ؔؔ܎Λߟྀͨ͠ڑ཭ؔ਺ ಉҰ෼෍ʹै͏ϕΫτϧ ʹର͢ΔϚϋϥϊϏεڑ཭͸ҎԼͰఆٛ: ͜͜Ͱ ͸ڞ෼ࢄߦྻ (൒ਖ਼ఆ஋ߦྻ). x, x′  ∈ ℝD dM (x, x′  ∣ Σ−1) = (x′  − x)⊤Σ−1(x′  − x) Σ ∈ ℝD×D ϚϋϥϊϏεڑ཭ (Mahalanobis Distance; MD) [Mahalanobis 36] ๏ ಛ௃ྔؒͷεέʔϧࠩͱ૬ؔؔ܎Λߟྀͨ͠ڑ཭ؔ਺ ✓ ೋ࣍ܗ߲ࣜͷఆࣜԽʹ ݸ ͷิॿม਺ͱ੍໿͕ࣜඞཁ 𝒪 (I2) (I ≫ D)
  20. 2025/10/17 K. Kanamori, Fujitsu Ltd. ํ਑ ൒ਖ਼ఆ஋ߦྻͷ෼ղ ʹ ج͍ͮͯۙࣅ͢Δ -MD

    Λಋೖ → ‣ ݸͷิॿม਺ͱ੍໿ࣜͰදݱՄೳ Σ−1 = U⊤U ℓ1 d2 M (x, x + a ∣ Σ−1) = ∥Ua∥2 2 = ∑ D d=1 (U⊤ d a) 2 ̂ dM (x, x + a ∣ Σ−1) := ∥Ua∥1 = ∑ D d=1 U⊤ d a 𝒪 (D) ϚϋϥϊϏεڑ཭ͷ MILO ఆࣜԽ (2/2) 23 ൒ਖ਼ఆ஋ߦྻͷ෼ղʹجͮۙ͘ࣅʹΑΓޮ཰Α͘ఆࣜԽ dM (x, x′  ∣ Σ−1) = (x′  − x)⊤Σ−1(x′  − x) ϚϋϥϊϏεڑ཭ (MD) MD dM -MD ℓ1 ̂ dM ‣ۙࣅ཰ D
  21. 2025/10/17 K. Kanamori, Fujitsu Ltd. LOF ͷ MILO ఆࣜԽ (1/2)

    24 LOF ͸ۙ๣఺ू߹ͷີ౓ൺʹجͮ͘֎Ε஋ݕग़είΞ ๏ -ۙ๣఺ू߹ ʹ͓͚Δฏۉີ౓ൺʹجͮ͘֎Ε஋ݕग़είΞ ✓ -ۙ๣఺ू߹ ͷఆࣜԽʹ ݸͷิॿม਺͕ඞཁ k Nk (x) k Nk (x + a) 𝒪 (N2) ڑ཭ۭؒ ্ͷ఺ू߹ ʹ͓͚Δ -LOF ͸ҎԼͰఆٛ: ͜͜Ͱ ͸ -ۙ๣఺ू߹, ͸ -ہॴີ౓. (ℝD, d) X = {x1 , …xN } k qk (x ∣ X) = 1 k ∑x′  ∈Nk (x) lrdk (x′  ) lrdk (x) Nk (x) ⊆ X k lrdk (x) = k−1 ⋅ Σx′  ∈Nk (x) d−1(x, x′  ) k Local Outlier Factor (LOF) [Breunig+ 00]
  22. 2025/10/17 K. Kanamori, Fujitsu Ltd. LOF ͷ MILO ఆࣜԽ (2/2)

    25 -ۙ๣఺ू߹Λ Ͱݻఆ͢Δ͜ͱͰޮ཰Α͘ఆࣜԽ k k = 1 ํ਑ -ۙ๣఺ू߹Λ Ͱݻఆ ‣ ݸͷิॿม਺ͰදݱՄೳ k k = 1 𝒪 (N) qk (x ∣ X) = 1 k ∑x′  ∈Nk (x) lrdk (x′  ) lrdk (x) Local Outlier Factor (LOF) x x + a x + a a q1 (x + a ∣ X) ≫ 1 q1 (x + a ∣ X) ≈ 1 (x∘ := arg minx′  ∈X d(x + a, x′  )) q1 (x + a ∣ X) = lrd1 (x∘) lrd1 (x + a) = ∑ x′  ∈X lrd1 (x′  ) ⋅ d(x + a, x′  ) ⋅ 𝕀 [x′  = x∘] ఆ਺߲ ૒ઢܗܗࣜ (Big-M ๏ͰදݱՄೳ) = lrd1 (x∘) ⋅ d(x + a, x∘)
  23. 2025/10/17 K. Kanamori, Fujitsu Ltd. MILO ໰୊ͱͯ͠ͷఆࣜԽ 26 ୅ཧؔ਺ͷಋೖʹΑΓม਺ͱ੍໿ࣜͷ૯਺ͷ࡟ݮʹ੒ޭ Overall

