The Application Of Transfer Learning On Vertical E-commerce Recommender Systems

Nowadays recommender systems occupy an increasingly important position in people’s lives.Recommender systems are widely applied in many areas, they discover users’potential consuming habits by analyzing their behaviors, and then recommend users with what they may purchase. However, recommender systems on vertical e-commerce sites facing the problem of data sparsity. Data sparsity may cause inaccurate recommendations, thereby reducing the user’s shopping satisfaction.User-based collaborative filtering recommendation recommends users products according to their similar users’opinion. In this case, if we can get a more accurate similarity between users, the accuracy of recommender systems will be enhanced. However, the data on vertical e-commerce websites is very sparse.And the data sparsity would cause the inaccurate similarity between users.In this paper, we firstly propose a new method with joint purchase data and click data from e-commerce sites as input data for alleviating the data sparsity problem.When calculating the similarity between users with considering the user’s buying behavior and click behavior at the same time, this method can measure the similarity between users more accuratly. In order to alleviate the data sparsity problem in the vertical e-commerce site, we then propose a new approach based on the idea that combining user-based collaborative filtering techniques with transfer learning.The method alleviate the data sparsity problem by tranfer the knowledge learned from dense data sets to sparse ones.At last we conduct our experiments in data sets collected from vertical e-commerce websites in practical application.We use the order data and the click data collected from underware site to evaluate the method of the calculation of the similarity.We also use the data from a glasses site as the dense data set and the data from underware site as the sparse one, then experiments are conducted for evaluating proposed methods in this paper.Results show that our methods can alleviate the data sparsity problem and improve the effect of user-based collaborative filtering method.