280 likes | 442 Views
Ad Heat : An Influence-based Diffusion Model for Propagating Hints to Match Ads. -- WWW 2010, Hongji Bao, Edward Y. Chang Google Research, Beijing, China Presented by Flame Wang. Outline. Social Network Ad Model Relevance Model Ad Heat : Influence Model Algorithms
E N D
AdHeat: An Influence-based Diffusion Model for Propagating Hints to Match Ads --WWW 2010,Hongji Bao, Edward Y. Chang Google Research, Beijing, China Presented by Flame Wang
Outline • Social Network Ad Model • Relevance Model • AdHeat: Influence Model • Algorithms • Experiment Results • Conclusion
Content-based Ad Model Ads • Analyzing webpage content relevance • Not personalized
Social Network Ad Model Ads • User-targeting • Targeting Ads at Social Network Users • Mining Profiles, Friends & Activities for Relevance
Social Network Ad Model Activities Profiles Friends
Disadvantage for Social Network Ad Relevance Model • Observation: • Influential users do not click on relevant ads.
Disadvantage for Social Network Ad Relevance Model • Observation: • Non-influential user data are too sparse for relevance analysis. • Influential users attract and are followed by many non-influential users, like a heat source.
AdHeat Model • Considering both relevance & influence • To solve data sparsity. • Propagating influence by heat diffusion • To improve ad matching relevance.
AdHeat Model Social Graph Construction Relevance Analysis (Hint word generation) Influential User Ranking Influence Propagation
AdHeat Model Influence
AdHeat Model • Hint Word Generation • Construct a latent layer for better semantic matching in relevance, • Each user is presented by some hint words, • Latent Dirichlet Allocation (LDA).
AdHeat Model • Influential User Ranking • Social graph construction • A weighted directed graph G(U,E), U: users, E: edges, : weight of user i dependence to user j • Edges built by the frequency and quality of users’ interactions. • BBS/Forum: no. of post’s page view • Q&A: some features of answers
Q&A User • Features: [Xiance Si,09]
Influential User Ranking • Influential User Ranking • Influential users defined by level of activity & authority. • HITS algorithm to determine influence score • Hub score: activity to propagate information • Authority score: authority to provide contents
HITS • HITS algorithm • Iteratively update till convergence • Influence score: W: adjacency matrix of social graph, : W with its rows normalized to sum to one, : reset probability
Influence Propagation • Illustrative Example
Heat Diffusion Model • Heat equation: • Heat diffusion on directed graph ,
Experiment Results • Dataset • Google Confucius (Q&A community), • Half a million registered users’ data in one month for conducting AdHeat, • 5,000 active users for targeting ads, • Ads data from Google AdSense.
Evaluation • Metric:
Evaluation • Influence-based model without propagation and content-based model
Evaluation • Influence model with and without propagation
Conclusions • AdHeat utilizes both relevance and influence. • Improve ad matching performance on CTR. • Solve the data sparsity problem.