基于图谱理论几何空间结构变换的大数据核聚类算法

现有核聚类算法需要学习完整的核矩阵,计算效率较低,仅适用于小规模数据,对此提出了基于图谱理论的核函数分类算法。首先,基于图谱理论建立无标记数据的相似度图；然后,计算其拉普拉斯矩阵,并选取拉普拉斯矩阵的部分特征向量来学习；最终,利用核k-means算法实现数据分类。对比试验结果证明,本算法在具有较好聚类性能的前提下,计算效率明显优于其他同类型算法,并适合中型及大型规模数据分类处理。

Spectral graph geometric transform based kernel clustering approach for big scale data with high computer efficiency

For the problem that the existing kernel clustering approaches need to learn the entire kernel matrix with the low compute efficiency, aimed at it proposed a clustering approach based on the spectral. Firstly, base on spectral construct the similarity graph; then, compute the Laplacian matrix for the graph and select little part of the matrix to learn; lastly, realize the classification with kernel k-means approach. The simulating results prove that the proposed approach has better compute efficiency than the other kernel clustering approaches with a comparable clustering performance and works for big scale data.