研究多模态表征学习与跨模态语义对齐，探索视觉与语言信息的有效融合，以及视觉-语言模型在复杂场景理解任务中的应用。
研究时间序列表示学习、多尺度动态建模与跨分布泛化，重点探索模型在不同采样频率、未见数据域及分布偏移条件下的预测能力与泛化性能。
研究从高维、时序、多源异构数据中挖掘潜在模式与知识，构建面向复杂场景的预测、分类与决策模型。
IEEE Transactions on Geoscience and Remote Sensing
Multimodal vision–language modeling has emerged as a promising paradigm for remote sensing (RS) image understanding. However, existing methods are limited by fixedresolution visual processing and single-scale vision–language alignment, making it difficult to simultaneously preserve finegrained details and maintain semantic consistency across different spatial granularities. To address these challenges, we propose RAH-VLA, a Resolution-Adaptive Hierarchical Vision– Language Alignment framework for multimodal remote sensing understanding. Specifically, a Dynamic Resolution Input Strategy (DRIS) is developed to enable resolution-adaptive visual representations, while a Multi-scale Vision–Language Alignment Mechanism (MS-VLAM) is introduced to establish hierarchical semantic correspondence across object-level, region-level, and global-level representations. Extensive experiments on multiple remote sensing benchmarks demonstrate that RAH-VLA consistently improves image captioning, visual grounding, and crossmodal reasoning performance while reducing computational redundancy. Qualitative analyses further illustrate the effectiveness of the proposed resolution-adaptive perception and hierarchical vision–language alignment mechanisms, as well as the crossmodality generalization capability of the proposed framework. Overall, RAH-VLA provides an effective and scalable solution for multimodal remote sensing interpretation, offering a practical pathway toward efficient and semantically robust RS vision– language models.
The IEEE International Conference on Data Mining 2026 (ICDM 2026)
Cross-frequency financial forecasting is challenging because temporal sampling and aggregation substantially reshape volatility structures and dependency patterns, resulting in sampling-induced temporal distribution shift. Existing methods mainly rely on multi-frequency joint training or implicit multi-scale modeling, while lacking explicit modeling of transferable temporal dynamics across heterogeneous sampling resolutions. To address this issue, we propose TransDyn, a Transformer-driven framework for robust cross-frequency financial forecasting. Trained only on single-asset, single-frequency 5-minute K-bar data, TransDyn directly generalizes to unseen daily-frequency assets without fine-tuning. Specifically, TransDyn collaboratively integrates Multi-Scale Temporal Extraction, Cross-Scale Interaction, Adaptive Temporal Modulation, and Information Bottleneck Regularization to learn resolution-invariant temporal dynamics, while auxiliary objectives further enhance cross-resolution consistency. Extensive experiments on real-world financial markets demonstrate that TransDyn consistently outperforms representative forecasting baselines in zero-shot cross-frequency and cross-asset forecasting. Qualitative analysis further shows that TransDyn effectively alleviates prediction collapse, unstable oscillations, and long-term temporal drift under severe distribution shift.
In Proceedings of the 2025 International Conference on Digital Education and Information Technology
The research aims to study the efficacy of random forest algorithm modelling student grades in software engineering education and compares the predictions made by the model with actual student grades. The data are collected from 88 students from Sanda University in the course "Software Testing Practice" including both subjective ratings by teachers and team leaders across five dimensions: testing requirements, testing plans, testing cases, defect discovery, and testing reports; and students' actual grades in the course "Software Engineering and Project Management" which are objective assessment criteria. It will be compared between two models, Random Forest versus AdaBoost, to determine which is better predicting the students' grades. The overall model designed in this study with subjective ratings from a single course to predict future grades in related courses can be demonstrated to show different performance using their respective different performance models. It has been shown through correlation studies that the random forest model exhibits a greater correlation with actual scores (0.629) than does the AdaBoost model (0.475). More evaluation with measurements such as MAE, MPE, MAPE, as well as R² confirmed the advantages of the random forest model in terms of accuracy and reliability. The model can be further enhanced and made adaptable to various learning environments to more accurately reflect and inform better decision-making in the future and ongoing software engineering education improvements.
2025 6th International Conference on Education, Knowledge and Information Management（ICEKIM 2025）
In response to the transition in higher education from teacher-centered knowledge transfer to student-centered competency building, this work proposes an outcome-based education (OBE) approach to university physics laboratory instruction. Using the classical experiment of determining the gravitational acceleration by a simple pendulum as an example, the teaching process is re-conceptualized with real-life physical situations, fundamental concept comprehension and competency-oriented learning outcomes as its core. The instructional design includes the non-linear modelling of the period of a simple pendulum at large angles, which leads the students to examine the applicability of models, carry out error corrections and pursue innovations via extended topics. A multi-dimensional assessment system consisting of four dimensions and six indicators based on Blooms taxonomy and the CDIO engineering education model was developed. Results show that the students have made significant progress in modeling skills, error analysis and interdisciplinary integration. The suggested OBE approach to experiments promotes higher-order thinking and meaningful learning outcomes effectively, offering a reproducible model for the reform of physics laboratory teaching under the new engineering education paradigm.
《探索科学》
本文通过大学物理实验中的经典实验项目“用单摆法测量重力加速度”来测量上海本地的重力加速度g，并探讨小摆角近似条件下修正公式的适用性。实验首先使用固定摆长和改变摆长两种方法测量小角度的摆动的g，然后使用固定摆长进行大角度摆动（＞5°）的g测量，并对大角度摆动的修正公式在不同摆角下的测量结果进行数值上的验证。实验结果表明，小角度条件下单摆法测得重力加速度g的误差较小；在大摆角情况下，摆动周期的修正公式有效降低摆动周期的相对误差，并使g更加接近理论值。