基于可扩展分解机器的搜索广告点击率预估

发布时间：2018-02-09 16:50

本文关键词： 搜索广告点击模型可扩展分解机器特征体系 CTR预估　出处：《浙江大学》2013年硕士论文　论文类型：学位论文

【摘要】：搜索广告通过用户的检索行为触发广告的生成,是目前互联网流量变现的主要模式之一。搜索引擎一般按照广告的点击率(Click-Through-Rate, CTR)和广告的出价之积来筛选广告,其中CTR预估是核心问题之一,它对搜索引擎的收入和用户的体验都有重大影响,点击模型是预估CTR的主要方式。但是,由于搜索广告数据量大,特征维度高且长尾现象明显,目前大多数点击率预估系统无法高效的在如此大量的稀疏且高维的数据上准确预估CTR。因此设计恰当的特征体系,建立高效可扩展的点击模型,并且利用在线优化算法快速迭代成为急需解决的三个问题。本文的主要工作如下： 1.提出点击率预估特征设计原则,基于此提出五组特征,包括从用户,广告,以及环境三个维度,由单特征到组合特征,统计特征等多粒度的特征集合； 2.设计了用作点击模型的可扩展分解机器SFM,利用维度树结构将分解机器所基于的切片张量分解重构成层次切片分解。由于利用了层次结构,不仅模型的空间复杂度大大下降而且方便了在线算法的设计,此外,借助于分解模型,模型能在长尾数据上更准确的预估参数,利用proximal gradient方法,在线工作集优化算法能更快收敛。 3.基于真实搜索广告日志数据进行了三组实验来模拟线上CTR的预估,结果表明,CTR特征体系能够有效提升模型对搜索广告点击率预估的准确度,SFM相比FM能够有效降低模型的存储空间,online策略的工作集算法比batch策略能更快的收敛到局部最优解,在点击率预估方面,SFM在高频和长尾搜索日志数据上的CTR预估准确度以及排序能力均于好于作为基线模型的分解机器,Logistic Regression模型和User Browsing模型。
[Abstract]:Search advertising triggers the generation of advertisements through users' retrieval behavior, which is one of the main modes of Internet traffic realization at present. Search engines usually screen advertisements according to the click rate of advertisements and the product of advertising bids. CTR estimation is one of the core problems, which has a significant impact on search engine revenue and user experience. Click model is the main way to predict CTR. However, because of the large amount of search advertising data, the feature dimension is high and the phenomenon of long tail is obvious. At present, most click rate prediction systems can not predict CTRs accurately on such a large amount of sparse and high-dimensional data. Therefore, the appropriate feature system is designed to build an efficient and scalable click model. And the rapid iteration of online optimization algorithm has become the urgent need to solve three problems. The main work of this paper is as follows:. 1. The principle of feature design for the prediction of click rate is proposed. Based on this, five groups of features are proposed, including the multi-granularity feature set from user, advertisement and environment, from single feature to combinatorial feature and statistical feature. 2. The scalable decomposition machine SFMused as click-model is designed. The slice Zhang Liang based on the decomposition machine is decomposed into hierarchical slices by using dimension tree structure. Not only the space complexity of the model is greatly reduced, but also the design of the online algorithm is convenient. In addition, with the help of the decomposition model, the model can predict the parameters more accurately on the long tail data and use the proximal gradient method. The online working set optimization algorithm can converge faster. 3. Based on the real search advertising log data, three groups of experiments are carried out to simulate the online CTR prediction. The results show that the CTR feature system can effectively improve the accuracy of the model in predicting the click rate of search advertisements. Compared with FM, the working set algorithm of the model can effectively reduce the storage space of the model and converge to the local optimal solution faster than the batch strategy. The accuracy and sorting ability of CTR prediction on high frequency and long tail search log data are better than that of decomposition machine logistic Regression model and User Browsing model.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.3

【共引文献】