Graph Neural Network And Temporal Information Based Dynamic Network Representation Learning

In the era of big data,everyone has become a content producer.How to effectively use the massive amounts of browsing,comment,text and other useful data has become an issue of great concern to industry and academia.By forming them as networks,the connections between individuals can be visualized intuitively,such as social networks,text networks,traffic networks and so on.Analyzing these networks with some new technology to explore the individual feature has broad prospects like personalized recommendations,traffic prediction,etc.However,in real life,the nodes and edges in the network usually change continuously over time.Making full use of the temporal properties can achieve high-quality node feature and network information.The current network research is inspired by the field of natural language processing.It uses emerging technologies,such as deep learning,matrix decomposition,etc.,to map the highdimensional sparse structure to the low-dimensional vector space,which can be applied to node classification,link prediction,and node clustering subsequently.However,most of the current researches are mainly focusing on static networks and do not take into account the structural changes in the network.Our paper focus on studying dynamic networks.In other words,the goal is exploring the historical information and embedding the time properties into the final node representation as much as possible to increase the robustness and predictive of the representations.Therefore,this paper proposes a dynamic network representation learning method based on graph neural network and a dynamic network representation learning method based on temporal information exploring.We show the main work as follows.Dynamic representation learning method based on graph neural network: This method divides the dynamic network into several snapshots according to the edge time,following by the graph neural networks to obtain the low-dimensional node representation at each snapshot.The paper proposed a new weighted aggregator to aggregate the neighbor information effectively.In order to reduce the time complexity,the graph network at each snapshot share the same weight matrix.Finally the model apply a LSTM layer to capture the evolution of the network.Experimental results on three data sets demonstrate that the proposed method achieves better performance than state-of-the-art methods.Dynamic representation learning method based on temporal information exploring: This method also divides the dynamic graph into multiple time periods,but limits the number of nodes of each subgraph to be consistent.In order to reduce the time and space complexity caused by high-dimensional sparse vectors,this method first uses a static method to initialize the node a structural feature,and then continuously aggregates the neighbor information over time.The paper proposed a new attention mechanism and a sampling strategy to embed the historical information into node representations more accurately.The LSTM layer will then be applied to model the contribution of each snapshot to final representations.The extensive experimental results demonstrate that the algorithm achieves significant improvements over several state-of-the-art baselines.