The growing demand for intelligent, adaptive resource management in next-generation wireless networks has underscored the importance of accurate and scalable wireless traffic prediction. While recent advancements in deep learning and foundation models such as large language models (LLMs) have demonstrated promising forecasting capabilities, they largely overlook the spatial dependencies inherent in city-scale traffic dynamics. In this paper, we propose TIDES (Traffic Intelligence with DeepSeek-Enhanced Spatial-temporal prediction), a novel LLM-based framework that captures spatial-temporal correlations for urban wireless traffic prediction. TIDES first identifies heterogeneous traffic patterns across regions through a clustering mechanism and trains personalized models for each region to balance generalization and specialization. To bridge the domain gap between numerical traffic data and language-based models, we introduce a prompt engineering scheme that embeds statistical traffic features as structured inputs. Furthermore, we design a DeepSeek module that enables spatial alignment via cross-domain attention, allowing the LLM to leverage information from spatially related regions. By fine-tuning only lightweight components while freezing core LLM layers, TIDES achieves efficient adaptation to domain-specific patterns without incurring excessive training overhead. Extensive experiments on real-world cellular traffic datasets demonstrate that TIDES significantly outperforms state-of-the-art baselines in both prediction accuracy and robustness. Our results indicate that integrating spatial awareness into LLM-based predictors is the key to unlocking scalable and intelligent network management in future 6G systems.

Due to the limited energy, communication bandwidth and computing ability of edge devices in Industrial Internet of Things (IIoT) networks, it is incredibly challenging to compress and transmit those massive manufacturing data collected at the edge, thus greatly degrading the transmission and computing efficiency and finally results in long latency. To address this, we propose a quantized deep compressed sensing network (QDCS-Net) for both linear and nonlinear measurements to help better compress the industrial data to reduce the transmission volume of data and achieve good reconstruction performance. The joint design of customized quantization layers, dual-path structures, and swish activation function in QDCS-Net is adopted to achieve high-precision data reconstruction at high compression ratios. The latency is analyzed for different transmission deployment schemes to get a better edge-cloud collaboration strategy. We evaluate QDCS-Net by using real-world datasets collected from a vibration signal acquisition system. Experimental results demonstrate that the proposed QDCS-Net performs better in recovering industrial signals even at extremely low compression ratios of 1/128, thus can effectively improve data reconstruction accuracy and communication efficiency.

Sensing human activity via WiFi Channel State Information (CSI) has considerable application prospects in future intelligent interaction scenarios such as virtual reality, intelligent games, metaverse, etc. Recently, many Deep Learning-based WiFi sensing schemes have been proposed in the literature, which gained high accuracy for a wide range of simple activities such as standing, squatting, and bending. However, the performance will be suffered when existing approaches are used to recognize drastic activities, such as actions in vigorous sports. This is mainly due to the reason that the spatiotemporal information of these actions is not well utilized. To overcome this drawback, we propose a novel DL-based WiFi sensing method for drastic activity recognition by combining the Convolutional Neural Network (CNN) and the Bidirectional Long Short-Term Memory (BiLSTM) network. The designed CNN-BiLSTM architecture is in parallel with feature extraction, which can simultaneously extract sufficient spatiotemporal features of action data and establish the mapping relationship between actions and CSI streams, thereby improving the accuracy of activity recognition. The CNN is used to extract information on the spatial dimension, while the BiLSTM extracts information on the time dimension. To verify the performance of the proposed scheme, we build a hardware experiment platform and constrain a dataset with 1400 pieces of records for 7 classes of basketball actions. After training over the dataset, the proposed CNN-BiLSTM scheme achieves 96% experimental accuracy on the test set, which is better than the benchmark methods.

