Click Through Rate Estimation For Sponsored Advertising Based On Scalable Factorization Machine

Search advertising is one of the major traffic monetization methods, is triggered by the queries. Search advertising is charged by clicks, hence, search engines choose ads according to the product of Click Through Rate (CTR) and bid provided by the advertisers. Consequently, CTR estimation becomes one of the key problems which poses significant influence on revenue of search engines and users’experiences. Click models are major tools for estimating CTR.However, Search ads log is a massive dataset with extremely high dimension features which includes a large proportion of long tail data; unfortunately, most of previous click models were unable to effectively estimate CTR on such sparse and high-dimensional data. Thus, designing appropriate feature system, building scalable click model and applying efficient online optimization algorithm to accelerate the convergence become three key problems. Our contribution about these problems are the follows:1. It proposes the CTR feature designing principle, and based on it, five sets of features are constructed from user, ads and context aspects. These features have multiple hierarchical granularity which includes single features, combinational features and statistical features;2. It designs scalable factorization machine for click model which mortifies the slice tensor factorization which Factorization Machine relied to hierarchical slice factorization based on dimension tree. Thanks to hierarchical structure, the storage complexity is greatly reduced and the online algorithm is easier to be devised, moreover, due to the factorization properties, the model’s parameters are easier to be estimated precisely on long tail dataset; It leverages the path-coding regularization term and hierarchical structure to combine features effectively, and online working-set algorithm to optimize the model which based on proximal gradient method.3. Three real world experiments shows that CTR feature system promotes the estimation accuracy dramatically, SFM has much lower storage complexity than FM, and the online strategy is more helpful to convergent to local optima than batch ones, besides, in CTR estimation, SFM outperforms all the other click models, including Factorization Machine, Logistic Regression and User Browsing Model both in long tail and high-frequency search ads log data with regard to the accuracy and ranking criteria.