Semi-supervised Learning Applied to Large Data Sets with Very Few Labeled Examples

A semi-supervised classification approach, SS-LFL, is proposed. In SS-LFL, some weak binary classifiers, each of which can identify instances of one particular class, are firstly trained on the labeled data, and the whole data set is then clustered into partitions until they are tight and pure enough. SS-LFL alternates between assigning ¿imperfect-classes¿ to the unlabeled data in these partitions and constructing the next weak binary classifiers using both the labeled and ¿imperfect¿ data. It works well in large data sets with very few labeled examples, moreover, it neither requires known parametric distributions of data nor participation of an expert. Experimental results carried out on some public datasets collected from the UCI machine learning repository show that SS-LFL is a promising method.