Learning to Impress in Sponsored Search Xin Supervisors

Learning to Impress in Sponsored Search Xin Supervisors: Prof. King and Prof. Lyu 1

Sponsored Search query sponsored search results organic results sponsored search results Pay-per-click model: advertisers will pay an amount of money to search engine company once their ads shown in sponsored search results are clicked by users. 2

Sponsored Search is Important • Online advertising has obtained more and more money in recent years. • Sponsored search takes the main contribution for online advertising. 3

Impression in Sponsored Search 1 2 3 The research task: How many ads should be shown to users given a query? Ad Impression Number: The number of the advertisements shown to a user for a certain query in sponsored search. 4

Why learning the Ad impression number is important? 5

The Roles in Sponsored Search Advertiser Balance the utilities of advertisers, users, and publishers Advertisers wants ROI and volume Ad network wants the amount of users and advertisers Agency dual roles User wants relevance User Publisher wants revenue per impressions/search 6

If we do not control the Ad impression number… 7

Hidden Commercial Intrusion –Users will not be satisfied. • Some users will not use the search engine • Existing users will be trained to ignore the ads. –Advertisers will have less ROI. –Publishers will get less money. –The whole network will become smaller. 8

Evidences of Commercial Intrusion • Users have shown bias against sponsored search results [Marable 2003]. • More than 82% of users will see organic results first [Jansen 2006]. –Learning to impress is an important • Organic results have obtained much research issue in sponsored search higher clickthrough rate [Danescu. Niculescu-Mizil 2010]. • Irrelevant ads will train the users to ignore ads [Buscher 2010] 9

Motivation • Previous work is mainly about estimating the expected revenue of a query-ad pair. • This paper is focusing on the strategy to impress ads 10

Our Contributions • Formulate the task of learning to impress in sponsored search • We show that the revenues of query-ads pairs are not stable and therefore static method cannot work well. • We propose a novel dynamic approach for the unstable problem. • We combine the dynamic and static methods for another improvement. 11

Problem Definition 12

Evaluation Metric • The smaller the better. 13

Dataset • Sogou click-though data. 14

Static Method 15

Experimental Result 16

Unstable problem (1) 17

Unstable Problem (2) 18

Unstable Problem (3) 19

Unstable Problem (4) 20

Dynamic Method 21

Algorithm 22

Theoretically Analysis 23

Experimental Analysis (1) 24

Experimental Analysis (2) 25

Experimental Analysis(3) 26

Combination method 27

Real Value in Application 28

Conclusion • We formulate the problem of learning to impress task in sponsored search • We demonstrate the unstable problem • We propose a dynamic algorithm that can well solve the unstable problem • The combination method outperform static methods significantly 29