摘要:
与SOFM,最大熵聚类,K均值聚类相比,Neural-Gas网络算法具有收敛速度快、代价误差小等优点。但Neural-Gas网络用于非均匀分布的线性或非线性数据集进行降维或可视化时,输出空间上固定有序的神经元表现出极不理想的距离信息。为此,该文根据归一化概率自组织特征映射的基本思想,提出混合Neural-Gas网络和Sammon映射的新方法来解决此问题,通过Neural-Gas网络算法进行特征聚类以降低计算复杂度,通过Sammon映射保持输入空间和输出空间上神经元间的距离相似性。仿真结果表明,该混合算法对合成数据集或现实数据集的可视化能够取得较理想的效果,从而验证了该混合算法的可行性和有效性。
Abstract:
Compared with Self-Organizing Feature Map(SOFM), maximum-entropy clustering and K-means clustering, the Neural-Gas network algorithm has advantages of faster convergence, smaller cost distortion errors, etc. However, the fixed and regular neurons on the output space represent worse distance information when the neural gas network algorithm is used for dimension reduction and visualization of linear or nonlinear data sets with nonuniform distribution. Therefore, according to the basic idea of the probabilistic regularized SOFM, a new visualization method for hybridizing neural gas network and Sammons mapping is proposed to overcome this problem, and it reduces the computational complexity with using neural gas network algorithm for feature clustering and preserves the interneuronal distances resemblance from input space into output space by using Sammons mapping. Simulation results show that the proposed hybridizing algorithm can obtain the better visualization effect on the synthetic and real data sets, thus demonstrating the feasibility and effectiveness of the hybridizing algorithm.
下载:
