基于光滑表示的半监督分类算法

Smooth Representation-based Semi-supervised Classification

近年来,基于图的半监督分类是机器学习与数据挖掘领域的研究热点之一。该类方法一般通过构造图来挖掘数据中隐含的信息,并利用图的结构信息来对无标签样本进行分类。因此,半监督分类的效果严重依赖于图的质量。文中提出了一种基于光滑表示的半监督分类算法。具体来说,此方法通过应用一个低通滤波器来实现数据的平滑,然后将光滑数据用于半监督分类。此外,所提方法将常见的图构造和标签传播集成到一个统一的优化框架中,使它们互相促进,从而避免低质量图导致的次优解。对人脸和物品数据集进行大量实验,结果表明,所提SRSSC算法在大部分情况下都优于其他算法,从而证明了光滑表示的重要性。

Graph-based semi-supervised classification is a hot topic in machine learning and data mining.In general,this method discovers the hidden information by constructing a graph and predicts the labels for unlabeled samples based on the structural information of graph.Thus,the performance of semi-supervised classification heavily depends on the quality of graph.In this work,we propose to perform semi-supervised classification in a smooth representation.In particular,a low-pass filter is applied on the data to achieve a smooth representation,which in turn is used for semi-supervised classification.Furthermore,a unified framework which integrates graph construction and label propagation is proposed,so that they can be mutually improved and avoid the sub-optimal solution caused by low-quality graph.Extensive experiments on face and subject data sets show that the proposed SRSSC outperforms other state-of-the-art methods in most cases,which validates the significance of smooth representation.