TY - GEN
T1 - Weighted item ranking for pairwise matrix factorization
AU - Zhang, Haiyang
AU - Ganchev, Ivan
AU - Nikolov, Nikola S.
AU - O'Droma, Mairtin
N1 - Publisher Copyright:
© 2017 TEI OF WESTERN MacEdonia.
PY - 2017/10/26
Y1 - 2017/10/26
N2 - Recommendation systems employed on the Internet aim to serve users by recommending items which will likely be of interest to them. The recommendation problem could be cast as either a rating estimation problem which aims to predict as accurately as possible for a user the rating values of items which are yet unrated by that user, or as a ranking problem which aims to find the top-k ranked items that would be of most interest to a user, which s/he has not ranked yet. In contexts where explicit item ratings of other users may not be available, the ranking prediction could be more important than the rating prediction. Most of the existing ranking-based prediction approaches consider items as having equal weights which is not always the case. Different weights of items could be regarded as a reflection of items' importance, or desirability, to users. In this paper, we propose to integrate variable item weights with a ranking-based matrix factorization model, where learning is driven by Bayesian Personalized Ranking (BPR). Two ranking-based models utilizing different-weight learning methods are proposed and the performance of both models is confirmed as being better than the standard BPR method.
AB - Recommendation systems employed on the Internet aim to serve users by recommending items which will likely be of interest to them. The recommendation problem could be cast as either a rating estimation problem which aims to predict as accurately as possible for a user the rating values of items which are yet unrated by that user, or as a ranking problem which aims to find the top-k ranked items that would be of most interest to a user, which s/he has not ranked yet. In contexts where explicit item ratings of other users may not be available, the ranking prediction could be more important than the rating prediction. Most of the existing ranking-based prediction approaches consider items as having equal weights which is not always the case. Different weights of items could be regarded as a reflection of items' importance, or desirability, to users. In this paper, we propose to integrate variable item weights with a ranking-based matrix factorization model, where learning is driven by Bayesian Personalized Ranking (BPR). Two ranking-based models utilizing different-weight learning methods are proposed and the performance of both models is confirmed as being better than the standard BPR method.
KW - Bayesian Personalized Ranking
KW - collaborative filtering
KW - implicit feedback
KW - item recommendation
KW - matrix factorization
UR - http://www.scopus.com/inward/record.url?scp=85041320128&partnerID=8YFLogxK
U2 - 10.23919/SEEDA-CECNSM.2017.8089996
DO - 10.23919/SEEDA-CECNSM.2017.8089996
M3 - Conference Proceeding
AN - SCOPUS:85041320128
T3 - South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference, SEEDA-CECNSM 2017
BT - South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference, SEEDA-CECNSM 2017
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2017 South-East Europe Design Automation, Computer Engineering, Computer Networks and Social Media Conference, SEEDA-CECNSM 2017
Y2 - 23 September 2017 through 25 September 2017
ER -