Abstract: Link prediction aims to utilize the known nodes or network topology to predict the likelihood of connection between two unlinked nodes in a network. Traditional link prediction methods often extract transition matrices from the original graph, which leads to sparse information. So, a link prediction algorithm based on graph neural networks and random walks called LP-GNRW is proposed in this paper. Firstly, Bert, an attention-based graph neural network, is introduced to obtain the multiple embedding representations for nodes. Then, the high-order structural information of the graph is captured by combining these representations with random walks. Finally, we transform the link prediction task into a binary classification problem and the graph neural network is adopted to classifying the obtained high-order structural information to perform the link prediction. Experimental results demonstrate that LP-GNRW can effectively learn the graph structural features. Compared with other heuristic methods based on random walks, it achieves a superior AUC score and enhances the performance of link prediction.