Implicit neural representations (INRs) use neural networks to provide continuous and resolution-independent representations of complex signals with a small number of parameters. However, existing INR models often fail to capture important frequency components specific to each task. To address this issue, in this paper, we propose a Fourier Kolmogorov-Arnold network (FKAN) for INRs. The proposed FKAN utilizes learnable activation functions modeled as Fourier series in the first layer to effectively control and learn the task-specific frequency components. The activation functions with learnable Fourier coefficients improve the ability of the network to capture complex patterns and details, which is beneficial for high-resolution and high-dimensional data. Experimental results show that our proposed FKAN model outperforms four state-of-the-art baseline schemes, and improves the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) for the image representation task and intersection over union (loU) for the 3D occupancy volume representation task, respectively. The code is available at github.com/Ali-Meh619/FKAN.

Stock markets play an important role in the global economy, where accurate stock price predictions can lead to significant financial returns. While existing transformer-based models have outperformed long short-term memory (LSTM) networks and convolutional neural networks (CNNs) in financial time series prediction, their high computational complexity and memory requirements limit their practicality for real-time trading and long-sequence data processing. To address these challenges, in this paper, we propose SAMBA, an innovative framework for stock return prediction that builds on the Mamba architecture and integrates graph neural networks (GNNs). SAMBA achieves near-linear computational complexity by utilizing a bidirectional Mamba block to capture long-term dependencies in historical price data and employing adaptive graph convolution to model dependencies between daily stock features. Our experimental results demonstrate that SAMBA significantly outperforms state-of-the-art baseline models in prediction performance, maintaining low computational complexity. The code and datasets are available at github.com/Ali-Meh619/SAMBA.

Terahertz (THz) wireless systems aim to support content-rich applications with ultra-high data rate. Due to high molecular absorption, THz signals experience severe path loss. Adaptive sub-band bandwidth (ASB) allocation can mitigate absorption attenuation by allocating THz sub-bands with variable bandwidth to the users. However, in ASB allocation, since the bandwidth of sub-bands may not be known a priori, accurate channel estimation is challenging. To overcome this issue, in this paper, we propose a heterogeneous graph-transformer network (HGTN) to bypass the channel estimation phase. We formulate a sum-rate maximization problem with quality-of-service (QoS) constraints in a multiuser multiple-input multiple-output (MU-MIMO) THz system to optimize the precoding and ASB allocation. The proposed HGTN parameterizes the mapping from input features (e.g., location information, users' minimum data rate) to the optimized system parameters via unsupervised learning. The proposed HGTN can be applied to systems with different number of users once it is trained. Simulation results show that our proposed HGTN achieves a higher system sum-rate with faster convergence when compared with the unsupervised deep neural network learning algorithm.

Terahertz (THz) wireless systems aim to support content-rich applications with ultra-high data rate. Due to high molecular absorption, THz signals experience severe path loss over long distance. To alleviate distance limitation, reconfigurable intelligent surface (RIS) can improve the coverage range. Adaptive sub-band bandwidth (ASB) allocation can mitigate absorption attenuation by allocating THz sub-bands with variable bandwidth to the users. However, in ASB allocation, since the bandwidth of sub-bands may not be knowna priori, accurate channel estimation is challenging. To overcome this issue, in this paper, we propose a metapath-based heterogeneous graph-transformer network (MHGphormer) to bypass the channel estimation phase. We formulate a sum-rate maximization problem with quality-of-service (QoS) constraints in a RIS-aided multiuser multiple-input multiple-output (MU-MIMO) THz system to optimize the precoding, phase shifts, and ASB allocation. The proposed MHGphormer parameterizes the mapping from input (e.g., location information, users’ minimum data rate) to the optimized system parameters via unsupervised learning. The proposed MHGphormer has the permutation invariance/equivariance property. It can be applied to systems with different number of users. Simulation results show that our proposed MHGphormer achieves a higher system sum-rate when compared with the homogeneous graph neural network, unsupervised deep neural network, and alternating optimization baseline algorithms.

Predictive analysis of wireless cellular traffic plays an important role in network resources provisioning. Accurate traffic prediction is a challenging task due to the dynamic spatial-temporal nature of wireless traffic. Most of the existing approaches do not consider spectral domain information for wireless traffic prediction. Some of the approaches cannot capture the spatial dependencies between neighbouring and distant cells. In this paper, we propose a dynamic Bernstein graph recurrent network for traffic prediction in wireless cellular networks. First, we design a spectral dynamic graph construction (SDGC) method to model the spatial dependencies between cells as a dependency graph in a data-driven fashion. A dynamic Bernstein polynomial filtering (DBPF) scheme based on the$K$-order Bernstein polynomial approximation is then developed to capture the spatial correlations and infer the cell-specific parameters. To predict the spatial-temporal traffic demands, we propose a dynamic Bernstein graph recurrent network (DBGRN), which integrates the proposed DBPF module with a gated recurrent unit (GRU) network. We evaluate the performance of our proposed model using a real-world dataset. Results show that our proposed model outperforms four state-of-the-art baseline schemes, and achieves up to 8% and 10% improvements in terms of the root mean squared error (RMSE) and mean absolute error (MAE), respectively.

Execution traces can be overwhelmingly large. To reduce their size, sampling techniques, especially the ones based on random sampling, have been extensively used. Random sampling, however, may result in samples that are not representative of the original trace. We propose a trace sampling framework based on stratified sampling that not only reduces the size of a trace but also results in a sample that is representative of the original trace by ensuring that the desired characteristics of an execution are distributed similarly in both the sampled and the original trace.