Most of the existing influence maximization problems assume that k user promotion targets are selected to the entire mobile social networks (MSN) under the complete network structure. However, in reality, it is unrealistic to acquire the complete network structure. Therefore, it is our motivation to maximizing influence under partially observable networks. Firstly, we propose a new model named Variational Graph Auto-Encoder with Network Gravity (VGAE-WNG) which combined VGAE with a new effective decoder to obtain the link structure that was not presented before. Secondly, we propose a novel Similarity Decreasing Transfer Algorithm (SDTA) to evaluates the reachability of a node’s influence on other nodes, by according to the transfer of similarity between nodes and the distance of information spread on the path between nodes. Finally, we performed experiments on three different scale networks. The results show that our model outperforms other algorithms by about 2% in link prediction, and our method achieves similar or even better propagation performance in the absence of partial network structures than state-of-the-art algorithms with full network structures.