Research On Membership Payment Prediction Method Of Video User Based On Transfer Learning

With the rapid development of Internet,users' demand for entertainment is becoming stronger and stronger.More and more online video websites are emerging,such as iQIYI,Tencent Video,etc.,whose user scale and video resources are gradually expanding.Compared with traditional TV,online video websites give users more choices,so they can watch video according to their interests and preferences.However,not all video resources are free,a considerable number of TVs or movies require users to pay for membership before they have the right to watch.Users buy membership service in order to enjoy better video resources and better viewing experience.However,whether users pay for membership is often based on certain background conditions.When the videos provided by the membership service are highly attractive to users,they will have a strong willingness to pay.Users with high willingness to pay are potential membership users,and the website can promote the payment of them in time.The existence of this membership mode not only improves the income of video website,but also enhances the viscosity of users.Therefore,for video websites,it is very important to promote the conversion of users from non-membership to membership,and the most important key is to identify potential members with high payment probability.This paper focuses on the problem of video user membership payment prediction,mines user interest preferences by analyzing the behavior data of users,and predicts the probability of users paying for video membership service.At present,there are relatively few researches on the prediction of video users' payment for membership,and the prediction methods related to this problem include advertising click through rate prediction algorithms and prediction algorithms in recommendation system.The difference is that the prediction problem of video users' membership payment needs to predict the probability of users' payment for membership service according to their interest in paid videos,not the probability of being interested in a particular video.Traditional prediction methods based on machine learning or deep learning are according to user portrait features or user behavior sequence,which will face the problem of data sparsity when applied to our problem.For non-memberships,to predict their payment probability needs to predict whether they are interested in the paid videos provided by membership service.However,non-memberships do not have the right to watch paid videos(only partial trial permission),which leads to the scarcity of paid videos data in their historical behavior,so the prediction effect is affected.However,such users usually have a large number of free video viewing records,and the same user's preference for free video and paid video is related.For example,the user who likes free comedy movies will naturally be interested in paid comedy movies.Therefore,we can learn the user's paid video preferences according to their rich free video preferences,effectively transfer knowledge between the two preferences,make up for the lack of paid video preferences,and then improve the prediction effect.Transfer learning is one of the common methods to solve the problem of data sparsity.It applys the knowledge learned from a certain field to the target field to improve the learning effect.Most of the existing methods rarely carry out personalized knowledge transfer according to different user preferences.To alleviate the impact of data sparsity,we face the following two challenges:first,how to express users' video preferences?Second,how to share and transfer knowledge among video preferences?In view of the above challenges,this paper proposes a prediction model based on transfer learning named Deep Cross-Attention Network(DCAN),which uses preference transfer learning to enhance the user preference representation and improve the prediction effect.The main works of this paper are summarized as follows:1.In order to alleviate the impact of data sparsity and enhance the feature representation,this paper proposes a video vector representation method(Video2Vec).In addition to the conventional category features and numerical features of video,Video2Vec also considers the co-occurrence feature and text feature of video.We utilize Word2Vec to process user video sequences,learn the relationship between videos,and generate co-occurrence feature vectors.For the text content of video,this paper uses BERT to generate text vector.Video2Vec combines multiple video features representation to get video representation vector,which enhances its feature representation ability.2.In order to realize the transfer learning between user preferences,this paper proposes a deep cross-attention autoencoder(DCAA).In DCAA,the cross-attention unit connects each layer of the encoder of the two preference domains to realize information sharing and knowledge transfer.In the process of extracting user preferences,transfer learning between the two preferences is realized.Attention mechanism makes the content transferred more personalized,and gets the most concerned content in different users' preferences,thus enhancing the sparse preference representation of paid video.Based on the enhanced preference representation,this paper uses Multi-Layer Perceptron(MLP)to learn the prediction function,which makes the prediction results more accurate.3.In order to verify the effectiveness of DC AN model,a large number of comparative experiments are carried out with real datasets.In this paper,we select real Hisense play dataset and MovieLens dataset.The experimental results show that DC AN model can effectively achieve preference extraction and preference transfer learning,so as to alleviate the problem of data sparsity and improve the effect of video user membership payment prediction.It also proves that our model can be extended to other cross domain prediction problems.In addition,this paper also makes a visual analysis to show the effectiveness of each component in the model.