    Formulation minimize ∑ D d=1 δd + λ ⋅ ∑ N n=1 ln ⋅ ρn ∑ D d=1 ∑ Id i=1 (cn,d,i − cn′  ,d,i) πd,i ≤ Cn (1 − νn ), ∀n, n′  ∈ [N] ρn ≥ dn ⋅ νn , ∀n ∈ [N] ρn ≥ ∑ D d=1 ∑ Id i=1 cn,d,i πd,i − Cn (1 − νn ), ∀n ∈ [N] ∑ N n=1 νn = 1 πd,i ∈ {0,1}, ∀d ∈ [D], i ∈ [Id ] δd ≥ 0,∀d ∈ [D] νn ∈ {0,1}, ρn ≥ 0,∀n ∈ [N] subject to ∑ Id i=1 πd,i = 1,∀d ∈ [D] ∑ D d=1 wd (xd + ∑ Id i=1 ad,i ⋅ πd,i) ≥ 0 −δd ≤ ∑ D d′  =1 Ud,d′  ∑ Id′  i=1 ad′  ,i πd,i ≤ δd , ∀d ∈ [D] ม਺ͱ੍໿ࣜͷ૯਺Λ ࡟ݮ͠ٻղͷߴ଎Խʹ੒ޭʂ 𝒪 (N2 + I2) 𝒪 (N2 + I2) 𝒪 (N2 + D) 𝒪 (N + I) Exact Proposed #Variables #Constraints : σʔλ਺ɹ : ಛ௃ྔ਺ : ΞΫγϣϯ૯਺ ( ) N D I I ≫ D ‣ MILO ໰୊ͳͷͰ਺ཧܭըιϧόʔΛ༻͍ͯٻղՄೳ
  24. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ࣮ݧʳطଘख๏ͱͷൺֱ 27 ಛ௃ྔ૬ؔͱ֎Ε஋ϦεΫΛߟྀͨ͠ΞΫγϣϯΛఏࣔͰ͖ͨ Dataset Method

    Logistic Regression Random Forest Multilayer Perceptron MD 10-LOF MD 10-LOF MD 10-LOF FICO TLPS 5.44 ± 4.6 1.49 ± 1.0 2.12 ± 1.1 1.33 ± 0.63 4.25 ± 4.2 1.97 ± 1.3 MAD 8.49 ± 5.8 1.49 ± 1.1 2.11 ± 1.2 1.39 ± 0.73 1.02 ± 0.79 1.42 ± 0.61 PCC 9.34 ± 5.1 1.5 ± 1.1 2.45 ± 1.4 1.29 ± 0.26 8.11 ± 1.1e+01 1.5 ± 0.96 DACE 2.5 ± 1.5 1.32 ± 0.43 1.9 ± 0.97 1.24 ± 0.23 0.819 ± 0.72 1.4 ± 0.5 German TLPS 6.61 ± 2.0 1.23 ± 0.39 8.25 ± 1.9 1.23 ± 0.38 9.1 ± 0.37 1.27 ± 0.47 MAD 9.51 ± 4.9 1.24 ± 0.4 9.02 ± 3.1 1.48 ± 0.86 9.95 ± 2.9 1.29 ± 0.55 PCC 6.52 ± 2.5 1.23 ± 0.4 8.07 ± 3.3 1.53 ± 0.92 11.9 ± 4.0 1.29 ± 0.55 DACE 3.76 ± 1.1 1.18 ± 0.34 3.91 ± 1.1 1.09 ± 0.23 3.73 ± 1.1 1.04 ± 0.24 WineQuality TLPS 2.03 ± 1.6 4.54 ± 11 1.3 ± 0.89 1.55 ± 1.1 1.45 ± 1.5 1.82 ± 2.4 MAD 2.06 ± 1.7 1.44 ± 0.99 1.3 ± 0.94 1.46 ± 1.0 1.09 ± 0.97 1.77 ± 2.4 PCC 8.43 ± 5.5 1.45 ± 0.98 7.27 ± 4.1 1.47 ± 0.98 4.94 ± 3.1 1.43 ± 0.62 DACE 1.18 ± 0.92 1.33 ± 0.81 0.84 ± 0.5 1.39 ± 0.86 0.919 ± 0.69 1.35 ± 0.66 Diabetes TLPS 26.4 ± 13 5.34 ± 4.1 1.71 ± 1.5 1.24 ± 0.28 26.4 ± 15 6.46 ± 5.8 MAD 3.45 ± 2.9 1.33 ± 0.42 1.6 ± 1.3 1.27 ± 0.44 3.38 ± 2.6 1.36 ± 0.4 PCC 5.12 ± 3.6 1.24 ± 0.4 4.09 ± 3.0 1.28 ± 0.44 6.97 ± 3.9 1.27 ± 0.45 DACE 1.1 ± 0.89 1.18 ± 0.24 0.963 ± 0.66 1.22 ± 0.28 1.11 ± 0.83 1.18 ± 0.22 ‣ ఏҊख๏ (DACE) ͸طଘख๏ͱൺֱͯ͠࠷΋ྑ͍ϚϋϥϊϏεڑ཭ͱ 10-LOF Λୡ੒
  25. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ࣮ݧʳطଘख๏ͱͷൺֱ 28 ಛ௃ྔ૬ؔͱ֎Ε஋ϦεΫΛߟྀͨ͠ΞΫγϣϯΛఏࣔͰ͖ͨ Dataset Method

