分散システムの性能異常に対する機械学習の解釈性に基づく原因診断手法 / A Method for Diagnosing the Causes of Performance Issues in Distributed Systems Based on the Interpretability of Machine Learning

by tsurubee

Slide 1

Slide 1 text

Slide 2

Slide 2 text

2 1. എܠͱ໨త 2. ػցֶशͷղऍੑ 3. ఏҊ͢ΔݪҼ਍அख๏ 4. ධՁͱߟ࡯ 5. ·ͱΊͱࠓޙͷల๬ ໨࣍

Slide 3

Slide 3 text

1. എܠͱ໨త

Slide 4

Slide 4 text

4 γεςϜͷෳࡶԽͱӡ༻ͷ՝୊ γεςϜͷෳࡶԽ ӡ༻ͷ՝୊ γεςϜ؅ཧऀͷೝ஌ෛՙ͕ߴ·Δ ϝτϦοΫͷܥྻ਺ͷ૿େ ϝτϦοΫͷܥྻ਺ͷมԽ γεςϜ؅ཧऀ͸ɼ֤ߏ੒ཁૉͷੑೳΛఆྔతʹࣔ͢ࢦඪ (Ҏ߱ɼϝτϦοΫ) Λऩू͠ɼ γεςϜͷঢ়ଶͷ؂ࢹʹ׆༻͍ͯ͠Δɽ ߏ੒ཁૉ਺ͷมԽ ߏ੒ཁૉ਺ͷ૿େ • Ϋϥ΢υ্ͷγεςϜ͸ɼෳ਺ͷҟͳΔߏ੒ཁૉ (αʔό΍ίϯςφͳͲ) ͕૬ޓʹωοτϫʔΫ௨৴ ͢Δ෼ࢄγεςϜͱͯ͠ߏ੒͞ΕΔɽ • ͜ΕΒͷγεςϜ͸ɼར༻ऀ͔Βͷଟ༷ͳཁٻʹԠ͑ΔͨΊʹҎԼͷ܏޲͕ݟΒΕΔɽ • γεςϜߏ੒͕ෳࡶԽ͠ɼߏ੒ཁૉ਺΍छྨ͕૿େ͍ͯ͠Δɽ • γεςϜ΁ͷมߋස౓͕ߴ͘ͳΓɼߏ੒ཁૉ਺΍ෛՙಛੑͷมԽ͕଎͘ͳ͍ͬͯΔɽ γεςϜʹҟৗ͕ൃੜͨ͠ࡍʹɼͦͷݪҼಛఆʹཁ͢Δ͕࣌ؒ௕͘ͳΓɼαʔϏεͷఀࢭ࣌ؒͷ૿େ ΍ػձଛࣦͳͲʹܨ͕ΔͨΊɼҟৗͷݪҼΛࣗಈత͔ͭਝ଎ʹ਍அͰ͖Δख๏͕ඞཁͰ͋Δɽ

Slide 5

Slide 5 text

5 ઌߦख๏ɿػցֶशΛ༻͍ͨΞϓϩʔν • طଘख๏Ͱ͸ɼֶशʹ௕͍࣌ؒΛཁ͢Δਂ૚ֶशϕʔεͷϞσϧ͕༻͍ΒΕ͍ͯΔͨΊɼࣄલʹϞσϧ Λֶश͓ͯ͘͠ඞཁ͕͋Δɽ • Ϟσϧͷֶशͷࡍͷೖྗಛ௃ྔͷ਺͸ݻఆͰ͋ΔͨΊɼϝτϦοΫͷܥྻ਺͕૿ݮͨ͠৔߹͸ɼ࠶౓ɼ ৽ͨͳϞσϧΛֶश͠ͳ͚Ε͹ͳΒͳ͍ɽ ΑͬͯɼγεςϜͷมߋʹਝ଎ʹ௥ैͨ͠ݪҼ਍அΛߦ͏͜ͱ͕೉͍͠ɽ ※3 L. Wu et al, Performance Diagnosis in Cloud Microservices using Deep Learning, International Conference on Service-Oriented Computing, 2020. ※4 H. Zhao et al, Multivariate Time-series Anomaly Detection via Graph Attention Network, arXiv:2009.02040 2020. MTAD-GATͷΞʔΩςΫνϟ※3,4 2. ҟৗ౓ͷࢉग़ 3. ߩݙ౓ͷࢉग़ ࣮ߦͷྲྀΕ 2. ҟৗ౓ͷࢉग़ ֶशޙͷϞσϧͷ༧ଌ஋ͱ࣮ଌ஋ͱͷޡࠩΛҟৗ౓ ͱͯ͠ࢉग़ 3. ߩݙ౓ͷࢉग़ ҟৗ౓ʹର͢Δೖྗಛ௃ྔ (ϝτϦοΫ)ͷߩݙ౓Λࢉग़ 1. Ϟσϧͷֶश ਖ਼ৗ࣌ͷϝτϦοΫͷܥྻΛػցֶशϞσϧΛ༻͍ͯ ֶश ػցֶशΛ༻͍ͯҟৗͷݪҼͱͳΔϝτϦοΫΛ਍அ͢Δख๏͕੝Μʹݚڀ͞Ε͓ͯΓɼߴ͍਍அਫ਼౓ Λୡ੒͍ͯ͠Δ※3,4ɽ

