AI를 설명하면서 속도도 빠르게 할 순 없을까? SHAP 가속화 이야기 (feat. 산학협력)

9"*ۆ ஠஠য়ߛ௼੄ 9"* ࢑೟ഈ۱ਸ ాೠ 9"* оࣘച োҳ प ࢲ࠺झ
੸ਊ

F9QMBJOBCMF "SUJGJDJBM*OUFMMJHFODF ࢸݺоמೠ ੋҕ ૑מ

"SUJGJDJBM*OUFMMJHFODF ੋҕ ૑מ

G উ֞ )FMMP "* ߣ৉ ١ ࢤࢿഋ ݽ؛

G Ѣې੉۱ 'SBVE '%4 ".-੉࢚ఐ૑ ݽ؛

G ঴ҷ ࠄੋৈࠗ "*উݶ ੋૐ

G Y Z ੋҕ૑מب ѾҴ ೣࣻ

F9QMBJO ࢸݺ

! 𝑌 = $ 𝑓(𝑋! , 𝑋" , … ,
𝑋# ) ੑ۱җ ୹۱੄ t࢚ҙҙ҅u

+ 𝑦 = ! β$ + ! β! 𝑥! +
! β" 𝑥" + ⋯ + ! β# 𝑥# ? ୹۱ী ؀ೠ ੑ۱੄ tӝৈبu

FY नਊ੼ࣻ = ! β$ + ! β! 𝐼(ࣗٙ) +
! β" 𝐼(ࠁਬ؀୹) + j + ! β# 𝐼(ো୓੉۱) ? ୹۱ী ؀ೠ ੑ۱੄ tӝৈبu

FY नਊ੼ࣻ = ! β$ + ! β! 𝐼(ࣗٙ) +
! β" 𝐼(ࠁਬ؀୹) + j + ! β# 𝐼(ো୓੉۱) ? ୹۱ী ؀ೠ ੑ۱੄ tӝৈبu à 5PCFDPOUJOVFEj

8IZ t9"*о যڌѱ ഝਊؼ ࣻ ੓ભ u

ࢿמ न܉ Ѿҗী ؀ೠ Ѩૐ ਗੋ ѐࢶ੼ ౵ঈ ࣗా بݫੋ
૑ध оࢸ ഝਊ

9"*ۆ ஠஠য়ߛ௼੄ 9"* ࢑೟ഈ۱ਸ ాೠ 9"*оࣘച োҳ प ࢲ࠺झ ੸ਊ

஠஠য়ߛ௼੄ 9"*

஠஠য়ߛ௼੄ "* ୁࠈ नਊಣо ੉࢚Ѣېఐ૑

஠஠য়ߛ௼੄ 9"* ࠂ੟ೞҊ ׮নೠ ࢎӝ ಁఢਸ ೞաೞա ଺ইࢲ ݽפఠ݂ೞӝ
য۰ਕਃ L بݫੋ ૑ध੉ա ࢤпೞח оࢸ ҃೷੸ਵ۽ ୓ٙೠ ֢ೞ਋ܳ ঌ۰઱दݶ ୭؀ೠ Ѩૐ оמೞب۾ ѐߊ೧ࠁѷणפ׮ э੉ ٜ݅যࠊਃ J ࣗ࠺੗ࠁഐप "*प

஠஠য়ߛ௼੄ 9"* è AI 이상거래 예측 확률 : 약 67%
è 실제 3일 뒤 사고 접수, 보상 처리 건 17세 남학생, mini 카드로 금요일 아침 6시 30분 아이스크림 할인점에서 약 7,000원 결제 AI

஠஠য়ߛ௼੄ 9"* è "* ੉࢚Ѣې ৘ஏ ഛܫ ড
è पઁ ੌ ٍ ࢎҊ ੽ࣻ ࠁ࢚ ୊ܻ Ѥ ࣁ թ೟ࢤ NJOJ ஠٘۽ Әਃੌ ইஜ द ࠙ ই੉झ௼ܿ ೡੋ੼ীࢲ ড ਗ Ѿઁ "* 8):

஠஠য়ߛ௼੄ 9"* ✔ ࣁ ✔ NJOJ஠٘ ✔ ই੉झ௼ܿ ೡੋ੼ ࣁ
թ೟ࢤ NJOJ ஠٘۽ Әਃੌ ইஜ द ࠙ ই੉झ௼ܿ ೡੋ੼ীࢲ ড ਗ Ѿઁ

஠஠য়ߛ௼੄ 9"* ژې ஘ҳٜ੉ ই੉झ௼ܿ ೡੋ੼ীࢲ ࣗ࠺ೞח ಁఢ ࣁ թ೟ࢤ
NJOJ ஠٘۽ Әਃੌ ইஜ द ࠙ী ই੉झ௼ܿ ೡੋ੼ীࢲ ড ਗ੉ Ѿઁؽ Ѿઁ दр

஠஠য়ߛ௼੄ 9"* ࣁ թ೟ࢤ NJOJ ஠٘۽ Әਃੌ ইஜ द ࠙ী
ই੉झ௼ܿ ೡੋ੼ীࢲ ড ਗ੉ Ѿઁؽ ژې ஘ҳٜ੉ ই੉झ௼ܿ ೡੋ੼ীࢲ ࣗ࠺ೞח ಁఢ Ѿઁ Әঘ

஠஠য়ߛ௼੄ 9"* ੋҕ૑מ ݽ؛ ੉࢚Ѣې ఐ૑ Ѿҗ ࢎӝഛܫ ড
"* അস ੹ޙо ੉࢚Ѣې ಁఢ ౵ঈ ,FSOFM4)"1 ӝ߈ ੉࢚Ѣې ఐ૑ Ѿҗ ੉ਬ ࢑୹

