Research On Prediction Model For Customers' Repeated Purchase

With the rapid development and popularization of the internet,the e-commerce achieves a greatly evolution,and the competition among these major e-commerce platforms is becoming increasingly fierce.In order to raise the market shares,many merchants and platforms conduct mass promotions on festival to attract new buyers.For higher investment conversion rate,these promotions should be conducted only for those who may become repeated buyers(potential loyalty customers).Therefore,it is of great practical significance to study how to use the behavioral logs to predict repeated purchases.This paper researches the repeated purchase prediction based on the behavioral data generated by new customers of many Tmall merchants on the "double 11" activity day and the previous six months.Firstly,the feature engineering is designed through analysing the affect factors of the repeated purchase.And then,a combined prediction method based on support vector machine(SVM)and random forest algorithm is proposed,which achieves a high performance in the experiment.In addition,a repeated purchase prediction algorithm based on deep learning is proposed in this paper,which uses deep neural network(DNN)to automatically learn effective features from the distribution of a large number of training samples.In this algorithm,the word embedding layer is firstly used to encode the original data,and then the encoded word vector is inputted into the DNN model to predict the probability of repeated purchase.The mainly work in this paper includes the following aspects:(1)Designing feature engineering.By analyzing the influence factors of repeated purchase,we extract many dimensions features,which include customer features,merchant features,interaction features between customers and specific merchants,age and gender,etc.Meanwhile,for the DNN method,the feature vectors are generated by the feature submodule models and the word embedding layer is utilized to encode the original data.(2)Researching the application of single model and combined model in feature engineering.Firstly,the SVM models and random forest models are respectively trained,and then a combined model are achieved by combining these models.Finally,the experiments show that the combined model can take advantage of all models to achieve better performance.(3)Proposing a repeated purchase prediction algorithm based on deep learning.Firstly,a multilayer neural network is elaborately constructed.At the bottom,the word vector is encoded through the word embedding layer.And then these feature sub-module are utilized to generate the features,Finally the prediction is made at the top of the model.