NCSGOI: A Node Classification Method for a Streaming Graph Using Lightweight Variable Graph Neural Network

With the rapid development of mobile internet in recent years, a large scale of continuous arrival correlative data, namely dynamic streaming graph, are extensively generated in various application fields. Analyzing, mining and making good use of such a streaming graph face great challenges. In dynamic graph analytics and identification, node classification of graph is one of the most important works and has become a hot research filed. It has a wide range of application scenarios, such as malicious users detection in social networks and shopping recommendation in E-commerce. Due to the large scale and variable structure of the streaming graph, node classification become very difficult. To this end, this paper proposes a node classification method for a streaming graph based on a lightweight variable graph neural network, named NCSGOI. We use the time window mode for processing a streaming graph in batches. In the first time window, a graph neural network pre-training model is built and trained by contrastive learning. So, the optimal hyper-parameters and the seed nodes with labels can be obtained, and the cold start problem can be solved in this way. In subsequent time windows, the optimal hyper-parameters and classification results obtained in each previous time window are reused. Then, the reconstructed model is fine-tuned and further optimized, which can avoid a lot of repetitive training. Meanwhile, using the principle of “firefly light intensity variation”, we obtain a small-scale subgraph locally affected by newly added nodes. Treating such subgraph as the input graph for a lightweight reconstructed model, the size of the training graph neural network can be reduced, and the process of node classification can be accelerated. Extensive experiments indicate that our algorithm NCSGOI greatly improves the accuracy of node classification compared to two traditional algorithms. Even compared to the state-of-the-art node classification algorithm DynGNN, NCSGOI shows a better accuracy and faster speed in node classification.