摘要：In this paper, we present a novel semi-supervisedsmooth harmonic transductive learning algorithm that canget closed-form solution. Our method introduces the unlabeledclass information to the learning process and triesto exploit the similar configurations shared by the labeldistribution of data. After discovering the property ofsmooth harmonic function based on spectral clusteringin classification task, we design an adaptive thresholdingmethod for smooth harmonic transductive learning basedon classification error. The proposed adaptive thresholdingmethod can select the most suitable thresholds flexibly.Plentiful experiments on data sets show our proposed closedformsmooth harmonic transductive learning framework getexcellent improvement compared with two baseline methods.
