Accurate traffic forecasting plays a critical role in modern transportation systems by enabling effective congestion management and route optimization. Although recent deep learning-based models have shown substantial progress in modeling complex spatiotemporal dependencies, most existing methods overlook the challenges posed by anomalous traffic conditions. To address this gap, we propose a two-stage anomaly-aware forecasting correction framework. The first stage employs an unsupervised auto-regressive anomaly detector that combines representation learning and spatiotemporal attention to capture normal traffic patterns and filter anomalous inputs through error thresholding. In the second stage, an anomaly optimization predictor leverages a memory module and a sparsity regularization learning strategy to enhance the representation of normal patterns and suppress noise. It models spatiotemporal dependencies using an information propagation layer composed of sequential small-kernel temporal convolutions and memory-guided graph convolutions. Extensive experiments on real-world traffic flow datasets demonstrate that our framework outperforms state-of-the-art models under anomalous conditions.

Next Point of Interest (POI) recommendation primarily predicts future activities based on users’ past check-in data and current status, providing significant value to users and service providers. We observed that the popular check-in times for different POI categories vary. For example, coffee shops are crowded in the afternoon because people like to have coffee to refresh after meals, while bars are busy late at night. However, existing methods rarely explore the relationship between POI categories and time, which may result in the model being unable to fully learn users’ tendencies to visit certain POI categories at different times. Additionally, existing methods for modeling time information often convert it into time embeddings or calculate the time interval and incorporate it into the model, making it difficult to capture the continuity of time. Finally, during POI prediction, various weighting information is often ignored, such as the popularity of each POI, the transition relationships between POIs, and the distances between POIs, leading to suboptimal performance. To address these issues, this article proposes a novel next POI recommendation framework called Graph Disentangler with POI Weighted Module (GDPW) . This framework aims to jointly consider POI category information and multiple POI weighting factors. Specifically, the proposed GDPW learns category and time representations through the Global Category Graph and the Global Category-Time Graph. Then, we disentangle category and time information through contrastive learning. After prediction, the final POI recommendation for users is obtained by weighting the prediction results based on the transition weights and distance relationships between POIs. We conducted experiments on two real-world datasets, and the results demonstrate that the proposed GDPW outperforms other existing models, improving performance by 3% to 11%.

Spatio-temporal forecasting is critical for urban applications such as epidemic control and crime prevention, yet many existing methods assume dense and consistent sensor data, which is often unavailable due to infrastructural or cost constraints. This work explores the challenge of forecasting in sensor-less regions, where direct temporal observations are missing. Building on two prior studies: Multi-View Graph Fusion Approach with Approximation Module (MVGAM) for disease risk prediction and Graph Disentangler with POI Weighted Module (GDPW), a contrastive learning framework for enhancing POI embedding, we outline a new research direction. Our framework integrates large language models (LLMs), gated recurrent units (GRUs) and multi-layer perceptron (MLP) to encode multi-modal signals, with contrastive learning aligning heterogeneous representations. A dual-stage graph propagation mechanism consolidates knowledge in sensor-rich areas and transfer it to sensor-less regions via localized subgraphs. We anticipate using crime forecasting in Chicago as a case study, this work lays the foundation for robust and interpretable forecasting in data-scarce urban settings.

Dengue fever is an emergency disease spread by mosquitoes. The most direct way to prevent the disease is to predict risky areas and bolster mosquito preventive strategies. Risk is usually evaluated by monitoring the number of eggs in the ovitraps set up by the government. However, areas without sensors still need to be checked and managed for dengue risk. In this study, we focus on forecasting each region’s fine-grained dengue fever risk, especially in regions without sensor coverage. The paucity of historical data makes this endeavor challenging. Furthermore, determining how to effectively blend different features is another important research challenge and practical issue. We propose a Multi-View Graph Fusion Approach with Approximation Module (MVGAM) to address these two issues. For the regions that have no sensor coverage, MVGAM first uses a feature extractor to learn their representation based on their dynamic and static features. Then, we use a graph constructor to formulate the relationship between sensors from different perspectives and a multi-view graph fusion module to learn the embedding of sensors. Finally, we use an approximation module to deal with the lack of historical data. We conducted experiments using a real-world dataset from the urban area of Tainan, Taiwan. The results show that the proposed MVGAM outperforms the state-of-the-art methods and baselines. The ablation study also shows that every component in MVGAM has a significant impact on boosting the prediction effectiveness.

