The Design Of High-Potential User Mining Algorithm Based On Machine Learning And System Construction

With the trend of the information age swept up,hardware devices and software technologies have also developed rapidly.A large amount of data that has been traded offline before has been able to store in online servers processes and queries efficiently.By timely and accurately analyzing a large number of user-based data and behavioral data,obtained the useful information and knowledge through mining user behavior patterns,which supporting the promotion and improvement of commercial services.User profile have been widely used in various industries and fields as an important tool for depicting target users and related user appeals and product design.This paper takes the networked car industry as the background and solves the problem of high potential users mining across business lines.For users of different lines of business,according to the whose historical behavior patterns,to judge trends of future.Using the user's basic information and behavioral information,abstracting them into features,exploring the performance of the three models of ordinary machine learning algorithms,deep learning and transfer learning,combined with the design and implementation of the crowd profile analysis system,which can let the user tags produced by the model systematically and visually displayed to internal members of the company.Experiments show that these methods can accurately determine the high-potential user groups,enabling enterprises to conduct targeted marketing activities,effectively reducing the cost of promotion,and improving the return on investment.The main work and innovations are as follows:1)In the XGBoost model,this paper proposes a feature processing method that uses woe values to encode non-numeric features and ranks,then maps sparse and indefinite long non-numeric features to a fixed number feature.Then the new map type features are generated.The experimental results show that this non-numeric feature processing method can significantly promote the effect of the model,the model has been deployed on-line and become a powerful tool to choose high-quality group.2)In the neural network experiment,this paper uses deep neural network(DNN)and deep&cross network(DCN)for experimental comparison,and constantly adjusts the network structure and selected features until optimal.3)This paper adopts feature-based transfer learning method and parameter-based transfer method,proposes transfer deep&cross network(TDCN),pretraining several sparse categorical features,updates embedded matrix parameters to optimal,and transfer the weights matrix and the embedded matrix into the new deep cross network,and the methods of frozen and fine-tuning the embedded matrix parameters are tried respectively.The experiment shows that the model used the pretrained features and fine-tune raise by 3 percent of area under roc curve(AUC)compared with the normal neural network.4)Participate in the design and implementation of the crowd profile analysis system,which support for precision marketing and data analysis,and the system has been successfully launched within the company.