一种基于MapReduce的半监督近邻传播算法

近邻传播算法(Affinity Propagation)是一种具有较高准确度的聚类算法，但是其具有较高的时间复杂度，且无法有效聚类结构松散数据，针对这两个问题，提出了一种基于MapReduce的半监督近邻传播算法(MR-SAP)。算法首先利用MapReduce编程框架，在各个数据节点上运行AP算法，得到局部的聚类中心，以及代表每一个局部聚类中心成为全局聚类中心可能性的决策系数，然后综合局部聚类中心进行全局的AP聚类，其中初始参考度的选取依据输入的决策系数，最后通过引入IGP聚类评价指标比较聚类效果，引导算法向结果最优方向运行。实验结果表明该算法在处理不同大小、不同类型数据集时均具有良好的效率和扩展性，且具有较高的聚类精度。

An Semi-supervised affinity propagation algorithm based on MapReduce

affinity propagation algorithm is a high accuracy clustering algorithm, but it has high time complexity, and can not effectively cluster loosely structured data, in order to solve these two problems, proposed an Semi-supervised affinity propagation algorithm based on MapReduce (MR-SAP). Firstly, used the MapReduce programming framework to run the AP algorithm in each data node, clustering centers obtained locally, and also obtained decision coefficient which represent each local clustering center became the possibility of global clustering center, then combined local AP clustering center for running global AP, the initial preference of selecting decision coefficient based on the input, then through the comparison of clustering results based on IGP clustering evaluation index, make the algorithm runs in the best direction. experiments show that MR-SAP has good efficiency and scalability in dealing with different sizes and different types of data sets, and has high clustering accuracy.