Recommendation Model Research Based On User Behavior And Comment Data

As online shopping has now become one of the most common and convenient ways of shopping. The growth rate of online goods has become more quickly and the number of goods in online sales has reached more than one billion. An accurate and efficient personalized recommendation system for either the user to find and meet their own needs or the e-commerce websites has become increasingly important. The main recommendation of the system currently used in most e-commerce websites, essentially based on collaborative filtering(item CF or user CF), or based on the Content of the auction description. Both have advantages and disadvantages, the accuracy of collaborative filtering based recommendation system has been proved the accuracy in both research and industry area, however it is difficult to explain why a specific product recommendation is made to a specific user; conversely content-based recommendation system in terms of good interpretability but accuracy is poor. It is a big difficult problem to make the recommendation system more accuracy and more interpretability.Most recommendation systems only consider the various user behavior data, while some data(such as the review) after the transaction are rarely considered. With the formation of the good habits, many users will write a review after purchase, which contain some descriptive information with personal feelings and characteristics of the goods. These comments are helpful for us to explain why people buy these goods. In this paper, we propose the Collaborative Aspect and Sentiment Model to model both behavioral data and reviews, which provide more accurate and interpretable recommendations. The CASM model uses a review specific topic model models the text data, and uses a similar probabilistic matrix factorization model to model user behavior data. The item latent variable was generated from the item reviews' topic distribution and sentiment distribution, which makes the topic learned from the topic model can explain each dimension of the item latent variable. Finally we can predict the rating by computing the inner product of user latent vector and item latent vector. Extensive experimentation based on a large set of transaction logs from one of worldwide largest online shopping website confirms the effectiveness and interpretability of the proposed method compared to several state-of-the-art recommendation models.