Realtime Bid Optimization with Smooth Budget Delivery in Online Advertising

Transcript

1. 3FWJFX
   
 l3FBM
   5JNF
   #JE
   0QUJNJ[BUJPO
   XJUI
   4NPPUI
   #VEHFU
   %FMJWFSZ
   JO
   0OMJOF
   "EWFSUJTJOHz JUJU
2. 1. Introduction 2. Background and Related Work 1. Problem Setup 3. Online Bid Optimization 1. Smooth Delivery of Budget 2. Selection of High Quality Ad Requests - Flat CPM Campaigns 3. Selection of High Quality Ad Requests - Dynamic CPM Campaigns 4. Estimation of CTR and AR (CVR) 4. Practical Issue 1. Cold Start Problem 2. Prevention of Overspending 5. Experimental Results 1. Comparison of Pacing Strategies 2. Evaluation of Real Campaign Performance 6. Conclusions Agenda
3. 1. Introduction • RTBʹ͓͚Δ࠷దԽϓϩηε “༧ࢉεέδϡʔϧ಺ͷ༧ࢉ੍໿Լʹ͓͍ͯɺCampaignͷύϑΥʔϚϯεͷ࠷େԽΛਤΔ”͜ͱ • ύϑΥʔϚϯεͷܭଌࢦඪ • CPCͱCPAͷ࠷খԽ • CTRͱAR(CVR)ͷ࠷େԽ • ༧ࢉ഑৴੍໿ͷฏ׈Խ • ઃఆ࣌ؒΑΓલʹɺڵຯͷ͋ΔimpͷҰ෦Ҏ্Λߪೖ͠ͳ͍Α͏ʹ͢Δ • εέδϡʔϧ్தʹ༧ࢉ͕ਚ͖Δ͜ͱ΍ɺա৒ͳߪೖΛ๷͙ • ຊݚڀʹ͓͚ΔఏҊɾݕ౼ • ֤CampaignͷͨΊͷฏ׈ͳ഑৴੍໿Λຬ଍ͤ͞Δؒͷ
 ύϑΥʔϚϯεMetricsΛ࠷దԽ͢ΔͨΊͷΦϯϥΠϯΞϓϩʔνΛఏҊ • ฏ׈Խ഑৴੍໿Λ՝ͨ͢Ίͷະདྷͷফඅ཰Λ܁Γฦ͠ਪఆͨΊʹ
 ίϯτϩʔϧɾϑΟʔυόοΫɾϧʔϓΛݕ౼
4. 2. Background and Related Work 1. Problem Setup time slot ຖͷ༧ࢉͷ՝୊ goal ࢦඪͱ੍໿৚݅ ֫ಘͨ͠Ձ஋ ೖࡳճ਺ ޿ࠂओͷίετ time slot ຖͷ༧ࢉ time slot t Ͱফඅ͞Εͨ༧ࢉ eCPMͷ্ݶ஋ “ε-δ࿦๏”
5. 3. Online Bid Optimization 1. Smooth Delivery of Budget ফඅ༧ࢉ time slot ຖͷ౤ࢿ཰ͷߋ৽ time slot t ͷ޿ࠂϦΫΤετ૯਺ ࣍ͷtime slot t+1ͷad request਺ͱwinning rateͷ༧ଌ஋ Feed Back Control ΛՄೳʹ͢Δ time slot t Ͱফඅ͞Εͨ༧ࢉ
6. time slot t ͷ1ظઌͷ༧ࢉͷࢉग़ ༧ࢉ͕Ұ༷഑෼ͷ৔߹ ༧ࢉ͕Ұ༷഑෼Ͱ͸ͳ͍৔߹ - U ͕UʹΑͬͯ
 ҟͳΔ৔߹ - U ͕֤UͰ
 Ұ༷ͳ৔߹ time slot t ͷ࣌ؒతͳ௕͞ʢsecʣ Click or CVͷ༧ଌ֬཰ ೔༧ࢉͷ࢒ֹ time slotͷ௕͞ͷൺ click or conversion֬཰͕
