Resumen de A Method to Identify Graphs Containing a Specific Sub-graph
Graph structure represents a variety of adjacencies. A subgraph is a portion of a graph, and a particular subgraph can obtain the characteristics of an entire graph. In this paper, we propose a method to infer a graph that contain a target subgraph. A graph convolutional network (GCN) was used to turn the graphs into embedding vectors, which represent tendency of their structure. In the loss function used for training, we used Simpson coefficient to measure the overlapping of graphs. The degree of inclusion was measured as cosine similarity of the obtained vectors. To evaluate the method, we used card images of competitive Karuta. The adjacency of letters on the card was represented as a graph, and lack of recognized letters under the realistic environment such as under sunlight or shadow gives a subgraph of the graph for the card. We used our model to identify card with the subgraph, and found our method achieved over 90% of accuracy in this setting.