    Logistic Regression Random Forest Multilayer Perceptron MD 10-LOF MD 10-LOF MD 10-LOF FICO TLPS 5.44 ± 4.6 1.49 ± 1.0 2.12 ± 1.1 1.33 ± 0.63 4.25 ± 4.2 1.97 ± 1.3 MAD 8.49 ± 5.8 1.49 ± 1.1 2.11 ± 1.2 1.39 ± 0.73 1.02 ± 0.79 1.42 ± 0.61 PCC 9.34 ± 5.1 1.5 ± 1.1 2.45 ± 1.4 1.29 ± 0.26 8.11 ± 1.1e+01 1.5 ± 0.96 DACE 2.5 ± 1.5 1.32 ± 0.43 1.9 ± 0.97 1.24 ± 0.23 0.819 ± 0.72 1.4 ± 0.5 German TLPS 6.61 ± 2.0 1.23 ± 0.39 8.25 ± 1.9 1.23 ± 0.38 9.1 ± 0.37 1.27 ± 0.47 MAD 9.51 ± 4.9 1.24 ± 0.4 9.02 ± 3.1 1.48 ± 0.86 9.95 ± 2.9 1.29 ± 0.55 PCC 6.52 ± 2.5 1.23 ± 0.4 8.07 ± 3.3 1.53 ± 0.92 11.9 ± 4.0 1.29 ± 0.55 DACE 3.76 ± 1.1 1.18 ± 0.34 3.91 ± 1.1 1.09 ± 0.23 3.73 ± 1.1 1.04 ± 0.24 WineQuality TLPS 2.03 ± 1.6 4.54 ± 11 1.3 ± 0.89 1.55 ± 1.1 1.45 ± 1.5 1.82 ± 2.4 MAD 2.06 ± 1.7 1.44 ± 0.99 1.3 ± 0.94 1.46 ± 1.0 1.09 ± 0.97 1.77 ± 2.4 PCC 8.43 ± 5.5 1.45 ± 0.98 7.27 ± 4.1 1.47 ± 0.98 4.94 ± 3.1 1.43 ± 0.62 DACE 1.18 ± 0.92 1.33 ± 0.81 0.84 ± 0.5 1.39 ± 0.86 0.919 ± 0.69 1.35 ± 0.66 Diabetes TLPS 26.4 ± 13 5.34 ± 4.1 1.71 ± 1.5 1.24 ± 0.28 26.4 ± 15 6.46 ± 5.8 MAD 3.45 ± 2.9 1.33 ± 0.42 1.6 ± 1.3 1.27 ± 0.44 3.38 ± 2.6 1.36 ± 0.4 PCC 5.12 ± 3.6 1.24 ± 0.4 4.09 ± 3.0 1.28 ± 0.44 6.97 ± 3.9 1.27 ± 0.45 DACE 1.1 ± 0.89 1.18 ± 0.24 0.963 ± 0.66 1.22 ± 0.28 1.11 ± 0.83 1.18 ± 0.22 ‣ ఏҊख๏ (DACE) ͸طଘख๏ͱൺֱͯ͠࠷΋ྑ͍ϚϋϥϊϏεڑ཭ͱ 10-LOF Λୡ੒ 55 60 65 70 75 80 85 90 ExternalRiskEstimate 0 20 40 60 80 100 PercentInstallTrades TLPS DACE (ours) 0 50 100 150 200 250 MSinceOldestTradeOpen 0 20 40 60 80 100 AverageMInFile TLPS DACE (ours) ఏҊख๏ ఏҊख๏ طଘख๏ طଘख๏
  26. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ·ͱΊʳσʔλ෼෍Λߟྀͨ͠ΞΫγϣϯઆ໌ 29 ಛ௃ྔؒͷ૬ؔͱ֎Ε஋ϦεΫΛߟྀͨ͠ΞΫγϣϯઆ໌ΛఏҊ 1. ϚϋϥϊϏεڑ཭ͱ

    Local Outlier Factor ʹجͮ͘ίετؔ਺Λಋೖ ‣ ಛ௃ྔؒͷ૬ؔͱ֎Ε஋ϦεΫΛߟྀͯ͠ΞΫγϣϯͷ࣮ݱՄೳੑΛධՁ 2. ಋೖͨ͠ίετؔ਺ʹରͯ͠ MILO ʹجͮ͘ޮ཰ྑ͍࠷దԽํ๏ΛఏҊ ‣ ม਺ͱ੍໿ࣜͷ૯਺Λେ෯ʹ࡟ݮ͠ߴ଎Խ 3. ࣮σʔλ্ͷܭࢉػ࣮ݧͰ༗ޮੑΛ֬ೝ ‣ ΑΓ࣮ݱՄೳੑͷߴ͍ΞΫγϣϯͷఏࣔʹ੒ޭ 0 50 100 150 200 250 MSinceOldestTradeOpen 0 20 40 60 80 100 AverageMInFile TLPS DACE (ours) ఏҊख๏ طଘख๏
  27. 2025/10/17 K. Kanamori, Fujitsu Ltd. 30 ʲ༨ஊʳͦͷଞͷݚڀ੒Ռ (1/2) ஞ࣍త൓࣮Ծ૝આ໌ (PRICAI-24)

    ‣ ΞΫγϣϯͷ్தܦաΛදݱ͢Δ ಛ௃ྔύεΛ࠷దԽͯ͠ఏࣔ طଘ ఏҊ طଘख๏ͱൺֱͯ͠׈Β͔ͳύεΛಘΒΕͨʂ S. Yamao, K. Kobayashi, K. Kanamori, T. Takagi, Y. Ike, K. Nakata: Distribution-Aligned Sequential Counterfactual Explanation with Local Outlier Factor. PRICAI2024. ॱং෇͖൓࣮Ծ૝આ໌ (AAAI-21) ‣ ಛ௃ྔؒͷҼՌؔ܎Λߟྀͯ͠ ΞΫγϣϯͷ࣮ߦॱং΋࠷దԽ Education JobSkill Income WorkPerDay HealthStatus 1 Order Feature Value 2 3 HealthStatus WorkPerDay Income +3 +1 +4 ҼՌDAG ॱং෇͖ΞΫγϣϯ K. Kanamori, T. Takagi, K. Kobayashi, Y. Ike, K. Uemura, H. Arimura: Ordered Counterfactual Explanation by Mixed-Integer Linear Optimization. AAAI2021.
  28. 2025/10/17 K. Kanamori, Fujitsu Ltd. 31 ʲ༨ஊʳͦͷଞͷݚڀ੒Ռ (2/2) ௕ظతվળͷͨΊͷΞϧΰϦζϜతঈؐ (ICML-25)