Slide 6

Slide 6 text

6 ݚڀͷ໨త ຊൃදͰ͸ɼҟৗͷݕ஌ޙʹߴ଎ʹֶशՄೳͰ͋ΔܰྔͳػցֶशϞσϧͱɼػցֶशͷղऍख๏ ͱͯ͠஫໨͞Ε͍ͯΔγϟʔϓϨΠ஋Λ༻͍ͯɼγεςϜߏ੒ͷมߋʹର͢Δ௥ैੑͱߴ͍਍அਫ਼ ౓Λཱ྆ͨ͠ݪҼ਍அख๏ΛఏҊ͢Δɽ γεςϜߏ੒ͷมߋʹର͢Δ௥ैੑͱߴ͍਍அਫ਼౓Λཱ྆ͨ͠ݪҼ਍அख๏ͷఏҊΛ໨తͱ͢Δɽ • ߴ଎ʹֶश͕ՄೳͰ͋ΔܰྔͳػցֶशϞσϧΛ ༻͍ͯɼҟৗͷݕ஌ޙʹϞσϧͷֶशΛߦ͏ɽ • γεςϜͷมߋʹ௥ैͯ͠ϞσϧΛֶश͠௚͢ඞ ཁ͕ͳ͘ɼҟৗൃੜ࣌ͷγεςϜͷߏ੒Λ൓ө͠ ͨݪҼ਍அ͕Ͱ͖Δ • ܰྔͳػցֶशϞσϧ͸ਂ૚ֶशΛ༻͍ͨϞσϧ ΑΓ΋දݱྗ͕௿͍ͨΊɼݪҼ਍அͷਫ਼౓͕௿͘ ͳΔՄೳੑ͕͋Δɽ • ػցֶशͷղऍੑͷ෼໺Ͱ༗༻ੑ͕ࣔ͞Ε͍ͯΔ γϟʔϓϨΠ஋Λ༻͍ͯ਍அਫ਼౓ΛߴΊΒΕΔ͔ ݕ౼͢Δɽ ఏҊͷ֓ཁ

Slide 7

Slide 7 text

2. ػցֶशͷղऍੑ

Slide 8

Slide 8 text

8 ػցֶशͷղऍੑ • ਂ૚ֶशΛ͸͡Ίͱ͢ΔػցֶशϞσϧ͸ͦͷܭࢉաఔ͕ෳࡶͰ͋ΔͨΊɼϞσϧͷ༧ଌ΍൑அ ͷࠜڌΛਓ͕ؒཧղ͢Δ͜ͱ͕Ͱ͖ͳ͍ϒϥοΫϘοΫεͱͳΔ͜ͱ͕໰୊ࢹ͞Ε͓ͯΓɼػց ֶशϞσϧͷղऍੑʹ͍ͭͯͷݚڀ͕஫໨͞Ε͍ͯΔ※5ɽ • AIར׆༻ݪଇҊ (૯຿লɼ2018೥) • ಁ໌ੑͷݪଇɼΞΧ΢ϯλϏϦςΟ(આ໌੹೚) ͷݪଇ • DARPA (ถࠃ๷ߴ౳ݚڀܭըہ) • Explainable Artiﬁcial Intelligence (XAI) ϓϩδΣΫτ ※5 A. Adadi and M. Berrada, Peeking Inside the Black-Box: A Survey on Explainable Artiﬁcial Intelligence (XAI), IEEE Access, 2018. ػցֶशͷղऍੑɾઆ໌ੑʹؔ͢Δ࿦จ਺ͷਪҠ※5 • ಛʹɼػցֶशϞσϧͷ༧ଌʹର͢Δೖྗಛ௃ྔͷ ߩݙ౓Λఏࣔ͢Δख๏͕਺ଟ͘ఏҊ͞Ε͍ͯΔɽ

