Time series data plays a pivotal role in a wide variety of fields but faces challenges related to privacy concerns. Recently, synthesizing data via diffusion models is viewed as a promising solution. However, existing methods still struggle to capture long-range temporal dependencies and complex channel interrelations. In this research, we aim to utilize the sequence modeling capability of a State Space Model called Mamba to extend its applicability to time series data generation. We firstly analyze the core limitations in State Space Model, namely the lack of consideration for correlated temporal lag and channel permutation. Building upon the insight, we propose Lag Fusion Mamba and Permutation Scanning Mamba, which enhance the model's ability to discern significant patterns during the denoising process. Theoretical analysis reveals that both variants exhibit a unified matrix multiplication framework with the original Mamba, offering a deeper understanding of our method. Finally, we integrate two variants and introduce Diffusion Mamba for Time Series (DiM-TS), a high-quality time series generation model that better preserves the temporal periodicity and inter-channel correlations. Comprehensive experiments on public datasets demonstrate the superiority of DiM-TS in generating realistic time series while preserving diverse properties of data.

Next POI recommendation plays a crucial role in delivering personalized location-based services, but it faces significant challenges in capturing complex user behavior and adapting to dynamic interest distributions. Most methods often provide insufficient modeling of implicit features in user trajectories, such as directional transitions and latent edge relationships, which are essential for understanding user behavior. Moreover, existing diffusion models, constrained by Gaussian priors, struggle to handle the diverse and evolving nature of user preferences. The lack of a unified scheduling for noise and sampling also limits the flexibility of diffusion models. In this paper, we propose a Unified Bridge-based Diffusion model (UB-Diff) for the next POI recommendation. UB-Diff incorporates a direction-aware POI transition graph learning, which jointly captures spatio-temporal and directional features. To overcome the limitations of Gaussian priors, we introduce a bridge-based diffusion POI generative model. It can achieve distribution translation from the user's historical distribution to the target distribution by learning a bridge to associate user behavior with POI recommendation, adapting to dynamic user interests. In the end, we design a novel intermediate function to unify the diffusion process, enabling precise control over noise scheduling and modular optimization. Extensive experiments on five real-world datasets demonstrate the superiority of UB-Diff over advanced baseline methods. Our code is available athttps://github.com/JKZuo/UBDiff.

The contemporary urban intelligent transportation system (ITS) generates an enormous amount of trajectory data daily, serving as an essential reflection of traffic dynamics. Accurate estimation of arrival time by mining spatio-temporal features and semantic relationships from historical trajectories has become increasingly vital. However, most existing works overlook the joint features between links (i.e., road segments) and crossroads in trajectories. Additionally, they often treat all links uniformly without considering the semantics of critical links, leading to deficiencies in captured representation. To address these issues, this study proposes a novel deep encoder learning framework called the Triple Feature Encoder-based Dual-Granularity Graph Learning Network (TriDGNet) for enhanced travel time estimation. Specifically, we design a triple feature learning encoder to explore the spatio-temporal correlations of trajectories from three perspectives: Depth, Ensemble, and Sequence. Meanwhile, we introduce a consistent modeling method to integrate both links and crossroads. Furthermore, we construct two graph learning modules at different scales. One is an edge-enhanced graph attention network (E-GAT) to capture global spatial dependencies across the entire road network. The other is a backtracking-based subgraph representation network (BackNet) to learn local contextual information from bustling links. Our proposed TriDGNet model has been evaluated on three extensive datasets. The experimental results demonstrate that it outperforms state-of-the-art approaches.

Effective Point-of-Interest (POI) recommendation systems play a pivotal role in modern location-aware applications and human mobility, facilitating customized suggestions for users’ upcoming exploration destinations. Understanding the intricate dynamics of user movement, which are often influenced by a multitude of factors, remains a formidable task. Moreover, discrepancies between the acquired representation distribution and the authentic target distribution of user interests also present a notable obstacle. To tackle these problems, we make an attempt to bridge the gap by introducing diffusion models and propose a diffusion-based dual graph multiattention network (Diff-DGMN). Specifically, we have constructed two types of graphs: one is a user-oriented local POI transition graph and the other is a global-based POI distance graph. Subsequently, we put forward two graph learning representation modules to capture the sequential encoding of users and the geographic representations of nodes, respectively. Furthermore, an attention-based location prototype generation module is introduced to merge the captured sequential encoding and geographic representation, yielding richer semantic interaction features. In the end, we obtain the final results by leveraging the forward diffusion process and corresponding its reverse-time generation to sample users’ future preferences from the posterior distribution. Our Diff-DGMN model demonstrates its remarkable recommendation performance through extensive experimentation on five real-world data sets. Compared with the most state-of-the-art methodologies, Diff-DGMN has improved performance in accuracy, normalized discounted cumulative gain (NDCG), and mean reciprocal rank (MRR) by 8.04%, 8.63%, and 9.09%, respectively. Our codes are available athttps://github.com/JKZuo/Diff-DGMN.

In recent years, the travel time prediction has been receiving sustained attention because of the prevalence of location based applications, smart city engineering and online car-hailing. In this paper, a Self-organizing Graph Embedding Deep Network (SGED-Net) model is proposed to address the challenging Origin-Destination(OD) based travel time estimation (TTE). Specifically, SGED-Net comprises four modules, involving travel feature extraction, spatial association graph generation, graph embedding, and travel time deep learning prediction modules. First, we comprehensively extract three travel characteristics and feed them into a modified LightGBM component for learning the importance sort of features for different trips. Second, considering the lack of intermediate trajectories in the OD-based TTE, we employ a self-organizing feature mapping (SOM) approach to obtain the prominent nodes among pick-up and drop-off locations. Meanwhile clustering algorithm is used to divide a city into n-clusters districts of variant sizes. The topology learned by SOM is combined with the districts divided by clustering algorithm to obtain a district-based spatial association graph. Moreover, an improved SDNE algorithm is leveraged to gain a low-dimension spatial association representation while preserving the global and local structure of the graph. Then, we design a deep neural network for learning the captured spatial association representation. Finally, a series of experiments in two real-world large-scale datasets demonstrate the SGED-Net's excellent performance.

As an important information fusion method in video surveillance forensics, infrared and visible image fusion has received extensive attention and research. In order to reduce the sampling number of the original video surveillance image information in the fusion process, improve the processing efficiency and the background quality of the fused image, a frequency domain fusion algorithm of infrared and visible image based on compressed sensing for video surveillance forensics is proposed. Firstly, the compressed sensing is used for sampling, and then the subspace pursuit (SP) algorithm is applied to reconstruct the sparse coefficient. Secondly, the sparse coefficient is transformed in frequency domain, and the fusion is completed in frequency domain, and finally obtains the fused image. A large number of experiments show when the number of compressed sensing samples is 70%, the fusion quality is better than the other four comparison algorithms that take advantage of full sampling. Compared with similar algorithms, this algorithm has obvious advantages in the presentation of image background information and gradient structure information.