Start Date

11-12-2016 12:00 AM

Description

Recommender systems have been widely used to provide personal and convenient services for users. As one of successful recommendation methods, collaborative filtering explores users’ interests from item consumptions. However, it suffers from the data sparsity problem that most users have interacted with a small number of items. Particularly, data sparsity causes the discontinuous user activities over time, which limits the time-dependent recommendation methods for tracking users’ changing interests. In this paper, we extend existing methods and propose an inhibited hidden Markov model to alleviate the sparsity problem. The model considers the statuses of users’ interests at each time unit and allows for capturing users’ dynamic interests under idle status. We derive an EM algorithm to estimate the model parameters and predict users’ actions. We perform a comprehensive experiment on the datasets of various sparsity levels. The results show our model has been consistently and significantly better than the state-of-the-art algorithms.

Share

COinS
 
Dec 11th, 12:00 AM

Modeling Idle Customers to Tackle the Sparsity Problem in Time-dependent Recommendation

Recommender systems have been widely used to provide personal and convenient services for users. As one of successful recommendation methods, collaborative filtering explores users’ interests from item consumptions. However, it suffers from the data sparsity problem that most users have interacted with a small number of items. Particularly, data sparsity causes the discontinuous user activities over time, which limits the time-dependent recommendation methods for tracking users’ changing interests. In this paper, we extend existing methods and propose an inhibited hidden Markov model to alleviate the sparsity problem. The model considers the statuses of users’ interests at each time unit and allows for capturing users’ dynamic interests under idle status. We derive an EM algorithm to estimate the model parameters and predict users’ actions. We perform a comprehensive experiment on the datasets of various sparsity levels. The results show our model has been consistently and significantly better than the state-of-the-art algorithms.