Yan Ma

Trajectory-User Linking (TUL) aims to identify the owner of an anonymous trajectory from a set of candidate users, supporting human mobility analysis and personalized location-aware services. TUL is a complex real-world classification problem because sparse and overlapping trajectories contain heterogeneous temporal, semantic, and transition information. Existing methods commonly learn these features independently, underuse structural knowledge shared across trajectories, and compress structural and sequential information before classification. To address these limitations, we propose Multi-Relational Knowledge Graph Embedding for Check-in Trajectory-User Linking (MakeTUL), a unified neural framework that integrates multi-relational knowledge representation, hypergraph learning, and sequential modeling. To the best of our knowledge, MakeTUL is the first attempt to introduce knowledge graph representation learning into TUL. It organizes visit-time, POI-category, and transfer-speed information as typed relations in a multi-relational mobility knowledge graph, allowing heterogeneous mobility semantics to jointly constrain the learned embeddings. A trajectory hypergraph further incorporates high-order POI co-occurrence patterns as structural priors, improving the representation of sparse trajectories. A sequence learning module captures ordered mobility dependencies, while a dual-branch classifier combines global structural and sequential evidence at the decision level. We evaluate MakeTUL on Foursquare-NYC, Foursquare-TKY, and Foursquare-JKT using five metrics across six user-scale settings. MakeTUL consistently achieves the best results. In the most challenging setting with 1,400 users, it outperforms the strongest baseline by 10.67% in ACC@1 and 12.70% in Macro-F1. These results demonstrate that integrating relational knowledge with complementary structural and sequential neural learning provides robust and discriminative representations for complex mobility data. The source code is available at https://github.com/superior-DL/MakeTUL.

Accurate traffic flow forecasting requires modeling both stable macroscopic dependencies and abrupt local fluctuations in complex road networks. Existing spatiotemporal forecasting models usually learn spatial structures from raw time-domain traffic signals, where low-frequency trends and high-frequency fluctuations are entangled. Although decomposition-based and frequency-aware methods have shown the benefit of separating heterogeneous traffic components, how frequency decomposition can support reliable high-order topology learning remains less explored. To address this issue, we propose FEDHNet, a Frequency-Guided Dynamic Hypergraph Network for traffic flow forecasting. FEDHNet first performs adaptive spectral decomposition on the hidden representation to obtain low-frequency and complementary high-frequency latent components. The low-frequency branch constructs dynamic hyperedges from the relatively smooth latent representation to model non-local high-order dependencies, while the high-frequency branch employs a lightweight 2D Inception module with GLU-based gated denoising to model rapidly varying latent responses. A low-frequency-anchored residual fusion module then adaptively integrates high-frequency residual information into the low-frequency latent representation for multi-step prediction. Experiments on four public PeMS datasets show that FEDHNet achieves competitive forecasting accuracy and multi-horizon performance, together with favorable computational efficiency compared with recent spatiotemporal forecasting baselines. Further analyses examine the effects of topology-source selection and controlled high-frequency residual modeling, revealing that the benefit of low-frequency hypergraph construction is dataset-dependent.

Multimodal large language models (MLLMs) are increasingly used to evaluate educational materials, but a holistic rating says little about whether a model responds to a particular instructional-design construct. CFES-P24 is a theory-grounded counterfactual benchmark for auditing educational slides. Its registry specifies 504 planned pairs derived from 24 author-created, three-slide micro-lessons in six disciplines. Of these, 432 target pairs instantiate six multimedia learning principles at three parameterized levels, and 72 are visual-equivalence sham controls. This paper reports the benchmark design and a pilot validation rather than the full 504-pair evaluation. Twenty-one pairs had been generated for this version. They passed 100/100 transformation-rule tests, 21/21 inverse-restoration checks, and 21/21 independent disk-level quality checks. A frozen candidate gate then tested five pairs once with each of two MLLMs, for 10 calls in total. Every response satisfied the structured-output schema on the first attempt. On the eight target calls, observable operation, principle, minimal repair, and evidence-anchor matches were each 8/8; strict direction selection was 6/8, and severity exact match was 0/8. Both sham calls were correctly classified as having no material difference. The preregistered all-or-none gate consequently failed one of 14 rules, so 16 holdback calls were not released. These results suggest that construct recognition, comparative judgment, and severity calibration should be evaluated separately. They also show the importance of distinguishing a registered design from completed evidence when MLLMs are used to audit educational materials.

