Traffic prediction is vital to traffic planning, control, and optimization, which is necessary for intelligent traffic management. Existing methods mostly capture spatiotemporal correlations on a fine-grained traffic graph, which cannot make full use of cluster information in coarse-grained traffic graph. However, the flow variation of clusters in the coarse-grained traffic graph is more stable compared with nodes in the fine-grained traffic graph. And the flow variation of a fine-grained node is generally consistent with the trend of the cluster to which the node belongs. Thus information in the coarse-grained traffic graph can guide feature learning in the fine-grained traffic graph. To this end, we propose a Spatiotemporal Multiscale Graph Convolutional Network (SMGCN) that explores spatiotemporal correlations on a multiscale graph. Specifically, given a fine-grained traffic graph, we first generate a coarse-grained traffic graph by graph clustering, and extract spatiotemporal correlations on both fine-grained and coarse-grained traffic graphs. Then we propose a cross-scale fusion (CF) to implement information diffusion between the fine-grained and coarse-grained traffic graphs. Moreover, we employ an adaptive dynamic graph convolution network to mine both static and dynamic spatial features. We evaluate SMGCN on real-world datasets and obtain a $1.18\% -3.32\%$ improvement over state-of-the-arts.

As one of the most significant components of Intelligent Transportation Systems (ITS), traffic prediction has gained much popularity given its enormous application value in vehicular communications, traffic management, and traffic control. As such, many traffic prediction models have been proposed. The current methods commonly adopt graph convolution networks (GCNs) to capture spatial correlations. GCN-based approaches mainly focus on pair-wise interactions between road vertices (i.e. dyadic relations). However, the interactions between road vertices are not necessarily dyadic, but also can be high-order (i.e., multivariate relations). Further, few existing works focus on how to mine correlations between different types of traffic data. To this end, we develop STGHCN, a spatial-temporal gated hypergraph convolution network, that not only captures pair-wise and high-order spatial patterns between vertices but also characterizes the type-aware traffic data influences. Specifically, we adopt a spatial gated graph convolution and a hypergraph convolution to explore pair-wise and high-order interactions. More importantly, we design a spatial-temporal aware channel attention mechanism to mine hidden patterns between cross-type traffic features. Extensive experiments conducted on four real-world traffic datasets validate the effectiveness of STGHCN, with STGHCN's MAE, RMSE, and MAPE at between 6.91%-10.00%, 5.42%-8.90%, and 2.54%-5.03% lower than the state-of-the-art methods.

Cellular traffic prediction provides significant support for the management of intelligent networks. Existing models commonly combine recurrent neural networks (RNNs) with attention mechanisms, convolutional neural networks (CNNs), or graph convolutional networks (GCNs) to capture spatial-temporal correlations of cellular traffic. However, attention mechanisms lack sensitivity to local information; CNNs ignore the interaction among distant regions with similar semantics; GCNs exhibit limitations in exploring high-order (beyond pairwise) spatial correlations. To this end, we develop a hypergraph attention recurrent network (HARN) that exploits locality, semantics, and high-order correlations for cellular traffic prediction. Specifically, we first propose a spatial trend-aware attention to perceive local trends, thus easing the mismatching problem of attention mechanisms. Then, we construct a hypergraph to characterize the interactions between distant regions with similar semantics, and leverage a hypergraph convolution network to extract high-order correlations. More importantly, to extract heterogeneous and varying spatial patterns, we further enhance the hypergraph convolution network by incorporating spatial-temporal representations. Last, extensive experiments on three real-world datasets demonstrate the superiority of HARN over state-of-the-art baselines in terms of mean absolute error and root mean square error, with specific improvements of 1.83% and 5.79% on SMS (short message service) dataset, 3.05% and 11.27% on Call dataset, and 1.36% and 1.65% on Internet dataset, respectively.

