Zhao et al. introduce a product recommendation model for social media users who have never shopped online before. The most important premise for a successful product recommendation is that the user has bought something before and has shown interest in some products while browsing. But how can an e-commerce website recommend a product to a user who has never shopped online? That is, how can it recommend a product to a user who has never left any trace on the shopping website?
The authors’ research attempts to fill this gap. The main idea is to learn product information embedded in the users’ social media communications such as microblogging, and then use this information to make online product recommendations. The system first collects features in microblogging to represent the users. These features include demographic information, blog text, network attributes showing preferences among group members, and temporal attributes indicating the user’s daily routines. The next step is to learn product information embedding and user information embedding from the features collected. The neural network word2vec method is used to learn product embeddings. The user embeddings are learned through the combination of para2vec and word2vec methods. The system then generates mappings between user interests shown in their microblogs and the product information to make a shopping recommendation. “The key idea is to use a small number of linked users across [social media sites and e-commerce] sites.”
Real data is used for testing the system. The microblogging data from Sina WeiBo contains 1.7 billion tweets from 5 million active users during the first half of 2013. The JingDong e-commerce data contains about 139 million transactions from 12 million users on 200,000 products. The system identified 23,917 links from the 5 million Sina WeiBo users to the JingDong site. Eight different methods of product recommendation are compared, including various configurations of cold-start. The results show that the proposed model significantly improves the recommendations measured by common metrics of Precision@k, Recall@k, the mean average precision (MAP), the mean reciprocal rank (MRR), and the area under the receiver operating characteristic (ROC) curve (AUC).
The authors’ approach of recommending e-commerce products to users who have never shopped before through social media data is very interesting. The paper is well organized and well written. The presentation is thorough and accurate. The authors clearly articulate their arguments. The results are convincing.