自组织映射拓扑保持的增强

Enhancement of topology preservation of self-organizing map

在自组织映射(SOM)中,网格各单元的权值向量仅仅是根据各单元和最佳匹配单元(BMU)之间的距离进行更新的,因而输入数据间的拓扑关系不能得到很好的保持。为此提出了两种改进方案。在第一种改进方案中,各单元的权值向量根据各单元和BMU之间对应各坐标的差进行更新。实验结果表明,这种改进方案可以很好地保持拓扑关系,但输入数据的分布密度却不能得到较好的体现。在第二种改进方案中,各单元的权值向量同时根据各单元和BMU之间对应各坐标的差与距离进行更新。实验结果表明,这种改进方案不仅能使拓扑关系得到比SOM更好的保持,而且较好地体现了输入数据的分布密度,并加快了训练的收敛速度。

In the Self-Organization Map （ SOM）, the weight vectors of the units in the grid are updated only according to the distance between the units and the Best Matching Unit （BMU）, so the topological relationship between input data can not be preserved very well. Therefore, two improved schemes were proposed. In the first improved scheme, the weight vectors of the units were updated according to the differences of the corresponding coordinates between the units and the BMU. Experimental results show that this improved scheme can preserve topological relationship very well, but the distribution density of the input data can not be reflected quite well. In the second improved scheme, the weight vectors of the units were updated both according to the differences of the corresponding coordinates and the distance between the units and the BMU. Experimental results show that this improved scheme can not only preserve topological relationship better than SOM, but also reflect the distribution density of the input data quite well and accelerate the convergence speed of the training.