"* അস ੹ޙо ੉࢚Ѣې ಁఢ ౵ঈ ,FSOFM4)"1 ӝ߈ ੉࢚Ѣې ఐ૑ Ѿҗ ੉ਬ ࢑୹ ࣗਃदр _ ୡ

"* അস ੹ޙо ੉࢚Ѣې ಁఢ ౵ঈ ,FSOFM4)"1 ӝ߈ ੉࢚Ѣې ఐ૑ Ѿҗ ੉ਬ ࢑୹ ୡ ޷݅ оࣘച ೙ਃ

"* അস ੹ޙо ੉࢚Ѣې ಁఢ ౵ঈ ,FSOFM4)"1 ӝ߈ ੉࢚Ѣې ఐ૑ Ѿҗ ੉ਬ ࢑୹ ୡ ޷݅ оࣘച ೙ਃ 3%.PUJWBUJPO

प ࢲ࠺झ ੸ਊ j

प ࢲ࠺झ ੸ਊ .POJUPSJOH 4ZTUFN j %PNBJO&YQFSU

प ࢲ࠺झ ੸ਊ .POJUPSJOH 4ZTUFN %BUB4DJFOUJTU j %PNBJO&YQFSU

प ࢲ࠺झ ੸ਊ .POJUPSJOH 4ZTUFN "*.PEFM %BUB4DJFOUJTU j %PNBJO&YQFSU

.-0QT प ࢲ࠺झ ੸ਊ .POJUPSJOH 4ZTUFN "*.PEFM %BUB4DJFOUJTU j %PNBJO&YQFSU

.-0QT प ࢲ࠺झ ੸ਊ .POJUPSJOH 4ZTUFN "*.PEFM %BUB1SPD %BUB4DJFOUJTU j
%PNBJO&YQFSU 5SBOTBDUJPO &WFOU

.-0QT प ࢲ࠺झ ੸ਊ .POJUPSJOH 4ZTUFN %PNBJO&YQFSU Ѿҗ Ѩૐ
بݫੋ ૑ध ҃೷ "*.PEFM 5SBOTBDUJPO &WFOU %BUB1SPD %BUB4DJFOUJTU j

بݫੋ ૑ध ҃೷ "*.PEFM 5SBOTBDUJPO &WFOU %BUB1SPD ழޭפா੉࣌ ࠺ਊ ֫਺ j %BUB4DJFOUJTU ݽ؛ ٣ߡӦ XՐ੐হח ࢿמ ѐࢶ X֢য়۱

بݫੋ ૑ध ҃೷ "*.PEFM 5SBOTBDUJPO &WFOU %BUB1SPD оࣘച 9"* %BUB4DJFOUJTU j

प ࢲ࠺झ ੸ਊ 5SBOTBDUJPO &WFOU .-0QT "*.PEFM %BUB1SPD оࣘച 9"*
.POJUPSJOH 4ZTUFN %PNBJO&YQFSU Ѿҗ Ѩૐ بݫੋ ૑ध ҃೷ %BUB4DJFOUJTU j

प ࢲ࠺झ ੸ਊ 5SBOTBDUJPO &WFOU .-0QT "*.PEFM %BUB1SPD %PNBJO&YQFSU
Ѿҗ Ѩૐ بݫੋ ૑ध ҃೷ %BUB4DJFOUJTU ݽ؛ ٣ߡӦ X9"* ࢿמ ѐࢶ X9"* ਗഝೠ ழޭפா੉࣌ *5*OUFSOBM"VEJU5FBN "*Ѣߡքझ ળࣻ ੼Ѩ ݽ؛ ಞೱࢿ ੉ग ੼Ѩ оࣘച 9"* .POJUPSJOH 4ZTUFN j

प ࢲ࠺झ ੸ਊ 9"* ঌҊ્ܻ 9"*Ѿҗ ࢑୹ ࣘب MBUFODZ ߛ௼
"*ݽ؛ ߛ௼ "* ݽ؛ ߛ௼ "*ݽ؛ ,FSOFM4)"1 CBTFMJOF Y Y Y оࣘച ,FSOFM4)"1 3% ױ҅ 4".4)"1 3%ױ҅

"*ݽ؛ ߛ௼ "* ݽ؛ ߛ௼ "*ݽ؛ ,FSOFM4)"1 CBTFMJOF Y Y Y оࣘച ,FSOFM4)"1 3% ױ҅ Y Y Y 4".4)"1 3%ױ҅

प ࢲ࠺झ ੸ਊ XAI 알고리즘 XAI 결과 산출 속도 (latency)
뱅크 AI 모델 (1) 뱅크 AI 모델 (2) 뱅크 AI 모델 (3) KernelSHAP1) (baseline) 01.00 x 01.00 x 01.00 x 가속화 KernelSHAP (R&D 1단계) 02.38 x (± 0.12) 02.82 x (± 0.39) 02.97 x (± 0.44) SAMSHAP (R&D 2단계) 10.35 x (± 1.36) 04.81 x (± 0.79) 07.70 x (± 1.94)

"*ݽ؛ ߛ௼ "* ݽ؛ ߛ௼ "*ݽ؛ ,FSOFM4)"1 CBTFMJOF Y Y Y оࣘച ,FSOFM4)"1 3% ױ҅ Y Y Y 4".4)"1 3%ױ҅ Y Y Y )PX t"*ܳ ޖट ߑߨਵ۽ ࢸݺ೮ਵݴ যڌѱ оࣘച ೠѤоਃ u

"*ݽ؛ ߛ௼ "* ݽ؛ ߛ௼ "*ݽ؛ ,FSOFM4)"1 CBTFMJOF Y Y Y оࣘച ,FSOFM4)"1 3% ױ҅ Y Y Y 4".4)"1 3%ױ҅ Y Y Y 4)"1 4)BQMFZ "EEJUJWF FY1MBOBUJPOT

4)"1 ਷ ഈ۱ѱ੐੉ۿ DPBMJUJPOBMHBNFUIFPSZ ઺ীࢲ п ѱ੐ ଵৈ੗ QMBZFS ীѱ
ࣻ੊੄ ߓ׼ QBZPVU ਸ tҕ੿ೞѱ ߓ࠙uೞח ߑߨਵ۽ ੜ ঌ۰૓ 4IBQMFZ7BMVFܳ ӝ߈ਵ۽ ೠ׮ ݠन۞׬ ޙઁܳ ഈ۱੸ ѱ੐ ޙઁ۽ ࠺ਬೞݶ ৘ஏч 𝑦̂ ª QBZPVU ਸ ࢑୹ೞח ঌҊ્ܻ 𝑓 ª DPBMJUJPOBMHBNF ী ߸ࣻо ੑ۱ 𝑋𝑝 ª QMBZFS عਸ ٸ п ੑ۱߸ࣻо ৘ஏчী ӝৈೠ ੿بܳ ҕ੿ೞѱ ஏ੿ೠ Ѿҗ 𝛽𝑝 ª 4IBQMFZ 7BMVF ܳ ଺ח Ѫ ੉୊ۢ 4IBQMFZWBMVFߑߨۿਸ 9"*ޙઁী 1SBDUJDBMೞѱ ਽ਊೠ ঌҊ્ܻ੉ 4)"1 4)BQMFZ "EEJUJWFFY1MBOBUJPOT ੉׮ 4)BQMFZ "EEJUJWFFY1MBOBUJPOT 4)"1 ಪ ֢੉݅ ੉ۿ࢚ ࢚ࣻ ֢߰ ҃ઁ೟࢚ ࢚ࣻ -MPZE4IBQMFZ _