    ‣ Ϟσϧͷ༧ଌ݁Ռ͚ͩͰͳ࣮͘ੈքͰͷ݁Ռ΋վળ͢ΔΞΫγϣϯΛఏࣔ͢Δ K. Kanamori, K. Kobayashi, S. Hara, T. Takagi: Algorithmic Recourse for Long-Term Improvement. ICML2025. For each round , 1. Receive an instance and candidate valid actions 2. Suggest an action based on the past observations 3. Sample a reward and delay , where is the probability that executes and 4. Observe feedback on the past rewards Goal Maximize the mean expected reward t = 1,2,…, T xt 𝒜 t at ∈ 𝒜 t Rt ∼ ℬ(pt ) Dt ∼ 𝒟 pt xt at h*(xt + at ) = y* {Rs ∣ s + Ds = t}t−1 s=1 RT = 1 T ∑ T t=1 𝔼 [Rt ] Problem 1. (AR for Long-Term Improvement; ARLIM) or repayment default Outcome Instance (disagree to execute) Instance (agree to execute) AR Agent No feedback Action Suggestion Delayed Feedback
  29. 33 Learning Decision Trees and Forests with Algorithmic Recourse Kentaro

    Kanamori1, Takuya Takagi1, Ken Kobayashi2, Yuichi Ike3 1Fujitsu Limited, 2Tokyo Institute of Technology, 3Kyushu University 41st International Conference on Machine Learning (ICML2024) Spotlight Paper
  30. 2025/10/17 K. Kanamori, Fujitsu Ltd. ՝୊ 34 ͦ΋ͦ΋Ϣʔβ͕࣮ݱՄೳͳΞΫγϣϯ͕ଘࡏ͢Δͱ͸ݶΒͳ͍ ϩʔϯਃ੥ऀ X͞Μ

    (48ࡀ / ݄ऩ50ສԁ) :FT /P :FT /P :FT /P :FT /P ೥ྸ ≤ 33ࡀ ݄ऩ ≤ 28ສԁ ݄ऩ ≤ 54ສԁ ঝೝ ঝೝ ൱ೝ ൱ೝ ൱ೝ ೥ྸ ≤ 42ࡀ ೥ྸ ݄ऩ ‣ ֶशࡁΈϞσϧʹର࣮ͯ͠ݱՄೳͳΞΫγϣϯ͕ଘࡏ͢Δอূ͸ͳ͍ ঝೝ͞ΕΔʹ͸ Ͳ͏ͨ͠Β͍͍ʁ ೥ྸΛݮΒ͞ͳ͚Ε͹ঝೝ͞Εͳ͍ (࣮ߦෆՄೳ) -15ࡀ -6ࡀ & +4ສԁ ۜߦͷ༩৴৹ࠪϞσϧ (ver. 1.0) ༧ଌਫ਼౓: 90%
  31. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ໨ඪʳΞΫγϣϯ഑ྀܕֶश 35 ΞΫγϣϯͷଘࡏΛอূͭͭ͠ߴਫ਼౓ͳϞσϧΛֶश͍ͨ͠ ϩʔϯਃ੥ऀ X͞Μ

    (48ࡀ / ݄ऩ50ສԁ) ೥ྸ ݄ऩ ݄ऩΛ૿΍ͤ͹ঝೝ͞ΕΔ (࣮ߦՄೳ) ۜߦͷ༩৴৹ࠪϞσϧ (ver. 2.0) +6ສԁ ໨ඪ ༧ଌਫ਼౓Λҡ࣮࣋ͭͭ͠ݱՄೳΞΫγϣϯΛߴ֬཰ͰอূͰ͖ΔϞσϧΛֶश͢Δ :FT /P :FT /P :FT /P :FT /P ೥ྸ ≤ 33ࡀ ݄ऩ ≤ 28ສԁ ݄ऩ ≤ 56ສԁ ঝೝ ঝೝ ൱ೝ ൱ೝ ൱ೝ ೥ྸ ≤ 50ࡀ ঝೝ͞ΕΔʹ͸ Ͳ͏ͨ͠Β͍͍ʁ ༧ଌਫ਼౓: 88%
  32. 2025/10/17 K. Kanamori, Fujitsu Ltd. ܾఆ໦ɾܾఆ໦Ξϯαϯϒϧ 36 දܗࣜσʔλͷ༧ଌλεΫʹ͓͚ΔσϑΝΫτελϯμʔυϞσϧ ‣࣮ࣾձͰසग़͢Δදܗࣜσʔλʹରͯ͠ ߴ͍ਫ਼౓ͱֶशޮ཰Λ࣋ͭ

    [Grinsztajn+ 22] ܾఆ໦ [Breiman+ 84] “if-then-else” ܗࣜͷ༧ଌϧʔϧΛ ೋ෼໦Ͱදݱͨ͠ղऍՄೳͳϞσϧ ܾఆ໦ΞϯαϯϒϧϞσϧ ෳ਺ͷܾఆ໦ͷ߹ٞͰ༧ଌ஋Λܾఆ :FT /P :FT /P :FT /P ໨త = ৽ंߪೖ ݄ऩ ≤ 50ສԁ ঝೝ ঝೝ ൱ೝ ൱ೝ ֶྺ ≤ ֶ࢜ … Ξϯαϯϒϧ ܾఆ໦1 ܾఆ໦2 ܾఆ໦T
  33. 2025/10/17 K. Kanamori, Fujitsu Ltd. ঈؐଛࣦ (recourse loss) 37 ϞσϧͷΞΫγϣϯอূ཰ͷධՁࢦඪͱͯ͠ঈؐଛࣦΛಋೖ

