Research On The Algorithm Of Click-Through Rate Prediction

The explosion of Internet development has ushered in the era of online advertising. Highlighted by its popularity and trackable effect,online advertising undisputedly stands out as the main form of advertising. As the carrier of advertisements, sites, in most cases,adopt Cost-Per-Click(CPC) strategy as their major method of settlement. To gain more profits, realizing the maximization of clicks numbers within limited slots serves as a vital issue that requires careful contemplation. Click-Through Rate(CTR) prediction is a task to predict the probability of a certain user clicking some certain advertisements. Theoretically speaking, only if the sites accurately predict the CTR, putting the right advertisements in the right slots,users will naturally be attracted to click the advertisements.Therefore, accurate prediction is a necessary prerequisite for increasing the sites' profits, guaranteeing satisfactory advertisement effect as well as enhancing the user experience, which plays a key role in realizing the tripartite benefits of the sites, advertisers and users.CTR prediction has long been a concerned issue both in academia and industry, common solutions of which are machine learning and statistic analysis of data. Based on logistic regression,the CTR prediction algorithm is the optimized baseline algorithm that I employ to analyze data. In this paper, the design and optimization of the algorithm is accomplished, aiming to solve the problems that baseline algorithm may encounter in the process of display advertising and sponsored search. Specifically, in order to reduce the time and labor cost of artificial feature engineering and to improve the accuracy of prediction in display advertising, a CTR prediction algorithm based on GBDT and factorization machine is designed and applied, realizing the automation of feature engineering and the fitting of non-linear relationship in data.Additionally, in sponsored search, the textual similarity between a query and an advertisement title sometimes fails to represent users'real intent; therefore this paper also introduces a CTR prediction algorithm based on attention model, enhancing accuracy via the data mining of the similarity between the query and the advertisement title in the level of "search intent". It is attested by experimental result that the optimized CTR prediction algorithm outperforms baseline algorithm in time and accuracy, demonstrating a relatively high performance and enriching the research of CTR prediction.