Slide 1

Slide 1 text

ࡾ୐༔հ / Pepabo R&D Institute, GMO Pepabo, Inc. 2019.05.24 ୈ45ճ ৘ใॲཧֶձ Πϯλʔωοτͱӡ༻ٕज़ݚڀձ Synapse: ར༻ऀͷจ຺ʹԠͯ͡ ܧଓతʹਪનख๏ͷબ୒Λ ࠷దԽ͢ΔਪનγεςϜ

Slide 2

Slide 2 text

1. ͸͡Ίʹ 2. എܠ 3. ఏҊख๏ 4. ධՁ 5. ·ͱΊ 2 ໨࣍

Slide 3

Slide 3 text

1. ͸͡Ίʹ

Slide 4

Slide 4 text

• ECαΠτͰ͸঎඼૿େʹ൐͏৘ใաଟ໰୊Λղܾ͢ΔͨΊ঎඼Λࣗಈతʹఏ Ҋ͢ΔػೳʢਪનγεςϜʣ͕ಋೖ͞ΕΔɽ • ਪનख๏͸ਪનࠜڌͱͳΔ৘ใݯ΍ํࣜʹΑͬͯબఆ͢Δ঎඼͕ҟͳΔ • ར༻ऀͷཁٻΛຬͨ͢঎඼Λબఆ͢ΔՄೳੑͷߴ͍ਪનख๏Λબ୒͢Δ͜ͱ͕ ӡӦऀʹͱͬͯॏཁ 4 ݚڀͷ໨త

Slide 5

Slide 5 text

• Which is the best ? • ಺༰ϕʔεܕਪન • ྨࣅը૾ • ྨࣅςΩετ • ڠௐܕਪન • ஌ࣝϕʔεܕਪન • ϋΠϒϦουܕਪન • ฒྻܕɼ௚ྻܕ… 5 ਪનख๏ͷબ୒

Slide 6

Slide 6 text

• Which is the best • ಺༰ϕʔεܕਪન • ྨࣅը૾ • ྨࣅςΩετ • ڠௐܕਪન • ஌ࣝϕʔεܕਪન • ϋΠϒϦουܕਪન • ฒྻܕɼ௚ྻܕ… 6 ಛఆͷ৚݅Ͱͷਪનख๏ͷબ୒ • In the case ? • ਫ਼౓΍଎౓ • ঎඼ͷಛੑ • ಺༰ͷදݱྗ • ίʔϧυελʔτ • ۙࣅ • จ຺ • ར༻ऀͷঢ়گͱཁٻ

Slide 7

Slide 7 text

• Which is the best • ಺༰ϕʔεܕਪન • ྨࣅը૾ • ྨࣅςΩετ • ڠௐܕਪન • ஌ࣝϕʔεܕਪન • ϋΠϒϦουܕਪન • ฒྻܕɼ௚ྻܕ… 7 ECαΠτͷಛఆͷ৚݅Ͱͷਪનख๏ͷબ୒ • In the case • ਫ਼౓΍଎౓ • ঎඼ͷಛੑ • ಺༰ͷදݱྗ • ίʔϧυελʔτ • ۙࣅ • จ຺ • ར༻ऀͷঢ়گͱཁٻ • On a EC site ? • ܧଓతͳվળ • ػೳ௥Ճ • ར༻ऀ૿Ճ • அଓతͳվѱ • ෆ۩߹ • γεςϜෛՙ • ݱࡏͷ࠷ળखͷ௥ٻ

Slide 8

Slide 8 text

ਪનख๏ͷ༏ྼ͸ଟ͘ͷ৚݅ʢจ຺ʣʹΑͬͯࠨӈ͞ΕΔ 8 ݚڀͷഎܠͱఏҊͷࠎࢠ ࣄલʹఆΊͨจ຺͝ͱʹਪનख๏ͷબ୒Λࣗಈత͔ͭܧଓతʹ࠷దԽ͢Δਪન γεςϜͷఏҊ ༗ޮͳਪનख๏Λػձଛࣦ͕ͳ͍Α͏ʹจ຺ʹԠͯ͡࢖͍෼͚͍ͨ

