| With the development of mobile Internet,e-commerce platforms have entered a period of rapid development.So far,each e-commerce platform has accumulated a huge amount of user behavior data.From the dimension of whether it can directly reflect users' preferences,the data of user behavior can be divided into two categories: one is explicit feedback data,the other is implicit feedback data.The study of implicit feedback is less,but the implicit feedback data of the electric business platform accounted for the largest,based on the implicit feedback of user behavior data mining can provide electricity to achieve further precise marketing may,at the same time analysis of implicit feedback data mining are also a good supplement to explicit feedback data,so for electric business platform for implicit feedback is very important to the analysis of user behavior mining.Therefore,this paper proposes a series of methods based on data visualization,feature engineering and machine learning model to process the implicit feedback data of ecommerce,so as to realize the prediction of users' purchasing behavior and better push products for users.Firstly,this paper reviews the literature on implicit feedback,user behavior,feature engineering and purchase prediction.This paper summarizes the research contents of many literatures and analyzes the shortcomings of these literatures.Then the research framework and method are put forward.Based on the large-scale implicit feedback data of e-commerce,this paper transforms the purchase prediction problem into a machine learning dichotomy problem.First of all,the original data was cleaned and sorted to remove brush users and crawler users,and the distribution of purchase conversion rate in time was observed through data visualization.Then based on the observed pattern to build four characteristic index,which in view of the implicit feedback data can't clear said user preferences in this design the user average weighted selection tendency characteristics to solve this problem,and achieved good results,all the features according to the characteristics of the group is divided into six groups,a total of 988 features.Then the random forest algorithm is used to demonstrate the feature's importance,then the feature is screened and the 784 dimensional feature is left.Finally,Logit regression,CNN convolutional neural network and Inception network are used to conduct purchase prediction on the processed data,in which the training of CNN and Inception is to transform the 784 dimensional features into a 28*28 grayscale image for training.Because the sample data is unbalanced,F1 index is used to evaluate the prediction effect in the experiment.The final Logit regression F1 value is 11.25,CNN F1 value is 11.41,Inception F1 value is 12.42.Experiments show that the two main design ideas based on Inception network: multi-scale convolution extraction of multidimensional features and hidden layer can also output results,the Inception network designed in this paper has achieved good prediction effect on the data set of tianchi user behavior.This paper takes the implicit feedback data of e-commerce users' behavior as the center,completes the prediction of users' purchase behavior by observing user behavior,conducting feature engineering and designing model.The user average weighted propensity feature and Inception model constructed in this paper can improve the effect of the final prediction.It provides certain reference value for e-commerce platform to better push commodities to users. |