Slide 9

Slide 9 text

9 γϟʔϓϨΠ஋Λ༻͍ͨղऍ ※6 S. Lundberg and S. I. Lee, A Uniﬁed Approach to Interpreting Model Predictions, Advances in Neural Information Processing Systems 30(NIPS 2017), 2017. ※ https://github.com/slundberg/shap ڠྗήʔϜͱػցֶशͷରൺ ήʔϜ ϓϨΠϠʔ རಘ ಛ௃ྔ ༧ଌ஋ Ϟσϧ ػցֶश ϓϨʔϠʔΛಛ௃ྔɼརಘΛ༧ଌ஋ʹஔ͖׵͑ ը૾ͷղऍͷྫ※ • ୅දతͳղऍख๏ͱͯ͠ɼڠྗήʔϜཧ࿦ͷγϟʔϓϨΠ஋Λ༻͍ͨղऍख๏͕͋Δ※6ɽ • γϟʔϓϨΠ஋ͱ͸ɼڠྗήʔϜʹ͓͍ͯෳ਺ϓϨΠϠʔͷڠྗʹΑͬͯಘΒΕͨརಘΛ֤ϓϨΠϠʔͷ ߩݙ౓ʹԠͯ͡ެਖ਼ʹ෼഑͢ΔͨΊͷखஈͷҰͭ • γϟʔϓϨΠ஋Ͱ͸ɼ͋ΒΏΔϓϨΠϠʔͷ૊Έ߹ΘͤͰಘΒΕΔརಘΛݩʹɼଞͷϓϨΠϠʔͱͷڠྗ ʹΑΓൃੜ͢ΔརಘͷӨڹΛഉআͨ͠ಛఆͷϓϨΠϠʔͷߩݙ౓Λࢉग़

Slide 10

Slide 10 text

10 ղऍੑΛ༻͍ͨҟৗͷݪҼ਍அ • Ϟσϧͱͯ͠ɼೖྗσʔλʹର͢Δҟৗ౓Λग़ྗ͢ΔϞσϧΛ༻͍ͨ৔߹ɼγϟʔϓϨΠ஋Λ༻͍ͯɼ ҟৗ౓ʹର͢Δ֤ಛ௃ྔͷߩݙ౓ΛٻΊΒΕΔͨΊɼҟৗͷݪҼಛఆʹ׆༻Ͱ͖Δɽ • γϟʔϓϨΠ஋Λ༻͍ͨղऍख๏͸ɼҟৗͷݪҼ਍அʹ༗༻Ͱ͋Δ͜ͱΛࣔ݁͢Ռ͕ใࠂ͞Ε͍ͯΔ※7,8ɽ ಛ௃ྔ ༧ଌ஋ Ϟσϧ ػցֶश ϝτϦοΫ (ex. CPU usage) ҟৗ౓ • γϟʔϓϨΠ஋͸ɼಛ௃ྔͷ਺͕૿͑ΔʹͭΕͯܭࢉྔ͕๲େʹͳΔɽ • ଈ࣌ੑ͕ٻΊΒΕΔҟৗͷݪҼ਍அʹ͓͍ͯɼγϟʔϓϨΠ஋ͷܭࢉ͕ ࣮༻తͳ࣌ؒ಺ʹܭࢉՄೳͰ͋Δ͔ͷٞ࿦͸ߦΘΕ͍ͯͳ͍ɽ γϟʔϓϨΠ஋ΛݪҼ਍அʹ༻͍Δࡍͷ՝୊ ※7 L. Antwarg et al., Explaining Anomalies Detected by Autoencoders Using SHAP, arXiv:1903.02407, 2019. ※8 N. Takeishi and Y. Kawahara, On Anomaly Interpretation via Shapley Values, arXiv:2004.04464, 2020.

Slide 11

Slide 11 text

3. ఏҊ͢ΔݪҼ਍அख๏

Slide 12

Slide 12 text

12 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌

Slide 13

Slide 13 text

13 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌ • ػցֶशϞσϧ͸ɼݱࡏҟৗݕ஌ʹ޿͘ར༻͞Ε͍ͯΔ͕ɼఏҊख๏Ͱ͸ɼػցֶशϞσϧΛ༻͍ͯ ҟৗݕ஌ΛߦΘͣɼҟৗ͕ݕ஌͞ΕͨޙͷݪҼ਍அΛߦ͏͜ͱʹয఺Λ౰͍ͯͯΔɽ • ҟৗݕ஌ͷͨΊʹ͸ɼαʔϏεͷ৴པੑʹؔ͢Δ໨ඪ஋Ͱ͋ΔSLOʢService Level Objectiveʣ΍ ϝτϦοΫͷܥྻ͝ͱʹઃఆͨ͠ᮢ஋ͳͲΛ༻͍Δɽ

