Considering that the existing community search algorithms represent nodes as high-dimensional sparse vectors and have high time complexity, a community search algorithm based on node embedding representation learning (CSNERL) is proposed. Node embedding techniques can learn low-dimensional vectorial representation of nodes from network structure directly, and provide a new solution to community search problems. Firstly, in view of the problem that the existing node embedding algorithm has a high probability to walk back and forth between the closest neighbors, a node embedding model based on closest-neighbor biased random walk with non-immediately revisiting (NECRWNR) is proposed. Based on this model, vectorial representation of nodes is learned and used as feature vectors of nodes in the downsteam data mining task. Then, vectorial representation of a community is defined as the average of the vectors for nodes in the community, and a new community search algorithm is designed by choosing those nodes which are nearest to the current community. The proposed algorithm is tested on both real-world and synthetic network datasets with the related community search algorithms. The experimental results show that the CSNERL algorithm is more effective at community search than baselines.