Applying oCPC algorithm for production

by yoppi

Slide 1

Slide 1 text

OCPCΛຊ൪؀ڥʹద༻͢ Δ·Ͱ @YOPPIBLOG ADOKEN#1

Slide 2

Slide 2 text

ࣗݾ঺հ IMPORT “GITHUB.COM/YOPPI" ▸ ٛా തҰ @yoppiblog ▸ Speee, Inc. Engineer ▸ ࠷ۙ͸ΞυςΫք۾ʹ͍ͯɺࠓ͸UZOUͷϨίϝϯυΤϯδ ϯͱ͔࡞͍ͬͯ·͢ ▸ Go͕޷͖

Slide 3

Slide 3 text

OCPCΛຊ൪؀ڥʹద༻͢Δ·Ͱ ࠓ೔ͷ͓࿩ ▸ oCPCʹ͍ͭͯ ▸ ௐࠪ ▸ Ϟσϧͷ࡞੒ ▸ γϛϡϨʔγϣϯʹΑΔݕূ ▸ ຊ൪؀ڥͰͷ࣮૷ ▸ A/BςετͰͷϞχλϦϯάͱ෼ੳ ▸ ݱঢ়ͷ՝୊ͱࠓޙͷํ਑

Slide 4

Slide 4 text

Slide 5

Slide 5 text

OCPCʹ͍ͭͯ OCPC ▸ Optimize CPCͷݺশ ▸ ೖߘ͞Εͨ޿ࠂͷ໨ඪCPAͷୡ੒ͱ഑৴ֹۚͷ࠷େԽΛ໨ࢦͨ͢ΊʹࣗಈͰCPCೖ ࡳ͢Δػೳ ▸ جຊతʹCPA͕ѱԽ͢ΔͱͦͷΞυωοτϫʔΫͰͷ഑৴͕ࢭ·Δ ▸ CPAΛୡ੒ͤ͞Α͏ͱCPCΛखಈௐ੔΋Ͱ͖ͳ͘͸ͳ͍͕഑৴໘͕ଟྔͳͨΊਓख Ͱͷௐ੔͸ඇݱ࣮త ▸ ސ٬ͷຬ଍౓Λ޲্͢ΔͨΊʹ΋ϓϩμΫτͷ࣭ʢརӹʣΛ޲্͢ΔͨΊʹ΋ػೳͱ͠ ࣮ͯ૷͞Ε͍ͯΔ΂͖ ▸ ͱ͸͍͑໨ඪCPAΛݱ࣮ʹ͙ͦΘͳֹ͍ۚͰઃఆ͞Εͨ΋ͷΛͳΜͱ͔͢ΔػೳͰ ͸ͳ͍͜ͱ͸લఏ

Slide 6

Slide 6 text

OCPCʹ͍ͭͯ OCPCΛ࣮૷͍ͯ͠Δڝ߹ଞࣾ ▸ جຊతʹCPAͷվળٴͼӡ༻޻਺ͷ࡟ݮͱ͍͏Ґஔ͚ͮͰ UZOU΋ͦ͜Λΰʔϧͱ͢Δͷ͸มΘΒͳ͍ ▸ Outbrain ▸ Logly ▸ Smart News ▸ LINE Ads

Slide 7

Slide 7 text

OCPCΛຊ൪؀ڥʹద༻͢Δ·Ͱ ࠓ೔ͷ͓࿩ ▸ oCPCʹ͍ͭͯ ▸ ௐࠪ ▸ Ϟσϧͷ࡞੒ ▸ γϛϡϨʔγϣϯʹΑΔݕূ ▸ ຊ൪؀ڥͰͷ࣮૷ ▸ A/BςετͰͷϞχλϦϯάͱ෼ੳ ▸ ࠓޙͷ՝୊ͱ·ͱΊ

Slide 8

Slide 8 text

ௐࠪ ௐࠪ ▸ طଘͷϓϩμΫτ΍࿦จͷௐࠪ ▸ ࿦จΛยͬ୺͔ΒಡΜͰ͍͘ ▸ “Optimal Real-Time Bidding for Display Advertising” ͸ ͔ͳΓے͕ྑ͍ํ๏ͰࠓճͷϞσϧͷϕʔεʹͳ͍ͬͯΔ ▸ http://wnzhang.net/papers/ortb-kdd.pdf

