Research On Click-Through Rate Prediction Algorithms Based On Deep Learning

The Internet has profoundly changed people's living and working methods and significantly impacted many traditional industries.For example,online advertising has replaced TV,newspapers,billboards and other traditional media as the primary advertising mode.Especially,the accurately targeted advertising combined with the big data analysis has been widely deployed to various e-commerce,video and social networking sites,bringing substantial economic benefits to both platform and product sides.As the key to achieving accurate advertisement delivery,the click-through rate prediction algorithm has received unceasing attention from academia and industry.In this thesis,various existing click-through rate prediction algorithms are studied,especially the click-through rate prediction algorithm based on deep learning.The advantage of this kind of click-through rate prediction algorithm is that it can effectively capture the high-order feature information of each advertisement display opportunity,which makes the click-through rate prediction more accurate,but there are also some problems to be solved urgently: First,the depth model implicitly learns the high-order feature information in a "black box" way,which brings challenges to the interpretability of the model;Secondly,the current mainstream depth models are insufficient to mine user behavior data,and user behavior is the carrier of interest,so it is necessary to model user behavior to improve user experience and prediction accuracy.Given the above problems,the research work of this thesis is as follows:Firstly,this thesis applies attention mechanism to the click-through rate prediction task and proposes a shallow model HAN based on hierarchical attention mechanism.In the modeling stage of shallow model,two levels of attention mechanism are introduced to model finite order feature combinations explicitly.This method improves the efficiency of feature combination and improves the interpretability of the model.In this thesis,the performance of HAN is verified on three real-world data sets,and the experimental results show that HAN is superior to the existing typical prediction schemes.Secondly,based on the shallow model HAN,to fully explore the high-order nonlinear relationship between features,this thesis also introduces deep neural network and shallow structure HAN for joint learning and proposes a click-through rate prediction fusion model DHAN based on hierarchical attention mechanism.Experiments on three published data sets show that DHAN has the best prediction performance compared with the baseline model,and its AUC index is 0.037%(Criteo)and 0.052%(Avazu)higher than the latest x Deep FM.Finally,to better model the data of user behavior sequence and mine the potential interest of users,this thesis proposes a click-through rate prediction model UAIN based on user behavior sequence.This model is divided into two parts: interest extraction module based on user behavior sequence and user-candidate advertisement interaction module based on attention mechanism.Interest extraction module includes behavior refinement layer and interest development layer,which extracts the user's initial interest representation from the original user behavior and introduces the GRU unit with an attention mechanism to model the user's interest evolution process.According to the user portrait information,the user-candidate advertisement interaction module further excavates the interaction relationship between the candidate advertisement and the user portrait information,which solves the problem of less user behavior when the system is cold started.