ログから生まれる施策 / actions born from logs

by Hiroka Zaitsu

Slide 1

Slide 1 text

ϖύϘͷϩάz׆༻zج൫ʮ#JHGPPUʯ ࡒ௡େՆ(.01FQBCP *OD 8&#%*3&$50344&44*0/7PM ϩά͔Βੜ·ΕΔࢪࡦ

Slide 2

Slide 2 text

σʔλαΠΤϯςΟετσΟϨΫλʔ ࡒ௡େՆ![BJNZ NJOOFࣄۀ෦

Slide 3

Slide 3 text

ϋϯυϝΠυϚʔέοτNJOOF IUUQTNJOOFDPN

Slide 4

Slide 4 text

໨࣍ w8FCαʔϏεͷߦಈϩά wϖύϘͷϩάz׆༻zج൫ʮ#JHGPPUʯ wࢪࡦ΁ͷ׆༻ྫ

Slide 5

Slide 5 text

8FCαʔϏεͷߦಈϩά

Slide 6

Slide 6 text

8FCαʔϏεͷߦಈϩά wϢʔβʔ͕αʔϏεΛར༻ͨ͠ࡍͷཤྺ wʮ͍ͭʯʮ୭͕ʯʮԿΛʯߦͬͨͷ͔ w݁Ռ͚ͩͰ͸ͳͦ͘͜ʹࢸΔաఔ΍ɺ݁Ռ్தͰఘΊͯ͠·ͬͨϢʔ βʔͷߦಈ΋෼͔Δ

Slide 7

Slide 7 text

ߦಈϩάΛ࢖ͬͯ ΑΓྑ͍αʔϏε΁!

Slide 8

Slide 8 text

׆༻·Ͱͷஈ֊ wऩूߦಈϩά͕ग़ྗ͞ΕɺऔΓ·ͱΊΒΕ͍ͯΔঢ়ଶ w෼ੳऔΓ·ͱΊͨߦಈϩάΛࢹ֮Խɺ෼ੳͰ͖Δঢ়ଶ w׆༻෼ੳͨ͠ߦಈϩάΛ΋ͱʹܧଓతͳαʔϏεվળ͕ߦ͍͑ͯΔঢ়ଶ

Slide 9

Slide 9 text

#JHGPPU IUUQTJDPOTDPN

Slide 10

Slide 10 text

#JHGPPU wϖύϘͷϩάz׆༻zج൫ w෯޿͍δϟϯϧͷ8FCαʔϏε wϋϯυϝΠυϚʔέοτNJOOF wωοτγϣοϓ࡞੒Χϥʔϛʔγϣοϓ wϩϦϙοϓʂϨϯλϧαʔόʔ wϩάऩू͔Β׆༻·Ͱͷ֤ஈ֊ʹ͓͍ͯ൚༻తʹར༻Ͱ͖Δશࣾج൫

Slide 11

Slide 11 text

#JHGPPU rack-bigfoot Service Request Activity log Services DB Attribute Big Cube Cube https://speakerdeck.com/monochromegane/pepabo-log-infrastructure-bigfoot Bandit algorithm/ Recommendation Re-targeting Feedback Name identiﬁcation BI/Visualize

Slide 12

Slide 12 text

ऩू

Slide 13

Slide 13 text

໊دͤ wαʔϏεͷΞΧ΢ϯτͱΫϥΠΞϯτʢ୺຤ɾϒϥ΢βʣΛඥ͚ͮ wैདྷ͸ΫϥΠΞϯτ͝ͱͷܭଌ wະϩάΠϯঢ়ଶͷΞΧ΢ϯτ΋ͦͷޙϩάΠϯͨ͠λΠϛϯάͰաڈʹ Ḫͬͯඥ͚ͮΒΕΔ wαʔϏεΛލ͍ͩඥ͚ͮ΋Մೳ

Slide 14

Slide 14 text

෼ੳ

Slide 15

Slide 15 text

#JH$VCFͱ$VCF w#JHGPPU্ͷશͯͷϩάΛ#JH$VCFʹू໿ w෼ੳͷ੾ΓޱʢσΟϝϯγϣϯͱϝδϟʔʣ͕ܾ·ͬͨΒ$VCFʹूܭ wྫʣ࣌ؒ͝ͱ஫จֹۚ wߴ଎ͳࢀর͕Մೳ Activity Big Cube Cube HiveQL SQL Dashboard Ad-hoc query Analyst Managers, Product owners, Promotion groups

