Research And Implementation Of User Portrait System

With the birth of various artificial intelligence products,users' behavior data has changed greatly in breadth,depth and precision,and data has become an important resource.The technology of user data at the level of simple statistical query has been mature,and now it tends to use deep learning,machine learning and other artificial intelligence related technologies to deeply understand and explore the data.The user portrait is obtained by deep mining the user data and getting the user's label.Merchants can fully understand the user distribution based on the portrait,provide personalized services,enhance the user experience of the product,and the user portrait can provide factual support for some decisions.Therefore,user portraits are essential in many business intelligence systems.Compared with previous researchers and related practitioners to construct user portraits based on simple query statistics,this paper applies the big data machine learning model to complete the construction of population attribute tags according to the actual business needs of enterprises,and innovatively introduces deep learning models into users.The construction of interest tags.This paper first collects the user's APP behavior or text data,then cleans the original data from the data warehouse,extracts the training samples,and the population attribute label is selected to be 2.53 million,of which the gender label is 270,000 and the academic label is 2.26 million.,the negative sample 60931 in the interest tag,positive sample 110388.Then,in the classification prediction of the population attribute label,the APP behavior data is used,and the effects of different feature construction methods and LR and SVM models on the results are compared.In the intention identification of the interest tags,the text data is used,and the current NLP is compared.The effects of several of the most popular depth models Bert,Transformer,Bi-LSTM+Attention,and CNN.The final result shows that for the population attribute tag construction,the index value of the feature using the app name is mapped to 0 or 1,the sample sampling ratio is set to 1.0,and the model selection logistic regression + L2 is the best,and the overall accuracy of the age prediction can be achieved.85.2%,the overall accuracy of academic forecasts can reach 74.5%.For the construction of interest tags,the pre-trained Bert model has great advantages in downstream classification tasks,and can accurately determine user intent.The accuracy rate of interest tag prediction reaches 99.0%.