Analysis Of Search Behavior Based On User Profiling

In recent years,the number of Internet users has grown rapidly.It brings the amount of data in various neighborhoods on the Internet.How to analysis these data has become a difficulty.The user profiling can be obtained by analyzing the user's data and building a labeled model of user to reflect the user's personal attributes and characteristics.The search engine is one of the basic application of in the Internet.Users leave a large amount of rich and time-effective data in the search engine,such as query terms and access records on web page.It provides sufficient data resources to analyze the user's hobbies and personal attribute information,and to construct the user profiling.Enterprises can use these information to portray user portraits of various groups on the Internet efficiently,which helps companies achieve precise marketing and personalized services.However,users often leave incomplete data in the internet,which brings great inconvenience to data analysis.Therefore,it is necessary to use appropriate machine learning algorithms to predict the unknown attributes of users,and to mine the users' hidden information.At present,there is still a lot of deficiencies in the commonly used algorithm models.The effect of high-dimensional sparse feature prediction still needs to be improved.Fusion algorithms can often combine the advantages of each algorithm to make up shortcomings and improve prediction capabilities to a certain extent.This article analyzed the search records of users,took the task of predicting the user's static attribute tags,analyzed and compared the models' construction of user profiling.The main research as follows:1)The incompleteness of the user's data in the search engine brings challenge to constructing the user profiling.The stacking ensemble for constructing the user profiling of search engine is proposed in this article.The research predicts three labels of the user portrait and compare with the different models.This model has two-layer's structure.In the first layer,we trained the TF-IDF features with multi-class classifiers,and the doc2 ve features with Back Propagation NeuralNetwork.In the second layer,we trained the features from first layer with the support vector machine,than we predicted the user's label.And we compared this model with the conventional model.The results of the experiment showed that the model we proposed has higher accuracy in predicting labels of users.2)The experiment obtained the prediction results through data pre-processing,feature engineering construction and feature training fusion,and analyzed the user's search behavior with charts based on the complete experimental data.Finally,a visual example of the search engine user portrait was constructed.