Graph Convolutional Network with Time-based Mini-batch for Information Diffusion Prediction
Graph Convolutional Network with Time-based Mini-batch for Information Diffusion Prediction
复制标题
用于信息扩散预测的基于时间小批量的图卷积网络
DOI：
10.1007/978-3-030-65351-4_5
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Tsuyoshi Murata
中科院分区：
文献类型：
--
作者：
Hajime Miyazawa;Tsuyoshi Murata
信息传播预测是理解信息传播现象的基础性工作。以前的许多作品使用静态社交图或级联数据进行预测。相比之下，最近提出的深度学习模型DyHGCN [20]通过使用动态图新考虑了用户的动态偏好，并实现了更好的性能。然而，由于多个图卷积计算，DyHGCN的训练阶段在计算上是昂贵的。更快的训练对于快速反映用户的动态偏好也很重要。因此，我们提出了一种新的基于时间的小批量图卷积网络模型（GCNTM），以提高训练速度，同时建模用户的动态偏好。基于时间的小批量是一种新的输入形式，可以有效地处理动态图形。使用此输入，我们在每个小批处理中仅减少一次图卷积计算。三个真实世界数据集上的实验结果表明，我们的模型与基线模型的结果相当。此外，我们的模型学习速度比DyHGCN快5.97倍。
Information diffusion prediction is a fundamental task for understanding information spreading phenomenon. Many of the previous works use static social graph or cascade data for prediction. In contrast, a recently proposed deep leaning model DyHGCN [20] newly considers users’ dynamic preference by using dynamic graphs and achieve better performance. However, training phase of DyHGCN is computationally expensive due to the multiple graph convolution computations. Faster training is also important to reflect users’ dynamic preferences quickly. Therefore, we propose a novel graph convolutional network model with time-based mini-batch (GCNTM) to improve training speed while modeling users’ dynamic preference. Time-based mini-batch is a novel input form to handle dynamic graphs efficiently. Using this input, we reduce the graph convolution computation only once per mini-batch. The experimental results on three real-world datasets show that our model performs comparable results against baseline models. Moreover, our model learns about 5.97 times faster than DyHGCN.