Label Efficient Semi-Supervised Learning via Graph Filtering

• DOI: 10.48550/arxiv.1901.09993
• Publication Date: 2019-01-01
• Authors: Qimai Li; Xiao-Ming Wu; Han Liu; Xiaotong Zhang; Zhichao Guan
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Statistics - Machine Learning; 0202 electrical engineering, electronic engineering, information engineering; Machine Learning (stat.ML); 02 engineering and technology; Machine Learning (cs.LG)

Graph-based methods have been demonstrated as one of the most effective approaches for semi-supervised learning, they can exploit connectivity patterns between labeled and unlabeled data samples to improve learning performance. However, existing graph-based either are limited in their ability jointly model graph structures features, such classical label propagation methods, or require a considerable amount training validation due high complexity, recent neural-network-based methods. In this paper, we address efficient from filtering perspective. Specifically, propose framework that injects similarity into features by taking them signals on applying low-pass filter extract useful representations classification, where efficiency be achieved conveniently adjusting strength filter. Interestingly, unifies two seemingly very different -- convolutional networks. Revisiting under leads new insights modeling capabilities reduce complexity. Experiments various classification tasks four citation networks knowledge regression task zero-shot image recognition validate our findings proposals.