    ‣ ࣮ߦՄೳ͔࣮ͭߦίετ͕ ҎԼͰ͋ΔΞΫγϣϯ͕ଘࡏ͠ͳׂ͍߹ ε ίετ༧ࢉ෇͖ΞΫγϣϯू߹Λ ͱ͠, ঈؐଛࣦΛ ͰఆΊΔ ( : 0-1ଛࣦؔ਺). ·ͨ, αϯϓϧ ʹରͯ͠, ܦݧঈؐଛࣦΛҎԼͰఆٛ: 𝒜 ε (x) := {a ∈ 𝒜 (x) ∣ c(a ∣ x) ≤ ε} lrec (x; h) := mina∈ 𝒜 ε (x) l01 (+1,h(x + a)) l01 S = {(xn , yn )}N n=1 ̂ Ωε (h) := 1 N ∑ N n=1 lrec (xn ; h) ঈؐଛࣦ [Ross+ 21]
  34. 2025/10/17 K. Kanamori, Fujitsu Ltd. ֶश໰୊ͷఆࣜԽ 38 ܦݧঈؐଛࣦͷ੍໿ԼͰܦݧଛࣦΛ࠷খԽ͢Δܾఆ໦Λֶश Recourse-Aware Classification

    Tree (RACT) αϯϓϧ ͱύϥϝʔλ ʹରͯ͠, ࣍ͷ໰୊ͷ࠷దղͱͳΔܾఆ໦ ΛٻΊΔ: ͜͜Ͱ, ͸ܾఆ໦ͷू߹, ͸ܦݧଛࣦ. S = {(xn , yn )}N n=1 ⊆ 𝒳 × {±1} δ, ε > 0 h*: 𝒳 → {±1} minh∈ℋ ̂ R(h) subject to ̂ Ωε (h) ≤ δ ℋ ̂ R(h) = 1 N ∑ N n=1 l01 (yn , h(xn )) ‣ Ҏ্ͷׂ߹ͰΞΫγϣϯΛอূͭͭ͠ਫ਼౓Λ࠷େԽ 100 ⋅ (1 − δ) %
  35. 2025/10/17 K. Kanamori, Fujitsu Ltd. ֶशΞϧΰϦζϜͷ֓ཁ 39 “ᩦཉ๏ʹΑΔ෼ذ৚݅ͷֶश” ͱ “༧ଌϥϕϧΛमਖ਼͢Δޙॲཧ”

    • ঈؐଛࣦͷ੍໿͕ͳ͍৔߹Ͱ΋ɼܾఆ໦ͷݫີֶश͸ܭࢉྔతʹࠔ೉ ‣ ᩦཉ๏ʹجͮ͘τοϓμ΢ϯܕΞϧΰϦζϜ͕ओྲྀ (ྫ: CART [Breiman+ 84]) Step 1. Step 2. • ఏҊΞϧΰϦζϜ͸ 2 ͭͷεςοϓͰߏ੒: 1. ܦݧଛࣦ ͱܦݧঈؐଛࣦ Λߟྀͯ͠ தؒϊʔυͷ෼ذ৚݅Λτοϓμ΢ϯʹܾఆ 2. ֶशܾͨ͠ఆ໦ ੍͕໿ Λ ຬͨ͢Α͏ʹ༿ϊʔυͷ༧ଌϥϕϧΛฤू ̂ R ̂ Ωε h ̂ Ωε (h) ≤ δ
  36. 2025/10/17 K. Kanamori, Fujitsu Ltd. τοϓμ΢ϯܕֶशΞϧΰϦζϜ (1/3) 40 ܦݧঈؐଛࣦΛߟྀͨ͠෼ذ৚݅ͷܾఆ໰୊ΛఆࣜԽ •

    ݱࡏͷܾఆ໦ ͷ༿ϊʔυ ʹ৽͍͠෼ذ৚݅ Λ௥Ճ͢Δ ‣ ࠷΋ྑ͍෼ذ৚݅ ͱࢠϊʔυͷ༧ଌϥϕϧ ΛܾΊ͍ͨ • ෼ذ৚݅Λ௥Ճܾͨ͠ఆ໦Λ ͱͯ͠, ҎԼͷ෼ׂ఺ܾఆ໰୊Λղ͘: h i xd ≤ b (d, b) ̂ yL , ̂ yR ∈ {±1} h′  min d,b min ̂ yL , ̂ yR Φλ (d, b, ̂ yL , ̂ yR ) := ̂ R(h′  ) + λ ⋅ ̂ Ωε (h′  ) ੍໿ ͷ؇࿨ ̂ Ωε (h′  ) ≤ δ ̂ yi xd ≤ b ̂ yL ̂ yR ܾఆ໦ h ܾఆ໦ h′ 
  37. 2025/10/17 K. Kanamori, Fujitsu Ltd. τοϓμ΢ϯܕֶशΞϧΰϦζϜ (2/3) 41 ܦݧଛࣦͱܦݧঈؐଛࣦΛߟྀͯ͠෼ذ৚݅Λ࠶ؼతʹܾఆ min

