Backbone Kernels And PSO For The Graph Data Clustering

Kernel method is an effective way to solve nonlinear mode analysis problem, its core ideas are: first, expanding the original data to the appropriate high-dimensional feature space through some nonlinear mapping; Then, analizing and processing mode in this new space. Compaired with using generic nonlinear directly on the original data ,nuclear method has obvious advantages : firstly, nonlinear mapping of kernel method can integration problems which related to the prior knowledge. Moreover, kernel method can better ensure generalization capability. Also, as the computational complexity and high-dimensional feature space dimension is irrelevant in the kernel method, it reduces the computing complexity. Therefore, kernel method can be well study the figure data structure. The Figure learning mainly includes: frequent sub-graph mining, graph relationship learning, etc. Among them, the frequent sub -graph mining can be as graph clustering and figure classification foundation.Meanwhile, this paper proposes a particle swarm algorithm that could adapt population based on a lot researches.Combined with the traditional clustering algorithms, it forms discrete particles clustering algorithm which is adaptive. Through studying and analizing the aopology of particle swarm, the paper picks up the merits of Gbest model and Lbest model. In the early stage ,the paper uses Lbest, as the searching is going,it forms a multi-cluster structure. In clusters,we use fully connected topology, while between clusters the ring is topology used. In this way, it not only guarantees the global search ability of PSO, but also has a good control of local search capabilities of the particle swarm. Experiments show that the algorithm has a very good results in the cluster.In addition, the paper extract the high degree of sub-structure - the backbone Graph to analysis. Then, it uses the random path graph kernel function to define the core function of the trunk, and gives different weights to backbone graphs which have different orders. Finally, it studies the nuclear similar matrix through the adaptive population particle swarm algorithm. Experimental results show that the proposed method can well learn the graph structure data.