Research On Click-through Rate Prediction Method Based On Cross-Domain Recommendation

In a traditional recommender system(RS),click through rate(CTR)monitoring is often limited to a single domain,such as in news,advertising,movies,music and other domains,etc.,without considering the connections between the various domains.As a result,the design of CTR estimation methods in most existing recommender systems focuses mainly on predictions within a single-domain of media,while the design of cross-domain CTR prediction has been far less developed.One of the primary challenges for recommendation systems is the ability to provide nuanced preference predictions from a sparsity of useful data.Cross-domain recommendation(CDR)provides a potential solution by mining information from a multi-domain source to bolster CTR prediction performance on a target with sparse information,an ultimate improvement over Single-Domain CTR prediction.Additionally,since advertisements are usually displayed in conjunction with everyday content,cross-domain CTR prediction methods have an additional opportunity for problem-solving.From the results of previous CDR research,frequent problems have included poor extraction performance of user interests and insufficient feature combination in cross-domain recommendation models.The objective of this thesis is to address these challenges in CDR models of CTR prediction,through use of a deep learning Attention Mechanism and optimized neural network structures to build a corresponding recommendation model.Specifically,a cross-domain CTR prediction model based on user interest attention is proposed to improve user interest extraction,and another prediction model based on high-order feature interaction is implemented to increase the efficacy of feature combination for CDR.The main contents of this thesis are as follows:1.In order to address poor extraction performance and insufficient integration based on user interests in cross-domain learning,this work proposes an attentive user interest neural model(AUIN)for cross-domain CTR prediction.An AUIN draws a correlation between the impact of active user interest on user interest attention to solve the expression of user interest weight in cross-domain recommendation.Specifically,a smoothing Attention Mechanism,based on cross-domain user interest learning,is used to extract and fuse user interest,reducing excessive punishment to active users.Furthermore,an auxiliary classifier based on a residual network is used to enhance the resolution and expression ability of the model in the prediction layer.The offline experimental results show the effectiveness of this method in cross-domain CTR learning.2.A neural attentive user interest model(NAUI)is a new cross-domain recommendation method based on high-order feature interaction for cross-domain CTR prediction,and in this work is used to improve feature combination techniques.Though the AUIN algorithm has prominent advantages in terms of user interest learning and study,there are also many difficulties with its implementation.Due to the complexity and diversity of researching in advertisement,an AUIN faces certain restrictions on the model expression.However,with high-order feature interaction a larger volume of feature information can be learned,resulting in a positive impact on the cross-domain CTR prediction model.Therefore,this work further explores a NAUI model based on high-order feature interaction for cross-domain CTR prediction on the basis of AUIN research work.NAUI not only adopts the effective method of Attention Mechanism to extract user interest,but also combines a CrossNet module with a logarithmic multi-layer perceptron(MLP)component for joint learning,which effectively captures both the implicit and explicit high-order feature interaction in cross-domain recommendation.A larger number of nonlinear features and combined feature information are learned,which greatly improves the expression ability of the model in cross-domain recommendation.Similarly,as with the AUIN,an auxiliary classifier is added to the deep neural network to improve the recommendation performance.Our experiment results show that NAUI outperforms several frequently used state-of-the-art methods of CTR prediction,verifying the feasibility of the proposed model.