    ̂ yL , ̂ yR Φ1 (d1 , b1 , ̂ yL , ̂ yR ) = 3/6 + 0/6 = 3/6 ೥ྸ ݄ऩ ೥ྸ ݄ऩ min ̂ yL , ̂ yR Φ1 (d2 , b2 , ̂ yL , ̂ yR ) = 1/6 + 0/6 = 1/6 ೥ྸ ݄ऩ min ̂ yL , ̂ yR Φ1 (d3 , b3 , ̂ yL , ̂ yR ) = 1/6 + 2/6 = 3/6 ‣ ෼ׂͨ͠αϯϓϧʹରͯ͠΋࠶ؼతʹ෼ذ৚݅Λܾఆ͠໦ߏ଄Λֶश ͷͱ͖ h(x) = + 1 lrec (x; h) = 0 ೥ྸ ݄ऩ
  38. 2025/10/17 K. Kanamori, Fujitsu Ltd. τοϓμ΢ϯܕֶशΞϧΰϦζϜ (3/3) 42 ࣌ؒܭࢉྔ͸ඪ४తͳܾఆ໦ֶशΞϧΰϦζϜͱมΘΒͳ͍ •

    ௨ৗͷ෼ׂ఺ܾఆ໰୊ ( ) ͸ ࣌ؒܭࢉྔ Ͱղ͚Δ • Ұํ, ܦݧঈؐଛࣦ Λߟྀ͢Δ৔߹ ( ) ͸ඇࣗ໌…… λ = 0 𝒪 (D ⋅ N) ̂ Ωε λ > 0 min d,b min ̂ yL , ̂ yR Φλ (d, b, ̂ yL , ̂ yR ) := ̂ R(h′  ) + λ ⋅ ̂ Ωε (h′  ) ෼ׂ఺ܾఆ໰୊ ఆཧ (ఏҊΞϧΰϦζϜͷ࣌ؒܭࢉྔ) ఏҊΞϧΰϦζϜ͸෼ׂ఺ܾఆ໰୊ ( ) Λ࣌ؒܭࢉྔ Ͱղ͘ λ > 0 𝒪 (D ⋅ N) ‣ ద੾ͳલॲཧΛߦ͑͹, ܦݧଛࣦ ͱಉ༷ʹ ΋ͳΒ͠ఆ਺࣌ؒͰܭࢉՄೳ ̂ R ̂ Ωε
  39. 2025/10/17 K. Kanamori, Fujitsu Ltd. ू߹ඃ෴ʹجͮ͘༧ଌϥϕϧฤू (1/2) 43 ܦݧঈؐଛࣦʹؔ͢Δ੍໿Λຬͨ͢Α͏ʹ༧ଌϥϕϧΛमਖ਼ •

    ֶशܾͨ͠ఆ໦ ੍͕໿ Λඞͣ͠΋ຬͨ͢ͱ͸ݶΒͳ͍ h ̂ Ωε (h) ≤ δ ̂ y3 ̂ y2 ̂ y1 ̂ y4 ܾఆ໦ h ̂ R(h) = 0.18| ̂ Ωε (h) = 0.12 • ֶशࡁΈ෼ذ৚݅ͱ໦ߏ଄Λݻఆͯ͠ ༿ϊʔυͷ༧ଌϥϕϧͷΈΛमਖ਼͢Δ ҎԼͷ༧ଌϥϕϧฤू໰୊Λղ͘: ‣ ੍໿Λຬͨ͢༿ϊʔυमਖ਼ํ๏ͷதͰ ܦݧଛࣦ͕࠷খͷ΋ͷΛٻΊΔ minϕ∈{±1}I ̂ R(hϕ ) subject to ̂ Ωε (hϕ ) ≤ δ ̂ y3 ̂ y2 ̂ y1 ̂ y4 ̂ R(hϕ ) = 0.23 ∣ ̂ Ωε (hϕ ) = 0.05 ܾఆ໦ hϕ
  40. 2025/10/17 K. Kanamori, Fujitsu Ltd. ू߹ඃ෴ʹجͮ͘༧ଌϥϕϧฤू (2/2) 44 ࠷খඃ෴໰୊ʹؼண͞Εɼۙࣅ཰อূ͖ͭͰޮ཰Α͘ղ͚Δ •

    ༧ଌϥϕϧΛมߋ͢Δ༿ϊʔυͷ෦෼ू߹ΛબͿ૊߹ͤ࠷దԽ໰୊ • ॏཁͳ؍࡯ ܾఆ໦ͷܦݧঈؐଛࣦ͸Πϯελϯεू߹্ͷඃ෴ؔ਺ͰදݱՄೳ ‣ ۙࣅ཰อূ෇͖ͷଟ߲ࣜ࣌ؒΞϧΰϦζϜ͕ଘࡏ (ྫ: ᩦཉ๏ [Kearns 90]) ఆཧ (༧ଌϥϕϧฤू໰୊ͷؼண) ༧ଌϥϕϧฤू໰୊͸ॏΈ෇͖෦෼ඃ෴໰୊ʹؼண͞ΕΔ
  41. 2025/10/17 K. Kanamori, Fujitsu Ltd. ঈؐଛࣦͷ PAC ղੳ 45 ςετσʔλʹର͢Δঈؐଛࣦʹ΋֬཰తอূΛ༩͑Δ͜ͱ͕Մೳ