Slide 14

Slide 14 text

14 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌ • ҟৗ͕ݕ஌͞Εͨ৔߹ɼγεςϜͷ֤ߏ੒ཁૉ͔ΒಘΒΕΔϝτϦοΫΛݕ஌ͷ͔࣌ؒΒҰఆظؒ Ḫͬͯऔಘ͢Δɽ • औಘͨ͠ෳ਺ͷϝτϦοΫͷܥྻʹ͸ɼҟৗ࣌ͷϝτϦοΫͷܥྻͱɼͦͷ௚ۙͷਖ਼ৗ࣌ͷϝτ ϦοΫͷܥྻؚ͕·Ε͍ͯΔͨΊɼ্ਤͷΑ͏ʹϝτϦοΫͷܥྻΛ࣌ؒͷ࣠ʹରͯ͠ਖ਼ৗ࣌ͱҟ ৗ࣌ʹ෼ׂ͢Δɽ

Slide 15

Slide 15 text

15 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌ • ෼ׂͨ͠σʔλͷ͏ͪɼਖ਼ৗ࣌ͷϝτϦοΫͷΈΛֶशσʔλͱͯ͠ɼػցֶशϞσϧͷڭࢣͳ͠ ֶशΛߦ͏ɽ

Slide 16

Slide 16 text

16 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌ • ҟৗ࣌ͷϝτϦοΫΛֶशࡁΈͷϞσϧʹೖྗ͠ɼϞσϧ͔Βͷग़ྗͱͯ͠ҟৗ౓ΛಘΔɽ

Slide 17

Slide 17 text

17 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌ • ҟৗ౓ʹର͢Δ֤ϝτϦοΫͷߩݙ౓ΛγϟʔϓϨΠ஋Λ༻͍ͯࢉग़͠ɼࢉग़ͨ͠ߩݙ౓Λ߱ॱʹ ੔ྻ͢Δ͜ͱͰɼҟৗͷݪҼͱͳΔϝτϦοΫͷީิΛϥϯΫ෇͚͢Δɽ • ϥϯΫ෇͚ͨ͠ϝτϦοΫΛ਍அ݁Ռͱͯ͠ɼγεςϜ؅ཧऀʹఏࣔ͢Δɽ

Slide 18

Slide 18 text

18 ΞʔΩςΫνϟ֓ཁ ఏҊ͢ΔݪҼ਍அख๏ͷ֓ཁਤ 4. γϟʔϓϨΠ஋ʹΑΔ਍அ ࣮ߦͷྲྀΕ 3. ҟৗ౓ͷࢉग़ 2. ػցֶशϞσϧͷֶश 1. ϝτϦοΫͷऩूͱ෼ׂ 0. ҟৗͷݕ஌ 2ͭͷεςοϓʹ͍ͭͯ ΑΓৄࡉʹઆ໌͢Δ

Slide 19

Slide 19 text

19 ػցֶशϞσϧͷֶश • ਝ଎ʹ਍அ݁ՌΛఏࣔ͢ΔͨΊʹɼϞσϧͷֶशΛߴ଎ʹ࣮ߦՄೳͰ͋Δ͜ͱɽ • Ϟσϧ͸ೖྗσʔλʹରͯ͠ɼҟৗ౓Λฦ͢Α͏ͳϞσϧͰ͋Δ͜ͱɽ ఏҊख๏Ͱ͸ɼγεςϜߏ੒ͷมߋʹର͢Δ௥ैੑͷཁ݅Λຬͨͨ͢ΊʹɼػցֶशϞσϧͷֶश ΛࣄલʹߦΘͣʹҟৗݕ஌ޙʹߦ͏ΞʔΩςΫνϟΛ࠾༻͢Δɽ Ϟσϧʹର͢Δཁ݅ Ϟσϧͷީิ • ্هͷཁ݅Λຬͨ͢Ϟσϧͱͯ͠ɼओ੒෼෼ੳ (PCA) ϕʔεͷϞσϧ͕ڍ͛ΒΕΔɽ • PCA͸ɼߴ࣍ݩͷσʔλΛΑΓ௿࣍ݩͷσʔλʹѹॖ͢Δ࣍ݩ ѹॖख๏ͷҰͭͰ͋Δɽ • ਖ਼ৗ࣌ͷσʔλͷΈΛ༻͍ͯɼ௿࣍ݩ΁ͷѹॖํ๏Λֶश͠ɼ ֶशޙʹ༩͑ΒΕΔಛఆͷσʔλʹରͯ͠ɼҰ౓࣍ݩΛѹॖ͠ ͨޙʹ࠶౓ݩͷ࣍ݩʹσʔλΛ෮ݩͨ͠ࡍͷޡࠩ (࠶ߏ੒ޡࠩ) Λҟৗ౓ͱͯ͠ղऍͰ͖Δɽ ѹॖ ෮ݩ ೖྗ ϝτϦοΫ m1 m2 mN ⋅ ⋅ ⋅ m′ 1 m′ 2 m′ N ⋅ ⋅ ⋅ e1 = m′ 1 − m1 ࠶ߏ੒ޡࠩ e2 eN

