Network Representation Learning Based on Topological Structure and Vertex Attributes

Network Representation Learning (NRL) is an essential task in the field of network data analysis, which tries to learn the distributed representation of each vertex in the network for downstream vector-based data mining tasks. NRL is helpful in solving the computationally expensive or intractable problems of large-scale network analysis. Most related NRL methods only focus on encoding the network topology information into vertex representation. However, vertices may contain rich attributes that directly impact the network formation and measure the attribute-level similarity between vertices. Additionally, encoding the vertex attributes information into the representation vector may improve the performance of the representation. This paper proposes a general NRL framework TAFNE that can effectively retain both network topology and vertex attributes information. For complex types of vertex attributes, we design two different information fusion methods that take both training efficiency and generality into account. The proposed TAFNE framework is extensively evaluated through various data analysis tasks, including clustering, visualization and node classification, and achieves superior performance compared with baseline methods.