Prediction Of E-commerce Online Consumer Purchase Behavior Based On Machine Learning

In recent years,people's daily consumption patterns have undergone earth-shaking changes.Due to the popularity of the Internet,people began to use PCs and mobile devices for online shopping,which broke the time and geographical restrictions.The variety of online products is complete and the price is lower,which can better meet the diverse needs of consumers.However,while a large number of products are presented to consumers,it also requires consumers to spend a lot of energy to select products.More and more merchants have refined their consumer demand in order to better meet the consumer's consumption needs,and developed online shopping platforms that meet the needs of consumers to refine and bring them to the market,making the online retail market more competitive.How to accurately understand the consumer's consumer demand and provide more targeted vertical services is an important part of the ecommerce's subsequent development process.With the continuous advancement of data technology,big data has become an emerging topic in recent years,and there has been a big improvement in the level of big data storage computing,which has led to the development of blockchain technology.Every transaction data of the consumer is recorded in the server,and then the consumer's online behavior and transaction data can be analyzed through machine learning and various intelligent calculation methods to predict the consumer's future consumption behavior.This paper uses the real data of the desalination process from Taobao shopping platform provided by Alibaba Tianchi Big Data Platform to statistically mine the behavior of consumers and predict which products consumers will buy.The proposed consumer purchasing behavior prediction model is divided into four steps: The first step is the processing of data outliers.The raw data is removed from noise,the default values are removed,and the consumer behavior is preliminarily statistically derived to obtain a basic distribution,which is prepared for further feature selection and extraction and selection of machine learning methods.The second step is feature selection.From the dimension of the commodity,three characteristics groups of consumer characteristics,commodity characteristics and behavioral characteristics of consumer-commodity interaction are constructed.The consumer behavior is connected in the chronological order of occurrence as a sequence of interaction behaviors,and various transformations are used to find other different combinations of features that are more in line with the characteristics of the data,and are added to the feature set.Then,with the positive sample set size as a reference,the negative samples are randomly sampled without returning;since the positive samples are too low in the overall data set,the positive samples are all sampled.The third step is to filter the statistical behavior data.In the original data,there are most records with too few operational behaviors,which will affect the accuracy of the model during training.The problem is to filter the data by qualitative analysis of consumer behavior,delete the records with suspected impulsive consumption and too few behaviors,and layer the data according to the length of different behavior sequences.The fourth step is model training and prediction.This paper attempts to apply the recurrent neural network algorithm(RNN)to study the consumer behavior sequence,and uses the N vs 1 structure RNN to classify the behavioral sequence behavior tendency to obtain the consumer behavior tendency score.The score is then taken as a new feature,and the new data set is further predicted using the naive bayesian algorithm.The results were compared with those obtained by modeling with a single naive Bayesian algorithm.The experimental results of the model using the training set show that the model using RNN and Bayesian fusion is more stable,which can reduce the influence of time series length on prediction accuracy.The prediction accuracy is relative to a single naive bayesian model.There is also a certain improvement,and the model results have an AUC value of 0.92.Finally,this paper puts forward the application direction and ideas of the model in the actual trading scene of e-commerce,and analyzes the shortcomings of the model itself,and discusses the further research direction of this topic in more detail.