Training urban spatio-temporal foundation models that generalize well across diverse regions and cities is critical for deploying urban services in unseen or data-scarce regions. Recent studies have typically focused on fusing cross-domain spatio-temporal data to train unified Transformer-based models. However, these models suffer from quadratic computational complexity and high memory overhead, limiting their scalability and practical deployment. Inspired by the efficiency of Mamba, a state space model with linear time complexity, we explore its potential for efficient urban spatio-temporal prediction. However, directly applying Mamba as a spatio-temporal backbone leads to negative transfer and severe performance degradation. This is primarily due to spatio-temporal heterogeneity and the recursive mechanism of Mamba's hidden state updates, which limit cross-domain generalization. To overcome these challenges, we propose Damba-ST, a novel domain-adaptive Mamba-based model for efficient urban spatio-temporal prediction. Damba-ST retains Mamba's linear complexity advantage while significantly enhancing its adaptability to heterogeneous domains. Specifically, we introduce two core innovations: (1) a domain-adaptive state space model that partitions the latent representation space into a shared subspace for learning cross-domain commonalities and independent, domain-specific subspaces for capturing intra-domain discriminative features; (2) three distinct Domain Adapters, which serve as domain-aware proxies to bridge disparate domain distributions and facilitate the alignment of cross-domain commonalities. Extensive experiments demonstrate the generalization and efficiency of Damba-ST. It achieves state-of-the-art performance on prediction tasks and demonstrates strong zero-shot generalization, enabling seamless deployment in new urban environments without extensive retraining or fine-tuning.

Underwater wireless optical communication (UWOC) has emerged as a high-bandwidth, low-latency technology for marine applications, such as sensor networks, autonomous underwater vehicles (AUVs), and the Internet of Underwater Things (IoUT). However, the underwater channel faces severe impairments from absorption, scattering, turbulence due to temperature and salinity gradients, and bubbles from wave breaking or biological activity. These cause beam wandering, scintillation, beam spreading, link misalignment, inter-symbol interference (ISI), and signal loss, limiting effective ranges in turbid waters. Accurate modeling of these combined effects is vital for optimizing modulation, error correction, and system reliability. This study combines experimental and numerical approaches to evaluate the performance of UWOC channels under the influence of bubbles, turbulence, absorption, and scattering. Experimentally, small bubbles are produced via microporous ceramics, while turbulence is induced by propeller rotation to mimic real disturbances. Numerically, a composite model employs Monte Carlo simulations and ray tracing, integrating phase screen methods and Mie theory to simulate scattering from multi-size particles in turbulent conditions. UWOC systems are tested with intensity modulation/direct detection (IM/DD) using OOK, BPSK, QPSK, and 16-QAM modulations. Findings show synergistic degradation: turbulence exacerbates scattering by changing photon angular distributions, heightening scintillation and path distortion. Bubbles have a more pronounced impact than turbulence, leading to higher channel unreliability. The study quantifies how parameters like bubble count, turbulence strength, scattering level, distance, and aperture size affect scintillation index and bit error rate (BER). These results provide critical insights for enhancing underwater optical networks, wake tracking, and IoUT in harsh environments.

Predicting and understanding student learning performance has been a long-standing task in learning science, which can benefit personalized teaching and learning. This study shows that the progress towards this task can be accelerated by using learning record data to feed a deep learning model that considers the intrinsic course association and the structured features. We proposed a multi-source sparse attention convolutional neural network (MsaCNN) to predict the course grades in a general formulation. MsaCNN adopts multi-scale convolution kernels on student grade records to capture structured features, a global attention strategy to discover the relationship between courses, and multiple input-heads to integrate multi-source features. All achieved features are then poured into a softmax classifier towards an end-to-end supervised deep learning model. Conducting insights into higher education on real-world university datasets, the results show that MsaCNN achieves better performance than traditional methods and delivers an interpretation of student performance by virtue of the resulted course relationships. Inspired by this interpretation, we created an association map for all mentioned courses, followed by evaluating the map with a questionnaire survey. This study provides computer-aided system tools and discovers the course-space map from the educational data, potentially facilitating the personalized learning progress.

