Slide 1

Slide 1 text

Copyright @ Liberal Arts Community. All Rights Reserved. XGBoost: A Scalable Tree Boosting System 1 SOK@LiberalArtsCommunity

Slide 2

Slide 2 text

Copyright @ Liberal Arts Community. All Rights Reserved. ໨࣍ • ࣗݾ঺հ • ࿦จ֓؍ • ࿦จৄࡉ • ܾఆ໦(CART)ͷ෮श • Tree Boosting in a Nutshell • Split Finding Algorithms • System Design • ࢀߟจݙ 2

Slide 3

Slide 3 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࣗݾ঺հ 3

Slide 4

Slide 4 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࣗݾ঺հ twitter: @sokei14 ౦ژେֶେֶӃ਺ཧՊֶݚڀՊम࢜՝ఔमྃɻઐ໳͸ෳૉزԿֶɻ ͦͷޙɺϝΨόϯΫͰΫΦϯπͱͯ͠ࢢ৔ϦεΫ؅ཧۀ຿ʹैࣄɻ ݱࡏ͸ϕϯνϟʔͰAI༥ࢿ৹ࠪϞσϧͷ։ൃʹܞΘΔɻAIͰۚ༥αʔϏεͷ มֵΛເݟΔػցֶशΤϯδχΞɻ 4

Slide 5

Slide 5 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จ֓؍ 5

Slide 6

Slide 6 text

Copyright @ Liberal Arts Community. All Rights Reserved. ABSTRACT • ͜ͷ࿦จͰ͸XGBoostͱݺ͹ΕΔεέʔϥϒϧ͔ͭend-to-endͳπϦʔϒʔεςΟϯάΞϧΰϦζϜΛ ঺հ͢Δɽ • ఏҊ͢Δख๏ͱͯ͠ҎԼ͕ڍ͛ΒΕ͍ͯΔɽ 1. sparcity-aware-algorithm, weighted quantile sketch → 3ষͰઆ໌ 2. cache-aware access, data compression and shading → 4ষͰઆ໌ 6

Slide 7

Slide 7 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จ֓؍ 6ͭͷষͰߏ੒͞Ε͍ͯ·͢ɻ 1. INTRODUCTION 2. TREE BOOSTING IN A NUTSHELL 3. SPLIT FINDING ALGORITHMS 4. SYSTEM DESIGN 5. RELATED WORKS 6. END TO END EVALUATIONS ͕͜͜ϝΠϯ 7

Slide 8

Slide 8 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จ֓؍ 2. TREE BOOSTING IN A NUTSHELL XGBoostͷίΞͱͳΔΞϧΰϦζϜʹ͍ͭͯ·ͱΊΒΕ͍ͯΔɽ • tree boostingͷΞϧΰϦζϜͷղઆ • ςΠϥʔల։ʹΑΔϩεؔ਺ͷۙࣅ • Shrinkage • Column Subsampling 8

Slide 9

Slide 9 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จ֓؍ 3. SPLIT FINDING ALGORITHMS XGBoostʹ͓͚Δ෼ׂ఺୳ࡧͷ޻෉఺ʹ͍ͭͯड़΂ΒΕ͍ͯΔɽ • جຊͱͳΔExact Greedy Algorithmͷղઆ • ෼ׂީิ఺Λߜͬͯ୳ࡧʢApproximate Algorithmʣ • ॏΈ෇͖෼Ґ఺ͷ࠾༻ • ॏΈ෇͖෼Ґ఺ͷࢉग़ͷߴ଎ԽʢWeighted Quantile Sketchʣ • ܽଛσʔλʹରͯ͠͸default directionΛ࠾༻ʢSparcity-aware Split Findingʣ 9

Slide 10

Slide 10 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จ֓؍ 4. SYSTEM DESIGN XGBoostʹ͓͚ΔγεςϜଆͷ޻෉఺ʹ͍ͭͯड़΂ΒΕ͍ͯΔɽ • ιʔτͷܭࢉίετͷ࡟ݮʢColumn Block for Parallel Learningʣ • CSCʹΑΔεύʔεߦྻσʔλѹॖ • σʔλͷϒϩοΫԽ • ܭࢉྔͷൺֱ • ޯ഑৘ใͷϓϦϑΣονʢCache-aware Accessʣ • ϒϩοΫαΠζͷ࠷దԽ • σΟεΫIOͷεϧʔϓοτ޲্ʢϒϩοΫѹॖɾ ϒϩοΫஅยԽʣ 10

Slide 11

Slide 11 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉ 11

Slide 12

Slide 12 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श ■ ܾఆ໦ʢCARTʣͱ͸ ͋Δಛ௃࣠ͱᮢ஋ͷେখؔ܎ͷ൑அͷ૊Έ߹ΘͤͰ෼ྨ໰୊΍ճؼ໰୊Λղ ͘ΞϧΰϦζϜͷ͜ͱɽ ܾఆ໦ͷ͏ͪɼԼਤͷΑ͏ʹඞͣೋ෼͞ΕΔ΋ͷΛCARTͱ͍͏ 12

Slide 13

Slide 13 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श ■ ܾఆ໦ʢCARTʣͱ͸ ϊʔυͱϊʔυΛ݁ͿϦϯΫ͔Βߏ੒͞Ε͍ͯΔɽϊʔυʹ͍ͭͯ͸໦ͷͲ ͷ෦෼ʹҐஔ͢ΔʹΑͬͯ࣍ͷΑ͏ʹ۠ผ͞Ε͍ͯΔɽ ໊લ ҙຯ ࠜϊʔυ ໦ͷҰ൪্ʹ͋Δϊʔυ ༿ϊʔυʢϦʔϑʣ ໦ͷҰ൪Լʹ͋Δϊʔυ ಺෦ϊʔυ ࠜϊʔυͱ༿ϊʔυҎ֎ͷϊʔυ ༿ϊʔυ ࠜϊʔυ ಺෦ϊʔυ ϦϯΫ 13

Slide 14

Slide 14 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श • 14

Slide 15

Slide 15 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श • 15

Slide 16

Slide 16 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श • 16

Slide 17

Slide 17 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श • 17

Slide 18

Slide 18 text

Copyright @ Liberal Arts Community. All Rights Reserved. ࿦จৄࡉɹͦͷલʹ…ܾఆ໦ͷ෮श • 18

Slide 19

Slide 19 text

Copyright @ Liberal Arts Community. All Rights Reserved. TREE BOOSTING IN A NUTSHELL 19