Personalized Recommendation For E-Commerce Users Based On Ensemble Learning

Thanks to the continuous development of information technology and the market advantage brought by Chinese population base,Chinese Internet economy has been rapidly developed.With domestic efficient logistics distribution system,the market share of online shopping to retail consumption goods is increasing.Among these segment fields,mobile E-commerce grows fastest and becomes the most promising,mainly due to the lower price and convenient operation of mobile devices.However,the rapid development of information technology has also brought negative impact,like information overload,which make it difficult for seller to get to precise marketing to target users in the online shopping while users become more and more disgusted with all kinds of push information.Personalized recommendation is regarded as an effective way to solve the above problems,Alibaba believes that user's behavior data can be used to build a recommendation model of the users to commodity subset,which is used as a research topic.They shared user behavior data from December 18,2014 to November 18,2014 with scientific research workers.Based on the subject and data set analysis,the article will transform the problem into a bi-level classification problem,which predicts whether the user will buy a commodity or not,the results can help find out which products will attract users' interest and become the buying recommendation to mobile E-commerce users.Firstly,the paper design features from the original data,separately extracted from the user,product and category,and considering the alternating features between user and product and category as well as user and category.The behavioral features include direct statistical ones and indirect ratio features,which show as clicking,collecting,adding to cart and purchasing behavior's total number and the corresponding purchase conversion rate,and it divides into 3 dimensions as the past 1 day,3 days and 7 days.In accordance with the above three principles,the final amount of features is more than 100.After determining the feature,processing data by two mature ensemble learning algorithms as random forest and gradient boosting decision tree,F-value and ROC curve are used as evaluation indexes to realize algorithm parameter tuning and to determine the optimal purchase prediction model.Personalized recommendation are made base on the results.