4IBQMFZ7BMVFT 5FBN $POUSJCVUJPO Payout )PXUPtGBJSMZuEJTUSJCVUF UIFJSFBSOJOHT t(BNFu

4)BQMFZ "EEJUJWFFY1MBOBUJPOT 4)"1 *OQVU'FBUVSFT 1SFQSPDFTTJOH ! 𝒚 )PXUPtGBJSMZuEJTUSJCVUF UIFJSDPOUSJCVUJPOT !
𝒇(𝑥) 𝑥!, 𝑥", … , 𝑥# ! β 0VUQVU

4IBQMFZ7BMVFT ౱ ഈ۱ ࣻ੊ п ଵৈ੗੄ tӝৈuী ٮܲ ࣻ੊੄ ߓ࠙
ߑߨ ध׼

ഈ۱ 4IBQMFZ7BMVFT

4IBQMFZ7BMVFT .BSHJOBM7BMVF ౱
ഈ۱ "CTFOU

ഈ۱ 4IBQMFZ7BMVFT *OUFSBDUJPO&GGFDU FH ா޷ п ଵৈ੗੄
tӝৈuী ٮܲ ࣻ੊੄ ߓ࠙ ߑߨ

4IBQMFZ7BMVFT п ଵৈ੗੄ tӝৈuী ٮܲ ࣻ੊੄ ߓ࠙
ߑߨ ഈ۱ *OUFSBDUJPO&GGFDU FHா޷

$POUSJCVUJPO )PXUPtGBJSMZuEJTUSJCVUF UIFJSFBSOJOHT t(BNFu Payout 4VCTFUT Payout 1BZPVU 4IBQMFZ7BMVFT à
.BSHJOBM$POUSJCVUJPO

4IBQMFZ7BMVFT п ଵৈ੗੄ tӝৈuী ٮܲ ࣻ੊੄ ߓ࠙
ߑߨ ഈ۱ *OUFSBDUJPO&GGFDU FHா޷

4IBQMFZ7BMVFT ౱ ഈ۱ п ଵৈ੗੄ tӝৈuী ٮܲ ࣻ੊੄
ߓ࠙ ߑߨ

4IBQMFZ7BMVFT ౱ ഈ۱ .BSHJOBM7BMVF

! β!! 𝑓, 𝑥 = ' 4IBQMFZ7BMVFT
4VCTFU j 4VCTFU )JHIFS XFJHIU -PXFS XFJHIU 4VCTFU )JHIFS XFJHIU

! β!! 𝑓, 𝑥 = ' 4)BQMFZ "EEJUJWFFY1MBOBUJPOT 4)"1
4VCTFU j 4VCTFU )JHIFS XFJHIU -PXFS XFJHIU 4VCTFU )JHIFS XFJHIU 𝑥" 𝑥# 𝑥$ 𝑥" 𝑥# 𝑥$

4)BQMFZ "EEJUJWFFY1MBOBUJPOT 4)"1 ! β# 𝑓, 𝑥 = 3 %⊆'\
# 𝐹 1, 𝑆 , 𝐹 − 𝑆 − 1 )! 𝑓%∪ # (𝑥%∪ # ) − 𝑓% (𝑥% ) where : F = the set of all features S = the set of all feature subsets • Features are like simplified or aggregated data in case of unstructured dataset. • To exclude a specific feature, we just input random values from trainset. à SHAP is approximated Shapley Value with sampling feature subsets and fitting a surrogate model. XFJHIUFEBWFSBHFPGBMMQPTTJCMFEJGGFSFODFT XFJHIUJOHXNVMUJOPNJBM DPFGGJDJFOU EJGGFSFODFCXDPPQFSBUJPO PSOPU .BSHJOBM$POUSJCVUJPO

# 1! 𝑆 ! 𝐹 − 𝑆 − 1 ! 𝐹 ! 𝑓%∪ # (𝑥%∪ # ) − 𝑓% (𝑥% ) where : F = the set of all features S = the set of all feature subsets • Features are like simplified or aggregated data in case of unstructured dataset. • To exclude a specific feature, we just input random values from trainset. à SHAP is approximated Shapley Value with sampling feature subsets and fitting a surrogate model. XFJHIUFE BWFSBHFPGBMMQPTTJCMFEJGGFSFODFT XFJHIUJOHXNVMUJOPNJBM DPFGGJDJFOU EJGGFSFODFCXDPPQFSBUJPO PSOPU .BSHJOBM$POUSJCVUJPO

# 𝐹 1, 𝑆 , 𝐹 − 𝑆 − 1 )! 𝑓%∪ # (𝑥%∪ # ) − 𝑓% (𝑥% ) where : F = the set of all features S = the set of all feature subsets • Features are like simplified or aggregated data in case of unstructured dataset. • To exclude a specific feature, we just input random values from trainset. à SHAP is approximated Shapley Value with sampling feature subsets and fitting a surrogate model. XFJHIUFEBWFSBHFPGBMMQPTTJCMFEJGGFSFODFT XFJHIUJOHXNVMUJOPNJBM DPFGGJDJFOU EJGGFSFODFCXDPPQFSBUJPO PSOPU .BSHJOBM$POUSJCVUJPO

4)BQMFZ "EEJUJWFFY1MBOBUJPOT 4)"1 ϕ# 𝑓, 𝑥 = 3 %⊆'\ #
1! 𝑆 ! 𝐹 − 𝑆 − 1 ! 𝐹 ! 𝑓%∪ # (𝑥%∪ # ) − 𝑓% (𝑥% ) where : F = the set of all features S = the set of all feature subsets • Features are like simplified or aggregated data in case of unstructured dataset. • To exclude a specific feature, we just input random values from trainset. à SHAP is approximated Shapley Value with sampling feature subsets and fitting a surrogate model. XFJHIUFEBWFSBHFPGBMMQPTTJCMFEJGGFSFODFT XFJHIUJOHXNVMUJOPNJBM DPFGGJDJFOU EJGGFSFODFCXDPPQFSBUJPO PSOPU .BSHJOBM$POUSJCVUJPO

