ܦࡁֶऀʹ஌ͬͯ΄͍͠ػցֶश ~൓ࣄ࣮ϞσϧʹΑΔ༧ଌ~ Kazuki Taniguchi 

• ৬ྺ • 2014.4-2019.3 • גࣜձࣾαΠόʔΤʔδΣϯτΞυςΫຊ෦ AI Lab • 2019.4- • ITܥϕϯνϟʔ (ϓϩμΫτ։ൃ/ϚʔέςΟϯά) • ϑϦʔϥϯε (AI/MLͷݚڀ։ൃ) • ݚڀ෼໺ • Pattern Recognition / Image Restoration • Recommendation / Response Prediction • Counterfactual ML ࣗݾ঺հ ୩ޱ ࿨ً (@kazk1018) 

Introduction 

(લஔ͖) ػցֶशͷఆٛ • ҰൠతͳػցֶशͷఆٛΑΓ΋ڱٛ f(x) .PEFM Input: Կ͔͠Βͷσʔλ (਺஋, ը૾, ςΩετ, etc…) Output: ग़ྗ݁Ռ (ϥϕϧ, ਺஋, ը૾, etc…) ༧ଌ ਪఆ 

ػցֶशٕज़ͷٸ଎ͳීٴ ʮ/&$ͷإೝূɺճ࿈ଓͰੈքҐʹʯ ʮ"SUXPSL
1FSTPOBMJ[BUJPO
BU
/FUqJYʯ ʮϥʔϝϯ࣍࿠ͷళฮΛ"VUP.-Ͱ൑ผʯ ʮ.BDIJOF
-FBSOJOH
BU
6CFSʯ 

ϥʔϝϯ࣍࿠ͷళฮΛAutoMLͰ൑ผ
• 1170ຕ × 41ళฮ = ໿48,000ຕͷը૾ΛϥϕϧͱηοτͰ༻ҙ • Google AutoML VisionΛར༻ͯ͠95%ͷਫ਼౓Λ࣮ݱ • ػցֶशͷ஌͕ࣝͳ͍ਓؒͰ΋ֶश࣌ؒ͸18෼ͰϞσϧΛ֫ಘ 

Artwork Personalization at Netflix
ϢʔβA ϢʔβB ʮGood Will HuntingʯͷΞʔτϫʔΫͷྫ 

ػցֶश΁ͷߋͳΔظ଴ ҩྍ਍அ ࣗಈӡస Ϗδωεߩݙ ༧ଌͷਫ਼౓͚ͩΛ௥͍ٻΊΔͷͰ͸ͳ͘ ݸਓ΍ࣾձʹରͯ͠ϦεΫ͕͋ΔྖҬ΁ͷظ଴ 

ػցֶश΁ͷߋͳΔظ଴ Explainability Stability (Robustness, Generalization) [1] ֶशͨ͠Ϟσϧͷઆ໌Մೳੑ ֶशͨ͠Ϟσϧͷ҆ఆੑ (ಛఆͷσʔλ΍λεΫʹґଘͤͣ, ظ଴ͨ͠ಈ࡞͕҆ఆͯ͠ಘΒΕΔ) 

ػցֶशͷࠜຊతͳ՝୊ • ಠཱಉ෼෍ (i.i.d.) • ֤ඪຊ͕ಉ֬͡཰෼෍͔ΒಠཱʹಘΒΕΔ • ػցֶशͷଟ͘ͷΞϧΰϦζϜʹԾఆͱͯ͠༻͍Δ • ໬౓ • େ਺ͷ๏ଇ • த৺ۃݶఆཧ • … • ݱ࣮͸ֶशσʔλͱ༧ଌ͢Δσʔλͷ෼෍͸ҟͳΔ [1] 

ػցֶशͷࠜຊతͳ՝୊ ֶशσʔλ ༧ଌ͢Δσʔλ ઇͷ্͸େৎ෉ʁ ͓΋ͪΌ͸େৎ෉ʁ ̋ ʁ ʁ ֶशσʔλʹ͍ۙ 

ػցֶशͷࠜຊతͳ՝୊ • ༧ଌ͍ͨ͠σʔλͷ෼෍ͷఆٛ (Ϟσϧͷཁ݅) • ʮHorseʯ • ʮHorse on the grassʯ • ʮHorse Toyʯ • બ୒όΠΞε • ར༻͢ΔαϯϓϧΛબ୒͢Δ࣌఺Ͱൃੜ͢ΔόΠΞε • ֶश͢Δσʔλֶ͕श͢Δσʔλͦͷ΋ͷʹґଘͯ͠બ୒͞Ε͍ͯ Δ৔߹ʹൃੜ 

