http://www.paper.edu.cn
- 1 -
Cross-domain Recommendation Based on LS-SVR with
Global Constraints
#
Jie Zhang, Xin Xin
**
5
(School of Comp. Sci., Beijing Institute of Technology, Beijing,100081)
Foundations: Ph.D. Programs Foundation of Ministry of Education of China (No. 20131101120035).
Brief author introduction:Jie Zhang(1992-),Male,M.A.,main research recommender system,urban computing
Correspondance author: Xin Xin(1984-),Male,Ph.D.,main research information retrieval and mining,artificial
neural network and deep learning,natural language processing,urban computing. E-mail: xxin@bit.edu.cn
Abstract: The existing cross domain recommendation algorithm can solve the problem of sparse data
in the target domain by transferring the knowledge from the auxiliary domain. The key problem is how
to map the user feature in different domains. In traditional transfer learning algorithms, user's feature
vector is mapped to the target domain in a linear way, but the limitation of this method is that the real 10
data does not always follow the linear mapping relations. In our previous work, we utilize support
vector regression as the nonlinear function in mapping user feature across different domains, and
demonstrate its effectiveness in improving the recommendation performance for new users. However,
in the previous proposed framework, the mapping functions for different dimentions in a user feature
vector are learned independently. Consequently, the optimization objective is an indirect one, and 15
cannot reflect the error of rating prediction. In this paper, we propose a novel model that extends the
previous model from independently modeling the mapping functions to jointly modeling the mapping
functions. We utilize the rating prediction error as a bridge for learning different mapping functions.
Through the experimental analysis, it is proved that the proposed method consistantly outperforms
previous ones.20
Key words: Recommended System;Transfer Learning;Collaborative Filtering
0 Introduction
Most of the single domain collaborative filtering(CF)
[1]
algorithms are based on user history
records for users to recommend items purchase, if two users have similar items purchase 25
records ,their recommended goods should also be similar. Neighbor-based
[2]
algorithms follow
this idea, to identify similar users and recommend relevant items. However, these methods often
face the problem of data sparsity and cold start. If a user has very few ratings in the target domain,
it is difficult to predict his/her ratings for other items. For example, users may buy few movies, but
they may often buy books. Therefore, it is the key issue of this paper that how to use the ratings of 30
the book domain to predict the ratings of the movie domain. The task belongs to the transfer
learning research field in collaborative filtering (CF)
[3]
.
Typical transfer learning methods include CMF(Singh and Gordon, 2008)
[4]
, tensor model(Hu
et al., 2013)
[5]
, CSVD(Pan and Yang, 2013)
[6]
, the user's feature vector is linearly mapped across
multiple domains. In our previous work, we proposed an nonlinear cross domain collaborative 35
filtering algorithm based on LS-SVR
[7]
(least square support vector regression). But the mapping
functions of each feature dimention are learned independently. In this paper, we extend previous
model by globally traning the mapping functions. The experimental results demonstrate that the
prediction accuracy is higher than the existing linear and nonlinear methods.
1 Problem Definition 40
As it is defined in previous work
[7]
,users’ rating in two domains is shown in the Table.1,
denotes the rating information in the auxiliary domain, and denotes rating information in the