ݽٚ оמೠ ઑ೤ী ؀ೠ ӝৈب ҅࢑ਸ ਤ೧ ݒ਋ ݆਷ ҅࢑
࠺ਊ ࣗਃ 4)BQMFZ "EEJUJWFFY1MBOBUJPOT 4)"1 ϕ# 𝑓, 𝑥 = 3 %⊆'\ # 1! 𝑆 ! 𝐹 − 𝑆 − 1 ! 𝐹 ! 𝑓%∪ # (𝑥%∪ # ) − 𝑓% (𝑥% ) ೠ҅੼

,FSOFM4)"1

౱ ഈ۱ ࣻ੊ п ଵৈ੗੄ tӝৈuী ٮܲ ࣻ੊੄ ߓ࠙ ߑߨ
ध׼ ,FSOFM4)"1

.BSHJOBM7BMVF ౱ ഈ۱
"CTFOU ,FSOFM4)"1

,FSOFM4)"1 оמೠ Feature subset іࣻ 2!ѐ ⋮

,FSOFM4)"1 ⋮ оמೠ 'FBUVSF4VCTFU 2!ѐ 𝑋 ∈ 0(𝑎𝑏𝑠𝑒𝑛𝑡), 1(𝑝𝑟𝑒𝑠𝑒𝑛𝑡) !
Y = 𝜙! + 𝜙" 𝑋" + 𝜙# 𝑋# + 𝜙$ 𝑋$ + 𝜙% 𝑋%

,FSOFM4)"1 Approximation of Shapley value 4BNQMJOH ⋮ оמೠ 'FBUVSF4VCTFU 2!ѐ
𝑋 ∈ 0(𝑎𝑏𝑠𝑒𝑛𝑡), 1(𝑝𝑟𝑒𝑠𝑒𝑛𝑡) ! Y = 𝜙! + 𝜙" 𝑋" + 𝜙# 𝑋# + 𝜙$ 𝑋$ + 𝜙% 𝑋%

,FSOFM4)"1 Coalition 𝒛 Feature 𝒙
Coalition 𝒛! Feature 𝒉𝒙 (𝒛!) "CTFOU

ߓ҃ ؘ੉ఠ ߓ҃ ؘ੉ఠ۽ࠗఠ t"CTFOUu GFBUVSFчਸ ؀୓ ,FSOFM4)"1

,FSOFM4)"1 Coalition 𝒛 Feature 𝒙
⋮ 𝒇 𝒉𝒙 𝒛𝟏 # Coalition 𝒛𝟏 Coalition 𝒛𝟐 Coalition 𝒛𝒌 𝒇 𝒉𝒙 𝒛𝟐 # 𝒇 𝒉𝒙 𝒛𝒌 # ⋮ 𝐿 𝑓, 𝑔, π"" = 5 #"∈% 𝑓 ℎ" 𝑧& − 𝑔 𝑧& ' 𝜋"" 𝑧& -PTTGVODUJPOUPNJOJNJ[F 8FJHIUJOH,FSOFM 𝑔 𝑧H = 𝜙I + 0 JKL M 𝜙J 𝑧J H FYQMBOBUJPONPEFMUIBUJTBMJOFBSGVODUJPOPGCJOBSZWBSJBCMFT where : 𝑧" ∈ 0, 1 #, M is the number of simplified input features 𝜙 = Shapley value

,FSOFM4)"1 Exact Shapley Value KernelSHAP 장점 - 특징별 중요도를 정확히
계산 가능함 - 샘플 수가 많을수록 Exact Shapley Value와 근사함 - Exact Shapley Value보다 계산 시간 단축 - Model-Agnostic 한 방법론 단점 - 2& 만큼의 경우의 수를 고려해야하기 때문에 계산 속도가 느림 - Exact Shapley Value에 근사하기 위해서는 많은 샘플 수를 필요로 함 à 계산 시간 증가

,FSOFM4)"1 &YBDU4IBQMFZ7BMVF ,FSOFM4)"1 ੢੼ ౠ૚߹ ઺ਃبܳ ੿ഛ൤ ҅࢑ оמೣ
ࢠ೒ ࣻо ݆ਸࣻ۾ &YBDU 4IBQMFZ7BMVF৬ Ӕࢎೣ &YBDU 4IBQMFZ7BMVFࠁ׮ ҅࢑ दр ױ୷ .PEFM"HOPTUJDೠ ߑߨۿ ױ੼ - 𝟐𝐅 ݅ఀ੄ ҃਋੄ ࣻܳ Ҋ۰೧ঠೞӝ ٸޙী ҅࢑ ࣘبо וܿ &YBDU 4IBQMFZ 7BMVFী Ӕࢎೞӝ ਤ೧ࢲח ݆਷ ࢠ೒ ࣻܳ ೙ਃ۽ ೣ à ҅࢑ दр ૐо оࣘച ؀࢚ ݽ؛