Germination percentage is an endpoint measure and therefore does not describe when an individual seed begins visible growth or how rapidly its radicle and plumule expand. We developed a time-resolved phenotyping workflow to quantify rice seed germination continuously in shallow-water culture. A single industrial camera moved along a 1 m rail and imaged three culture boxes at 1 h intervals for up to 80 h. The archive comprised 1,062 full-frame images and 6,372 seed-level repeated observations under the six-seed field-of-view configuration. A physical grid maintained seed identity through time and enabled individual regions of interest to be extracted. Whole-seed foregrounds were obtained with a pretrained U ² -Net, and a masked RGB intensity rule separated newly emerging tissue from the darker hull. For each tracked seed, projected emerging-tissue area and interval growth rate were calculated. Three representative normally germinating seeds first showed measurable tissue at 48 h, yet subsequently followed distinct trajectories: final projected areas ranged from 2,605 to 4,700 pixels and peak interval growth rates ranged from 106.88 to 287.92 pixels h ⁻¹ . B-1 accumulated 63.71% of its final visible area during 72–80 h, whereas B-3 accumulated 73.51% during 60–72 h. Thus, seeds with the same observed emergence interval can differ substantially in the timing and magnitude of post-emergence expansion. The workflow converts repeated images into biologically interpretable temporal phenotypes and provides a basis for nondestructive studies of rice seed vigor and germination heterogeneity.

Clustering complex data structures remains a pivotal challenge in unsupervised learning, particularly when determining the optimal number of clusters in highly non-linear datasets. In this paper, we introduce RISM (Relative Density and Inter-Cluster Connectivity Degree-based Split-and-Merge), a novel clustering algorithm that integrates both density-based and connectivity-driven principles to automatically infer the optimal cluster configuration. RISM operates in two distinct phases: splitting and merging. During the splitting phase, we define a new relative density metric, which evaluates the local density of each data point in relation to its k-nearest neighbors, allowing for the identification of potential cluster centers. Subclusters are then formed based on both relative density and relative distance, effectively capturing dense regions in the data. In the merging phase, we propose a connectivity-aware inter-cluster distance measure that incorporates both inter-cluster distance factors and inter-cluster connectivity degrees, enabling a principled and iterative merging of clusters based on their proximity and structural relationships. The final number of clusters is determined by maximizing the difference in inter-cluster distances observed throughout the merging process. Extensive empirical evaluations on both synthetic and real-world datasets demonstrate the superiority of RISM over nine state-of-the-art clustering algorithms, particularly in terms of clustering accuracy, robustness to noise, and scalability.

Background: Skin pigment separation is a key task in the fields of medical aesthetics, clinical analysis, and dermatological diagnosis. Existing pigment separation methods suffer from the problem of unclear separation between melanin and hemoglobin. Methods: This study proposes a novel approach that combines a sub-block selection algorithm with local clustering for skin pigment separation. The sub-block selection algorithm sorts and filters sub-blocks based on the average pixel difference, reconstructing the input data to ensure accurate separation of melanin and hemoglobin. The local clustering method uses the Euclidean distance between sample points and their surrounding cluster centers to assign sample points to their respective clusters. A random sampling of sample points is performed from each cluster to ensure convergence of independent component analysis during the iteration process. Results: Experimental evaluations demonstrate that the proposed method achieves accuracy and precise separation in skin pigment separation. Furthermore, the average success convergence rate of this approach reaches 92%, surpassing the performance of existing methods.