Devices and servers in Federated Edge Learning (FEL) are self-interested and resource-constrained, making it critical to design incentives to improve model performance. However, dynamic network conditions raise energy consumption, while data heterogeneity undermines device cooperation. Current research overlooks the interplay between system efficiency and device clustering, resulting in suboptimal updates. To address these challenges, we develop BENCH, a bilateral pricing mechanism consisting of three core rules aimed at incentivizing participation from both devices and servers. Specifically, we first design a reward allocation rule, based on the Rubinstein bargaining model, which dynamically allocates rewards. Theoretically, we derive a closed-form solution for this rule, demonstrating BENCH achieves Nash equilibrium. Secondly, we design a device partitioning rule that leverages modularity to group similar devices, facilitating personalized edge aggregation to accelerate local data adaptation. Thirdly, we design an edge matching rule that employs the Kuhn-Munkres algorithm to balance the load at edge servers, thus minimizing the congestion. Together, these three rules enable hierarchical optimization of pricing and associations, effectively mitigating the impact of dynamic costs and device heterogeneity. Extensive experiments demonstrate BENCH’s effectiveness in increasing device participation by 28.81% and improving model performance by 2.66% compared to state-of-the-art baselines.

Although Federated Learning (FL) enables collaborative training across multiple decentralized devices, strategic clients may be reluctant to participate in FL unless sufficient incentives are available. Existing researches predominantly emphasize short-term incentives, while FL model training is typically a long-term process, and malicious clients may exhibit dishonest behavior during local training. Although designing a long-term incentive mechanism is crucial for FL, this task is challenging due to the heterogeneous nature of FL and the limitations of imperfect system monitoring. To this end, this paper designs the first sustainable incentive mechanism for FL called RATE, which aims to incentivize clients to continuously contribute more high-quality data. Specifically, by modeling the competition and cooperation relationship between servers and clients as a multi-server multi-client Stackelberg game, we prove the existence and uniqueness of the Stackelberg equilibrium (SE), and derive the unique SE through the cautious calculation of designed algorithms. Since the derived SE may not be optimal, we further utilize reputation to measure the long-term contribution of clients and build a function between their revenues and reputations to ensure the optimal revenue allocation, thereby maximzing the social welfare of RATE. Extensive experiments on both synthetic and real datasets demonstrate the superiority of RATE. Compared to the state-of-the-art baselines, RATE increases the total utility of the servers by up to 20%, reduces the reputation of malicious clients by up to 90%, and improves the average test accuracy overall.

Although Federated Learning (FL) offers significant potential for developing model marketplaces through collaborative training and privacy preservation, challenges such as insufficient training data and arbitrage issues severely impede the development of FL-based model marketplaces. Existing studies either lack satisfactory security guarantees or are too profit-driven to address potential arbitrage issues. In this paper, we propose a novel Personalized prIvacy-prEserving inCentive mEchanism named PIECE, with the aim of achieving social optimality while avoiding arbitrage. Specifically, we first formulate a dual-objective optimization problem to simultaneously maximize social utility and model performance while ensuring arbitrage-free conditions through differential privacy. Due to dynamic model training and heterogeneous privacy budgets that complicate the design of arbitrage-free properties, we model the transformation between local and global privacy requirements across scenarios as a privacy choice game. This game guarantees the identification of a constraint to generate desired model versions based on Nash equilibrium. Next, by generalizing the properties of different data-owner groups under equilibrium conditions, we prove that the dual-objective optimization problem is always conflict-free, thus allowing transformation into a social optimal problem without arbitrage. Furthermore, to tackle the significant difficulty in characterizing the model revenue and interpolating pricing, we propose a two-stage solution based on subadditivity relaxation. The first stage establishes a set of ideal prices as the target, while the second stage establishes polynomial-time solvability and provides rigorous arbitrage-free boundaries. Finally, comprehensive experiments on four real-world datasets validate the efficacy of PIECE. The results indicate a minimum 8% boost in model revenue within the specified marketplace scale, and a maximum 16.67% improvement in model performance compared to the state-of-the-art baselines.