ࢎ۹

࢑೟ഈ۱ਸ ాೠ ҕزোҳۆ ӝসҗ ೟Үо ഈডਸ ݛয ҕా੄ ݾ಴ܳ ਤ೧
ೣԋ োҳ੘সਸ ࣻ೯ೞח Ѫ

஠஠য়ߛ௼ ࢑೟ഈ۱ Әਲ਼ӝࣿোҳࣗ ؀ղ৻ ഈস

Әਲ਼࠙ঠ ࢸݺоמ ੋҕ૑מ ঌҊ્ܻ Ҋبച ஠஠য়ߛ௼ Әਲ਼ӝࣿোҳࣗ ஠੉झ౟ ࢸݺоמੋҕ૑מোҳࣃఠ 9"*$
࢑೟ഈ۱ ҕزোҳ

оࣘച ӏઁӝࣿ ࢸݺоמ ੋҕ૑מ ঌҊ્ܻ оࣘച ࢸݺоמ ੋҕ૑מ ӏઁӝࣿزೱ ઑࢎ
ҕزোҳ ઱ਃ ղਊ

оࣘച ߑউ ઑࢎ оࣘച ঌҊ્ܻ ѐߊ оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿
9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ߑউ ઑࢎ оࣘച ঌҊ્ܻ ѐߊ оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿
оࣘച ߑউ ઑࢎ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ߑউ ઑࢎ ূ૑פয݂ ӝ߈ ঌҊ્ܻ ӝ߈ ೞ٘ਝয ӝ߈ "NPSUJ[FE
/PO "NPSUJ[FE 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ߑউ ઑࢎ 9"*ঌҊ્ܻ оࣘച োҳ ੢੼ ױ੼ ূ૑פয݂ ӝ߈
ӝઓ 9"*ঌҊ્ܻҗ زੌೠ Ѿҗܳ ب୹ೞݴ ҅࢑दр ױ୷ оמ ҳഅػ ௏٘ী ٮۄࢲ оࣘച оמ ৈࠗ Ѿ੿ ঌҊ્ܻ ӝ߈ "NPSUJ[FE ҅࢑ दр ௼ѱ ױ୷ оמ न҃ݎ ݽ؛ਸ ࢎਊೞӝ ٸޙী ӝઓ 9"* ঌҊ્ܻ੄ ࢸݺ۱ ࠁઓ੉ য۰਑ /POBNPSUJ[FE "NPSUJ[FEߑߨۿ ؀࠺ ࢸݺ۱ਸ ࠁઓೡ ࣻ ੓਺ "NPSUJ[FEߑߨۿ ؀࠺ ҅࢑दр ѐࢶী ೠ҅о ੓਺ ೞ٘ਝয ӝ߈ ӝઓ 9"*ঌҊ્ܻ੄ ࢸݺ۱ਸ ࠁઓ оמೣ ೞ٘ਝয ੢࠺੄ ઁড੉ ઓ੤

оࣘച ߑউ ઑࢎ ੢੼ ױ੼ ূ૑פয݂ ӝ߈ ӝઓ 9"*ঌҊ્ܻҗ
زੌೠ Ѿҗܳ ب୹ೞݴ ҅࢑दр ױ୷ оמ ҳഅػ ௏٘ী ٮۄࢲ оࣘച оמ ৈࠗ Ѿ੿ ঌҊ્ܻ ӝ߈ "NPSUJ[FE ҅࢑ दр ௼ѱ ױ୷ оמ न҃ݎ ݽ؛ਸ ࢎਊೞӝ ٸޙী ӝઓ 9"* ঌҊ્ܻ੄ ࢸݺ۱ ࠁઓ੉ য۰਑ /POBNPSUJ[FE "NPSUJ[FEߑߨۿ ؀࠺ ࢸݺ۱ਸ ࠁઓೡ ࣻ ੓਺ "NPSUJ[FEߑߨۿ ؀࠺ ҅࢑दр ѐࢶী ೠ҅о ੓਺ ೞ٘ਝয ӝ߈ ӝઓ 9"*ঌҊ્ܻ੄ ࢸݺ۱ਸ ࠁઓ оמೣ ೞ٘ਝয ੢࠺੄ ઁড੉ ઓ੤ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ 9"*ঌҊ્ܻ оࣘച
োҳ

оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ ੿ഛب $PTJOF4JNJMBSJUZ 5PQ4JHOFE3BOL"HSFFNFOU ࣘب ࢸݺ दр
ױ୷ rୡ ޷݅s r'BTU4)"1 ؀࠺ ࢿמ ೱ࢚s 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ $PTJOF4JNJMBSJUZ 𝐶𝑜𝑠𝑖𝑛𝑒 𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝐴, 𝐵) = cos
Θ = 𝐴 ⋅ 𝐵 𝐴 𝐵 = ∑()* + 𝐴( 𝐵( ∑ ()* + 𝐴( ' ⋅ ∑ ()* + 𝐵( ' 9"*ঌҊ્ܻ оࣘച োҳ ف ߭ఠр пب੄ ௏ࢎੋчਸ ੉ਊ೧ ஏ੿ػ ߭ఠр ਬࢎب

оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ 5PQ,4JHOFE3BOL"HSFFNFOU 𝑇𝑜𝑝 − 𝑘 𝑆𝑖𝑔𝑛𝑒𝑑𝑅𝑎𝑛𝑘𝐴𝑔𝑟𝑒𝑒𝑚𝑒𝑛𝑡(𝐸, ,
𝐸- , 𝑘) = | ⋃.∈/ 𝑓 𝒇 ∈ 𝑻𝑭 𝑬𝒂 , 𝒌 ∧ 𝒇 ∈ 𝑻𝑭 𝑬𝒃 , 𝒌 ∧ 𝑺 𝑬𝒂 , 𝒇 = 𝑺 𝑬𝒃 , 𝒇 ∧ 𝑹 𝑬𝒂 , 𝒇 = 𝑹 𝑬𝒃 , 𝒇 }| 𝑘 ࢚ਤ ,ѐ ઺ ҕా ౠ૚ਸ ನೣೞח૑ ৈࠗ زੌ ࠗഐ ৈࠗ زੌ ࣽਤ ৈࠗ 9"*ঌҊ્ܻ оࣘച োҳ ف ࢸݺѾҗ੄ ࢚ਤ ,ѐ ౠ૚ GFBUVSF ૘೤ীࢲ زੌೠ ౠࢿ ӝৈب ࠗഐ ৬ ࣽਤо ҕా੸ਵ۽ ನೣػ ౠ૚ٜ੄ ࠺ਯ

оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ Rank feature Shapley value 1 Feature
A 0.9 2 Feature B 0.8 3 Feature C 0.7 4 Feature D 0.6 5 Feature E 0.5 ࢸݺ ݽ؛ 1੄ Ѿҗ(𝐸!) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ Rank feature Shapley value 1 Feature A 0.85 2 Feature B 0.8 3 Feature D 0.75 4 Feature E 0.6 5 Feature F 0.55 ࢸݺ ݽ؛ 2੄ Ѿҗ(𝐸" ) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ 𝑇𝑜𝑝 − 𝑘 𝑆𝑖𝑔𝑛𝑒𝑑𝑅𝑎𝑛𝑘𝐴𝑔𝑟𝑒𝑒𝑚𝑒𝑛𝑡(𝐸! , 𝐸" , 𝑘) = ⋃#∈% 𝒇|𝒇 ∈ 𝑻𝑭 𝑬𝒂 , 𝒌 ∧ 𝒇 ∈ 𝑻𝑭 𝑬𝒃 , 𝒌 ∧ 𝑺 𝑬𝒂 , 𝒇 = 𝑺 𝑬𝒃 , 𝒇 ∧ 𝑹 𝑬𝒂 , 𝒇 = 𝑹 𝑬𝒃 , 𝒇 𝑘 9"*ঌҊ્ܻ оࣘച োҳ