This paper proposes a representation learning model to identify task-state fMRIs for knowledge-concept recognition, which has the potential to model the human cognitive expression system. The traditional CNN-LSTM is usually employed to learn deep features from fMRIs, where CNN aims at extracting the spatial structure and LSTM accounts for the temporal structure. However, the manifold smoothness of the latent features caused by the fMRI sequence is often ignored, leading to unsteady data representation. In this paper, we model latent features as a hidden Markov chain and introduce a Markov-guided Spatio-Temporal Network (MSTNet) for brain image representation. Concretely, MSTNet has three parts: CNN that aims to learn latent features from 3D fMRI frames where a Markov Regularization enforces the neighborhood frames to have similar features, LSTM integrates all frames of an fMRI sequence into a feature vector and fully connected network (FCN) that is to implement the brain image classification. Our model is trained towards minimizing the cross entropy (CE) loss. Our experiment is conducted on the brain fMRI datasets achieved by scanning college students when they were learning five concepts of computer science. The results show that the proposed MSTNet can benefit from the introduced Markov regularization and thus result in improved performance on the brain activity classification. This study not only shows an effective fMRI classification model with Markov regularization but also provides the potential to understand brain intelligence and help patients with language disabilities.

Student knowledge diagnosis (SKD) is a fundamental and crucial task in educational data mining (EDM). SKD aims to diagnose student latent knowledge which is inferred from student’s performance. The model used for SKD in EDM comes from two sources: variant classical psychometric approaches, and research on machine learning-based approaches. Tradition psychometric models and their variants diagnosis student knowledge state relying on the question-concept matrix (Q-matrix) empirically designed by experts. However, the expert concepts are expensive and inter-overlapping in their constructions, leading to ambiguous explanations. The recent model Meta-knowledge Dictionary Learning (MetaDL), a learning-based model, proposes a linear sparse dictionary method to mine Q-matrix without expert definition and student latent knowledge representation. MetaDL aims to learn a meta-knowledge dictionary from student responses, where any knowledge entity is a linear combination of a few atoms in the meta-knowledge dictionary. However, a linear model cannot capture complex features from the student learning process and MetaDL fails to solve the missing data. This paper proposes a novel Variational Student Knowledge Diagnosis (VarSKD) method that extends the linear sparse representation of student latent knowledge space into non-linear probabilistic sparse representation. This model based on variational sparse coding can obtain better student latent knowledge representation. Furthermore, extensive experimental results on real-world datasets demonstrate the prediction accuracy and effective power of VarSKD framework.

Prediction of undergraduate grades before their course enrollments is beneficial to the student’s learning plan on selective courses and failure warnings to compulsory courses in Chinese higher education. This study proposed to use a deep learning-based model composed of sparse attention layers, convolutional neural layers, and a fully connected layer, called Sparse Attention Convolutional Neural Networks (SACNN), to predict undergraduate grades. Concretely, sparse attention layers response to the fact that courses have different contributions to the grade prediction of the target course; convolutional neural layers aim to capture the one-dimensional temporal feature on these courses organized in terms; the fully connected layer is to complete the final classification based on achieved features. We collected a dataset including grade records, student’s demographics and course descriptions from our institution in the past five years. The dataset contained about 54k grade records from 1307 students and 137 courses, where all mentioned methods were evaluated by the hold-out evaluation. The result shows SACNN achieves 81% prediction precision and 85% accuracy on the failure prediction, which is more effective than those compared methods. Besides, SACNN delivers a potential explanation to the reason of the predicted result, thanks to the sparse attention layer. This study provides a useful technique for personalized learning and course relationship discovery in undergraduate education.

As one of the most well-known artificial feature sampler, the sliding window is widely used in scenarios where spatial and temporal information exists, such as cv, nlp, data stream, and time series. Among which time series is common in many scenarios like credit card payment, user behavior, and sensors. General feature selection for features extracted by sliding window aggregate calls for time-consuming iteration to generate features, and then traditional feature selection methods are employed to rank them. The decision of the period of sliding windows depends on the domain knowledge and calls for trivial. Currently, there is no automatic method to handle the sliding window aggregate features selection. As the time consumption of feature generation with different periods and sliding windows is huge, it is very hard to enumerate them all and then select them. In this paper, we propose a general framework using Markov Chain to solve this problem. This framework is very efficient and has high accuracy, such that it is able to perform feature selection on a variety of features and period options. We show the detail by 2 common sliding windows and 3 types of aggregation operators. And it is easy to extend more operators in this framework by employing existing theory.