Content area
Abstract
Search engines use sophisticated auction mechanisms to determine the positioning of text ads on their result pages. Advertisers seeking to promote their products or services must compete in sealed auctions to establish the order in which their ads will appear. More specifically, ad ranks are calculated based on the amounts the advertisers are willing to pay for each click, as well as the relevance of the ads and their related websites. Generally, ads placed at the top of the results list get more visibility and generate more clicks than the ads at the bottom of the list. Therefore, advertisers try to find the optimal bids that will allow them to obtain sufficient visibility and clicks, while maintaining reasonable costs.
Our main objective is to develop automatable algorithms that can increase the performance of search engine text ad campaigns. In order to achieve this, we first present a summary explaining the various aspects of this relatively new research field. Thereafter, based on several studies that have been published, we model the problem using a linear programming approach. Assuming an ad’s performance is measured by the number of clicks it generates, our linear program is designed to determine an optimal way to allocate the text ads to the various available positions while respecting a budget constraint. In order to estimate the number of clicks and the average cost per click for each position, the use of prediction functions is necessary.
Since keywords do not all provide the same quality of regressions, we develop a classification algorithm that allows us to identify which processing can be used, depending on each individual keyword’s characteristics. We aim to classify the keywords in a way that should reduce, as much as possible, the prediction errors. This should also maximize the scope of the optimization model, allowing us to estimate the clicks and costs per click of almost every single keyword in a campaign.
When we are able to obtain adequate regressions using the keywords’ historical data, it is relatively easy to predict clicks and cost per click as a function of position. However, our analysis shows that a large proportion of keywords do not provide statistically acceptable regressions, which makes it necessary to develop alternative prediction methods. We then study the possibility of using generic prediction functions. Such a prediction approach suggests that the relative decrease rate of the click and cost per click functions are almost constant from one keyword to another. After several tests performed on our databases, we conclude that the generic functions provide estimations that are precise enough to be used for prediction purposes. While comparing a linear function approach with an exponential function approach, we determine that the exponential version offers more advantages.
We then seek to refine our method, in order to ultimately improve the quality of predictions we are able to generate. We find that the dynamic repositioning of our prediction functions and the use of an exponentially decreasing weight in order to prioritize recent observations allow us to achieve much better results. Once these improvements are added to the method, we are quite satisfied with the prediction errors that are obtained. However, we still suggest several research areas that could potentially further improve the quality of the predictions.
Overall, this research project has allowed us to acquire a better understanding of the mechanisms that are used by search engines to manage their publicity networks. The optimization model, the classification algorithm and the prediction methods we suggest should eventually be integrated in a software platform, in such a way as to form a complete and functional optimization algorithm that can be used on a daily basis. Such an algorithm would allow advertisers to manage their text ad campaigns more efficiently and should increase their profitability.