 ߴ͍Campaignʹଟ͘ͷ༧ࢉΛׂ͘ ˎ ͜ͷϩδοΫ͚ͩͰ͸ɺclick or conversion֬཰͕θϩͷCampaignʹ͸༧ࢉׂ͕͔Εͳ͍ͨΊɺ2ͭͷઓུΛ྆ํѻ͏ *
7. 3. Online Bid Optimization 2. Selection of High Quality Ad Requests - Flat CPM Campaigns goalɿ
 ݱࡏͷslotͷpacing rate Λߟྀ͠ɺ౤ࢿ͢΂͖ϦΫΤετΛબ୒͢Δ͜ͱ ݻఆ͞ΕͨೖࡳՁ֨ ʹ
 time slot ຖͷ c_i ͷ౷ܭྔΛࢉग़͢ΔͨΊʹ࢖༻
 σʔλͷ೔਺ͷฏۉ஋ աڈͷ࣮੷σʔλ͔Βࢉग़ ʢࢀߟʣ
8. ๲େͳad requestྔ͔ΒɺClick or CV͠қ͍ad requestΛબ୒͢Δඞཁੑ CTR, CVRʹج͍ͮͨtime slot tʹ͓͚Δad requestͷ෼෍ CTR, CVRΛج४ͱͨ͠ϑΟϧλʔͰߟྀ͠ɺબ୒͢Δ q_t(s)͕աڈͷ࣮੷σʔλ͔Βࢉग़͞ΕΔͨΊɺਖ਼֬ͳσʔλͱ͸ݴ͍೉͍ ରࡦͱͯ͠ɺτ(t)ʹ৴པ۠ؒΛઃ͚Δ time slot t ʹ͓͚Δᮢ஋
9. 1.96
 (95% confidence interval) • τ(t)ʹ৴པ۠ؒʢᮢ஋ʣ աڈͷ࣮੷σʔλ͔Β౷ܭྔΛܭࢉ͢ΔͨΊͷσʔλ೔਺ Ad request ࠾୒্ݶ Ad request ࠾୒Լݶ ৴པ۠ؒΛݩʹbid priceΛࢉग़ µ(t) + ⌧ (t) p d µ(t) ⌧ (t) p d τ(t)͸Ψ΢ε෼෍ΛԾఆ τ(t)͕৴པ۠ؒ಺Ͱ͋Ε͹ɺpacing_rate(t)ͷ֬཰Ͱೖࡳ ˎɹ৴པ۠ؒͷᮢ஋ܭࢉࣜ ҎԼͷ਺ࣜΛ༻͍ͯɺτ(t)Λܭࢉ ˎ “ t -1 ” ͱ͍͏աڈͷtime slotΛ༻͍ͯࢉग़
10. 3. Online Bid Optimization 3. Selection of High Quality Ad Requests - Dynamic CPM Campaigns 1. bid price c* ͷࢉग़ 2. base bid priceͷࢉग़ ui = AR ⇥ G base bid price ༧ଌ஋ (CVR) ໨ඪͱ͢ΔCPA 3. pacing_rate(t)ʹΑͬͯೖࡳՁ֨Λௐ੔ ݻఆ͞ΕͨೖࡳՁ֨ ʹ
 time slot ຖͷ c_i ͷ౷ܭྔΛࢉग़͢ΔͨΊʹ࢖༻
 σʔλͷ೔਺ͷฏۉ஋ CPA CampaignʹϑΥʔΧε pacing rate(t)͸ɺbiddingͷස౓Λίϯτϩʔϧ͢Δ
11. ˆ ci = ✓⇤ui ✓ = ci ˆ ci bidprice ˆ ci = ui ˆ ci = ⇢⇤ui (1) ɹSafe region : pacing rate(t) <= β1 ώετάϥϜθͷԼҐ1ʙ2ˋ མࡳՁ֨ σʔλΛ஝ੵ͠ɺ
 ώετάϥϜΛ࡞੒ (2) ɹCritical region : β1 <= pacing rate(t) <= β2 (3) ɹDanger region : β2 <= pacing rate(t) 1. ϦΫΤετΛϦδΣΫτ͍͗ͯ͢͠Δ 2. ೖࡳՁ͕֨௿͗͢Δ λʔήςΟϯάͷ੍໿͕ݫ͗͢͠Δ ҎԼͷࣜΛద༻ ೖࡳՁ֨ͷ্ݶ 3. ద੾ͳೖࡳՁ֨Ͱೖࡳ͢ΔͨΊɺpacing_rate(t)Λ3ྖҬʹ෼͚ͯɺೖࡳ ೖࡳՁ֨
12. 3. Online Bid Optimization 4. Estimation of CTR and AR (CVR)
