Safety-aware Graph-based Semi-Supervised Learning

摘要: Abstract In machine learning field, Graph-based Semi-Supervised Learning (GSSL) has recently attracted much attention and many researchers have proposed a number of different methods. GSSL generally constructs k nearest neighbors graph to explore manifold structure which may improve performance GSSL. If one uses an inappropriate learn semi-supervised classifier, the classifier be worse than that supervised (SL) only trained by labeled samples. Hence, it is worthy design safe version broaden application area this paper, we introduce Safety-aware (SaGSSL) method can adaptively select good graphs simultaneously. The basic assumption high quality if sample margin obtained with larger SL. By identifying high-quality setting corresponding weights large, predictions our algorithm will approach those graphs. Meanwhile, low-quality should small close Hence degeneration probability reduced expected realize goal exploitation Experimental results on several datasets show simultaneously implement selection safely exploit unlabeled