Slide 20

Slide 20 text

20 γϟʔϓϨΠ஋ʹΑΔ਍அ • ਝ଎ʹ਍அ݁ՌΛఏࣔ͢ΔͨΊʹɼγϟʔϓϨΠ஋ͷܭࢉΛߴ଎ʹߦ͏͜ͱɽ • γϟʔϓϨΠ஋͸ɼಛ௃ྔͷ਺͕૿େ͢ΔʹͭΕͯݫີͳܭࢉ͕ෆՄೳͰ͋ΔͨΊɼγϟʔϓϨΠ஋ Λߴ଎͔ͭਖ਼֬ʹۙࣅ͢Δख๏͕ඞཁɽ γϟʔϓϨΠ஋ͷܭࢉʹٻΊΒΕΔཁ݅ γϟʔϓϨΠ஋ͷۙࣅख๏ • ্هͷཁ݅Λຬͨ͢γϟʔϓϨΠ஋ͷۙࣅख๏ͱͯ͠ɼSHAP͕ڍ͛ΒΕΔɽ • SHAPʹ͸ɼ͍͔ͭ͘ͷۙࣅΞϧΰϦζϜ͕ଘࡏ͢Δ͕ɼͦͷதͰ΋͋ΒΏΔػցֶशϞσϧʹ ద༻ՄೳͳKernel SHAP※6͕༗ޮͰ͋Δɽ • Kernel SHAP͸ɼॏΈ෇͖ͷ࠷খೋ৐ઢܗճؼʹΑΓγϟʔϓϨΠ஋ΛۙࣅͰ͖Δɽ ఏҊख๏Ͱ͸ɼߴ͍਍அਫ਼౓Λୡ੒͢ΔͨΊʹɼҟৗ౓ʹର͢Δ֤ϝτϦοΫͷߩݙ౓ΛγϟʔϓϨΠ ஋Λ༻͍ͯࢉग़͠ɼҟৗͷݪҼͱͳΔϝτϦοΫΛಛఆ͢ΔΞʔΩςΫνϟΛ࠾༻͢Δ ※6 S. Lundberg and S. I. Lee, A Uniﬁed Approach to Interpreting Model Predictions, Advances in Neural Information Processing Systems 30(NIPS 2017), 2017.

Slide 21

Slide 21 text

4. ධՁͱߟ࡯

Slide 22

Slide 22 text

࣮ݧͷ؀ڥ 22 ※ ௶಺༎थ, ੨ࢁਅ໵, MeltriaɿϚΠΫϩαʔϏεʹ͓͚Δҟৗݕ஌ɾݪҼ෼ੳͷͨΊͷσʔληοτͷಈతੜ੒γεςϜ, Πϯλʔωοτͱӡ༻ٕज़γϯϙδ΢Ϝ࿦จू, 2021, 63-70 (2021-11-18), 2021೥11݄ Fron-end Catalogue Orders Carts User Payment Shipping Sock Shop Locust Prometheus ϚΠΫϩαʔϏεΫϥελ ੍ޚαʔό ֎෦ෛՙͷੜ੒ ϝτϦοΫͷ ऩूɾอଘ ղੳαʔό ϝτϦοΫ औಘϞδϡʔϧ ղੳϞδϡʔϧ 4core, 24GB • KubernetesΫϥελͷࣗಈ؅ཧαʔϏεͰ͋ΔGKE (Google Kubernetes Engine)্ʹϚΠΫϩαʔϏεͷ ϕϯνϚʔΫΞϓϦέʔγϣϯͰ͋ΔSock ShopΛߏஙͨ͠ɽ • Sock ShopΛߏ੒͢Δίϯςφ͔ΒcAdvisorΛ༻͍ͯCPU࢖༻཰ͳͲͷϝτϦοΫΛ15ඵִؒͰऩूͨ͠ɽ • ੑೳҟৗΛٖࣅతʹੜ੒͢ΔͨΊʹɼCPUաෛՙ͓ΑͼϝϞϦϦʔΫΛಛఆͷίϯςφ্Ͱ࠶ݱͨ͠ɽ ࠶ݱํ๏ͷৄࡉ͸ɼIOTS2021ͰൃදͷMeltriaͷ࿦จ※Λࢀর