A 0.9 2 Feature B 0.8 3 Feature C 0.7 4 Feature D 0.6 5 Feature E 0.5 Rank feature Shapley value 1 Feature A 0.85 2 Feature B 0.8 3 Feature D 0.75 4 Feature E 0.6 5 Feature F 0.55 ࢚ਤ ,ѐ ઺ ҕా ౠ૚ਸ ನೣೞח૑ ৈࠗ ࢸݺ ݽ؛ 1੄ Ѿҗ(𝐸!) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ ࢸݺ ݽ؛ 2੄ Ѿҗ(𝐸" ) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ 𝑇𝑜𝑝 − 𝑘 𝑆𝑖𝑔𝑛𝑒𝑑𝑅𝑎𝑛𝑘𝐴𝑔𝑟𝑒𝑒𝑚𝑒𝑛𝑡(𝐸! , 𝐸" , 𝑘) = ⋃#∈% 𝒇|𝒇 ∈ 𝑻𝑭 𝑬𝒂 , 𝒌 ∧ 𝒇 ∈ 𝑻𝑭 𝑬𝒃 , 𝒌 ∧ 𝑺 𝑬𝒂 , 𝒇 = 𝑺 𝑬𝒃 , 𝒇 ∧ 𝑹 𝑬𝒂 , 𝒇 = 𝑹 𝑬𝒃 , 𝒇 𝑘 9"*ঌҊ્ܻ оࣘച োҳ

A 0.9 2 Feature B 0.8 3 Feature C 0.7 4 Feature D 0.6 5 Feature E 0.5 Rank feature Shapley value 1 Feature A 0.85 2 Feature B 0.8 3 Feature D 0.75 4 Feature E 0.6 5 Feature F 0.55 زੌ ࠗഐ ৈࠗ ࢸݺ ݽ؛ 1੄ Ѿҗ(𝐸!) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ ࢸݺ ݽ؛ 2੄ Ѿҗ(𝐸" ) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ 𝑇𝑜𝑝 − 𝑘 𝑆𝑖𝑔𝑛𝑒𝑑𝑅𝑎𝑛𝑘𝐴𝑔𝑟𝑒𝑒𝑚𝑒𝑛𝑡(𝐸! , 𝐸" , 𝑘) = ⋃#∈% 𝒇|𝒇 ∈ 𝑻𝑭 𝑬𝒂 , 𝒌 ∧ 𝒇 ∈ 𝑻𝑭 𝑬𝒃 , 𝒌 ∧ 𝑺 𝑬𝒂 , 𝒇 = 𝑺 𝑬𝒃 , 𝒇 ∧ 𝑹 𝑬𝒂 , 𝒇 = 𝑹 𝑬𝒃 , 𝒇 𝑘 9"*ঌҊ્ܻ оࣘച োҳ

A 0.9 2 Feature B 0.8 3 Feature C 0.7 4 Feature D 0.6 5 Feature E 0.5 Rank feature Shapley value 1 Feature A 0.85 2 Feature B 0.8 3 Feature D 0.75 4 Feature E 0.6 5 Feature F 0.55 ࢸݺ ݽ؛ 1੄ Ѿҗ(𝐸!) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ ࢸݺ ݽ؛ 2੄ Ѿҗ(𝐸" ) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ زੌ ࣽਤ ৈࠗ 𝑇𝑜𝑝 − 𝑘 𝑆𝑖𝑔𝑛𝑒𝑑𝑅𝑎𝑛𝑘𝐴𝑔𝑟𝑒𝑒𝑚𝑒𝑛𝑡(𝐸! , 𝐸" , 𝑘) = ⋃#∈% 𝒇|𝒇 ∈ 𝑻𝑭 𝑬𝒂 , 𝒌 ∧ 𝒇 ∈ 𝑻𝑭 𝑬𝒃 , 𝒌 ∧ 𝑺 𝑬𝒂 , 𝒇 = 𝑺 𝑬𝒃 , 𝒇 ∧ 𝑹 𝑬𝒂 , 𝒇 = 𝑹 𝑬𝒃 , 𝒇 𝑘 9"*ঌҊ્ܻ оࣘച োҳ

A 0.9 2 Feature B 0.8 3 Feature C 0.7 4 Feature D 0.6 5 Feature E 0.5 Rank feature Shapley value 1 Feature A 0.85 2 Feature B 0.8 3 Feature D 0.75 4 Feature E 0.6 5 Feature F 0.55 = 𝟐/𝟓 ࢸݺ ݽ؛ 1੄ Ѿҗ(𝐸!) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ ࢸݺ ݽ؛ 2੄ Ѿҗ(𝐸" ) ࢚ਤ 5ѐ੄ ౠ૚җ Ӓ ࣽਤ 𝑇𝑜𝑝 − 𝑘 𝑆𝑖𝑔𝑛𝑒𝑑𝑅𝑎𝑛𝑘𝐴𝑔𝑟𝑒𝑒𝑚𝑒𝑛𝑡(𝐸! , 𝐸" , 𝑘) = ⋃#∈% 𝒇|𝒇 ∈ 𝑻𝑭 𝑬𝒂 , 𝒌 ∧ 𝒇 ∈ 𝑻𝑭 𝑬𝒃 , 𝒌 ∧ 𝑺 𝑬𝒂 , 𝒇 = 𝑺 𝑬𝒃 , 𝒇 ∧ 𝑹 𝑬𝒂 , 𝒇 = 𝑹 𝑬𝒃 , 𝒇 𝑘 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ 9"*ঌҊ્ܻ оࣘച োҳ 정확도 $PTJOF4JNJMBSJUZ 5PQ4JHOFE3BOL"HSFFNFOU
속도 ࢸݺ दр ױ୷ rୡ ޷݅s r'BTU4)"1 ؀࠺ ࢿמ ೱ࢚s

оࣘച ঌҊ્ܻ ѐߊ оࣘച ঌҊ્ܻ ಣо૑಴ ࢶ੿ оࣘച ঌҊ્ܻ ѐߊ
9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ োҳѐߊ 9"*ঌҊ્ܻ оࣘച োҳ ஠஠য়ߛ௼ Әਲ਼ӝࣿোҳࣗ ஠੉झ౟
ࢸݺоמੋҕ૑מোҳࣃఠ 9"*$ ੸ਊ ߂ ഛੋ