We present LINKin-PARK, an innovative system that seamlessly merges geographic visualization with an advanced Dual Attention Double Channel Convolutional Neural Network with Multilayer Perceptron (Dual Attention-DCCNN+MLP) to facilitate the efficient analysis of land valuation. LINKin-PARK provides robust visualization capabilities for intuitive comprehension. Our model outperforms traditional methods, e.g., linear regression, multilayer perceptron (MLP), Extreme Gradient Boosting (XGBoost), and the combination of CNN (Convolutional Neural Network) with MLP. An ablation study further evaluates the influence of specific components within the model, revealing that spatial and channel-wise attention mechanisms and the integration of DCCNN and skip connections are crucial for capturing spatial details and improving prediction accuracy. Users have the flexibility to explore and predict developable land valuation based on their specific requirements and provide their feedback to minimize errors in model prediction. For instance, this system can forecast future development potential and market demand for everywhere in an urban space, enabling users to make informed decisions before purchasing a property. Similarly, retailers can anticipate future revenues to aid in strategic decisions, such as selecting optimal locations for establishing new retail outlets. In summary, LINKin-PARK effectively combines geographic visualization and Dual Attention-DCCNN+MLP to assist users in analyzing and predicting land valuation and other scenarios.

We introduce ACCEPT, a geospatial decision support system that merges robust, intuitive visualization with grid-based data processing and neural networks to enhance spatial data analysis and interpretation in context-sensitive scenarios. It offers versatile machine learning modules with multiple prediction models, tailored to specific requirements with user-defined configurable parameters and flexible predictive target selection. The system serves as an accessible introduction to geographic information systems (GIS) for the general public. The system maps Points of Interest (POIs) to grids, simplifying processes like weighting, intersection, and interpolation, enhancing data accessibility and manipulation. Our case studies show effective handling of spatial data, reflecting similar distribution patterns of POIs, spatial separation, local feature sensitivity, and proximity to infrastructure and kernel size affect evaluations. The extensible and user-friendly web interface includes geospatial data inquiries, overlay, import/export, statistic, and multiple map views, facilitating informed decisions in resource distribution and urban planning. It supports urban planners, analysts, and policymakers in achieving equitable resource distribution and enhancing residential justice, while also providing non-experts an introduction to advanced geospatial analyses, promoting wider engagement and understanding in spatial decision-making.

Parking behaviors near popular destinations often exhibit a preference for proximity, resulting in poor habits, limited parking availability, and a range of consequential issues such as traffic chaos, economic challenges due to congestion, and imbalanced parking utilization. Taiwan has also faced escalating challenges in this regard. To effectively address these issues, the Government of Taiwan has initiated the Smart City program, encompassing various initiatives to enhance urban functionality. One notable solution implemented under this program is the Smart Parking Meter System (SPMS), designed to enhance the overall parking experience. The SPMS incorporates intelligent billing and secure parking data transmission, ensuring a safer and improved parking environment. In this paper, we propose ParkFlow, a comprehensive software-based solution that seamlessly integrates with smart parking hardware, presenting a holistic approach to tackling these challenges. ParkFlow intelligently disperses parking shortages in highly frequented areas and addresses the problem from multiple perspectives, including user, engineering, and government scenarios. By exploring and addressing these scenarios, we aim to provide valuable insights and inspiration to regions worldwide grappling with similar parking-related difficulties. Based on historical data analysis, the implementation of ParkFlow in resolving the parking imbalance problem is anticipated to lead to a significant increase of up to 10% to 20% in available parking hours in popular areas of Tainan, Taiwan. ParkFlow is in the process of being integrated into the Tainan City Government's Parking application, indicating its potential to address real-world parking challenges.

Dengue fever is an emergency disease spread by mosquitoes. The most direct way to prevent the disease is to predict risky areas and increase mosquito preventive strategies. Risk is evaluated by monitoring the sensors set up by the government. However, areas without sensors still need to be managed for dengue risk. In this study, we focus on forecasting each region's fine-grained dengue fever risk, especially in regions without sensor coverage. The lack of historical data makes this endeavor challenging. Furthermore, determining how to effectively blend different features is another challenge. We propose a Multi-View Graph Fusion Recurrent Neural Network (MVGFRNN), which consists of a multi-view graph constructor, graph fusion module, and an approximation module to address these two issues. We conducted experiments using a real-world dataset from the urban area of Tainan, Taiwan. The results show that MVGFRNN outperforms state-of-the-art methods.

Self-supervised contrastive learning has achieved promising results in computer vision, and recently it also received attention in the medical domain. In practice, medical data is hard to collect and even harder to annotate, but leveraging multi-modality medical images to make up for small datasets has proved to be helpful. In this work, we focus on mining multi-modality Magnetic Resonance (MR) images to learn multi-modality contrastive representations. We first present multi-modality data augmentation (MDA) to adapt contrastive learning to multi-modality learning. Then, the proposed cross-modality group convolution (CGC) is used for multi-modality features in the downstream fine-tune task. Specifically, in the pre-training stage, considering different behaviors from each MRI modality with the same anatomic structure, yet without designing a handcrafted pretext task, we select two augmented MR images from a patient as a positive pair, and then directly maximize the similarity between positive pairs using Simple Siamese networks. To further exploit multi-modality representation, we combine 3D and 2D group convolution with a channel shuffle operation to efficiently incorporate different modalities of image features. We evaluate our proposed methods on liver MR images collected from a well-known hospital in Taiwan. Experiments show our framework has significantly improved from previous methods.