Slide 9

Slide 9 text

3. എܠ

Slide 10

Slide 10 text

• ᶃ ਪનख๏ͷ༏ྼ͕จ຺ʹΑͬͯมԽ͢Δ • ᶄ ༏ྼͷมԽ͢Δ൚༻తͰ໌֬ͳج४͕ͳ͍ • ᶅ ༏ྼ͕ܧଓతʹมԽ͢Δ 10 ਪનख๏ͷબ୒ʹ·ͭΘΔ՝୊ ECαΠτͷӡӦऀ͸ར༻ऀͷཁٻΛຬͨ͢Մೳੑͷߴ͍ਪનख๏Λબ୒͍ͨ͠ ͕ɼҎԼͷ՝୊ͷͨΊಋೖઌ͝ͱͷධՁͱௐ੔͕ߦΘΕ͍ͯΔɽ

Slide 11

Slide 11 text

11 ՝୊ᶃ ਪનख๏ͷ༏ྼ͕จ຺ʹΑͬͯมԽ͢Δ • ਪનख๏ͷ༏ྼ͸ଟ͘ͷ৚݅ʢจ຺ʣʹΑͬͯࠨӈ͞ΕΔ • ਪનख๏ͷલఏ΍ํࣜʹىҼ͢Δ੍໿ • ਪનର৅ͱͳΔ঎඼ͷಛੑ • Ԡ౴଎౓΍දࣔॱংͳͲͷ࣮૷ཁҼ • ΫϦοΫ཰΍ߪೖ཰ͳͲͷධՁࢦඪ • ਪન݁ՌΛධՁ͢Δར༻ऀଆͷঢ়گ

Slide 12

Slide 12 text

12 ՝୊ᶃ ਪનख๏ͷ༏ྼ͕จ຺ʹΑͬͯมԽ͢Δ ਪનର৅ͱͳΔ঎඼ͷಛੑɼධՁࢦඪʹΑΔ༏ྼࠩͷྫ

Slide 13

Slide 13 text

13 ՝୊ᶃ ਪનख๏ͷ༏ྼ͕จ຺ʹΑͬͯมԽ͢Δ ਪનख๏ͷ༏ྼ͸ଟ͘ͷ৚݅ʢจ຺ʣʹΑͬͯࠨӈ͞ΕΔ ৚݅ʢจ຺ʣΛఆΊͯɼจ຺͝ͱʹదͨ͠ਪનख๏Λ࢖͍෼͚Δ͜ͱͰਪનγ εςϜશମͰར༻ऀͷཁٻΛຬͨ͢঎඼Λબఆ͢ΔՄೳੑΛߴΊ͍ͨ

Slide 14

Slide 14 text

14 ՝୊ᶄ ༏ྼͷมԽ͢Δ൚༻తͰ໌֬ͳج४͕ͳ͍ • ༏ྼࠩͷج४ʹ͍ͭͯͷҰൠతͳ܏޲͸͜Ε·Ͱͷใࠂ΍ධՁߟ࡯͕͋Δ΋ͷ ͷจ຺ґଘͷͨΊ൚༻తͰ໌֬ͳج४͸ଘࡏ͠ͳ͍ • Ұൠతͳ܏޲΍ߟ࡯ʹج͖ͮͭͭ΋ɼ࣮؀ڥͰͷධՁʹΑͬͯ࢖͍෼͚ͷج४ ΛٻΊΔඞཁ͕͋Δ

Slide 15

Slide 15 text

• ϋΠϒϦουܕਪન • ਪનख๏Λ૊Έ߹Θ֤ͤͯख๏ͷ୹ॴΛิ͏ • ߪങཤྺ͕஝ੵ͞ΕΔ·Ͱ͸಺༰ϕʔεΛॏࢹɼ஝ੵޙʹڠௐܕΛॏࢹ౳ 15 ՝୊ᶄ ༏ྼͷมԽ͢Δ൚༻తͰ໌֬ͳج४͕ͳ͍ $53 )JTUPSZDPVOU -PX .JEEMF )JHI $POUFOUCBTF $PDPSBCPSBUFCBTF ڠௐܕਪન͕ߪങཤྺͷ஝ੵʹΑͬͯਫ਼౓͕޲্͠಺༰ϕʔεܕ ਪનͷਫ਼౓ͱٯస͢Δ͜ͱΛදݱͨ͠໛ࣜਤ ߪങཤྺͷ஝ੵ͕۩ମతʹԿ݅ʹୡͨ࣌͠ʹ಺༰ϕʔεܕਪનͱ ੾Γସ͑Δ΂͖͔͸࣮؀ڥͰͷධՁ͕ඞཁ

