Distributed Graph Matching Network for Deep Situation Recognition

In order to recognize battlefield situation or envi-ronment, one of the most important issue is to merge semantic in-formation obtained by multiple agents. In this paper, we propose a distributed graph matching network to classify multiple agents based on graph similarity analysis. To this end, we construct a hypergraph to analyze a high-order relationship between agents (i.e., graphs). Then a hypergraph incidence matrix is used to sample a tuple of (anchor, positive, negative) agents, where triplet loss is minimized to train agent embedding vector such that the trained vectors are similar for graphs with similar topological and semantic information. To tackle the over-fitting problem caused by fixed samples of agent, we propose a hypergraph induced sample validation for online update of hyperedges in training step. Experiment results demonstrate the performance of our proposed model.