All Dates/Times are Australian Eastern Standard Time (AEST)

Technical Program

Paper Detail

Paper IDD3-S2-T1.1
Paper Title Coded Alternating Least Squares for Straggler Mitigation in Distributed Recommendations
Authors Siyuan Wang, City University of Hong Kong, Hong Kong SAR of China; Qifa Yan, University of Illinoise at Chicago, United States; Jingjing Zhang, Fudan University, Hong Kong SAR of China; Jianping Wang, Linqi Song, City University of Hong Kong, Hong Kong SAR of China
Session D3-S2-T1: Distributed Computation I
Chaired Session: Wednesday, 14 July, 22:20 - 22:40
Engagement Session: Wednesday, 14 July, 22:40 - 23:00
Abstract Recommender systems are adopted by many companies and have a strong impact on everyone's life in modern society. Collaborative filtering, which involves large amount of matrix decomposition, is of great importance in recommender system. Alternating Least Squares(ALS) is an effective method to tackle the matrix decomposition problem. With the growing of user's data, the power of distributed learning is needed that computation tasks should be divided into subtasks and be distributed on multiple compute node. In such distributed system, however, because of hardware configurations or network delays, some workers that cannot return the results on time, also known as stragglers, could affect the efficiency of distributed algorithms. To tackle this, in this work, we design an distributed implementation of ALS algorithm and a stragglers-mitigate coding scheme using coding schemes to eliminate the impacts of stragglers. This approach divide the data matrix into $p^2$ batches and encodes the batches as $w$ new batches to be processed on $w$ workers.