஠஠য়ߛ௼ पؘ੉ఠ оࣘചػ 9"*ঌҊ્ܻ ѐߊ '%4ݽ؛ оࣘചػ 9"*ঌҊ્ܻ ੸ਊ ஠஠য়ߛ௼
9"*$ '%4ݽ؛ ஠஠য়ߛ௼ ۄ੉࠳ ݽ؛ ੸ਊद 9"*ঌҊ્ܻ оࣘച ഛੋ оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ #BOL"DDPVOU'SBVE #"' %BUB

оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ оࣘചػ 9"*ঌҊ્ܻ ѐߊ 9"*$
'%4ݽ؛ #BOL"DDPVOU'SBVE #"' %BUB

ৡۄੋ ਷೯ ҅ઝ ѐࢸ ࢎӝఐ૑ܳ ਤೠ ҕѐ ؘ੉ఠࣇ ؘ੉ఠ ࠛӐഋ
߸ࣻ ઙܨ ١ पઁ ؘ੉ఠ৬ ݒ਋ ਬࢎೠ ഋక оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ #BOL"DDPVOU'SBVE #"' %BUB

োҳѐߊ ݾ಴ 9"*ঌҊ્ܻ оࣘച োҳ #BOL"DDPVOU'SBVE #"' %BUB оࣘചػ
9"*ঌҊ્ܻ ѐߊ 9"*$ '%4ݽ؛

஠஠য়ߛ௼੄ ੉࢚ఐ૑ ࢎਊ ݽ؛ ঌҊ્ܻ ,FSOFM4)"1 ঌҊ્ܻ ੉ਊ ஠஠য়ߛ௼੄ पؘ੉ఠܳ
ࢸݺೞחؘ _ୡ ࣗਃ ,FSOFM4)"1 ঌҊ્ܻ ੉ਊ #BOL"DDPVOU'SBVE ؘ੉ఠܳ ࢸݺೞחؘ ୡ ࣗਃ оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ '%4ݽ؛

оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ #BOL"DDPVOU'SBVE #"' %BUB
оࣘചػ 9"*ঌҊ્ܻ ѐߊ 9"*$ '%4ݽ؛

,FSOFM4)"1੄ ੿ഛبܳ ਬ૑ೞݴ оࣘചػ ঌҊ્ܻ ѐߊ ӝઓ ,FSOFM4)"1਷ ҅࢑ী ࢎਊೞח
ࢠ೒ ࣻী ٮۄ ੿ഛب৬ ҅࢑ࣘبী ௾ ৔ೱਸ ޷ஜ ,FSOFM4)"1਷ .PEFM"HOPTUJDߑߨۿ оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ оࣘചػ 9"*ঌҊ્ܻ ѐߊ

оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ #BOL"DDPVOU'SBVE #"' %BUB
оࣘചػ 9"*ঌҊ્ܻ ѐߊ 9"*$ '%4ݽ؛ ஠஠য়ߛ௼ पؘ੉ఠ '%4ݽ؛ оࣘചػ 9"*ঌҊ્ܻ ੸ਊ ஠஠য়ߛ௼ ஠஠য়ߛ௼ ۄ੉࠳ ݽ؛ ੸ਊद 9"*ঌҊ્ܻ оࣘച ഛੋ

оࣘച ঌҊ્ܻ ѐߊ ূ૑פয݂ ӝ߈ ঌҊ્ܻ ӝ߈ 'BTU4)"1 ѐࢶ ,FSOFM4)"1
ѐࢶ ੿ഛب ࣘب ӝઓ ,FSOFM4)"1 ࢸݺ۱ ࠁઓ оࣘച ୭؀ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ ূ૑פয݂ ӝ߈ оࣘച оࣘച ,FSOFM4)"1 9"*ঌҊ્ܻ оࣘച
োҳ

оࣘച ঌҊ્ܻ ѐߊ 'FBUVSF(SPVQ4BNQMJOH 4IBQMFZ7BMVF&TUJNBUJPO ӝઓ ,FSOFM4)"1 ঌҊ્ܻ 9"*ঌҊ્ܻ оࣘച
োҳ

оࣘച ঌҊ્ܻ ѐߊ • ,FSOFM4)"1੄ ࢸݺ۱ ࠁઓ • ,FSOFM4)"1 ؀࠺
҅࢑ ࣘب ࡅܴ • ੿ഛب ഛࠁܳ ਤ೧ ݆਷ ࢠ೒ ࣻо ೙ਃ ੢੼ ױ੼ оࣘച ,FSOFM4)"1 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ ঌҊ્ܻ ӝ߈ оࣘച 'BTU4)"1 4".4)"1 9"*ঌҊ્ܻ оࣘച
োҳ

оࣘച ঌҊ્ܻ ѐߊ 'BTU4)"1 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ ߓؘ҃੉ఠ 𝒇 𝒉𝒙 𝒛𝟏 ! 𝒇 𝒉𝒙
𝒛𝟐 ! 𝒇 𝒉𝒙 𝒛𝒌 ! ⋮ 9"*ঌҊ્ܻ оࣘച োҳ • ӝઓ ,FSOFM4)"1 "QQSPYJNBUJPO 𝑓 ℎ! 𝑧"

оࣘച ঌҊ્ܻ ѐߊ ҅࢑ ࣘب ੷ೞ "QQSPYJNBUJPO 𝑓 ℎ! 𝑧"
ߓؘ҃੉ఠ 𝒇 𝒉𝒙 𝒛𝟏 ! 𝒇 𝒉𝒙 𝒛𝟐 ! 𝒇 𝒉𝒙 𝒛𝒌 ! ⋮ • ӝઓ ,FSOFM4)"1 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ ߓؘ҃੉ఠ 𝒇 𝒉𝒙 𝒛𝟏 ! 𝒇 𝒉𝒙
𝒛𝟐 ! 𝒇 𝒉𝒙 𝒛𝒌 ! ⋮ ҅࢑ ࣘب ೱ࢚ • 'BTU4)"1 ؀ܻ ݽ؛ о۰૓ ੑ۱чী ؀೧ࢲ ࢸݺ ؀࢚ ݽ؛җ э਷ чਸ ୹۱ೞب۾ ೟णೠ न҃ݎ ݽ؛ ӝ߈੄ ؀ܻ ݽ؛ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ ؀ܻ ݽ؛ ࢸݺ ؀࢚ ݽ؛ ؀ܻݽ؛ਸ ੉ਊೞৈ
4IBQMFZWBMVFܳ ৘ஏೞب۾ ೟णೠ न҃ݎ ݽ؛ ӝ߈੄ ࢸݺ ݽ؛ ࢸݺ ݽ؛ о۰૓ ੑ۱чী ؀೧ࢲ ࢸݺ ؀࢚ ݽ؛җ э਷ чਸ ୹۱ೞب۾ ೟णೠ न҃ݎ ݽ؛ ӝ߈੄ ؀ܻ ݽ؛ 9"*ঌҊ્ܻ оࣘച োҳ • 'BTU4)"1