    ظ଴ঈؐଛࣦΛ ͰఆΊΔ. ೚ҙͷ , ෼ྨث , αϯϓϧ ʹରͯ͠, ֬཰ ͰҎԼ͕੒ཱ: Ωε (h) := ℙx [∃a ∈ 𝒜 ε (x) : h(x + a) = + 1] α > 0 h ∈ ℋ S 1 − α Ωε (h) ≤ ̂ Ωε (h) + ln|ℋ| − ln α 2 ⋅ |S| ఆཧ (ঈؐଛࣦͷ PAC ղੳ) ‣ ϥϕϧฤू໰୊ͷ Λ ʹஔ͖׵Δ͜ͱͰ ςετσʔλʹର͢ΔΞΫγϣϯอূ཰΋੍ޚͰ͖Δ δ δ − (I ⋅ ln 2 − ln α)/(2 ⋅ N)
  42. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ࣮ݧʳϕʔεϥΠϯͱͷൺֱ 46 ਫ਼౓ͱ଎౓Λେ͖͘ଛͳΘͣʹߴ͍ΞΫγϣϯอূ཰Λ࣮ݱ ‣ ఏҊख๏

    (RACT) ͸༧ଌਫ਼౓ͱֶश଎౓Λҡ࣋ͭͭ͠ΞΫγϣϯอূ཰ͷ޲্ʹ੒ޭ Vanilla (0.711) OAF (0.689) RACT (0.714) 0.68 0.70 0.72 Accuracy FICO Vanilla (0.669) OAF (0.666) RACT (0.644) 0.66 0.68 Accuracy COMPAS Vanilla (0.807) OAF (0.779) RACT (0.804) 0.78 0.80 Accuracy Credit Vanilla (0.693) OAF (0.65) RACT (0.677) 0.650 0.675 0.700 Accuracy Bail Vanilla (0.757) OAF (0.889) RACT (0.973) 0.6 0.8 1.0 Recourse FICO Vanilla (0.842) OAF (0.841) RACT (0.941) 0.8 0.9 1.0 Recourse COMPAS Vanilla (0.956) OAF (1.0) RACT (0.982) 0.96 0.98 1.00 Recourse Credit Vanilla (0.32) OAF (0.455) RACT (0.788) 0.2 0.4 0.6 0.8 Recourse Bail Vanilla (0.483) OAF (0.324) RACT (0.429) 0.3 0.4 0.5 Time [s] FICO Vanilla (0.0979) OAF (0.0428) RACT (0.0358) 0.05 0.10 Time [s] COMPAS Vanilla (1.83) OAF (1.37) RACT (1.59) 1.4 1.6 1.8 2.0 Time [s] Credit Vanilla (0.337) OAF (0.0545) RACT (0.2) 0.1 0.2 0.3 Time [s] Bail ΞΫγϣϯอূ཰ ༧ଌਫ਼౓ ࣮ߦ࣌ؒ
  43. 2025/10/17 K. Kanamori, Fujitsu Ltd. ʲ·ͱΊʳΞΫγϣϯ഑ྀܕͷܾఆ໦ֶश 47 ༧ଌਫ਼౓ͱΞΫγϣϯอূΛཱܾ྆ͨ͠ఆ໦ͷֶशํ๏ΛఏҊ 1. ΞΫγϣϯอূ཰Λߟྀܾͨ͠ఆ໦ͷτοϓμ΢ϯܕֶश๏ΛఏҊ

    ‣ ࣌ؒܭࢉྔ͸ඪ४తͳܾఆ໦ֶशΞϧΰϦζϜͱಉ౳ 2. ΞΫγϣϯอূ཰ͷ੍໿Λຬͨ͢Α͏ʹ ֶशࡁΈܾఆ໦Λฤू͢Δޙॲཧ๏ΛఏҊ ‣ ඃ෴໰୊΁ͷؼணʹΑΓۙࣅ཰อূ෇͖Ͱղ͚Δ 3. ࣮σʔλ্ͷܭࢉػ࣮ݧͰ༗ޮੑΛ֬ೝ ‣ ਫ਼౓ͱ଎౓ΛଛͳΘͣʹΞΫγϣϯอূ཰Λ޲্ 0.67 0.68 0.69 0.70 0.71 0.72 0.73 0.5 0.6 0.7 0.8 0.9 1.0 Baseline 1 Baseline 2 RACT (ours) ΞΫγϣϯอূ཰ ༧ଌਫ਼౓ ఏҊख๏
  44. 2025/10/17 K. Kanamori, Fujitsu Ltd. 48 ʲ༨ஊʳͦͷଞͷݚڀ੒Ռ (1/2) ΞϧΰϦζϜతঈؐΛߟྀͨ͠ޯ഑ϒʔεςΟϯά໦ (NeurIPS-25)

    ‣ ༧ଌਫ਼౓ͱΞΫγϣϯอূΛཱ྆ͨ͠ޯ഑ϒʔεςΟϯά໦ (GBDT) Λֶश͢Δ ՝୊ɹঈؐଛࣦͷඍ෼ෆՄೳੑ Ұൠతͳ GBDT ͷ࿮૊ΈͰ͸ରԠෆՄೳ ղܾࡦɹঈؐଛࣦͷඍ෼Մೳͳ্քͷಋग़ ඪ४తͳֶश๏ͱಉ࣌ؒ͡ܭࢉྔͰֶशՄೳ K. Kanamori, K. Kobayashi, T. Takagi: Learning Gradient Boosted Decision Trees with Algorithmic Recourse. NeurIPS2025. Vanilla OAF RABIT 0.70 0.72 0.74 Accuracy FICO - Accuracy Vanilla OAF RABIT 0.6 0.8 Recourse FICO - Recourse Vanilla OAF RABIT 0.66 0.68 0.70 Accuracy COMPAS - Accuracy Vanilla OAF RABIT 0.80 0.85 0.90 0.95 Recourse COMPAS - Recourse 0.83 0.84 0.85 0.86 Accuracy 0.3 0.4 Recourse ༧ଌਫ਼౓Λҡ࣋ͭͭ͠ΞΫγϣϯอূ཰Λ޲্ʂ
  45. 2025/10/17 K. Kanamori, Fujitsu Ltd. 49 ʲ༨ஊʳͦͷଞͷݚڀ੒Ռ (2/2) :FT /P

