Utilizing Matrix Factorization Methods To Model Users’ Sparse Collaborative Behaviors

The recent unprecedented proliferation of social media makes the information extremely easy to be produced, transformed and shared. People all around the world are connected closely because of the daily basis millions of micro-blog posts, tweets and status update of the social network. With the popularity of E-commerce, the online consumption is becoming an essential part of people’s daily life, and is also transforming the structure of traditional industries. However, people are suffering from a serious challenge from the explosive growth of information: how to acquire quality resources from the numerous information pool? The emergence of recommender systems is exactly to deal with such issue, and progressively becomes a fundamental function for a variety of applications via automatically delivering personalized information to fit users after analyzing users’ personal preferences.This work mainly explores the challenging sparsity problem along with the design of effective recommender systems, and discusses how to design corresponding algorithms to alleviate such problem for rating prediction and item ranking perspectives.The main contributions of this work includes:1. Exploring the possibility of modeling the interactions between users and each event associated with their final decisions, incorporating those interactions into matrix factorization models to alleviate the sparsity problem caused by the lack of rating information. Experiments are conducted to demonstrate efficiency of our proposed method.2. Studying how to utilize learning to rank to deal with a seriously sparse scenario where ternary relationship is considered for collaborative retrieval task, in contrast with traditional collaborative filtering. Experimental results in two real-world datasets show that our proposed approach could effectively improve the prediction precision, especially for those items with sparse information.3. Designing an item-based pairwise sampling construction for the application of Pairwise learning to rank in recommender systems, in addition, an adaptive samplingstrategy is defined to improve the sampling performance, meanwhile enhance the prediction accuracy in sparse datasets.