Research Focus

The ​CityMind Lab focuses on advancing research in ​Spatio-Temporal Data Mining, ​Time Series Analysis, and ​Multimodal Urban Understanding. Leveraging deep learning, graph neural networks, Transformers, and causal inference, we tackle complex challenges in urban computing, environmental monitoring, and predictive analytics. Our work also emphasizes building large-scale datasets and open-source tools to drive progress in the field.

Spatio-Temporal Data Mining
  • Core Focus: Modeling, prediction, and analysis of spatio-temporal data, with applications in urban computing and environmental monitoring.
  • ​Key Research Areas:
    • ​Spatio-Temporal Graph Prediction: Using Graph Neural Networks and Transformers for urban traffic flow, air quality, and crowd movement forecasting.
    • ​Spatio-Temporal Dynamic Systems: Modeling complex spatio-temporal systems via neural discrete learning and expert-level approaches.
    • ​Data Augmentation and Optimization: Enhancing spatio-temporal prediction efficiency and robustness through dynamic sparse training and graph data augmentation.
    • ​Causal Inference: Analyzing spatio-temporal data from a causal perspective to address out-of-distribution learning and environmental confounding.
    • ​Large-Scale Datasets: Building and releasing large-scale spatio-temporal datasets to advance the field.
Time Series Analysis
  • Core Focus: Modeling, forecasting, and anomaly detection in time series data, with a focus on long-term prediction and anomaly detection.
  • ​Key Research Areas:
    • Long-Term Time Series Forecasting: Exploring the potential of LSTM, Transformer, and other models for long-term time series forecasting.
    • Anomaly Detection: Improving anomaly detection efficiency through frequency-domain learning and multi-pattern normality modeling.
    • Foundation Models: Investigating scaling laws and federated learning frameworks for time series foundation models.
    • ​Physics-Informed Learning: Integrating physical knowledge with deep learning to enhance prediction accuracy and interpretability.
    • ​Data Augmentation: Enhancing model generalization through contrastive learning and data augmentation techniques.
Multimodal Learning
  • Core Focus: Leveraging multimodal data (e.g., images, text, trajectories) for urban sensing and prediction.
  • ​Key Research Areas:
    • Vision-Language Pretraining: Developing multi-granularity vision-language pretrained models for urban indicator prediction and region profiling.
    • ​Satellite Image-Text Retrieval: Enhancing retrieval accuracy through cross-domain adaptation techniques.
    • ​Multimodal Data Fusion: Combining satellite imagery, trajectory data, and sensor data for urban flow and air quality prediction.
    • ​Urban Region Profiling: Constructing semantic representations of urban regions using contrastive learning and language-image pretraining.
    • ​Trajectory Modeling and Prediction: Improving trajectory generation and prediction accuracy through causal learning and diffusion models.