    :FT /P ࢒ۀ = ༗ ۀ੷ ≥ " ࢒ۀ ༗ → ແ ෦ॺ Ӧۀ → ਓࣄ ೥ऩ + 12K$ ൓࣮Ծ૝આ໌໦ Πϯελϯεू߹ Ϟσϧ $& ཭৬༧๷ͷͨΊͷ ࢪࡦཱҊʹ׆༻Մೳ ൓࣮Ծ૝આ໌໦ (AISTATS-22) ‣ Πϯελϯεू߹ʹର͢ΔΞΫγϣϯΛܾఆ໦ͰେҬతʹཁ໿ͯ͠ղऍՄೳʹ ྫʣ཭৬ϦεΫ༧ଌ K. Kanamori, T. Takagi, K. Kobayashi, Y. Ike: Counterfactual Explanation Trees: Transparent and Consistent Actionable Recourse with Decision Trees. AISTATS2022. ͲͷΑ͏ͳΫϥελʹ ͲͷΑ͏ͳΞΫγϣϯ͕༗ޮ͔͕ ղऍՄೳʹ
  46. 2025/10/17 K. Kanamori, Fujitsu Ltd. 50 ൃදͷ໨࣍ ΞϧΰϦζϜతঈؐ ࣮ݱՄೳͳΞΫγϣϯੜ੒๏ ݚڀͷഎܠ

    ·ͱΊ ΞΫγϣϯ഑ྀܕֶश๏ զʑͷݚڀ੒Ռ ࠓޙͷల๬ɾൃදͷ·ͱΊ
  47. 2025/10/17 K. Kanamori, Fujitsu Ltd. ࠓޙͷల๬ (1/2) 51 Ϣʔβ͔ΒͷϑΟʔυόοΫΛ׆༻ͨ͠બ޷഑ྀܕΞΫγϣϯੜ੒ ՝୊ɹϢʔβબ޷

    [Tominaga+ 24] Ϣʔβͷਓ֨ಛੑ΍બ޷ʹΑͬͯ ΞΫγϣϯͷධՁج४͕มΘΔ ‣ બ޷Λߟྀͨ͠ίετઃܭ͕ඞཁ ࢿ֨XΛ औಘ͠·͠ΐ͏ʂ طଘղ๏ɹબ޷ֶश [De Toni+ 24] Ϣʔβ͔ΒͷϑΟʔυόοΫΛ༻͍ͯ ίετؔ਺ͷύϥϝʔλΛਪఆ ‣ ϑΟʔυόοΫ͕ҰରൺֱͷΈ, ཧ࿦อূΛܽ͘ͳͲ, ·ͩൃల్த cθ (a ∣ x) or
  48. 2025/10/17 K. Kanamori, Fujitsu Ltd. ࠓޙͷల๬ (2/2) 52 ෆ׬શͳ࣮ߦ΍Ϟσϧͷ࠶ֶशʹରͯ͠ϩόετͳΞΫγϣϯੜ੒ ՝୊ɹ༷ʑͳෆ࣮֬ੑ

    [Jiang+ 24] • ఏࣔͨ͠ΞΫγϣϯͷෆ׬શͳ࣮ߦ • ࠶ֶशʹΑΔϞσϧมಈ • Πϯελϯεͷೖྗ஋ͷઁಈ ͳͲ Ϟσϧมಈͷྫ [Jiang+ 25] طଘղ๏ɹϩόετ࠷దԽ ྫʣϞσϧมಈͷ৔߹ [Upadhyay+ 21] ‣ ద༻ՄೳͳϞσϧʹ੍ݶ͕͋Δ, ղ͕อकతͳͲ, վળͷ༨஍͋Γʁ mina∈ 𝒜 (x) c(a ∣ x) subject to h′  (x + a) = + 1, ∀h′  ∈ Δ(h) ϩόετΞΫγϣϯੜ੒໰୊ ෆ࣮֬ੑू߹
  49. 2025/10/17 K. Kanamori, Fujitsu Ltd. ·ͱΊ 53 ΞϧΰϦζϜతঈؐʹجͮ͘આ໌Մೳͳػցֶशʹ޲͚ͯ • ػցֶशϞσϧ͔Βߠఆతͳ༧ଌΛಘΔͨΊͷ

    “ΞΫγϣϯ” Λઆ໌ • ࣮ݱՄೳͳΞΫγϣϯΛఏࣔ͢ΔͨΊͷ৽͍͠ϑϨʔϜϫʔΫΛఏҊ • DACE (IJCAI-20): σʔλ෼෍ͷಛੑΛߟྀͯ͠ΞΫγϣϯͷ࣮ݱՄೳੑΛධՁ • RACT (ICML-24): ࣮ݱՄೳͳΞΫγϣϯͷଘࡏΛอূͭͭ͠ϞσϧΛֶश ΞΫγϣϯ આ໌ٕज़ ͋ͳͨ͸ ౶೘පϦεΫ͕ ߴ͍Ͱ͢ ʁ ৴པੑ ௿Լ Ϟσϧ Ϣʔβ ౶೘පϦεΫΛ Լ͛Δʹ͸ BMIΛݮΒͦ͏ ৴པੑ ޲্ Ϟσϧ Ϣʔβ