Slide 1

1 Learning Relevance from Heterogeneous Social Network and Its Application in Online Targeting Chi Wang (UIUC), Rajat Raina (FB), David Fong (Stanford), Ding Zhou (FB), Jiawei Han (UIUC), Greg Badros (FB) Slide 2 2 Facebook Social Graph Slide 3 3 Relevance Matters Slide 4 4 Opportunity Can we use the rich, social information flowing through online social network to personalize user recommendations? Slide 5 5 Ad Targeting Goal: Estimate CTR for a given user/ad pair. Importance For users: relevant information For advertisers: more consumers Slide 6 6 Process Overview Slide 7 7 Concept Extraction (CE) Slide 8 8 Concept Aggregation and Match Slide 9 9 Baseline for CA and CM Slide 10 10 Limitations Concept Aggregation Different source type should have different weights. E.g., greetings between friends may not be as useful as clicks on a previous ad. Concept Match Different concept type should have different weights. E.g., top level node Society VS. leaf node Society/Relationships/Dating. One user concept can match multiple ad concepts. E.g., the video game FIFA Series can match the computer game or the sport soccer User concept space and ad concept space are not necessarily identical. Slide 11 11 Solution: Idea1 Slide 12 12 Solution: Idea2 Slide 13 13 Proposed Better Model Slide 14 14 Algorithms-Gradient Method Slide 15 15 Experiments Slide 16 16 Logloss-gain Slide 17 17 Interpretations Slide 18 18 Conclusion