Predicting Click Through Rate And Conversion Rate In Online Advertising

With the proliferation of Internet and e-commerce,online advertising has become an effective way to promote products and gain revenue.In industry,user behavior is an important metric to measure the performance of online advertising.Therefore,plenty of researchers have studied user behavior.There are two basic problems of user behavior researches in online advertising,which are CTR(Click Through Rate)prediction and CVR(Conversion Rate)prediction.Predicting ad click through rate is an essential issue for Internet companies to gain revenue in online advertising ecosystem.To solve this issue,researchers proposed lots of approaches to improve the prediction accuracy of click through rate in sponsor search advertising and contextual advertising.However,in contrast with sponsored search ads and context ads,the special application scenario associated with display ads(e.g.,unavailable search query or contextual information)makes many state-of-the-art in the literature difficult to be utilized directly,therefore,CTR prediction in DA is also a big challenge.Furthermore,CVR is a key metric to measure the performance of online advertising,since it directly reflects the revenue of advertising.CVR refers to the proportion of audiences who take a predefined,desirable action(such as purchasing an item,adding to a cart,adding favorite items,etc.).From this perspective,CVR quantitatively describes the problem which is particularly concerned by advertisers,i.e.,the quality of users.Therefore,it is a critical issue to allocate ads-budget and increase advertisers' revenues.Our work can be divided into two aspects:(1)Focusing on CTR prediction,we first quantitatively justifies and proposes a method SMUP(Similarity Method base on User Pair)which incorporates the potential impact of the user similarity underlying user pair on CTR prediction so that improving the prediction accuracy of click through rate.Comprehensive experimental results on large-scale datasets have demonstrated the performance of our solution is better than those of other state-of-the-art algorithms.(2)Focusing on improving the accuracy of CVR prediction in online advertising,this paper firstly analyzes and reveals the correlation underlying creative associated with ads and CVR,which is excluded by most state-of-the-arts in this literature.Furthermore,we propose a novel LR+(Logistic Regression Plus)model to utilize the potential impacts of creatives on predicting CVR.Experimental results and analysis on two public real-world datasets(REC-TMALL dataset and Taobao Clothes Matching dataset)validate the effectiveness of the proposed LR+ and demonstrate that the proposed LR+ outperforms typical models(e.g.,LR,GBDT and linear SVR)in term of RMSE(Root Mean Square of Error).