New Factorization and Multi-Label Based Matrix Completion Methods for Heterogeneous Data and Emojis Recommender System

Project: Research

View graph of relations

Description

In the era of big data and sharing, customer buying behavior and retail business model have been undergoing major changes. Online recommender systems have become a highly popular technology that has been reshaping the landscape of many businesses especially e-commerce based market. Recommender system recommends products and services to individual customers according to their historical purchase/browse record, and personal information. It has become an integral technology for e-commerce business such as Amazon, and Netflix. For example, it was reported that in 2017 Amazon generated 35 percent of the revenue by the Amazon recommender engine.First, a recommender system forms a huge matrix with one dimension or rows describing users and another dimension or columns describing products of interest. For example, suppose there are 1 million products and 1 million customers, the size of the product-customer matrix is 10⁶x10⁶ but the matrix is highly incomplete because each customer would only have searched, or browsed a very few number of products. Matrix completion is the key computational technique for recovering missing entries of incomplete matrices. Each entry of a matrix indicates a relationship or link between a product and a customer. Such an entry can reveal whether the customer likes the product or not, or whether the product is suitable to the customer or not. Through finding the missing entries, recommendations can be made. Certain products may have similar features and certain customers may exhibit similar preferences. Thus, the customers and product matrices can often be regarded low-rank.Recently, emojis have been widely used in social media as a new manner to convey opinions. Apart from using traditional ways to comment on the products, such as ratings, more people have become preferring to use emojis to express their emotional feelings on products. However, the traditional matrix completion methods are not capable of dealing with emoji included matrices.In this proposal, the PI will develop new factorization matrix completion methods aiming to handle emojis and heterogeneous data, because there not only exist numerical data, but also non-numerical data (e.g., nominal and ordinal data) in real applications. We will extend the methods to online applications and out-of-sample extrapolations. It means we can recover the missing scores and/or emojis of each user sequentially and new users without updating the model. The project will also research on the relationship between the emotional emojis and the unemotional scores. 

Detail(s)

Project number9042996
Grant typeGRF
StatusActive
Effective start/end date1/01/21 → …