޿ࠂ഑৴γεςϜͷࣄྫ 

޿ࠂ഑৴γεςϜ • Real Time Bidding (RTB) imp click Request RTB click͕؍ଌͰ͖ͳ͍ = ϥϕϧ͕ଘࡏ͠ͳ͍ win = ޿ࠂ͕දࣔ(imp)͞ΕΔ lose = ޿ࠂ͕දࣔ(imp)͞Εͳ͍ SSP DSP 

޿ࠂ഑৴γεςϜ • ΫϦοΫ༧ଌϞσϧ [2] *% ޿ࠂ Ϣʔβ JNQ DMJDL  " B :FT   # E /P  $ F /P  # C :FT   $ F :FT  ഑৴γεςϜͷϩάσʔλ 

޿ࠂ഑৴γεςϜ • ΫϦοΫ༧ଌϞσϧ [2] *% ޿ࠂ Ϣʔβ JNQ DMJDL  " B :FT   # E /P  $ F /P  # C :FT   $ F :FT  ഑৴γεςϜͷϩάσʔλ ֶशʹར༻Ͱ͖Δσʔλ 

޿ࠂ഑৴γεςϜͷ໰୊ઃఆ • ΫϦοΫ༧ଌϞσϧ [2] *% ޿ࠂ Ϣʔβ JNQ DMJDL  " B :FT   # E /P  $ F /P  # C :FT   $ F :FT  ഑৴γεςϜͷϩάσʔλ ࣮ࡍʹΫϦοΫΛ༧ଌ͍ͨ͠σʔλ ൓ࣄ࣮ 

޿ࠂ഑৴γεςϜͷ໰୊ઃఆ Ϟσϧͷֶशʹ༻͍Δσʔλͷ෼෍ͱ ༧ଌ͍ͨ͠σʔλͷ෼෍͕ҟͳΔ P(x) = q(win|x)P′(x) P′(x) ༧ଌ͍ͨ͠σʔλͷ֬཰෼෍: ֶशσʔλͷ֬཰෼෍: P(x) ≠ P′(x) ※ ͸RTBͰ ͕͖ͨͱ͖ͷwin͢Δ৚݅෇͖֬཰ q(win|x) x 

ػցֶशʹΑΔΞϓϩʔν • ڞมྔγϑτ (Covariate Shift) • Unsupervised Domain Adaptation 

ڞมྔγϑτ (Covariate Shift) 

ڞมྔγϑτ [3] • ҎԼͷ৚݅ͷ໰୊Λѻ͏ p(x) ≠ p′(x) p(y|x) = p′(y|x) ͸ ʹಠཱʹै͍, ͸ ʹ ಠཱʹै͏ͱԾఆ͢Δ D = {(xi , yi )}n i=1 p(x, y) D′ = {x′ i }m i=1 ∫ p′(x, y)dy ͜ͷͱ͖, ࣍ͷΑ͏ͳ৚݅Λຬͨ͢ͱ͖Λڞมྔγϑτͱ͍͏ 

ڞมྔγϑτԼͷ༧ଌϞσϧ • ͱ Λ༻͍ͯ৽ͨͳೖྗ ʹର͢Δऔಘ Λ ༧ଌ͢ΔϞσϧ Λֶश͍ͨ͠ • ଛࣦؔ਺Λ ͱ͢Δͱ͖ڞมྔγϑτ͸ॏཁ౓ॏΈ෇ ͚Λ༻͍Δ͜ͱͰղܾ͢Δ͜ͱ͕஌ΒΕ͍ͯΔ [3] {(xi , yi )}n i=0 {x′ j }m j=0 x′ y fθ (x) loss(y, fθ ) minθ n ∑ i=0 w(xi )loss(yi , fθ (xi ))) w(xi ) = p′(xi ) p(xi ) 

