[1] Niepert, M., Ahmed, M., & Kutzkov, K. (2016, June). Learning convolutional neural networks for graphs. In International conference on machine learning (pp. 2014-2023).
[2] Wang, H., Wang, J., Wang, J., Zhao, M., Zhang, W., Zhang, F., … & Guo, M. (2018, April). Graphgan: Graph representation learning with generative adversarial nets. In Thirty-Second AAAI Conference on Artificial Intelligence.
[3] Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Yu, P. S. (2019). A comprehensive survey on graph neural networks. arXiv preprint arXiv:1901.00596.
[4] Tang, L., & Liu, H. (2009, June). Relational learning via latent social dimensions. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 817-826). ACM.
[5] Perozzi, B., Al-Rfou, R., & Skiena, S. (2014, August). Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 701-710). ACM.
[6] Wang, D., Cui, P., & Zhu, W. (2016, August). Structural deep network embedding. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 1225-1234). ACM.
[7] Wu, L., Sun, P., Hong, R., Fu, Y., Wang, X., & Wang, M. (2018). SocialGCN: An Efficient Graph Convolutional Network based Model for Social Recommendation. arXiv preprint arXiv:1811.02815.
与之对比的,是我们提出的SelectorGAN输出如下图所示,其中,左边是ground truth ,右边是模型的输出结果。最明显的区别是我们的模型不容易生成全黑或者全白的无意义图片,几乎所有文字都是清晰可辨的。
在介绍我们的模型之前,我们先分析了当前state of art模型的缺陷。EMD模型的图例如下图所示,将16张候选图片,输入到CNN的16个channel中。这里,由于16张图片是无序不可分的,而CNN的不同Channel是有序可分的,这里就会导致交换输入图片的顺序会产生不同的结果。这其实是一个不正确的做法,倘若我们将16*1的CNN作用于16张图片修改16个为1*1的CNN分别作用于16张图片,可以将模型的参数个数降低16倍,有效降低过拟合的出现。用时,原文的欧几里得距离损失函数,也是的模型在对于生成图片没有信心时,直接输出全部空白,或者全部黑色,而正常图片也存在模糊边界的问题。这一系列问题到这了EMD模型实际跑的并没有文章中说的那么厉害。