Effective influential spreaders prediction in online social networks is critical for a variety of network services such as viral marketing, online recommendation, and many more. The conventional machine learning methods mainly adopted various hand-crafted features in the prediction models. However, their effectiveness relies on the domain knowledge heavily, and thus it is difficult to generalize these models to different domains. In this paper, we propose an Influence Deep Learning (IDL) model to learn users' latent feature representation for predicting influence spread. Our IDL model is fully data-driven, and it takes sampling subnetworks as input to the deep neural networks for learning users' latent vector representation. Moreover, we design a strategy to incorporate user-specific features and network structure into the graph convolutional neural network to overcome the imbalance problem of labeled training data. The experimental results show that our model outperforms other baselines in online social networks, and our model is more practical in large-scale data sets.