оࣘച ঌҊ્ܻ ѐߊ • ,FSOFM4)"1 ؀࠺ ҅࢑ ࣘب ࡅܴ •
,FSOFM4)"1 ؀࠺ ࢸݺ۱ ੷ೞ ੢੼ ױ੼ 'BTU4)"1 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ 4".4)"1 4VSSPHBUFGPS"NPSUJ[FE.BSHJOBM4)"1 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ,FSOFM4)"1 'BTU4)"1 ূ૑פয݂ ӝ߈ оࣘച ঌҊ્ܻ ӝ߈ оࣘച 4".4)"1
4VSSPHBUFGPS"NPSUJ[FE.BSHJOBM4)"1 оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ 𝒇 × ≈ ߓ҃ ؘ੉ఠ ؘ੉ఠ о઺஖
𝒈 ؀ܻݽ؛ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ 9"*ঌҊ્ܻ оࣘച োҳ

оࣘച ঌҊ્ܻ ѐߊ • ,FSOFM4)"1 ؀࠺ ҅࢑ ࣘب ࡅܴ •
'BTU4)"1 ؀࠺ ࢸݺ۱ ֫਺ • 'BTU4)"1 ؀࠺ ҅࢑ ࣘب וܿ ੢੼ ױ੼ 4".4)"1 9"*ঌҊ્ܻ оࣘച োҳ

֤ޙ ౠೲ ࢲ࠺झ੸ਊ ࢑୹ޛ

,FSOFM4)"1 җ੄ r҅࢑ ࣘبs ࠺Ү 100% 42% ≪ 1% 10%
,FSOFM4)"1 CBTFMJOF оࣘച ,FSOFM4)"1 'BTU4)"1 4".4)"1 0VST 9"*؀࢚ ஠஠য়ߛ௼ "*ࢲ࠺झ ݽ؛ ࢑୹ޛ

,FSOFM4)"1 җ੄ rࢸݺ য়ରs ࠺Ү 1.00 1.00 0.24 0.92 ,FSOFM4)"1
CBTFMJOF оࣘച ,FSOFM4)"1 'BTU4)"1 4".4)"1 0VST $PTJOF4JNJMBSJUZ 5PQ4JHOFE3BOL"HSFFNFOU 1.00 1.00 0.02 0.69 ,FSOFM4)"1 CBTFMJOF оࣘച ,FSOFM4)"1 'BTU4)"1 4".4)"1 0VST 9"*؀࢚ ஠஠য়ߛ௼ "*ࢲ࠺झ ݽ؛ ࢑୹ޛ

݃ޖܻ

ࢸݺоמ ੋҕ૑מ F9QMBJOBCMF "*חj t#MBDLCPYuۄ ܻࠛח "*੄ ੑ
୹۱ ࢚ҙҙ҅ܳ ࢸݺೡ ࣻ ੓ח ߑߨۿ 9"*ܳ ా೧ ਋ܻחj ੋҕ૑מ੄ न܉ࢿਸ ѨૐೞҊ ࢿמਸ ѐࢶೞחؘ ഝਊ ೡ ࣻ ੓ਵݴ ಽҊ੗ ೞח ޙઁ੄ %PNBJO&YQFSU৬ ࣗాೞݴ ૑ध ֢ೞ਋ оࢸਸ ݽ؛ ѐߊী ੸ਊೡ ࣻ ੓ѱ ೣ पઁ "*ࢲ࠺झী ഝਊೞӝ ਤೠ ࢑೟ഈ۱ 3%ࣻ೯ ,FSOFM4)"1 ӝળ ࣘبח ୭؀ ߓ ࡅܰݶࢲ ࢶ೯ োҳੋ 'BTU4)"1 ࠁ׮ ੿ഛೠ 9"*ঌҊ્ܻ ѐߊ 정리

ࢿמ न܉ Ѿҗী ؀ೠ Ѩૐ ਗੋ ѐࢶ੼ ౵ঈ ࣗా بݫੋ
૑ध оࢸ ഝਊ

-VOECFSH 4 -FF 4 "6OJGJFE"QQSPBDIUP*OUFSQSFUJOH.PEFM1SFEJDUJPOT/FVS*14 4IBQMFZ
-4 "7BMVFGPSO1FSTPO(BNFT$POUSJCVUJPOTUPUIF5IFPSZPG(BNFT ". 7PMVNF** m $IVBOH :/ 8BOH ( :BOH ' -JV ; $BJ 9 %V . )V 9 &GGJDJFOUYBJ UFDIOJRVFT"UBYPOPNJDTVSWFZ BS9JW QSFQSJOUBS9JW +FTVT 4 1PNCBM + "MWFT % $SV[ " 4BMFJSP 1 3JCFJSP 3 j#J[BSSP 1 5VSOJOHUIF5BCMFT#JBTFE *NCBMBODFE %ZOBNJD5BCVMBS%BUBTFUTGPS.-&WBMVBUJPO /FVS*14 +FUIBOJ / 4VEBSTIBO . $PWFSU *$ -FF 4* 3BOHBOBUI 3 'BTU4)"13FBM5JNF4IBQMFZ7BMVF &TUJNBUJPO *$-3 ଵҊ ޙ೴

AI를 설명하면서 속도도 빠르게 할 순 없을까? SHAP 가속화 이야기 (feat. ...

AI를 설명하면서 속도도 빠르게 할 순 없을까? SHAP 가속화 이야기 (feat. 산학협력)

More Decks by kakao

Other Decks in Programming

Featured

Transcript