TY - GEN
T1 - Social Recommendation Terms
T2 - 19th Monterey Workshop on Challenges and Opportunity with Big Data, 2016
AU - Liu, Jie
AU - Zhang, Lin
AU - Sheng, Victor S.
AU - Laili, Yuanjun
N1 - Funding Information:
Acknowledgments. This work is partially supported by National Nature Science Foundation of China (No. 61374199, National High-tech R&D Program (No. 2015AA042101),) and Beijing Natural Science Foundation (No. 4142031).
Publisher Copyright:
© 2017, Springer International Publishing AG.
PY - 2017
Y1 - 2017
N2 - The Probabilistic Matrix Factorization (PMF) model has been widely studied for recommender systems, which outperform previous models with a solid probabilistic explanation. To further improve its accuracy by using social information, researchers attempt to combine the PMF model with social network graphs by adding social terms. However, existing works on social terms do not provide theoretical explanations to make the models well understood. The lack of explanations limits further improvement of prediction accuracy. Hence, in this paper we provide our explanation and propose a unified covariance framework to solve this problem. Our explanation, including regularization terms, factorization terms and an ensemble of them, reveals how most social terms work from a probabilistic view. Our framework shows that those terms could be optimized in a direct way compatible to PMF. We find out that accuracy improvements for existing works on regularization terms rely more on personalized properties, and that social information for factorization terms is helpful but not always necessary.
AB - The Probabilistic Matrix Factorization (PMF) model has been widely studied for recommender systems, which outperform previous models with a solid probabilistic explanation. To further improve its accuracy by using social information, researchers attempt to combine the PMF model with social network graphs by adding social terms. However, existing works on social terms do not provide theoretical explanations to make the models well understood. The lack of explanations limits further improvement of prediction accuracy. Hence, in this paper we provide our explanation and propose a unified covariance framework to solve this problem. Our explanation, including regularization terms, factorization terms and an ensemble of them, reveals how most social terms work from a probabilistic view. Our framework shows that those terms could be optimized in a direct way compatible to PMF. We find out that accuracy improvements for existing works on regularization terms rely more on personalized properties, and that social information for factorization terms is helpful but not always necessary.
KW - Factorization terms
KW - Probabilistic matrix factorization
KW - Regularization terms
KW - Social networks
UR - http://www.scopus.com/inward/record.url?scp=85028525353&partnerID=8YFLogxK
U2 - 10.1007/978-3-319-61994-1_15
DO - 10.1007/978-3-319-61994-1_15
M3 - Conference contribution
AN - SCOPUS:85028525353
SN - 9783319619934
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 155
EP - 167
BT - Challenges and Opportunity with Big Data - 19th Monterey Workshop 2016, Revised Selected Papers
A2 - Zhang, Lin
A2 - Ren, Lei
A2 - Kordon, Fabrice
PB - Springer-Verlag
Y2 - 8 October 2016 through 11 October 2016
ER -