Research On Click-through Rate Prediction Method Based On Transfer Learning

Advertisement click-through rate prediction is a key task in online advertising,which helps to understand the characteristics of system represented by the advertising platform.Online advertising operation is the process of calculating advertisements matching the search content through background and delivering them accurately.The advertisement delivery mechanism can quickly increase the click-through rate(CTR)of advertisements delivered by advertisers,helping users obtain high-quality resource information.With the rapid development of technology,in the face of huge data volume,high-dimensional sparse features and redundant data constitute more complex advertising data.The traditional advertising click rate prediction method has been unable to meet the increasingly complex advertising data.How to quickly and accurately construct an advertisement click-through rate prediction model is an urgent research topic.Advertising click-through rate prediction is devoted to the detection of CTR after online advertising are delivered.When the classic logistic regression algorithm is used to deal with imbalanced high-dimensional sparse data of advertisements,the method of advertisement click rate prediction fails to consider the relationship between features,which makes the prediction accuracy not ideal.In order to solve this problem,this thesis extracts the potential relationship between imbalanced data samples through the transfer learning method,which greatly improves the prediction results of advertising data.This thesis proposes two methods of CTR: Click-Through Rate Prediction Method based on Robust Integrated Locally Kernel Embedding under the Influence of Transfer Learning(RTILKE)and Click-Through Rate Prediction Method based on Multi-view Feature Transfer(MFT).The evaluation indicators of RTILKE and MFT in different scale data sets show that the model effect is compared with multiple classic methods,thereby further verifying the effectiveness and feasibility of the two models.The main work of this article is as follows:(1)Click-Through Rate Prediction Method based on Robust Integrated Locally Kernel Embedding under the Influence of Transfer Learning(RTILKE):Firstly,the method divides the data into source and target domains.The source domain isdivided into three groups: positive samples(completely clicked),negative samples(completely clicked),mixed positive and negative samples(MIX).This operation is to solve the problem of imbalance between sample data.It is proposed to use transfer learning to improve the unstable prediction caused by imbalanced data.Then the preprocessed data is expanded into a kernel function,on this basis,the expansion coefficient matrix and the label symmetric kernel matrix form a Robust Integrated Locally Kernel Embedding(RILKE)model,thereby making the algorithm more stable.Finally,the construction of the similarity function and non-negative embedding matrix adopts iterative more optimization operation that alternately fixes the similarity function and the non-negative embedding matrix,so as to obtain the locally optimized prediction value.A large number of experimental results show that RTILKE is effective in the prediction of advertising click-through rate,and has obvious advantages in the prediction of imbalanced advertising data sets.(2)Click-Through Rate Prediction Method based on Multi-view Feature Transfer(MFT):The RTILKE algorithm has obvious advantages in the prediction of imbalanced advertising data sets,but considering the complexity of advertising data and diversity of features,the transfer between data cannot fully discover the characteristic relationships within the advertising data.Therefore,we further propose an advertising CTR method based on multi-view feature transfer.Firstly,MFT divides the data into ordinary features and selected features during data preprocessing.The selected features are combined into a feature matrix.The adjacency matrix is constructed by the method of K-nearest neighbors and the laplacian matrix is constructed by constructing a graph to obtain the first k feature vectors.In order to extract the connection between features,the method of feature transfer is adopted,and the obtained first k important feature vector matrix is combined with the common feature matrix into each view,so that MFT makes full use of different attributes of the data.Experiments on different scale data sets show that MFT has achieved good prediction results on advertising data sets,and its performance has been better than many advertising click-through rate prediction methods.