Personalized Advertising Recommendation Based On DSP-Research On User Interest Model

In recent years, the advertising industry is going under fast development thanks to a new advertisement injecting mode featured with Real Time Bidding(RTB). The platform for injecting ads in RTB, Demand Side Platform(DSP) helps the advertiser find the suitable target audience, thus increasing the return on investment. DSP takes a wide range of targeting approaches, including semantic targeting, behavior targeting and retargeting. In behavior targeting, users' behavior and the content they browse is analyzed to come up with their interest and preferences, so that related ads can be injected to them. Therefore, it is an important research subject for improving the effect of ads in behavior targeting.Interest model is established based on the content in the web page browsed by users and the measurement of their behavior. Currently, related researches use vector space model to textually represent the web page, and users' browsing behavior is used as a key indicator to measure their interest in the web page. Vector space model and browsing behavior measurement are combined to indicate the extent to which users are interested in the web page. However, this approach fails to consider the dynamic change of users' interest, which affects the accuracy of ad injecting. To solve this problem, the thesis takes the following steps.Firstly, the thesis builds the user interest model based on the vector space model and measurement of browsing behavior. Users' behavior is analyzed, the actions including saving, printing and adding to favorite, browsing frequency and browsing time are used as standards for measuring their interest. Then the vector space model, which has been preprocessed, is used to establish the user interest model. The model is improved by taking into consideration the different interest degree of users at different time of the day.Secondly, the thesis comes up with the dynamic interest model by combing the user interest model and interest stream model. So far, researchers have come up with a variety of approaches to show the dynamic interest of users. However, they rely so much on the choosing of parameters that they can't adaptively show the dynamic change of users' interest. In this thesis, the interest stream model, which adds Ebbinghaus forgetting curve to the user interest model, is established, laying foundation for the dynamic interest model.Thirdly, the thesis applies the dynamic interest model to ad matching in personalized advertising recommendation. Most of the existing targeted ads are textual ones. However, DSP ads contain only key words, and the traditional textual representing method doesn't apply. So the thesis counts the weights of the key words with short text similarity method and matches the interest clusters extracted with the clustering method with ads.With users' browsing record and ads as the research data, this thesis researches on ad matching based on the approach of personalized advertising recommendation. The result shows that the dynamic interest model effectively captures the change of users' interest while improving the accuracy of ad matching.