Abstract Understanding what graph layout human prefer and why they prefer such graph layout is significant and challenging due to the highly complex visual perception and cognition system in the human brain. In this paper, we present the first machine learning approach for predicting human preference for graph layouts. Specifically, we propose a CNN-Siamese-based model to predict human preference from a pair of different layouts of the same graph. We employ a transfer learning method to overcome the insufficiency of the available ground truth human preference experiment data for training deep neural networks. Specifically, we exploit the quality metrics, which are correlated to human preference on graph layouts, to pre-train our model. Then, we fine-tune the model using the ground truth human preference experiment data. Experimental results using the ground truth human preference data sets show that our model M+HP can successfully predict human preference for graph layouts, achieving the average test accuracy of $92.28\%$ for large scale-free and mesh graphs. To our best knowledge, this is the first approach for predicting qualitative evaluation of graph layouts based on the ground truth human preference experiment data. Moreover, comparison experiments show that our model outperforms a simple baseline model and a previous Siamese-based model, demonstrating the importance of using graph layout images and the CNN-based model for predicting human preference.