Slide 9

Slide 9 text

ௐࠪ OPTIMAL REAL-TIME BIDDING FOR DISPLAY ADVERTISING ▸ ੍໿৚݅෇͖ඇઢܗ࠷దԽ໰୊Λద༻ͯ͠RTBΛղ͍͍ͯΔ ▸ ొ৔ਓ෺ ▸ Winning RateʢeCPM͕͍͘ΒͳΒଞͷ޿ࠂʹউͬͯ໘ʹͰΔͷ͔ʣ ▸ ༧ଌCTRʢະདྷͷCTRʣ ▸ ༧ଌCVRʢະདྷͷCVRʣ ▸ ϥάϥϯδϡະఆ৐਺ʢϥάϥϯδϡະఆ৐਺๏Ͱղ͘ࡍʹ෇༩͞Ε ͑Δม਺ʣ

Slide 10

Slide 10 text

OCPCΛຊ൪؀ڥʹద༻͢Δ·Ͱ ࠓ೔ͷ͓࿩ ▸ oCPCʹ͍ͭͯ ▸ ௐࠪ ▸ Ϟσϧͷ࡞੒ ▸ γϛϡϨʔγϣϯʹΑΔݕূ ▸ ຊ൪؀ڥͰͷ࣮૷ ▸ A/BςετͰͷϞχλϦϯάͱ෼ੳ ▸ ݱঢ়ͷ՝୊ͱ·ͱΊ

Slide 11

Slide 11 text

Ϟσϧͷ࡞੒ Ϟσϧ࡞੒ ▸ ੍໿෇͖ඇઢܗ࠷దԽ໰୊ͱͯ͠਺ཧϞσϧΛ࡞ͬͯղ͘ ▸ ໨తؔ਺ɺ੍໿৚݅Λઃఆ͠ɺ੍໿৚݅Λຬͨͭͭ͠໨ తؔ਺Λ࠷େԽ͢ΔʢCPCՁ֨Λ࠷దԽ͢Δʣ͜ͱ͕ ΰʔϧʹͳΔ f(bix ) = T N ∑ i=1 bix W(bix , ˜ θix ) ˜ θix pi T N ∑ i=1 bix W(bix , ˜ θix ) ˜ θix pi ≤ rx T N ∑ i=1 θix W(bix , ˜ θix ) ˜ θix pi ໨తؔ਺ ੍໿৚݅

Slide 12

Slide 12 text

Ϟσϧͷ࡞੒ ੍໿෇͖ඇઢܗ࠷దԽ໰୊ ▸ ࠷దԽ໰୊ͷ͏ͪඇઢܗͳ΋ͷΛඇઢܗ࠷దԽ໰୊ͱݺ Ϳ ▸ ϥάϥϯδϡະఆ৐਺๏ͰύϥϝʔλΛٻΊΔ͜ͱͰۙࣅ ղΛಘΒΕΔ ▸ ࠓճ͸eCPMউ཰ϞσϧΛ ͱஔ͍͍ͯΔͷͰತؔ ਺ͱͳΓχϡʔτϯ๏Ͱϥάϥϯδϡະఆ৐਺ΛٻΊ͍ͯ Δ y = x x + l

Slide 13

Slide 13 text

Ϟσϧͷ࡞੒ ECPMউ཰Ϟσϧ ▸ ͲͷeCPMͳΒ഑৴͞ΕΔʢΦʔΫγϣϯʹউͭʣͰ͋Ζ͏ Ϟσϧ ▸ ໨తؔ਺͕ತؔ਺ʹͳΔΑ͏ͳ୯७ͳඇઢܗؔ਺༻͍ͯ࠷খ ೋ৐๏Ͱ࡞੒ ▸ ತؔ਺͸ඍ෼ՄೳʹͳΓ࠷దԽܭࢉͰۃ஋ܭࢉʹ͓͍ͯ ߴ଎ʹऩଋͤ͞͞ΕΔχϡʔτϯ๏͕࢖͑ΔͨΊ

Slide 14

Slide 14 text

Ϟσϧͷ࡞੒ CTR༧ଌɾCVR༧ଌ ▸ ະདྷͷbidՁ֨Λೖࡳ͢ΔͨΊະདྷͷCTRɾCVRΛࢉग़͢Δඞཁ͕͋Δ ▸ ࣄલ෼෍͔Βࣄޙ෼෍Λβ෼෍༻͍ͯࢉग़ͦ͠ΕΛ༧ଌ஋ͱͯ͠࢖༻ ▸ ഑৴ΞϧΰϦζϜͷҰͭͰଟ࿹όϯσΟοτ(Thompson Sampling)ʹͯ CTR༧ଌΛ͍ͯ͠Δ΋ͷ͕ͱͯ΋ੑೳ͕͍͍ͷͰͦͷ··ྲྀ༻͍ͯ͠Δ ▸ https://tech.speee.jp/entry/2018/08/08/090000 ▸ ECαΠτͰ͋ΔλΦόΦʢΞϦόόʣ಺ͷ޿ࠂ͸ϦΞϧλΠϜʹ༧ଌͯ͠ ͍Δ͜ͱΛ঺հ͍ͯ͠Δ ▸ https://arxiv.org/pdf/1703.02091.pdf

