Label propagation classification based on semi-supervised affinity propagation algorithm

Different position of labeled samples will bring about diverse results of Label Propagation (LP) classification algorithm. Labeled samples which are in the border region of class tend to decrease the effectiveness of LP. This paper proposes an improved LP classification method based on semi-supervised affinity propagation (AP) algorithm named as AP-LP. AP-LP runs clustering through semi-supervised AP firstly, and propagates labels of exemplars instead of labeled samples based on LP, finally transmits labels of exemplars to unlabeled samples in each cluster. LP classification does well in the case of the sampling distribution with both local consistency and global consistency. The algorithm analysis and experimental results show that the performance of AP-LP classification method is superior to LP as a whole.