This paper addresses radio resource management (RRM) in highly dynamic device-to-device (D2D) networks. Existing heuristic and deep learning (DL) approaches often overlook the network’s time-varying nature and struggle with variable link counts and channel conditions. We propose a Continuous-Time Dynamic Graph (CTDG) model that captures network events (activations, updates, and deactivations) in real time, rather than relying on discrete snapshots. Our Temporal Graph Neural Network (TGNN) processes these events, updating node-wise and graph-wise memories to track historical context. The resulting temporal embeddings drive power and channel allocation decisions that adapt to changing network topologies and mobility. We evaluate this TGNN-based solution in a realistic D2D setting using mobility traces from SUMO. Results show near-optimal throughput under stringent constraints and significant performance gains over conventional DL networks and memoryless GNN-based methods. This work underscores the importance of continuous-time graph modeling for scalable, efficient RRM in next-generation wireless systems.

In this paper, we address the challenge of power control and spectrum allocation in Device-to-Device (D2D) networks with the aim to maximize the long-term average network rate. The inherent time-varying nature of wireless channels, caused by user mobility, complicates radio resource management (RRM) in such networks. address this concern, in this study, we propose a novel two-level solution that integrates Channel State Information (CSI) prediction with a Graph Neural Network (GNN)-based RRM model. On the first level, we develop an attention-based recurrent neural network (RNN) to predict future CSIs using historical data. On the second level, we transform the past, current, and predicted CSIs into multiple graph structures. We then design a GNN-based RRM model that captures the geometric properties of the CSIs across the following three dimensions: within resource blocks (RBs), between RBs, and over time (both forward and backward). Specifically, the proposed model employs message passing within RBs to capture local interference patterns, between RBs to optimize spectral resource allocation, and across temporal states to use temporal dependencies. The results of our simulation demonstrate that our method outperforms benchmark schemes across various wireless scenarios, achieving a higher network throughput while satisfying Quality of Service (QoS) constraints.

This research paper introduces an unsupervised scalable probabilistic approach for radio resource management in device-to-device (D2D) communication networks, essential for enhancing wireless data service capacity. We propose a joint optimization framework for spectrum allocation and power control, aiming to optimize the network’s mean rate while meeting minimum data rate requirements. Although deep learning (DL) models have been explored for this purpose, their scalability is constrained by the fixed sizes of their input/output features, and their effectiveness is often limited by an insufficient understanding of the network’s geometric structure. Consequently, Graph Neural Networks (GNNs) were introduced to integrate the wireless network’s topology into the learning process. However, GNNs typically lose spatial correlation data when converting the tensorized channel state information (CSI) into a graph structure. To overcome this limitation, our solution combines GNNs, convolutional neural networks (CNNs), and variational autoencoders to extract meaningful embeddings from the CSI, preserving spatial and geometric features. We also introduce an innovative graph attention mechanism that enhances the model’s focus on crucial node and edge features. Our holistic approach exploits the wireless network’s topological and spatial relationships, offering a scalable, unsupervised, and generalizable solution without the need for retraining or architectural adjustments across various wireless setups. Our findings confirm our method’s superior performance and adaptability to different wireless environments.

Tuning big data frameworks is a very important task to get the best performance for a given application. However, these frameworks are rarely used individually, they generally constitute a pipeline, each having a different role. This makes tuning big data pipelines an important yet difficult task given the size of the search space. Moreover, we have to consider the interaction between these frameworks when tuning the configuration parameters of the big data pipeline. A trade-off is then required to achieve the best end-to-end performance. Machine learning based methods have shown great success in automatic tuning systems, but they rely on a large number of high quality learning examples that are rather difficult to obtain. In this context, we propose to use a deep reinforcement learning algorithm, namely Twin Delayed Deep Deterministic Policy Gradient, TD3, to tune a fraud detection big data pipeline. We show through the conducted experiments that the TD3 agent improves the overall performance of the pipeline by up to 63% with only 200 training steps, outperforming the random search on the high-dimensional search space.