Slide 23

Slide 23 text

࣮ݧͷઃఆ 23 ධՁ಺༰ͱࢦඪ 1. ݪҼͷ਍அਫ਼౓ 2. ਍அʹཁ͢Δ࣌ؒ (Ҏ߱ɼ਍அ࣌ؒ) Top-kਫ਼౓ɿݪҼͱͳΔϝτϦοΫͷܥྻ্͕Ґk൪໨·Ͱʹ͋Δ֬཰ʢk=1͓Αͼk=3ʣ ਍அʹ༻͍ΔϝτϦοΫ ਍அʹ༻͍ΔίϯςφͷϝτϦοΫ ֤ίϯςφ͔Βӈදʹڍ͛ͨ6ݸͷϝτϦοΫΛऔಘ ҟৗΛ࠶ݱޙɼ5෼ܦաͨ͠ͷͪʹ௚ۙ30෼ؒͷίϯςφ ͷϝτϦοΫΛऔಘ ݪҼ਍அख๏ ఏҊख๏ɿػցֶशϞσϧʹPCAΛɼγϟʔϓϨΠ஋ͷܭࢉʹKernel SHAPΛ༻͍ͨ (PCA+SHAP) 1. Gaussian Based Thresholding (GBT)Λ༻͍ͨख๏ (GBT) 2. PCAͷ࠶ߏ੒ޡࠩΛ༻͍ͨख๏ (PCA+࠶ߏ੒ޡࠩ) ϝτϦοΫͷܥྻ͝ͱͷਖ਼ৗ࣌ͷฏۉ஋ͱҟৗ࣌ͷฏۉ஋ͷࠩ෼Λҟৗ ΁ͷߩݙ౓ͱͯ͠ݪҼ਍அΛߦ͏ɽ PCA͕ग़ྗ͢ΔϝτϦοΫ͝ͱͷ༧ଌ஋ͱ࣮ଌ஋ͱͷޡࠩͰ͋Δ ࠶ߏ੒ޡࠩΛҟৗ΁ͷߩݙ౓ͱͯ͠ݪҼ਍அΛߦ͏ɽ ϕʔεϥΠϯͱ͢Δख๏ɿ ѹॖ ෮ݩ ೖྗ ϝτϦοΫ ૯࿨ m1 m2 mN ⋅ ⋅ ⋅ m′ 1 m′ 2 m′ N ⋅ ⋅ ⋅ N ∑ i ei SHAP e1 = m′ 1 − m1 ࠶ߏ੒ޡࠩ e2 eN

Slide 24

Slide 24 text

24 ਍அਫ਼౓ͷධՁ ݁Ռͱߟ࡯ɿ਍அਫ਼౓ͷൺֱ ධՁ݁Ռ શͯͷ਍அظؒʹ͓͍ͯ਍அਫ਼౓͸ɼPCA +SHAP (ఏҊख๏) > PCA+࠶ߏ੒ޡࠩ > GBT Top-1ਫ਼౓ Top-3ਫ਼౓ ਍அظؒͷఆٛ ϝτϦοΫͷऔಘظؒͷ͏ͪɼҟৗ࣌ͷϝτϦοΫͷ ࣌ؒൣғʢҎ߱ɼ਍அظؒʣΛมԽͤͯ͞਍அਫ਼౓Λ ධՁͨ͠ɽ

Slide 25

Slide 25 text

25 ਍அਫ਼౓ͷධՁ ݁Ռͱߟ࡯ɿ਍அਫ਼౓ͷൺֱ ධՁ݁Ռ શͯͷ਍அظؒʹ͓͍ͯ਍அਫ਼౓͸ɼPCA +SHAP (ఏҊख๏) > PCA+࠶ߏ੒ޡࠩ > GBT • GBTͰ͸ɼϝτϦοΫ͝ͱʹฏۉ஋ͷมԽΛݟΔख๏Ͱ͋ΓɼϝτϦοΫͷܥྻؒͷ૬ؔΛऔΓѻΘͳ͍ɽ • Ұํɼίϯςφಉ࢜͸ɼ૬ޓͷωοτϫʔΫ௨৴΍ɼϗετͱͳΔϊʔυͷڞ༗ͷͨΊɼϝτϦοΫͷܥྻؒ ʹ૬ؔΛ࣋ͭɽ • PCAΛ༻͍Δ͜ͱͰɼϝτϦοΫͷܥྻؒͷ૬ؔΛߟྀͨ͠਍அ͕ՄೳͱͳΔɽ Top-1ਫ਼౓ Top-3ਫ਼౓ ਍அظؒͷఆٛ ϝτϦοΫͷऔಘظؒͷ͏ͪɼҟৗ࣌ͷϝτϦοΫͷ ࣌ؒൣғʢҎ߱ɼ਍அظؒʣΛมԽͤͯ͞਍அਫ਼౓Λ ධՁͨ͠ɽ

