Improved DeepWalk Algorithm Based on Preference Random Walk

Abstract

Network representation learning based on neural network originates from language modeling based on neural network. These two types of tasks are then studied and applied along different paths. DeepWalk is the most classical network representation learning algorithm, which samples the next hop nodes of the walker with an equal probability method through the random walk strategy. Node2vec improves the random walk procedures, thus improving the performance of node2vec algorithm on various tasks. Therefore, we propose an improved DeepWalk algorithm based on preference random walk (PDW), which modifies the single undirected edge into two one-way directed edges in the network, and then gives each one-way directed edge a walk probability based on local random walk algorithm. In the procedures of acquiring walk sequences, the walk probability of the paths that have been walked will be attenuated according to the attenuation coefficient. For the last hop node of the current node in the walk sequences, an inhibition coefficient is set to prevent random walker from returning to the last node with a greater probability. In addition, we introduce the Alias sampling method in order to obtain the next hop node from the neighboring nodes of current node with a non-equal probability sampling. The experimental results show that the proposed PDW algorithm possesses a stable performance of network representation learning, the network node classification performance is better than that of the baseline algorithms used in this paper.