Abstract: Targeting the Internet advertising accurately is an eye-catching problem in the field of computational advertising.As an important evaluation criteria for online advertising effect,the precision of prediction for click through rate (CTR)benefits publishers,advertisers and users.Without considering feature differentiation,mainstream approaches are extracting features and establishing click prediction model,which use a single weight to measure the effect of a feature for CTR.According to the idea divide and conquer,a hybrid model based on user similarity and feature differentiation was proposed.The model divides users into several groups depending on user similarity evaluated by mixture gaussian distribution.For each group,model was built respectively and they were combined to excavate the different effects of a feature to different groups and improve predict CTR prediction accuracy.Several experiments on advertising data sets of an Internet companies were made and the effectiveness of the approach through detailed comparative analysis was verified with the mainstream approaches.

Key words: Computational advertising,CTR prediction,User similarity,Feature differentiation,Hybrid model