Slide 16

Slide 16 text

16 ՝୊ᶅ ༏ྼ͕ܧଓతʹมԽ͢Δ ECαΠτ͸ɼ࣮؀ڥͰͷධՁʹΑͬͯಘΒΕͨ৚݅͝ͱͷޮՌࠩʹج͍ͮͨਪ નख๏ͷ࢖͍෼͚ʹΑͬͯਪનγεςϜͷޮՌΛܧଓతʹ࠷େԽ͍ͨ͠ 1. จ຺ͷ੾Γ෼͚ 2. ৽͍͠ਪનख๏ͷಋೖ 3. ಋೖޙͷ࠶ܭଌɼޮՌ൑ఆɼޮՌతͳख๏ͷద༻ ͜ΕΒΛఆظత͔ͭ࠷୹ͰߦΘͳ͚Ε͹ӡ༻؀ڥͰ͸ػձଛࣦ͕ൃੜ͢Δ

Slide 17

Slide 17 text

3. ఏҊख๏

Slide 18

Slide 18 text

• ᶃ ਪનख๏ͷ༏ྼ͕จ຺ʹΑͬͯมԽ͢Δ • จ຺ΛఆΊͯɼఆΊͨจ຺͝ͱʹਪનख๏Λ࢖͍෼͚Δ • ᶄ ༏ྼͷมԽ͢Δ൚༻తͰ໌֬ͳج४͕ͳ͍ • ఆΊͨจ຺͝ͱʹ࣮؀ڥͰͷධՁΛߦ͏ • ᶅ ༏ྼ͕ܧଓతʹมԽ͢Δ • จ຺΍ਪનख๏ͷಋೖޙʹ࣌ؒࠩͳ͘దԠ͢Δ 18 ՝୊ͷ੔ཧ

Slide 19

Slide 19 text

• ར༻ऀͷཁٻΛຬͨ͢঎඼Λબఆ͢ΔՄೳੑͷߴ͍ਪનख๏Λӡ༻؀ڥͰػձ ଛࣦ͕ൃੜ͠ͳ͍Α͏ࣗಈత͔ͭܧଓతʹར༻͍ͨ͠ 19 ఏҊख๏ 1. จ຺ͷఆٛͱಋೖ 2. ৽͍͠ਪનख๏ͷಋೖ 3. ಋೖޙͷ࠶ܭଌɼޮՌ൑ఆɼޮՌతͳख๏ͷద༻ • ࣄલʹఆΊͨจ຺͝ͱʹਪનख๏ͷબ୒Λࣗಈత͔ͭܧଓతʹ࠷దԽ͢Δਪન γεςϜΛఏҊ

Slide 20

Slide 20 text

20 ਪનख๏ͷಋೖ • ධՁର৅ͷਪનख๏ΛਪનγεςϜ͕౳Ձʹѻ͑ΔΑ͏ڞ௨ͷΠϯλʔϑΣʔ εΛ࣋ͭϞδϡʔϧͱͯ͠ఆٛ • ਪનॲཧ͸ڞ௨͢ΔϑΟϧλʹΑͬͯߏ੒͞ΕΔ • Profileʢར༻ऀͷ৘ใΛऩूʣ • AssociationʢϓϩϑΝΠϧͱ৚݅ͷඥ෇͚ʣ • Searchʢ৚݅ʹै͍ީิΛݕࡧɼฒସ͑ʣ • ϑΟϧλͷڞ௨ར༻ʹΑΓอकੑͷ޲্

Slide 21

Slide 21 text

• ਪન݁Ռʹର͢Δར༻ऀͷ൓ԠΛܭଌ͢ΔͨΊͷϩάઃܭ • ར༻ऀ͝ͱͷϦΫΤετΛه࿥ • ར༻ऀ͝ͱͷਪનϦΫΤετʹର͢Δਪન݁ՌΛه࿥ • ਪનϦΫΤετͷ௚ޙͷߦಈ͕ਪન݁Ռͷ঎඼ʹؔ͢Δߦಈ͔Λൺֱ 21 จ຺͝ͱͷܭଌͱධՁ 5JNF $POUFYU .FUIPE 6TFS 1BUI 1BSBNT 3FTQPOTF $IBJS JNBHF " SFDPNNFOE " TIPX $IBJS DG # SFDPNNFOE # TIPX ਪનͷडೖ

