A primary difference between web search and sponsored search (SS) is that the web corpus is significantly larger than the ads database with more diverse content making it more likely to find relevant documents on the web using keyword-match techniques. Additionally users' are more tolerant to bad search results than to bad ads  and are more willing to reformulate queries when no relevant search results are found. In SS, however, there are many queries for which there are no relevant ads and standard IR techniques can result in spurious matches to unimportant words in the query. In such cases it is much better not to show any ads. These issues get particularly exacerbated for tail queries where advance match plays a dominant role. Yet, when possible, we do want to find relevant ads since a significant proportion of the SS revenue lies in monetizing the tail well. In this context we think that good query analysis can help achieve better accuracy for sponsored search in the tail. We describe 2 key problems in this regard. The ideas are however not restricted to SS but apply to web search as well.
Derek Hao Hu, Evan Wei Xiang, Qiang Yang
Sponsored search has become increasingly important due to the rapid development of Web search engines and pay per click (PPC) is amongst one of the most important advertising models search engines currently use. One of the key questions in sponsored search is that: Given a query or a substituted keyword, which ads should search engines display to the users in order to maximize their revenue? In other words, given a keyword, how can we choose ads out of a candidate list that will have higher click-through rates (CTR)? Previous works have attempted to estimate the CTR of ads via a query-independent perspective. In this paper, instead of predicting the CTR of ads, we will propose a new ranking-based approach to select ads that would have higher click-through rates via a query-dependent perspective. We first analyze some manually constructed heuristic rules that could be used to distinguish good ads from bad ones and then show how we could combine these rules into our ranking-based approach to reach our aim. Experiments on real-world datasets have confirmed the effectiveness of our proposed approach.
Eni Mustafaraj, Panagiotis Takis Metaxas
We have collected a set of 1131 textual ads that appeared in the Google Search results when searching for a candidate name running in the 2008 US Congressional elections. We have categorized the advertisers in four different categories: commercial, partisan, non-affiliated, and media. By analyzing the content of the collected ads, we discovered that the majority of them (63%) are commercial ads that have no political message, while the partisan group contributed only 14% of the ads. Furthermore, only 21 out of 124 monitored candidates were actively participating in sponsored search, by providing their own political message. We describe the different ways in which the advertisements were used and several problems that damage the quality of sponsored search, providing some suggestions to avoid such issues in the future.
The advent of online advertising has brought with it the need for new designs for markets that succeed in attracting participants and becoming viable businesses. Real-world design has been guided by theoretical insights, practical experience, and continual feedback from experimentation and data analytics. Statistical models have dual roles: they are an important part of the technology that ranks and delivers ads, and they are also used to evaluate the performance of the marketplace. This talk will focus on the interaction between theory and statistical analysis in paid search. It will highlight theoretical models of advertiser bidding behavior and market design that incorporate consumer search, as well as empirical models that help evaluate advertiser incentives and behavior. There can be important feedback between the design and performance of algorithms for scoring ads, and the incentives faced by bidders in online auctions as well as the efficiency of these auctions.
The Effect of Some Sponsored Search Auction Rules on Social Welfare: Preliminary Results from an Exploratory Study in the Laboratory
Roumen Vragov, David Porter, Vernon Smith
When consumers search sponsored links provided by a search engine they interact with advertisers in two distinct but related markets: the market for ads, and the market for the advertised products. The purely theoretical exploration of such complex combinatorial markets is limited because it requires assumptions about consumer and advertiser behavior that are too strict. This study explores the effects of some sponsored search auction rules on consumer surplus, advertiser profits, and search engine revenues through the use of laboratory experiments with human subjects. We find that, from the options we explored, the best payment method is pay-per-click and the best way to rank ads is by past click-through rates. We also suggest ways to extend the experimental design further to explore other important parameter spaces.
Product search is quite similar to online advertising in that the target audience and content providers have significant commercial intent. However, to provide a rich shopping experience product search requires understanding products not just as text but as structured data. These user interfaces are unforgiving of underlying data errors. The text mining and machine learning techniques used in product search must thus have unusually high precision. This talk will provide an overview of these different challenges and present details on two such applications.
Sarah K Tyler, Yi Zhang, Dou Shen
Sponsored search seeks to align paid advertisements with interested individual search engine users. Existing sponsored search algorithms are based on advertisers’ bidding on individual search terms. Search terms, however, are not an accurate description of a user's information need or ads preferences. Additionally, advertisers are not always good at identifying all the search terms that are relevant to their products or services. On the other hand, collaborative recommendation systems assume users who have similar tastes on some items may also have similar preferences on other items, and thus make recommendations for one user based on the feedback from other similar users. In this paper, we explore whether collaborative filtering methods can help predict which users are likely to click on which advertisements. More specifically, we use the user supplied query as well as session based user information to build a better user profile for inferring the users' hidden information needs. We tried two basic collaborative filtering algorithms, a k-nearest neighbor, and a probabilistic factorization model, to determine whether collaborative filtering on sessions and queries can beneffit sponsored search. The experimental results on 100 million Microsoft search impression data set demonstrates the effectiveness of the collaborative filtering approach.
Maryam Karimzadehgan, Manish Agrawal, ChengXiang Zhai
Web advertising has become a financial backbone of business success nowadays. All major Web search engines such as Google, Microsoft and Yahoo! derive significant revenue from advertising. However, as a new area of research, online advertising has not yet reached its full potential. In particular, little research has been done on advertising on social networks. In this position paper, we present our review of some research issues related to advertising on social networks and some preliminary results in a related task of recommending news articles to users of Facebook.
Wen Ye, Wenjie Li, Furu Wei, Chunbao Li
Traditional ads keyword extraction approaches process a Web page as a whole. However, many current Web pages like Weblogs and discussion forums allow people to leave their comments, responses or follow-up questions on popular topics. Due to interaction among active participants, these pages often exhibit different focused topics in different places on the pages. In this paper, we emphasize on the linking relations that are built upon replies and quotations and propose a novel dynamic extraction approach for both inter-post and whole page ads keyword extraction. Preliminary evaluation results on Chinese forum data set demonstrate the effectiveness of the proposed approach.