A Vectorization Approach for Graph-Structured Data to Pattern Recognition

Graphs are popular data structures that intuitively represent structured information. With the expansion of the graph scale and the complexity of graph attribute information, it is necessary to generalize graphs to apply graph-based learning techniques accurately and effectively. Based on the previous methods of generating fixed-size vectors for attributed graphs as the input to convolutional neural networks, we propose a novel vector structure to represent the complex attribute and structure information for graphs. In this paper, we define a general method of attribute classification and aggregation functions for node and edge attributes on the graphs, which are applied to weaken the performance bottlenecks of the previous methods in the multi-graphs scenario. Experiments show that our works can greatly improve the computation efficiency across multiple graphs while ensuring the accuracy of graph-based learning results.