Slide 15

Slide 15 text

Slide 16

Slide 16 text

γϛϡϨʔγϣϯʹΑΔݕূ γϛϡϨʔγϣϯʹΑΔݕূ ▸ ࣮ࡍʹoCPCϞσϧΛ࡞੒ͯ͠ҎԼͷ߲໨Λݕূ͢ΔͨΊʹγ ϛϡϨʔγϣϯʹ͓͍ͯݕূ ▸ ໨ඪCPAΛୡ੒Ͱ͖Δ͔Ͳ͏͔ ▸ ࠓճͷҰ൪ղܾ͍ͨ͠໨త ▸ ഑৴ֹۚΛ࠷େԽͰ͖Δ͔Ͳ͏͔ ▸ ໨ඪCPAΛୡ੒͍ͯͯ͠΋഑৴͞Εͳ͚Ε͹རӹʹͳΒ ͳ͍

Slide 17

Slide 17 text

γϛϡϨʔγϣϯʹΑΔݕূ γϛϡϨʔγϣϯͷ࣮૷ ▸ γϛϡϨʔγϣϯϑϨʔϜϫʔΫ͸ಛʹ࠾༻ͤͣɺPythonͰ ϑϧεΫϥονͰ࣮૷ ▸ ࡞੒ͨ͠ϞσϧͰCPCΛೖࡳͯ͠ ▸ ֤छϨίϝϯυΞϧΰϦζϜͰ޿ࠂΛநબ͠഑৴ ▸ Λ܁Γฦ͢୯७ͳ΋ͷ

Slide 18

Slide 18 text

γϛϡϨʔγϣϯʹΑΔݕূ ࣗલ࣮૷ͷPROS/CONS ▸ PROS ▸ ಛʹϥΠϒϥϦΛ࢖Θͣʹ࡞ͬͨͷͰॊೈʹ࣮૷Ͱ͖ͨ ▸ ޿ࠂΛநબ͢Δ෦෼΍ͦͷޙͷclickͷൃੜΛ੍ޚ͢Δඞཁ͕͋Δͷ ͰUZOUͷγεςϜʹدΓఴͬͨ΋ͷΛ࡞ͬͨ΄͏͕ྑ͍΋ͷʹͳΔ ▸ CONS ▸ ൚༻ԽͰ͖Δ΋ͷʹͳ͍ͬͯͳ͍ ▸ ࠓճͷoCPCʹಛԽͨ͠࡞Γʹͳ͍ͬͯΔͷͰଞͷϓϩδΣΫτͰ͸ ͦͷ··࢖͑ͳ͍

Slide 19

Slide 19 text

ຊ൪؀ڥͰͷ࣮૷ ࣮૷؀ڥ ▸ γϛϡϨʔγϣϯͱಉ͘͡PythonͰ࣮૷ ▸ ϦΞϧλΠϜͰͷbid͸͓ͯ͠Βͣόονܗࣜ

Slide 20

Slide 20 text

ຊ൪؀ڥͰͷ࣮૷ ͳΔ΂࣮͘ߦ࣌ؒΛ୹͘͢Δ޻෉ ▸ σʔλιʔε͔ΒͦΕͳΓͷσʔλΛऔಘ͢Δඞཁ͕͋Δ ▸ SQLΛ޻෉ͯ͠PythonଆͰͳΔ΂͘ܭࢉͤ͞ͳ͍Α͏ʹ ▸ ਺ཧϞσϧΛܭࢉ͢Δͱ͖͸NumPy.arrayͰߴ଎ʹܭࢉͰ͖ΔΑ͏ʹ ▸ ѻ͏σʔλ͕ଟ͍ͷͰຊདྷͳΒ͹O(mn)͸ڐ༰͢Δͱ͜ΖΛO(n)ʹ ͢ΔΑ͏ʹఆ਺߲ΛͳΔ΂͘ഉআ͢ΔΑ͏ʹ ▸ ͦ΋ͦ΋χϡʔτϯ๏Λ࠾༻͍ͯ͠ΔͷͰऩଋ͕ͱͯ΋ߴ଎ʢ͍ͩ ͍ͨ਺ճͷΠςϨʔγϣϯͰऩଋ͍ͯ͠Δʣ

