摘要： 利用经典流形学习算法研究流量矩阵中的流形结构,能够获得流量矩阵的本征维度。然而局部线性嵌入(LLE)算法依赖于近邻点的选取,传统近邻选取个数往往通过实验试凑法得到最优解,不能完全揭示流量矩阵的流形结构。针对上述缺点,提出一种改进的局部线性嵌入算法,该算法利用改进的LLE算法探索流量矩阵的流形结构,并对实际骨干网络中的流量矩阵进行分析。实验结果证明,改进算法具有较小的重构误差,相对于标准LLE算法,能更为准确地揭示流量矩阵的低维流形结构。

关键词: 流量矩阵, 局部线性嵌入算法, 本征维度, 流形结构, 骨干网络

Abstract: The manifold structure of the traffic matrix is studied by using classical manifold learning algorithms,and the intrinsic dimension of the flow matrix can be got. However,the Locally Linear Embedding(LLE) algorithms rely on the selection of neighbor points. The traditional method is often obtained by experiment. Even so,this does not fully reveal the manifold structure of traffic matrix. Aiming at the above shortcomings,this paper improves the LLE algorithm by improving the LLE algorithm. It applies the improved LLE algorithm to real OD traffic matrix taken from the backbone network (Abilene).Experimental result shows the improved LLE algorithm has a smaller reconstruction error,compared with standard LLE algorithm,it can reveal a low-dimensional manifold structure exactly.

Key words: traffic matrix, Locally Linear Embedding(LLE) algorithm, intrinsic dimensionality, manifold structure, backbone network

中图分类号: 

• TP391