Research And Implementation Of Learning-based User Abnormal Behavior Analysis Method

In recent years,the application of the Internet and big data has continued to blossom in all walks of life.People's behaviors have been widely recorded and a large amount of data containing user behavior information has been generated.These data have great management significance and business value for enterprises and industries.At the same time,the current user behaviors in various fields include not only the normal behaviors of users,but also the abnormal behaviors of users.The abnormal behaviors of users often cause problems to varying degrees in production and life.How to analyze and find the abnormal behavior of users and take corresponding measures are the problems that need to be solved in many fields at present.The current hot machine learning technology and massive user behavior data provide a direction for solving the problem of user abnormal behavior analysis.The user abnormal behavior analysis method based on machine learning can make fuller use of user behavior Data to accurately identify the user's abnormal behavior and take timely measures.Through the research results of machine learning in user abnormal behavior analysis scenarios,it is found that there are currently two problems.One is that traditional machine learning solutions based on a single behavior model lack effectiveness and universality,and the second is that a large number of deep learning solutions for abnormal behavior analysis fail to fully consider the sequence characteristics of user behavior and lack the ability to treat sequence information of user behavior differently.In response to these problems,this paper proposes a learning-based user abnormal behavior analysis method.Specifically,the method in this paper includes an integrated learning module and a deep learning module,and a weighted calculation module that combines the advantages of both.The integrated learning module innovatively designed an integrated learning method based on the Stacking method.By stacking a training meta-classifier on top of the basic classifier,a variety of user behavior patterns were fused to overcome the problems in a single behavior pattern;the deep learning module designed A bidirectional gated recurrent unit network(BiGRU)with attention mechanism is added to model user behavior sequences,taking full account of long-term dependency problems and the status of the sequence before and after the sequence,and filtering the high-value information for classification in the sequence;weighting The calculation module performs weighted calculation on the classification probability results of the two learning modules,which fully combines the advantages of manual feature extraction and automatic feature extraction.In this paper,four data samples were generated on the original BPI Challenge 2011 data to conduct multiple sets of comparative experiments.The model evaluation indicators not only selected F1-score and AUC and other comprehensive indicators of classification models,but also underreported and misunderstood the analysis of user abnormal behavior.The cost situation reported defines the Cost index.By comparing the experimental results,the rationality and superiority of the integrated learning module,deep learning module and the overall modeling solution are fully verified.