Abstract: Most of the existing network representation learning algorithms are mainly based on the DeepWalk algorithm. The main improvement is to modify DeepWalk's three-layer neural network to multi-layer neural network or to introduce the network's external attributes into the DeepWalk algorithm for joint representation learning. In network representation learning, the random walk strategy can be considered as network data preprocessing of network representation learning tasks. Learning the random walk sequences, including most of the network structure features, is very important for the network representation learning algorithm, because the subsequent training procedure of neural networks is to continuously adjust the representation of each node in the network based on the co-occurrence of node pairs. Therefore, the purpose of this paper is to improve the random walk strategy of the DeepWalk algorithm. We propose a novel DeepWalk algorithm based on the improved random walk with equal probability (EPDW). Although the next node in the random walk of the DeepWalk algorithm is chosen with equal probability, one of the neighbouring nodes of the current node is chosen as the next node of the random walk by the pseudo-random number. The improvement of EPDW is to choose the next hop node in the random walk using the roulette method of cumulative probability sum of random walk nodes. Although this method also chooses the next hop node with equal probability, it can choose the next hop random walk node more reasonably and effectively.

Key words: network representation learning, network representation, network embedding, feature learning