On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems

Murat Yagci; Tevfik Aytekin; Fikret Gurgen

doi:10.1145/3109859.3109906

On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems

Murat Yagci, Tevfik Aytekin, Fikret Gurgen

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

10 Citations (Scopus)

Abstract

Learning to rank with pairwise loss functions has been found useful in collaborative filtering recommender systems. At web scale, the optimization is often based on matrix factorization with stochastic gradient descent (SGD) which has a sequential nature. We investigate two different shared memory lock-free parallel SGD schemes based on block partitioning and no partitioning for use with pairwise loss functions. To speed up convergence to a solution, we extrapolate simple practical algorithms from their application to pointwise learning to rank. Experimental results show that the proposed algorithms are quite useful regarding their ranking ability and speedup patterns in comparison to their sequential counterpart.

Original language	English
Title of host publication	RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems
Publisher	Association for Computing Machinery, Inc
Pages	37-41
Number of pages	5
ISBN (Electronic)	9781450346528
DOIs	https://doi.org/10.1145/3109859.3109906
Publication status	Published - 27 Aug 2017
Externally published	Yes
Event	11th ACM Conference on Recommender Systems, RecSys 2017 - Como, Italy Duration: 27 Aug 2017 → 31 Aug 2017

Publication series

Name	RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems

Conference

Conference	11th ACM Conference on Recommender Systems, RecSys 2017
Country/Territory	Italy
City	Como
Period	27/08/17 → 31/08/17

Keywords

Learning to rank
Pairwise loss
Parallel SGD
Personalization

Access to Document

10.1145/3109859.3109906

Cite this

Yagci, M., Aytekin, T., & Gurgen, F. (2017). On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems. In RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems (pp. 37-41). (RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems). Association for Computing Machinery, Inc. https://doi.org/10.1145/3109859.3109906

@inproceedings{f4c5791f227f403e99dd8986ea46cc81,

title = "On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems",

abstract = "Learning to rank with pairwise loss functions has been found useful in collaborative filtering recommender systems. At web scale, the optimization is often based on matrix factorization with stochastic gradient descent (SGD) which has a sequential nature. We investigate two different shared memory lock-free parallel SGD schemes based on block partitioning and no partitioning for use with pairwise loss functions. To speed up convergence to a solution, we extrapolate simple practical algorithms from their application to pointwise learning to rank. Experimental results show that the proposed algorithms are quite useful regarding their ranking ability and speedup patterns in comparison to their sequential counterpart.",

keywords = "Learning to rank, Pairwise loss, Parallel SGD, Personalization",

author = "Murat Yagci and Tevfik Aytekin and Fikret Gurgen",

note = "Publisher Copyright: {\textcopyright} 2017 ACM.; 11th ACM Conference on Recommender Systems, RecSys 2017 ; Conference date: 27-08-2017 Through 31-08-2017",

year = "2017",

month = aug,

day = "27",

doi = "10.1145/3109859.3109906",

language = "English",

series = "RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems",

publisher = "Association for Computing Machinery, Inc",

pages = "37--41",

booktitle = "RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems",

}

Yagci, M, Aytekin, T & Gurgen, F 2017, On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems. in RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems. RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems, Association for Computing Machinery, Inc, pp. 37-41, 11th ACM Conference on Recommender Systems, RecSys 2017, Como, Italy, 27/08/17. https://doi.org/10.1145/3109859.3109906

On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems. / Yagci, Murat; Aytekin, Tevfik; Gurgen, Fikret.
RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems. Association for Computing Machinery, Inc, 2017. p. 37-41 (RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems

AU - Yagci, Murat

AU - Aytekin, Tevfik

AU - Gurgen, Fikret

PY - 2017/8/27

Y1 - 2017/8/27

N2 - Learning to rank with pairwise loss functions has been found useful in collaborative filtering recommender systems. At web scale, the optimization is often based on matrix factorization with stochastic gradient descent (SGD) which has a sequential nature. We investigate two different shared memory lock-free parallel SGD schemes based on block partitioning and no partitioning for use with pairwise loss functions. To speed up convergence to a solution, we extrapolate simple practical algorithms from their application to pointwise learning to rank. Experimental results show that the proposed algorithms are quite useful regarding their ranking ability and speedup patterns in comparison to their sequential counterpart.

AB - Learning to rank with pairwise loss functions has been found useful in collaborative filtering recommender systems. At web scale, the optimization is often based on matrix factorization with stochastic gradient descent (SGD) which has a sequential nature. We investigate two different shared memory lock-free parallel SGD schemes based on block partitioning and no partitioning for use with pairwise loss functions. To speed up convergence to a solution, we extrapolate simple practical algorithms from their application to pointwise learning to rank. Experimental results show that the proposed algorithms are quite useful regarding their ranking ability and speedup patterns in comparison to their sequential counterpart.

KW - Learning to rank

KW - Pairwise loss

KW - Parallel SGD

KW - Personalization

UR - http://www.scopus.com/inward/record.url?scp=85030454523&partnerID=8YFLogxK

U2 - 10.1145/3109859.3109906

DO - 10.1145/3109859.3109906

M3 - Conference contribution

AN - SCOPUS:85030454523

T3 - RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems

SP - 37

EP - 41

BT - RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems

PB - Association for Computing Machinery, Inc

T2 - 11th ACM Conference on Recommender Systems, RecSys 2017

Y2 - 27 August 2017 through 31 August 2017

ER -

Yagci M, Aytekin T, Gurgen F. On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems. In RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems. Association for Computing Machinery, Inc. 2017. p. 37-41. (RecSys 2017 - Proceedings of the 11th ACM Conference on Recommender Systems). doi: 10.1145/3109859.3109906

On parallelizing SGD for pairwise learning to rank in collaborative filtering recommender systems

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this