ॏཁ౓ॏΈ෇͚Λ༻͍ͨΫϦοΫ༧ଌ P(x) ≠ P′(x) ※ ͸RTBͰ ͕͖ͨͱ͖ͷwin͢Δ৚݅෇͖֬཰ q(win|x) x ΛԾఆ͢Δͱ,
 ॏཁ౓ॏΈ෇͚ʹΑͬͯ༧ଌϞσϧΛֶश͢Δ͜ͱ͕Մೳ p(y|x) = p′(y|x) P(x) = q(win|x)P′(x) P′(x) ༧ଌ͍ͨ͠σʔλͷ֬཰෼෍: ֶशσʔλͷ֬཰෼෍: 

ॏཁ౓ॏΈ෇͚Λ༻͍ͨΫϦοΫ༧ଌ • ॏཁ౓͸RTBʹ͓͍ͯwin͢Δ֬཰ͷٯ਺ • treatmentΛwinͱͨ͠ͱ͖ͷInverse probability of treatment weighting (IPTW) [4]ͱಉ༷ • ਅͷ ͸؍ଌͰ͖ͳ͍ͨΊPropensity ScoreΛਪఆ q(win|x) w(x) = 1 q(win|x) 

Unsupervised Domain Adaptation 

Unsupervised Domain Adaptation • Domain Adaptation (υϝΠϯదԠ) • े෼ͳ৘ใͷ͋ΔυϝΠϯ(Source Domain)ͷ஌ࣝΛ৘ใ͕গͳ͍
 υϝΠϯ(Target Domain)ʹదԠ͢Δٕज़ • సҠֶशͷҰछ • Unsupervised Domain Adaptation • Target Domainʹڭࢣϥϕϧ͕ͳ͍υϝΠϯదԠ 

Domain Adaptation Neural Network (DANN) [5] labelͷ༧ଌʹ͓͚Δଛࣦ domainͷ༧ଌʹ͓͚Δଛࣦ ( ̂ θy , ̂ θf ) = arg min θy ,θf L(y, d, x) ̂ θd = arg max θd L(y, d, x) ಛ௃දݱ ͸υϝΠϯͷݟ෼͚͕͔ͭͳ͘ͳΔΑ͏ʹ ϥϕϧͷ༧ଌʹ͓͚ΔଛࣦΛ࠷খԽ͢Δ f 

CTR Prediction by using DANN [6] • DANN͸ը૾Λର৅ͱͨ͠CNNϕʔεͷϞσϧ • RTBͷςʔϒϧσʔλʹରԠ͢ΔΑ͏ʹϞσϧΛվྑ 

CTR Prediction by using DANN [6] • ࣮ݧ • σʔληοτ • Criteo CTR Prediction Contest (Kaggle) • ൺֱख๏ • Baseline: Deep Neural Network (Only source dataset) • Importance Sampling 

4. ·ͱΊ 

·ͱΊ • ػցֶशͷٸ଎ͳීٴͱظ଴ • ExplainabilityͱStability • ػցֶशͷ՝୊ • ޿ࠂ഑৴γεςϜͷࣄྫ • ޿ࠂ഑৴γεςϜͷ֓ཁ • ൓ࣄ࣮Λߟྀͨ͠ػցֶशʹΑΔΫϦοΫ༧ଌ • ڞมྔγϑτ • Unsupervised Domain Adaptation 

References 1. Kun Kuang, Peng Cui, Susan Athey, Ruoxuan Xiong, and Bo Li, “Stable Prediction across Unknown Environments”, KDD, 2018 2. Oliver Chapelle, Eren Manavoglu, Romer Rosales, “Simple and scalable response prediction for display advertising”, TIST, 2015 3. Hidetoshi Shimodaira, “Improving predictive inference under covariate shift by weighting the log-likelihood function”, JSPI, 2000 4. James M. Robins, Andrea Rotnitzky, Lue Ping Zhao, “Estimation of Regression Coefficients When Some Regressors Are Not Always Observed”, JASA, 1994 5. Yaroslav Ganin, Victor Lempitsky, “Unsupervised Domain Adaptation by Backpropagation”, JMLR, 2015 6. ୩ޱ ࿨ً, ҆Ҫ ᠳଠ, “Domain Adaptation Neural NetworksΛ༻͍ͨΫϦοΫ༧ଌ”, JSAI, 2019 

No content