Research On Heterogeneous Multi-scale Network Embedding Algorithms

Network Embedding(Graph Embedding)Algorithms aim to represent nodes of a given graph as low-dimensional vectors space,which can reveal the context information of graph nodes.The node vectors can be fed into existing machine learning algorithms,which enables the application of abundant machine learning algorithms to graph analysis tasks.Network embedding algorithms are widely used in recommendation systems,user personas,etc.Most existing network embedding algorithms can only consider local neighborhood as the context to build node vectors,while ignoring the global structure and hierarchical clustering property of large complex graphs.On the other hand,the existing heterogeneous network embedding algorithms model networks as static graphs,which may be very inefficient while during with dynamic evolving networks,and difficult to be applied to real-world scenarios.To solve the above problems,we propose an ant-colony based multi-level network embedding algorithm and a heterogeneous link-prediction based dynamic network embedding algorithm in this paper.We apply ant-colony walking to the graph to detect the multi-layer hierarchical structures,and blend the vectors of each layer using dimensional reduction techniques to get multi-level embedding vectors that reveal the hierarchical clustering property of nodes.At the same time,we also propose a heterogeneous link-prediction based dynamic network embedding algorithm.First apply link-prediction technique to the sparse static subnetworks to solve the sparsity problem,and then embed the nodes of the dynamic subnetworks incrementally,to achieve high embedding efficiency on dynamic graphs and improve the usability in real-world scenarios.Comprehensive experiments on real-world graph datasets show the superiority of the proposed algorithms in graph analysis tasks.Besides,to test the performance of the heterogeneous link-prediction based dynamic network embedding algorithm on practical scenarios,we apply the algorithm to an author disambiguation scenario,and make comparisons with commercial scholar search engines.The results show that the algorithm can achieve better scores than some commercial systems do.