Slide 22

Slide 22 text

• ෳ਺ͷਪનख๏͔ΒಘΒΕΔޮՌΛ࠷େԽ͢Δ • ଟ࿹όϯσΟοτ໰୊ͱͯ͠ղऍ͠ɼෳ਺ͷਪનख๏ͷޮՌʢΫϦοΫ཰΍ߪ ೖ཰ʣΛ࠷େԽ͢ΔͨΊEpsilon-GreedyΞϧΰϦζϜΛ࠾༻ • A/Bςετʹ͓͚Δ௿ධՁͷख๏ར༻࣌ͷػձଛࣦΛճආ͢ΔͨΊͷख๏ • ࠷ॳ͸A/BςετͷΑ͏ʹಉස౓Ͱ੾Γସ͑Δ͕ධՁͷ஝ੵʢใु:rewardʣ ʹ൐͍ར༻ස౓ʹॏΈ෇͚͕ͳ͞ΕΔ 22 จ຺͝ͱͷධՁ Џ ׆༻ ୳ࡧ &QTJMPO(SFFEZΞϧΰϦζϜ

Slide 23

Slide 23 text

23 ධՁ݁Ռͷ൓ө ਪનγεςϜ͸ɼଟ࿹όϯσΟοτ໰୊ͱͯ͠ѻͬͨจ຺͝ͱͷޮՌతͳਪનख ๏ͷબ୒݁ՌΛఆظతʹऔΓࠐΈɼEpsilon-GreedyΞϧΰϦζϜͷॏΈ෇͚ʹ ै͍׆༻ํ਑Λมߋ͢Δ

Slide 24

Slide 24 text

Synapse 24 Context Routing Context Context Method Method Method Method Matching Process 0.33 0.33 0.33 Search Result Bandit Activity log Rewards Algorithms Epsilon- Greedy Softmax Feedback

Slide 25

Slide 25 text

Synapse 25 Context Routing Context Context Method Method Method Method Matching Process 0.1 0.8 0.1 Search Result Bandit Activity log Rewards Algorithms Epsilon- Greedy Softmax Feedback

Slide 26

Slide 26 text

4. ධՁ

Slide 27

Slide 27 text

• ECαΠτͰར༻தͷਪનख๏ͷΫϦοΫ཰࣮੷ʹΑΔఏҊख๏ͷޮՌ༧ଌ • ΫϦοΫ཰࣮੷ɿӾཡதͷ঎඼ʹର͢Δਪન঎඼ΛఏҊ͢Δػೳ • ਪનख๏ɿྨࣅը૾ɼྨࣅςΩετɼڠௐܕਪનʢϓϦϛςΟϒɼLLRʣɼ σϞάϥϑΟοΫ • ϞϯςΧϧϩ๏ʹΑΔྦྷੵใु༧ଌΛൺֱ • ࣄલʹఆΊΔจ຺ͱͯ͠Ӿཡதͷ঎඼ͷΧςΰϦΛ࠾༻ • ࠷దԽͷޮՌଌఆɿ࠷దԽʹΑΔྦྷੵใु༧ଌͷมԽΛൺֱ • จ຺ͷޮՌଌఆɿจ຺͝ͱͷ࠷దԽͷ༗ແͰྦྷੵใु༧ଌͷมԽΛൺֱ 27 จ຺Λߟྀͨ͠ਪનख๏ͷબ୒ͷ࠷దԽ

Slide 28

Slide 28 text

จ຺͝ͱͷਪનख๏ͷޮՌͷࠩ 28 • จ຺͸ɼར༻ऀ͕Ͳͷ঎඼Χ ςΰϦΛݟ͍ͯΔ͔ • ਪનख๏͔ΒͷఏҊʹର͢Δ ΫϦοΫ཰Λൺֱ • ΧςΰϦ͝ͱʹਪનख๏ͷޮ Ռͷ͕ࠩ͋Δ͜ͱ͕ݟͯऔΕ Δ ঎඼ΧςΰϦ͸ར༻ऀͷจ຺ͷ͏ͪγεςϜ͕೺ ѲͰ͖Δ΋ͷͰଞͷECαΠτͰ΋ల։͠΍͍͢ɽ