Slide 21

Slide 21 text

Slide 22

Slide 22 text

A/BςετͰͷಈ͖ํ A/BςετͰͷϞχλϦϯάͱ෼ੳ ▸ A/BςετΛ։࢝ͯ͠ɺ͏·͍ͬͯ͘Δ޿ࠂओͱ͏·͍ͬ͘ ͍͔ͯͳ͍޿ࠂओ͕ൃੜ ▸ ΞϧΰϦζϜ͸ۜͷ஄ؙͰ͸ͳ͍ͷͰԿ͕Өڹ͍ͯ͠Δͷ ͔෼ੳ͢Δඞཁ͕͋Δʢ͔͜͜Β͕Ή͠Ζຊ൪ ▸ ϞχλϦϯά͸μογϡϘʔυΛ࡞Γ೔࣌Ͱ੒ՌΛ෼ੳ ▸ μογϡϘʔυҎ֎Ͱ΋adhoc(Jupyter Notebook)Ͱ΋෼ ੳ͍ͯ͠Δ

Slide 23

Slide 23 text

A/BςετͰͷಈ͖ํ ϞχλϦϯάπʔϧ ▸ dashΛ࢖༻ ▸ Python੡ʢﬂaskͷ্ʹಠࣗͷϑϨʔϜϫʔΫΛ৐͍ͤͯ Δʣ ▸ SPAͰಈ͘ʢReactͷίϯϙʔωϯτΛPythonͰॻ͚Δʣ SSR

Slide 24

Slide 24 text

Slide 25

Slide 25 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ A/Bςετ݁Ռ ▸ ͏·͍͍ͬͯ͘Δ޿ࠂओ΋͍Ε͹ ▸ ໨ඪCPAΛୡ੒͍ͯͯ͠഑৴ֹۚ΋Ͱ͍ͯΔ ▸ ͏·͍͍ͬͯ͘ͳ͍޿ࠂओ΋͍Δ ▸ ໨ඪCPA͸ୡ੒͍ͯ͠Δ͕഑৴ֹ͕ۚग़͍ͯͳ͍΋ͷ͕ଟ ͍

Slide 26

Slide 26 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ͏·͍͍ͬͯ͘Δ΋ͷͱͦ͏Ͱͳ͍΋ͷΛ෼ੳ ▸ جຊతʹ͏·͍͍ͬͯ͘ͳ͍΋ͷ͸഑৴ྔ͕গͳ͍ ▸ ຊདྷ͋Δ΂͖ਅͷCPCʹಧ͔ͣෛ͚ͯ͠·ͬͯ഑৴͞Εͳ ͍ঢ়ଶ

Slide 27

Slide 27 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ਅͷCPCͱBID CPCͷൺֱ

Slide 28

Slide 28 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ਅͷCPCͱBID CPCͷൺֱ

Slide 29

Slide 29 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ਅͷCTRͱ༧ଌCTRͷ෼෍

Slide 30

Slide 30 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ਅͷCVRͱ༧ଌCVRɾ഑৴ֹׂۚ߹ͷ෼෍

Slide 31

Slide 31 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ECPMউ཰Ϟσϧͷਫ਼౓෼ੳ

Slide 32

Slide 32 text

ݱঢ়ͷ՝୊ͱࠓޙͷํ਑ ϞσϧͷվળҊ ▸ ༧ଌCVRͱeCPMউ཰Ϟσϧ͕ѱͦ͏ͱݴ͑ΔͷͰ͜͜Λվળ͍ͯ͘͠ ▸ ༧ଌCVRͷࣄલ෼෍ͷվળ ▸ eCPMউ཰ϞσϧΛվળ ▸ StepModelͱConstantModelͷࠞ߹Ϟσϧ͸ΘΓͱྑͦ͞͏͕ͩ ತؔ਺Ͱͳ͘ͳΔͷͰχϡʔτϯ๏͕࢖͑͘ͳΓSGD౳ʹ੾Γସ ͑Δඞཁ͋Γ ▸ https://media.readthedocs.org/pdf/lmﬁt-py/0.9.3/lmﬁt-py.pdf

Slide 33

Slide 33 text

ڊਓͷݞʹ৐Δ ͦͷଞͷࢀߟจݙ ▸ ࠷దԽ௒ೖ໳ https://www.slideshare.net/tkm2261/ss-42149384 ▸ ͜ΕͳΒΘ͔Δ࠷దԽ਺ֶ https://www.amazon.co.jp/dp/ 4320017862/ ▸ ඇઢܗ࠷దԽͷجૅ https://www.amazon.co.jp/dp/4254280017/ ▸ ತؔ਺ʹ͍ͭͯ http://www2.kaiyodai.ac.jp/~yoshi-s/Lectures/ Optimization/2013/lecture_1.pdf