Research On User Profile Based On Deep Neural Networks

With the development of Internet technologies and smart devices,various mobile applications have emerged and penetrated into people's lives,and the information generated from these applications grows explosively,making it difficult for people to obtain the information they need efficiently.And it is also difficult for enterprises to push products or information to users accurately.The basis of the recommendation system is user profile.Constructing user profile efficiently will be beneficial for enterprises to achieve refined marketing and personalized recommendations.However,relying solely on traditional expert experience to label user profile is inefficient.Therefore,label prediction based on algorithm models has become a hot topic for user profile.However,the current mainstream shallow learning model cannot dig deeply into the complex relationship between features,especially in the face of high-dimensional sparse features,and its prediction effect still needs to be improved.The purpose of this paper is to explore user profile based on Deep Neural Networks from the perspective of feature extraction and propose a new prediction model.In this paper,through the induction and summary of the research of user profile,the definition and construction process of it are defined,and the problem existing in the shallow learning model based on high-dimensional sparse data is pointed out.Based on advantages of Deep Neural Networks' capacity to mine the complex relationships between features,we propose a fusion model for user profile prediction that Deep Neural Networks method as feature extraction,and the gradient boosting decision tree as a classifier.Meantime,based on the practical significance of ID features,this paper illustrates the necessity of ID features participating in modeling,and proposes a method to deal with ID features by discretization technique and one-hot coding.This paper predicts user conversion rate based on the data of Tencent social advertising algorithm contest.The experimental results show that compared with shallow learning,the Deep Neural Networks model has advantages in feature learning of high-dimensional sparse data,and can be used as a feature extractor to improve the performance of shallow learning model.And the combined model of Deep Back Propagation Neural Network and gradient boosting decision tree presented in this paper performs best,and the AUC value reaches 0.81,and provides a new idea for user profile prediction.In addition,the method proposed in this paper is of great significance for its participation in modeling,and it has a significant effect on the improvement of user profile prediction model.