20 March 2023 in ACM Transactions on Knowledge Discovery From Data
With the success of Graph Neural Network (GNN) in network data, some GNN-based representation learning methods for networks have emerged recently. Variational Graph Autoencoder (VGAE) is a basic GNN framework for network representation. Its purpose is to well preserve the topology and node attribute information of the network to learn node representation, but it only reconstructs network topology, and doesn’t consider the reconstruction of node features. This strategy will make node representation can not well reserve node features information, impairing the ability of the VGAE method to learn higher quality representations. To solve this problem, we arise a new network representation method to improve the VGAE method for well retaining both node features and network structure information. The method utilizes adversarial mutual information learning to maximize the mutual information (MI) of node features and node representations during the encoding process of the variational autoencoder, which forces the variational encoder to get the representation containing the most informative node features. The method consists of three parts: a variational graph autoencoder includes a variational encoder(MI generator (G)) and a decoder, a positive MI sample module (maximizing MI module), and an MI discriminator(D). Furthermore, we explain why maximizing MI between node features and node representation can reconstruct node attributes. Finally, we conduct experiments on seven public representative datasets for nodes classification, nodes clustering, and graph visualization tasks. Experimental results demonstrate that the proposed algorithm significantly outperforms current popular network representation algorithms on these tasks. The best improvement is \(17.13\% \) than the VGAE method.
21 August 2025 in ACM Transactions on Knowledge Discovery From Data
Knowledge Graphs (KGs) are extensively used in recommendation systems and information retrieval but often suffer from incompleteness. A popular solution to this problem is multi-hop inference through a reinforcement learning framework, which provides an interpretable path for predicting missing links in KGs. Most previous work focuses on improving the performance of multi-hop link prediction. However, it has been observed that many multi-hop paths generated by these methods are irrational; they often fail to reasonably explain the predicted answer entities. To address this challenge, we introduce the Joint Multi-hop Link Prediction (JMLP) framework. The framework consists of a relation attention network and an entity attention network, which collaboratively generate the reasoning paths. The relation attention module utilizes an induction network to encode historical paths and employs the graph self-attention mechanism to refine the interaction of relation contextual information. The entity attention module uses the graph attention mechanism to obtain the aggregated contextual features and leverages self-attention to strengthen the correlation between local and global contextual entity features. Extensive experiments on five datasets validate the effectiveness of our approach, demonstrating significant improvements both in predictive performance and interpretability compared to state-of-the-art methods.
9 May 2025 in ACM Transactions on Knowledge Discovery From Data
Representation learning on networks, mapping the network into a low-dimensional vector space, has received signification attention recently due to its widespread application in graph data mining tasks. With the success of representation learning in static networks, we push further for practical scenarios of dynamic networks. Existing methods model the dynamic network by dividing the dynamic network into sequences of network snapshots, see each network snapshot as a static network, and utilize the dynamic evolution between snapshots, they capture the discrete dynamic evolution of dynamic networks. However, a dynamic network continuously evolves over time. Capturing the continuously dynamic evolution of dynamic networks is important for dynamic network representation. In this article, we regard a dynamic network as a dynamic system, use the ordinary differential equation (ODE) to model the dynamic evolution of dynamic networks, and integrate the ODE over continuous-time to capture continuously dynamic evolution of dynamic networks; and design a new encoder-decoder model for dynamic networks representation. We improve the gated recurrent unit (GRU) module (only capturing the discrete dynamic evolution of the dynamic network and structure information) by combining an ODE and a GRU. The improved GRU as the encoder can learn the continuously dynamic evolution, and structure information of the dynamic network, where the ODE parameterized by a graph neural network models the continuously dynamic evolution of each network snapshot. Use the ODE and Inner-Productor as the decoder, where the ODE is integrated over continuous-time to learn the continuous dynamic evolution of the latent representation of the whole dynamic network, and the Inner-Productor reconstructs the topological structure of each snapshot by doing the inner-product between nodes representation, the reconstructing errors as the objective function of our method. To assess our model, we expand the experiment on several real-world dynamic networks, and results show that our method consistently outperforms existing baselines in three dynamic link prediction tasks; the best is up to 6.54 \( % \) improvement. To our knowledge, our method is the first work using the ODE to capture the continuously dynamic evolution of dynamic networks.