Slide 16

Slide 16 text

ࢹ֮Խ w5BCMFBV wIUUQXXXUBCMFBVDPN wΞυϗοΫͳ෼ੳ΍ෳࡶͳσʔλ݁߹ʹ w3FEBTI wIUUQTSFEBTIJP w୭΋͕ࢀরͰ͖ΔμογϡϘʔυʹ

Slide 17

Slide 17 text

ύε෼ੳ wίϯόʔδϣϯʹؔ͢Δ෼ੳͷ͏ͪɺܦ࿏ʢύεʣʹओ؟Λஔ͍ͨ΋ͷ wύεͷ௕࣌ؒ͞ྨܕ wΞτϦϏϡʔγϣϯϞσϧϖʔδͷग़ݱҐஔʹΑΓॏΈΛม͑Δ referrer landing last cv ??? Point Analytics referrer landing last Path Analytics cv

Slide 18

Slide 18 text

ࢪࡦ΁ͷ׆༻

Slide 19

Slide 19 text

཭୤Ϣʔβʔ΁ͷϦςϯγϣϯ wߦಈϩά͔Β཭୤ϢʔβʔΛநग़ wΧʔτʹ౸ୡ͕ͨ͠ങΘͳ͔ͬͨϢʔβʔʢ͍ΘΏΔΧʔτམͪʣ w୹ظؒʹಉ͡࡞඼ΛԿ౓΋ݟ͍ͯΔϢʔβʔ wಛఆͷ৚݅ʹ߹க͢Ε͹ϓογϡ௨஌΍ϝʔϧͰϦςϯγϣϯ wϨγʔτϝʔϧͱಉ౳ͷ։෧཰ HiveQL Re-targeting

Slide 20

Slide 20 text

޿ࠂ࿈ܞ wߦಈϩά͔ΒϢʔβʔΛηάϝϯτ wϢʔβʔʹؔ࿈ੑͷߴ͍޿ࠂΛදࣔ wطʹϦʔνͨ͠Ϣʔβʔʹ͸޿ࠂΛදࣔ͠ͳ͍ʢσϦλʔήςΟϯάʣ

Slide 21

Slide 21 text

όϯσΟοτΞϧΰϦζϜ w࡞඼ݕࡧը໘͔ΒͷརӹΛ࠷େԽ͍ͨ͠ w$53ͷҟͳΔύλʔϯͷ࡞඼ݕࡧΞϧΰϦζϜ wച্ʹର͢Δൺॏ͕ଟ͍ҝʹγεςϜมߋͷϦεΫ͕ߴ͍ w࠷ྑͷύλʔϯΛ࢖͍ͳ͕ΒΑΓྑ͍ύλʔϯΛಈతʹ୳͢

Slide 22

Slide 22 text

όϯσΟοτΞϧΰϦζϜ w&QTJMPO(SFFEZ"MHPSJUIN w֬཰ Џ Ͱͦͷ࣌఺ͷ࠷΋ظ଴஋͕ߴ͍ύλʔϯΛ࢖༻ʢ׆༻ʣ w֬཰ЏͰϥϯμϜʹબ୒ͨ͠ύλʔϯΛ࢖༻ʢ୳ࡧʣ Activity Epsilon-Greedy algorithm User 1-ε: exploitation ε/pattern: exploration Click / Not click Import

Slide 23

Slide 23 text

όϯσΟοτΞϧΰϦζϜ wτϨʔυΦϑͷղܾ w׆༻͹͔Γ͍ͯ͠Δͱݱࡏͷظ଴஋ΑΓ΋ྑ͍࿹Λݟ͚ͭΒΕͳ͍ w୳ࡧ͹͔Γ͍ͯ͠Δͱظ଴஋͕ߴ͍࿹͕ར༻͞Εͳ͍

Slide 24

Slide 24 text

Ϩίϝϯσʔγϣϯ wڠௐϑΟϧλϦϯάʹΑΔʮ͋ͳͨʹ͓͢͢Ίͷ࡞Ոʯ !NPOPDISPNFHBOFʮNJOOFNFFUT)JWFNBMMʯ IUUQTTQFBLFSEFDLDPNNPOPDISPNFHBOFQFQBCPNJOOFNBUSJYGBDUPSJ[BUJPOJOIJWFNBMM

Slide 25

Slide 25 text

ϩάz׆༻zج൫

Slide 26

Slide 26 text

ϩάz׆༻zج൫ wᶃਓ͕ؒ෼ੳ݁ՌΛݟͯߦ͏੩తͳʮ׆༻ʯ wࢪࡦͷਫ਼౓޲্ wᶄࣗಈԽ͞ΕͨϑΟʔυόοΫʹΑΔಈతͳʮ׆༻ʯ wͳΊΒ͔ͳγεςϜ