Mixture distribution modeling for scalable graph-based semi-supervised learning

Abstract

Graph-based semi-supervised learning (SSL) has been widely investigated in recent works considering its powerful ability to naturally incorporate the diverse types of information and measurements. However, traditional graph-based SSL methods have cubic complexities and leading to low scalability. In this paper, we propose to perform graph-based SSL on mixture distribution components, named Mixture-distribution based Graph Smoothing (MGS), to address this challenge. Specifically, the intrinsic distributions of data are captured by a mixture density estimation model. A novel mixture-distribution based objective energy function is further proposed to incorporate few available annotations, which ensures the model complexity is irrelevant to the number of raw instances. The energy function can be simplified and effectively solved by viewing the instances and mixture components as the point clouds. Experiments on large datasets demonstrate the remarkable performance improvements and scalability of the proposed model, which proves the superiority of the MGS model.