13. 4. Practical Issue 1. Cold Start Problem ✴ ղܾࡦ • ྨࣅͨ͠CampaignσʔλΛ࢖༻
14. 4. Practical Issue 2. Prevention of Overspending ✴ ༧ࢉރׇͷ๷ࢭࡦ • ֤time slotͷ༧ࢉΛ௒͑ͨ৔߹͸࣍ͷtime slot·ͰೖࡳΛதࢭ
15. ཧ૝஋ͱͷဃ཭
 ʢ༧ࢉͷҰ༷഑෼Λཧ૝ͱͨ͠৔߹ʣ dailyͷޡࠩ͸໿1% x : time slot y : total budget spend 5. Experimental Results 1. Comparison of Pacing Strategies the ideal uniform spend and the actual spend using our uniform pacing strategy ͷ৔߹
16. x : time slot y : total budget spend x : time slot y : relative performance
 in term of 1/eCPA dailyͷޡࠩ͸໿2.3% ཧ૝ͷ౤ࢿ഑෼ ཧ૝஋ͱͷဃ཭
 ʢ֤time slotͰ
 ༧ࢉ͕ҟͳΔ৔߹ʣ the ideal spend and the actual spend using performance based pacing strategy ͷ৔߹
17. 5. Experimental Results 2. Evaluation of Real Campaign Performance the entire bid optimization framework 1. Flat CPM campaign • CTRͷ࠷େԽ 2. Dynamic CPM campaign • CPC, CPAͷ࠷খԽ
18. • Compare with baseline model (in Flat CPM campaign) Baseline modelͱ͸ɺtime slot tʹ͓͚Δpacing rateʹجͮ͘CTR(t)ͷᮢ஋͔Βࢉग़͢Δ
 ؆୯ͳదԠϑΟʔυόοΫ੍ޚΞϧΰϦζϜͰ͋Δɻ ฏۉͰ123%ͷύϑΥʔϚϯεվળ͕ݟΒΕΔ • Detail of Evaluation • ҟͳΔ޿ࠂओΧςΰϦʔ͔ΒϥϯμϜʹબ୒ͨ͠7ͭͷFlat CPM campaignΛ࢖༻ • ͦΕΒ7ͭͷcampaignΛѻ͍ɺఏҊ͢ΔΞϓϩʔνͱطଘͷΞϓϩʔνΛධՁ • ֤Method͸ɺ1िؒՔಇͤͨ͞ޙɺ2िؒ෼ͷσʔλΛར༻ͯ͠ൺֱ
19. • Compare with baseline model (in Dynamic CPM campaign) Baseline modelͱ͸ɺpacing rateΛҰ༷ʹௐ੔ͨ͠طଘͷϞσϧ • Detail of Evaluation • ҟͳΔ޿ࠂओΧςΰϦʔ͔ΒϥϯμϜʹબ୒ͨ͠Dynamic CPM campaignΛ
 ໨ඪઃఆຖʹ഑ஔʢCPC/CPAʣ͠ɺධՁ • ఏҊ͢ΔΞϓϩʔνͱطଘͷΞϓϩʔνͷLiftΛධՁ ฏۉͯ͠CPC͸19.02%ɺCPA͸30.87%ͷ
 ύϑΥʔϚϯεվળ͕ݟΒΕΔ
20. 6. Conclusion • ϦΞϧλΠϜͰͷ༧ࢉ഑෼ͷ࠷దԽख๏ΛఏҊ • “഑৴ޮՌͷվળ”ͱ”༧ࢉ഑෼ͷฏ׈Խ”ͷ྆ํΛୡ੒ • million of ad request per second΋ͷσʔλྔΛϋϯυϧՄೳ • ༧ࢉͷunderpacing΍overpacingΛߟྀͨ͠ΞϧΰϦζϜΛߟҊ • ࠓޙ͸ɺ΋ͬͱॲཧͰ͖ΔσʔλྔΛ૿΍͢