Distributed Graph-based Semi-supervised Classifiers: Sampling and Interpolation

点击翻译按钮获取中文摘要

We study distributed classifier learning from a graph signal processing (GSP) perspective. In practical networked systems, big data often reside in geographically diverse regions scattered in local networks around the world. Leveraging on Cisco's mature and extensive network infrastructure and our in-house expertise in GSP, we develop an efficient distributed graph-based classifier learning framework, by extracting and disseminating knowledge from data located across a large network. Specifically, we focus on two core problems in graph-based semi-supervised classifier learning: distributed graph sampling and distributed signal interpolation. Distributed graph sampling is the problem of pre-selecting nodes on the graph based solely on local network information to acquire informative labels. Distributed graph interpolation is the problem of interpolating missing samples in the remaining nodes, given a set of observed labels in a local neighborhood. We approach both problems from a graph spectral perspective, resulting in a classifier system with theoretically derived and explainable performance guarantees.