属性图相似性阈值对类属超图(CSHG)模型的训练结果具有重要影响。在满足聚类准确性的条件下，利用定义的熵函数给出优化CSHG模型结构的相似性阈值，并得到初始优化的CSHG模型，进一步利用FTOG之间的相似性矩阵得到最简CSHG模型结构。另外，利用亲缘传播聚类(affinity propagation clustering)方法去除FTOG聚类中的冗余属性图，最终得到最优的CSHG模型。实验结果表明，本方法是有效的。

CSHG (class specific hyper graph) model is largely influenced by the threshold of the similarity measure between two graphs. By setting a constraint to fault tolerance of clustering, the similarity measure threshold for initial optimized CSHG model is obtained using the entropy function defined on CSHG model and the initial optimized CSHG model is thus constructed. The similarity matrix of FTOG is then created and the most simplified CSHG structure is obtained. In addition, the redundant graphs in FTOG are detected and deleted using affinity propagation method and the final optimized CSHG model is acquired. Experimental results demonstrate the reliability and effectiveness.