Slide 29

Slide 29 text

ྦྷੵใु༧ଌ 29 • ࠷దԽΛߦͳ͍ͬͯͳ͍΋ͷ(1. No optimization) ͱൺֱͯ͠ ࠷ద ԽΛ͓͜ͳͬͨ΋ͷʢ2. Overall optimization, 3. Category-wise optimizationʣͷྦྷੵใु༧ଌ͕ ߴ͍ • ࠷ऴతͳྦྷੵใु༧ଌ͸จ຺ߟྀ ͨ͠࠷దԽ(3. Category-wize optimization)͕࠷΋ߴ͍

Slide 30

Slide 30 text

ྦྷੵใु༧ଌʢ্ཱ͕ͪΓʣ 30 • ࠷ॳͷൺֱͷઌ಄1000ճͷΈΛൺ ֱͨ͠΋ͷ • จ຺ߟྀ(3. Category-wise optimization)ͷ৔߹ɼจ຺͝ͱʹ ֶश͕ߦΘΕΔ͜ͱ͔Β্ཱ͕ͪ Γʹ͸͕͔͔࣌ؒͬͨ

Slide 31

Slide 31 text

ྦྷੵใु༧ଌʢ༏ྼࠩͷগͳ͍΋ͷʣ 31 • ࠷ॳͷൺֱͷΫϦοΫ཰ͷࠩΛҙ ਤతʹ௿Ίͨ΋ͷ • ࠷ॳͷ࣮ݧͱಉ͡ॱҐ͚ͮʹͳΔ ͕༏ྼࠩͷ൑அ·Ͱʹଟ͘ͷࢼߦ ճ਺Λཁͨ͠

Slide 32

Slide 32 text

• ਪનख๏ͷ༏ྼΛॿ௕͢Δద੾ͳ৚݅Λબఆͯ͠ɼ͜ΕʹԠͨ͡ਪનख๏ͷબ ୒ͷ࠷దԽΛߦ͏͜ͱͰྦྷੵใुֹͷ޲্ʹͭͳ͕Δ • ຊݚڀ͸؀ڥͷมԽͷଟ͍ঢ়گΛҙਤ͍ͯ͠Δ͜ͱ͔Βɼগͳ͍ࢼߦճ਺Ͱͷ ࠷దԽΛਤΕΔΑ͏ͳํࡦʹΑΔվળ͕ظ଴Ͱ͖Δ • ࢼߦճ਺ͷ૿Ճʹ൐͍୳ࡧ཰Λ௿ΊΔʢΞχʔϦϯάʣ • ଞͷόϯσΟοτΞϧΰϦζϜʢUCB, softmax…) • จ຺෇͖όϯσΟοτ 32 ධՁ

Slide 33

Slide 33 text

5. ·ͱΊ

Slide 34

Slide 34 text

• ར༻ऀͷจ຺ʹԠͯ͡ਪનख๏ͷಘखෆಘख͕͋Δ͜ͱ͕Θ͔ͬͨ • ఏҊख๏ʹΑΓख๏ಋೖͱจ຺ʹԠͨ͡࠷దͳख๏બఆ͕༰қʹͳΓɼػձଛ ࣦͷͳ͍ਪનγεςϜΛޮ཰తͳߏங͕Մೳͱͳͬͨ • ݱࡏɼจ຺ΛࣄલʹఆΊΔඞཁ͕͋ΔͨΊɼಘखෆಘख͕ੜ͡Δจ຺ʹ͍ͭͯ ௐࠪΛਐΊΔ • ಘखෆಘखΛิ͍߹͑ΔΑ͏ʹͳͬͨ͜ͱͰɼݶఆతͰ͋ͬͯ΋ޮՌͷߴ͍ਪ નख๏ͷ༗༻ੑ͕૿͢͜ͱ͕ߟ͑ΒΕΔͨΊɼͦͷΑ͏ͳख๏ͷݕ౼ΛਐΊΔɽ 34 ·ͱΊ

Slide 35

Slide 35 text

No content