Research And Implementation Of User Behavior Prediction Based On Multi-Model

Good recommendation system not only can greatly improve the efficiency,but also can improve the customer experience,so the research of e-commerce has a good application value and practical significance.Predict whether the customer will return to the same store to purchase,that is repeated buyer.This is one of the problems in recommended area.This problem is essentially a two category(0/1)problem,the factors that affect the classification effect depends mainly on two aspects,one is that the features are good or bad;the second is the choice and integration of the model.The original data is usually only a record of customer behavior,scattered and not targeted,how to use a variety of manners to deal with the original data,extract and predict the problem associated with the features;and how to select the effective model and through the multi-model Convergence to improve the prediction accuracy is a problem that needs to be solved.In this paper,the extraction of features,as well as the improvement and integration of the model were studied,the main work is as follows:1)In the feature engineering area:from the customer,business,customer-business three characteristics of the expansion of the perspective.Calculate the statistical characteristics of the customer's four behaviors(click/buy/love/add to the cart)on a monthly basis.And then combine the different feature sets to get a richer subset of features,which can reflect the characteristics of the data more comprehensively.2)In addition to using several classical models,such as neural network,logistic regression,stochastic forest and GBDT,and make an improvement in Deep Stack Network(DSN).Based on RBM generation characteristics,DSN model,the experimental results in the e-commerce data set show that the proposed algorithm is effective and has certain advantages in the field of electricity providers.3)In the aspect of model fusion,we study the method of determining the weight of model fusion based on AUC diversity measure,and explore the fusion method based on greedy strategy model.Compared with the classical fusion method,such as simple average method,neural network and logistic regression learning method,the experimental results show that the weights are determined by the AUC-based diversity measure method,and the effect is better than other fusion methods.