GraphSense: a self-aware dynamic graph learning networks for graph data over internet

Abstract

Dynamic graph data learning is an important data analysis technique. In the age of big data, the volume of data produced daily is immense, the data types are varied, the value density is low, and the data continues to accumulate over time. These characteristics make data processing more challenging. In particular, unstructured data, unlike structured data, does not have a fixed format, and its volume is large and variable, which presents a significant challenge to traditional data processing techniques. Nowadays, researchers have been employing graph neural network models to analyze unstructured data. However, real-world graph structures are dynamic and time-varying, and the static graph neural network cannot effectively learn graph node embeddings and network structures. To address the challenges mentioned above, we propose a self-aware dynamic graph network structure learning model, called GraphSense. The algorithm consists of two modules: self-sensing neighborhood aggregation algorithm and dynamic graph structure learning algorithm based on RNN. GraphSense can make each node discover more valuable neighbors through the self-aware neighborhood aggregation algorithm in each epoch. The algorithm employs gated recurrent unit to dynamically aggregate the information of node neighbors to learn the high-order information. Next, in order to capture the temporal properties of graph structures, we employ dynamic graph structure learning algorithm based on RNN to replicate the time evolution process of dynamic graphs. Finally, we evaluate the performance of GraphSense on four publicly available datasets by two specific tasks(edge and node classification). The experimental results show that the proposed GraphSense model outperforms the baseline model by 2.0% to 25.0% on the Elliptic dataset, 2.5% to 27.0% on the Bitcoin-alpha dataset, 3.0% to 31.0% on the Bitcoin-otc dataset, and 0.9% to 26.0% on the Reddit dataset in terms of F1 scores. The results suggest that our model is effective in learning from dynamic graph data.