Tiantian He

Dear Editor, This letter presents a novel graph neural network, namely modularized graph convolution network (MGCN), to address the underexplored issue in graph convolution networks (GCNs), wherein the weights for neighbor aggregation are fixed, leading to the limited capability of capturing diverse relationships among nodes for representation learning. Conventional GCNs always learn node representations in the graph according to the weights computed from the graph Laplacian, consequently overlooking the similarity and group cohesiveness of node features. Motivated by this observation, the proposed MGCN is designed to learn expressive node representations by considering the following two aspects. First, we propose a novel Laplacian operator to generate weights for feature aggregation that can adaptively capture the feature similarity among nodes. Second, based on modularity optimization [1], we design a novel auxiliary loss that enables the weights for aggregating features to capture the cluster structure of the graph. The proposed MGCN, therefore, can learn expressive representations for each node in the graph by leveraging the modularized messages, i.e., predominately aggregating the features of neighbors in the same cluster. The proposed MGCN has been tested with several well-established datasets and compared with strong GCN-based models. The experimental results demonstrate that MGCN is more effective in various downstream tasks.

Graph contrastive learning (GCL) has emerged as a dominant paradigm for self‐supervised representation learning for attributed graph data. However, existing GCL methods heavily rely on empirical graph data augmentation, which may distort intrinsic graph semantics and produce poor generalisation without carefully chosen or designed augmentation techniques. Furthermore, most GCL approaches focus on same‐granularity contrastive learning (e.g., node vs. node), neglecting the hierarchical and multigranular properties inherent in real‐world networks, leading to suboptimal performance. To address these limitations, we propose HPoolGCL, a cross‐granularity GCL framework compatible with various hierarchical graph pooling methods to capture multigranularity information. Our framework eliminates the need for handcrafted augmentations, explicit negative sampling and complex multiencoder architectures by applying two novel loss functions in hierarchical graph pooling. The theoretical analysis is provided to explain the effectiveness of unified MGC and HiCR losses from three perspectives, namely, the information maximisation principle, the redundancy reduction principle and the information bottleneck principle. The experimental results demonstrate that HPoolGCL achieves state‐of‐the‐art performance across multiple downstream tasks on five benchmarks. Our codes are available at https://github.com/Heycen/HPoolGCL .

In multiagent reinforcement learning (MARL), agents engage in collaborative or competitive interactions, adapting their policies to maximize rewards. However, in environments with only team rewards, agents cannot accurately perceive the rewards obtained from their individual actions. To address this limitation, we propose a multiagent belief state computation method termed mutual awareness belief (MAB). MAB extends belief modeling to decentralized partially observable settings by incorporating agents' behavioral intentions and collaborative relationships. This allows each agent to infer the contribution of individual behaviors to team rewards during decision-making and supports team-advantageous joint decisions. In addition, to address the issue of information lag, we develop progressive graph construction (PGC), a novel module for computing MAB. PGC constructs a dynamic graph with progressively updated node features and cooperation-informed edge weights. Through interaction with PGC, agents iteratively update and retrieve graph information to derive the MAB. Empirical results show that our method achieves competitive performance against baseline approaches on the StarCraft unit micromanagement benchmark and proves effective in the nonmonotonic Pursuit environment.

Cryptocurrency services, as one of the most successful applications of blockchain technology, have recently garnered significant attention from the graph learning community. Its large-scale dynamic transaction records contain a variety of behavioral patterns and rich knowledge involving accounts, making the dynamic cryptocurrency transaction network embedding (DCTNE) a hot, yet thorny research topic. As the trading accounts increase and time accumulates, considerable transaction services are dispersed into various time slots, leading to very sparse transaction data within a time slot, that is, the transaction service data is high-dimensional and incomplete (HDI). To efficiently mine high-value knowledge from HDI data, this article proposes a proximal-ADMM-incorporated nonnegative latent-factorization-of-tensors (PNL) model for DCTNE that adopts threefold ideas: 1) incorporating the proximal terms into the alternating-direction-method-of-multipliers (ADMMs)-based learning scheme to reduce the oscillations for high estimation accuracy and fast convergence; 2) implementing a parallel training process with hyperparameter self-adaptation for high computational efficiency; and 3) proving that the proximal-incorporated learning scheme guarantees the convergence to a Karush–Kuhn–Tucker (KKT) stationary point. Experimental results on eight real-world DCTNs show that the PNL significantly outperforms several state-of-the-art (SOTA) models, demonstrating not only high efficiency and accuracy in performing DCTNE, but also strong potential to enhance the operational reliability and stability of cryptocurrency transaction systems.

In this article, we present the dynamic graph mixer (DGM), a novel model for learning spatiotemporal-individual coupled features from high-dimensional and incomplete (HDI) tensors, which frequently represent dynamic interactions among real-world data samples. In contrast to existing methods, the proposed DGM possesses the following three advantages when learning representations from HDI tensors. First, it performs light graph message passing based on the conjoint attentions learned by jointly modeling latent features and implicit structures to extract the high-order connectivity. Second, a multilayer nonlinear tensor neural network (TNN) is adopted to learn the intricate attribute features of node–node–time from different views. Third, it follows the Tucker decomposition paradigm in a data density-oriented modeling mechanism to integrate node representations, preserving the overall multidimensional interaction patterns. In addition, we provide theoretical evidence that the key components in DGM can significantly improve expressiveness. Extensive experiments conducted on eight testing datasets of HDI tensors demonstrate that DGM outperforms state-of-the-art methods in both learning accuracy and efficiency.