Coded Distributed Graph-Based Semi-Supervised Learning

Abstract

Semi-supervised learning (SSL) has been applied to many practical applications over the past few years. Recently, distributed graph-based semi-supervised learning (DGSSL) has shown to have good performance. Traditional DGSSL algorithms usually have the problem of the straggler effect that algorithm execution time is limited by the slowest node. To solve this problem, a novel coded DGSSL(CDGSSL) algorithm based on the Maximum Distance Separable (MDS) code is proposed in this paper. Specifically, the proposed algorithm is based on the Maximum Distance Separable (MDS) code. In general, the proposed coded distributed algorithm is straggler-tolerant. Moreover, we provide optimal parameters design for the proposed algorithm. The superiority of the proposed algorithm has been confirmed via experiments on Alibaba Cloud Elastic Compute Service.