User Portrait From Online Behavior Log Data

The rapid development and widespread use of Internet technology has facilitated the generation and accumulation of large-scale log data.With the rapid development of data mining technology and remarkable results,user portrait as a key technology,provides a more vivid and concrete method for mining the information hidden behind these large amounts of log data.The online behavior log data is a kind of "wisdom" data,which contains many basic user attributes and dynamic behavior attributes.Many researchers have conducted research on online behavior log data,such as user interest recommendation and precision marketing.However,these studies tend to construct a single model simply,which will bring difficulties to improvement the model's effect,due to the limitations of each single model.Therefore,aiming at this problem,this paper proposes a user portrait method based on online behavior log data,which fully absorbs the existing research results.Specifically,this paper mainly does the following work:(1)A multi-index fusion feature selection method is constructed,which fuses Pearson's correlation coefficient,ridge regression,chi-square test,random forest,stability selection based on random lasso,improves the redundancy of features more accurately;At the same time,this paper extracts binary features and related features based on single features.It solves the sparseness of features and enhances the correlation with target features.(2)Combining multiple single classifiers to build a user portrait model based on Stacking technology.First,this method implements cross-validation of feature data sets based on multiple single classifiers,outputs the prediction results and a single classifier with the best prediction effects,and then add the prediction results to the original feature data set,and perform the model training the second time to get the final prediction results based on the optimal single classifier.(3)This paper conducts user portrait based on campus network log behavior data in three dimensions of gender,grade,and age.The experiment results prove that the user portrait method improves the accuracy of gender,grade,and age prediction.