Generative Adversarial Networks (GANs) have been used in the field of speech enhancement due to their huge potentials in reducing the noise mixed in the signals. Most of existing GAN-based speech enhancement approaches either operate on time domain or exploit the magnitude spectra in time-frequency domain, but lack consideration of direct optimization of the phase. In this paper, we propose a GAN architecture for speech enhancement based on gated linear units (GLUs) and Dual-Path Transformers (DPTs), which simultaneously deals with the amplitude and phase information on the time-frequency domain. The generator of the proposed GAN architecture is designed following an autoencoder structure fed by the real and imaginary parts of the time-frequency frames. The encoder of the generator is constructed by multiple cascaded convolutional GLUs (ConvGLUs), while the decoder consists of two groups of cascaded deconvolutional GLUs (DeconvGLUs), one for the real part of the spectrogram and the other for the imaginary part. The GLUs are adopted since they are potential in avoiding the gradient vanishing issue dwelling in deep architectures by providing a linear path for the gradients while retaining non-linear capabilities. Aiming at capturing the long-range dependent features in speech, we place DPTs between the encoder and the decoder of the generator, which contains multi-head attention modules and Bi-directional Gated Recurrent Units (BiGRUs). Moreover, the DPT structure is also merged with multiple one-dimensional convolutional layers in the discriminator of the GAN. Such a design not only improves the speech enhancement performance of GAN by focusing on multiple features of speech, but also reducing the volume of model parameters of GAN. Experimental results suggest that the proposed GAN architecture outperforms the existing benchmark GANs in terms of both objective speech intelligibility and quality with less computational complexity.

Wireless traffic prediction is essential for cellular networks to realize intelligent network operations, such as load-aware resource management and predictive control. Existing prediction approaches usually adopt centralized training architectures and require the transferring of huge amounts of traffic data, which may raise delay and privacy concerns for certain scenarios. In this work, we propose a novel wireless traffic prediction framework named Dual Attention-Based Federated Learning (FedDA), by which a high-quality prediction model is trained collaboratively by multiple edge clients. To simultaneously capture the various wireless traffic patterns and keep raw data locally, FedDA first groups the clients into different clusters by using a small augmentation dataset. Then, a quasi-global model is trained and shared among clients as prior knowledge, aiming to solve the statistical heterogeneity challenge confronted with federated learning. To construct the global model, a dual attention scheme is further proposed by aggregating the intra-and inter-cluster models, instead of simply averaging the weights of local models. We conduct extensive experiments on two real-world wireless traffic datasets and results show that FedDA outperforms state-of-the-art methods. The average mean squared error performance gains on the two datasets are up to 10% and 30%, respectively.

The controversy and argument on the usefulness of the physical layer (PHY) academic research for wireless communications are long-standing since the cellular communication paradigm gets to its maturity. In particular, researchers suspect that the performance improvement in cellular communications is primarily attributable to the increases in telecommunication infrastructure and radio spectrum instead of the PHY academic research, whereas concrete evidence is lacking. To respond to this controversy from an objective perspective, we employ econometric approaches to quantify the contributions of the PHY academic research and other performance determinants. Through empirical analysis and the quantitative evidence obtained, albeit preliminary, we shed light on the following issues: 1) what determines the cross-national differences in cellular network performance; 2) to what extent the PHY academic research and other factors affect cellular network performance; 3) what suggestions we can obtain from the data analysis for the stakeholders of the PHY research.

Machine (deep) learning-enabled accurate traffic modeling and prediction is an indispensable part for future big data-driven intelligent cellular networks, since it can help autonomic network control and management as well as service provisioning. Along this line, this paper proposes a novel deep learning architecture, namely Spatial–Temporal Cross-domain neural Network (STCNet), to effectively capture the complex patterns hidden in cellular data. By adopting a convolutional long short-term memory network as its subcomponent, STCNet has a strong ability in modeling spatial–temporal dependencies. Besides, three kinds of cross-domain datasets are actively collected and modeled by STCNet to capture the external factors that affect traffic generation. As diversity and similarity coexist among cellular traffic from different city functional zones, a clustering algorithm is put forward to segment city areas into different groups, and consequently, a successive inter-cluster transfer learning strategy is designed to enhance knowledge reuse. In addition, the knowledge transferring among different kinds of cellular traffic is also explored with the proposed STCNet model. The effectiveness of STCNet is validated through real-world cellular traffic datasets using three kinds of evaluation metrics. The experimental results demonstrate that STCNet outperforms the state-of-the-art algorithms. In particular, the transfer learning based on STCNet brings about 4%~13% extra performance improvements.