Distributed Representation - based Recommender Systems in E - commerce Van - Thuy Phi †

Recommender system plays an important role in many e-commerce services, such as in Rakuten. In this paper, we focus on the item-to-item recommender and the user-to-item recommenders, which are two most widely used functions in online services for presenting relevant items given an item, or a particular user. We use a large amount of log data from one of Rakuten markets, and apply distributed representation method to that data for developing two types of recommender systems. The key idea of our approach is treating items as words, and users’ sessions as sentences, then training the Word2vec model and Doc2vec models based on those items and user’s information. Resulting item vectors from the Word2vec model can be used to calculate the cosine similarity between items, and find the similar items given an item. Similarly, Doc2vec model helps users find relevant items that might interest them using similarity between items and vectors. We also use the item vectors from both embedding models to build an additional user-to-item recommender, namely Item Vector-based system. The experiments show that our best system achieved a hit-rate of 24.17% for recommending items to users in testing data, which outperformed conventional approaches to a significant extent. Keyword Recommender System,Distributed Representation,Item Vector-based