关键词: 链路预测, 几何布朗运动, 随机游走算法, 网络表示学习算法

Abstract: The network is an important form of expressing complex relationships between objects and exists extensively. Link prediction, as an important method of network analysis, has great research significance and application value. However, the traditional link prediction algorithms are generally designed based on the sparse representation scheme of the adjacency matrix. Such algorithms often have low computational efficiency and poor scalability. This paper first introduces the concept of network representation learning and innovatively proposes a random walk algorithm called GbmRw based on geometric Brownian motion, then further designs a network representation learning algorithm GBMLA to achieve a more differentiated and expressive representation of network. Finally, the Euclidean distance between the representation vectors of nodes is used to characterize their similarity, so as to predict the likelihood of the link??s existence. The results of repeated experiments in multiple networks from different domains show that compared with the traditional algorithm designed based on the original network, the proposed algorithm has obviously improved the prediction performance, which further confirms the significance of network representation learning for link prediction work.

Key words: link prediction, geometric Brownian motion, random walk algorithm, network representation learning algorithm