Slide 26

Slide 26 text

26 ਍அਫ਼౓ͷධՁ ݁Ռͱߟ࡯ɿ਍அਫ਼౓ͷൺֱ ධՁ݁Ռ શͯͷ਍அظؒʹ͓͍ͯ਍அਫ਼౓͸ɼPCA +SHAP (ఏҊख๏) > PCA+࠶ߏ੒ޡࠩ > GBT • PCA+࠶ߏ੒ޡࠩͰ͸ɼҟৗͷݪҼͱͳΔϝτϦοΫͷܥྻ͕มಈͨ͠৔߹ɼਖ਼ৗ࣌ʹ૬͕ؔߴ͔ͬͨϝτ ϦοΫͷ࠶ߏ੒ޡࠩ΋େ͖͘ͳΓɼҟৗ΁ͷߩݙ౓͕ߴ͘ݟੵ΋ΒΕΔ͜ͱ͕͋Δɽ • SHAPΛ༻͍ͨख๏Ͱ͸ɼγϟʔϓϨΠ஋ͷੑ࣭ʹΑΓɼਖ਼ৗ࣌ͷ૬ؔͷߴ͞ʹΑΔӨڹΛഉআͯ͠ɼݸʑͷ ϝτϦοΫͷߩݙ౓ΛܭࢉͰ͖Δɽ Top-1ਫ਼౓ Top-3ਫ਼౓ ϝτϦοΫͷऔಘظؒͷ͏ͪɼҟৗ࣌ͷϝτϦοΫͷ ࣌ؒൣғʢҎ߱ɼ਍அظؒʣΛมԽͤͯ͞਍அਫ਼౓Λ ධՁͨ͠ɽ ਍அظؒͷఆٛ

Slide 27

Slide 27 text

27 ਍அਫ਼౓ͷධՁ ݁Ռͱߟ࡯ɿ਍அظؒͷӨڹ ධՁ݁Ռ • ਍அظؒͷઃఆ͕਍அͷਫ਼౓ʹେ͖͘ӨڹΛ༩͑Δ͜ͱ͕Θ͔ͬͨɽ • ਍அظؒ < 5෼ɿػցֶशϞσϧʹ༩͑Δֶशσʔλʹҟৗ࣌ͷσʔλ͕ ؚ·ΕΔͨΊɼϞσϧͷֶश͕͏·͘ߦ͑ͣʹਫ਼౓͕௿Լ͢Δɽ • ਍அظؒ > 5෼ɿϞσϧ͸ਖ਼ৗ࣌ͷσʔλͷΈͰֶशͰ͖Δ΋ͷͷɼҟৗ ΁ͷߩݙ౓ͷܭࢉʹਖ਼ৗ࣌ͷσʔλؚ͕·ΕΔ͜ͱʹΑΓਫ਼౓͕௿Լ͢Δ ਍அظؒΛ5෼ͱͨ͠৔߹͕ɼ֤ܥྻΛ࠷΋Α͘ਖ਼ৗ ࣌ͱҟৗ࣌ʹ෼ׂͰ͖Δ Top-1ਫ਼౓ Top-3ਫ਼౓ ਍அظؒͷఆٛ औಘͨ͠ϝτϦοΫ͔ΒదԠతʹ ਍அظؒΛઃఆ͢Δख๏ͷݕ౼͕ ࠓޙͷ՝୊ ϝτϦοΫͷऔಘظؒͷ͏ͪɼҟৗ࣌ͷϝτϦοΫͷ ࣌ؒൣғʢҎ߱ɼ਍அظؒʣΛมԽͤͯ͞਍அਫ਼౓Λ ධՁͨ͠ɽ

Slide 28

Slide 28 text

28 ਍அ࣌ؒͷධՁɿCPUίΞ਺ ධՁ݁Ռ • CPUίΞ਺͕૿Ճ͢ΔʹͭΕͯఏҊख๏ͷ਍அ͕࣌ؒ୹͘ͳͬͨɽ • PCAͷֶश͸2 msͰ͋Γɼ਍அ࣌ؒʹରͯ͠ͷӨڹ͕ແࢹͰ͖Δ΄Ͳখ͍͞ɽ͕ͨͬͯ͠ɼఏҊख๏ͷ਍அ࣌ؒ ͸ɼSHAPʹΑΔߩݙ౓ͷܭࢉʹཁ͢Δ࣌ؒͱ΄΅ಉ౳Ͱ͋Δɽ ਍அظؒɿ5෼ ϝτϦοΫͷܥྻ਺ɿ66 (11ίϯςφ×6ϝτϦοΫ) Kernel SHAPͷܭࢉɿϚϧνίΞͰฒྻܭࢉ͕Մೳͳ ShapPack※Λར༻ CPUίΞ਺ʹର͢Δ਍அ࣌ؒͷมԽ ਍அ࣌ؒͷ಺༁ (CPUίΞ਺ɿ4) ࣮ݧͷઃఆ ※ https://github.com/tsurubee/shappack ਍அ࣌ؒɿҟৗΛݕ஌͔ͯ͠ΒɼγεςϜ؅ཧऀʹ ɹɹɹɹɹ਍அ݁ՌΛఏࣔ͢Δ·Ͱͷ࣌ؒ

