Evolutionary Strategy for Graph Embedding

Abstract

Graph embedding is an important method for learning low-dimensional representations of vertices in graph data. The problem of graph embedding requires that a better embedding method be used to optimize the corresponding objective function. There are two challenges associated with graph embedding. First, the optimization algorithm is based on gradient descent and falls easily into the local optimum. Second, whether the objective function design is reasonable has a huge impact on the embedding results. To tackle this two challenges, evolutionary strategies are used as the optimization algorithm for graph embedding. Evolutionary strategies do not need to know the specific analytical form of the objective function, and can effectively overcome the challenge of the problem of optimum. In addition, to tackle the challenge of the objective function, this paper improves on the design of the objective function based on the previous research. To verify the effectiveness of the algorithm, experiments on multi-label classification tasks were carried out on four real network data sets. Experiments show the effectiveness and potential of evolutionary strategy for graph embedding.