Learning from Multi-User Activity Trails for B2B Ad Targeting - PowerPoint PPT Presentation

Learning from Multi-User Activity Trails for B2B Ad Targeting Shaunak Mishra*, Jelena Gligorijevic, and Narayan Bhamidipati Yahoo Research, Verizon Media 1

B2B products Business-to-business (B2B) versus business-to-consumer (B2C) ❖ Same advertiser can have B2B and B2C products ❖ small-medium enterprise WiFi (B2B) service (advertiser) home internet (B2C) 2 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Online buying behavior: B2C unaware p u r c aware c h o a s n e s u f consider u m n n e e r l j intent o s t u a r g n e e buy s y 3 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Online buying behavior: B2C unaware p u r c aware c h o a s n e s u f visiting websites, consider u m n forums, tech blogs n e e r l j intent o s t u a r g n e e buy s y online purchase 4 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Online buying behavior: B2C More details on B2C KDD Applied Data Science in poster # 76 unaware p u r c aware c h o a s n e s u f visiting websites, consider u m n forums, tech blogs n e e r l j intent o s t u a r g n e e buy s y online purchase 5 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Online buying behavior: B2C vs. B2B B2C B2B unaware aware analyst, IT manager consider intent influence buy owner 6 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Conversion (purchase) prediction Conversion prediction is crucial for ad targeting ❖ ❖ Will conversion prediction suffer from multi-user activities for B2B purchase? user 2 user 1 user 3 conversion prediction model 7 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Conversion (purchase) prediction Conversion prediction is crucial for ad targeting ❖ ❖ Will conversion prediction suffer from multi-user activities for B2B purchase? user 2 user 1 user 3 conversion prediction model Assumptions ❖ No declared profiles! Online activity trails available ❖ 8 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Conversion (purchase) prediction Conversion prediction is crucial for ad targeting ❖ ❖ Will conversion prediction suffer from multi-user activities for B2B purchase? user 2 user 1 user 3 conversion prediction model user 2 trail: act2, act4, search WiFi offers , act5, act6 user 1 trail: act1, act2, search WiFi offers , act3, act4, purchase user 3 trail: act1, act3, act5, purchase 9 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Conversion (purchase) prediction Conversion prediction is crucial for ad targeting ❖ ❖ Will conversion prediction suffer from multi-user activities for B2B purchase? user 2 Model cannot relate user 1 Model can relate “search WiFi offers” “search WiFi offers” to purchase user 3 to purchase conversion prediction model user 2 trail: act2, act4, search WiFi offers , act5, act6 user 1 trail: act1, act2, search WiFi offers , act3, act4, purchase user 3 trail: act1, act3, act5, purchase 10 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Conversion (purchase) prediction Conversion prediction is crucial for ad targeting ❖ ❖ Will conversion prediction suffer from multi-user activities for B2B purchase? Intuition: augmenting user trails user 2 may help! user 1 Model can relate “search WiFi offers” user 3 to purchase conversion prediction model user 2 trail: act2, act4, search WiFi offers , act5, act6 user 1 trail: act1, act2, search WiFi offers , act3, act4, purchase user 3 trail: act1, act3, act5, purchase 11 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Trail augmentation: relevant users and acts user 2 trail: user 3 trail: act2, act4, search WiFi offers , act5, act6 act1, act3, act5, … purchase Consider users in the same cluster/household 12 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Trail augmentation: relevant users and acts user 2 trail: user 3 trail: act2, act4, search WiFi offers , act5, act6 act1, act3, act5, … purchase relevant act (advertiser specific) 13 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Trail augmentation: relevant users and acts user 2 trail: user 3 trail: act2, act4, search WiFi offers , act5, act6 act1, act3, act5, … purchase relevant act relevant user 14 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Trail augmentation: relevant users and acts user 2 trail: user 3 trail: act2, act4, search WiFi offers , act5, act6 act1, act3, act5, … purchase relevant act Model can relate “search WiFi offers” to purchase! relevant user augmented user 3 trail: act1, act3, act5, [ act2, act4, search WiFi offers, act5, act6 ] , … purchase 15 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Proposed approach users in an org (commercial IP, household ID) relevant acts seed list 16 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Proposed approach users in an org Identify (commercial IP, relevant users household ID) relevant acts seed list 17 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Proposed approach users in an org identify augment user trails in (commercial IP, relevant users cluster by household ID) relevant users’ trails relevant acts + seed list use in conv. model 18 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Proposed approach users in an org identify augment user trails in (commercial IP, relevant users cluster by household ID) relevant users’ trails relevant acts + seed list use in conv. model 19 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Relevant acts seed list Start with high “conversion rate” activities ❖ advertiser + manual review specific (e.g., visiting advertiser website) list of Iteratively expand seed list using online activities ❖ activity2vec embeddings 20 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Activity2vec based expansion activity2vec = word2vec style training with activity trails user trail = document user i trail: act1, act3, act5, … act10, act1, act2 session = sentence act = word 21 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Activity2vec based expansion activity2vec = word2vec style training with activity trails user trail = document user i trail: act1, act3, act5, … act10, act1, act2 neighbor based session = sentence iterative seed act = word expansion 22 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Results initial seed list noise in expanded seed list vs. no augmentation 23 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Results initial seed list noise in expanded seed list vs. no augmentation 24 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Summary Trail augmentation helps! ❖ Seed list useful for targeting, insights! ❖ Uniform influence assumption ❖ 25 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”

Summary Trail augmentation helps! ❖ Seed list useful for targeting, insights! ❖ Uniform influence assumption ❖ Questions? ❖ 26 Mishra, S., Gligorijevic, J., Bhamidipati, N. “Learning from Multi-User Activity Trails for B2B Ad Targeting”