Research On User Profile Based On Ensemble Model

People will leave huge amounts of data and information when they use the Internet everyday.How to effectively use these fragmented data and provide users with more professional personalized services is a hot trend of current research.User profile technology refers to using the digital information left by users on the network for modeling and analysis,deeply mining the hidden valuable information,and realizing the extraction of user features and the prediction of labels,so as to establish a comprehensive and three-dimensional user image.User profile technology is now applied to a variety of scenes,such as business circle analysis,intelligent recommendation,precision marketing.Traditional user profile construction methods have many defects,such as poor feature quality,low model accuracy,poor generalization and so on.To solve the above problems,this paper proposes a user profile construction method that based on Stacking double-layer model and pre training model ZEN,which uses model ensemble technology to predict user multi-dimensional attribute tags.The main contents of this paper are as follows:At the stage of feature selection in user profile,aiming at the short text data left by users when using search engine,this paper proposes to extract feature information from different angles.Firstly,the traditional feature extraction algorithm TFIDF is selected.It only relies on word frequency statistics,ignores the distribution differences of feature terms among different categories,and has the problem of inaccurate feature extraction.Based on its inverse feature frequency and category frequency,this paper proposes TWCF algorithm to extract users' common word features.Secondly,combining the feature representation of PV-DBOW model training with the feature representation of Denoising Auto Encoder training,and a DBOW-DAE feature model is proposed,which can effectively retain the semantic information in the user's query words while reducing the text noise,and improve the quality of the feature representation.The experimental results show that the improved feature model is more accurate in predicting user attribute labeling task.At the construction stage of user profile model,an ensemble model S-ZEN is proposed by using model combination strategy,stacking learning method and pre-training model ZEN,which can combine the advantages of each learner and comprehensively learn the correlation information between user attributes.The experimental results show that the ensemble model has better accuracy and generalization for the task of predicting user multidimensional attribute labels than various single models.