Slide 29

Slide 29 text

29 ਍அ࣌ؒͷධՁɿϝτϦοΫͷܥྻ਺ ධՁ݁Ռ • ϝτϦοΫͷܥྻ਺ͷ૿Ճʹ൐͍ɼ਍அ͕࣌ؒ૿Ճͨ͠ɽ • ίϯςφ਺250ͷγεςϜͷ৔߹ɼ਍அ࣌ؒ͸3෼17ඵͰ͋Δɽ • ίϯςφ਺300ͷγεςϜͷ৔߹ɼ਍அ࣌ؒ͸5෼Λ௒͑Δ͜ͱ͕Θ͔ͬͨɽ ϝτϦοΫͷܥྻ਺ʹର͢Δ਍அ࣌ؒͷมԽ ਍அظؒɿ5෼ CPUίΞ਺ɿ4 ࣮ݧͷઃఆ • γεςϜ؅ཧऀ͕ҟৗݕ஌ͷ௨஌ʹؾ͖ͮɼखಈͰ෮چͷ࡞ۀ Λߦ͏࣌ؒΛߟྀ͢Δͱɼ਍அ࣌ؒΛඵ୯ҐͰ׬ྃͤ͞Δඞཁ ͸ͳ͍ɽ • Google CloudͷSLA (Service Level Agreement)※ʹΑΔͱɼ5෼ Ҏ಺Ͱ͋Ε͹αʔϏεͷఀࢭظؒͱΈͳ͍ͯ͠ͳ͍ ཧ૝తͳ਍அ࣌ؒΛ5෼Ҏ಺ʹઃఆ ※ https://cloud.google.com/compute/sla-20151016/ ίϯςφ਺ 0 83 166 250 333 ϝτϦοΫͷܥྻ਺ͷ૿Ճ͸γεςϜΛߏ੒͢Δίϯςφ਺ ͷ૿ՃʹରԠ͢Δɽ SHAPʹΑΔߩݙ౓ͷܭࢉ ͷߴ଎Խ͸ࠓޙͷ՝୊

Slide 30

Slide 30 text

5. ·ͱΊͱࠓޙͷల๬

Slide 31

Slide 31 text

31 ·ͱΊͱࠓޙͷల๬ • ຊൃදͰ͸ɼҟৗͷݕ஌Λى఺ʹܰྔͳػցֶशϞσϧͷֶशΛߦ͍ɼγϟʔϓϨΠ஋Λ༻͍ͯϞ σϧ͔ΒಘΒΕΔҟৗ౓ʹର͢Δ֤ϝτϦοΫͷߩݙ౓Λࢉग़͢Δ͜ͱͰɼҟৗͷݪҼ਍அΛߦ͏ ख๏ΛఏҊͨ͠ɽ • ϚΠΫϩαʔϏεͷςετϕου؀ڥʹͯҟৗΛ࠶ݱ͢Δ࣮ݧͷ݁Ռɼ਍அظؒͷઃఆ͕ద੾ʹߦ ͑ͨ৔߹ɼఏҊख๏͸ҟৗͷࠜຊݪҼͷϝτϦοΫͷܥྻΛ44.8%ͷਫ਼౓Ͱ্Ґ1Ґʹɼ82.3%ͷ ਫ਼౓Ͱ্Ґ3ҐҎ಺ʹಛఆͰ͖Δ͜ͱΛ֬ೝͨ͠ɽ • ఏҊख๏ͷ਍அਫ਼౓͸ɼ਍அظؒͷઃఆʹେ͖͘ґଘ͢Δ͜ͱ͕࣮ݧ͔Β໌Β͔ʹͳͬͨͨΊɼ ҟৗݕ஌ޙʹऔಘͨ͠ϝτϦοΫ͔ΒదԠతʹ਍அظؒΛઃఆ͢Δख๏Λݕ౼͢Δɽ • ఏҊख๏ΛΑΓߏ੒ཁૉ਺͕େ͖͍ن໛ͷγεςϜʹରԠͤ͞ΔͨΊʹɼSHAPʹΑΔߩݙ౓ͷܭ ࢉΛߴ଎Խ͢Δख๏Λݕ౼͢Δɽ ·ͱΊ ࠓޙͷల๬