- LooM indicator — a latent POI-sentiment measure of subjective emotion across 88 fine-grained urban place types.
- NLP pipeline — processed a 2.78-million-post corpus; Llama-3.1 zero-shot filtering + fine-tuned Chinese RoBERTa sentiment classifier.
- Spatial modelling — dual-stream deep architecture with cross-attention; urban scaling laws of depression risk across 302 cities.
Researching cities, designing spaces,
documenting everyday life.
Reading the city,
line by line.
ResearchHuman · environment
I study how the built environment affects human well-being.
Combining text analysis, spatial data and deep learning to map mental health at the scale of cities.
By treating millions of social-media posts as field notes, I try to uncover the systemic dynamics hidden in everyday human–environment interactions.
A
Urban studies
Geospatial analysis
Scaling laws · spatial heterogeneity
B
Mental health
Public sentiment measurement
Depression perception · well-being
C
Social data science
Machine learning · deep learning
LLM deployment
D
Architecture
Environmental psychology
Architectural design
Publications
Publications.
My current study offers the first large-scale, fine-grained quantification of the link between depressive states and perceived urban places in China.
Experience04 records
Experience.
- CBEES framework — LDA + BERT embeddings extracting and weighting 48 spatial features from unstructured reviews.
- Interpretable ML — Random Forest with SHAP & partial-dependence plots revealing non-linear threshold effects.
- Longitudinal analysis — GAM time-series across 12 commercial projects, capturing post-COVID reopening dynamics.
- Activity analysis — CV pipeline (YOLOv8, Faster R-CNN, DeepSORT) with BERT-based activity-duration extraction.
- GNN simulation — graph architecture linking built-environment factors to individual restorative outcomes.
- VR validation — spatial-scenario simulation with wearable biosensors (HRV, EDA) for physiological grounding.
- Data strategy — systematic collection of geotagged text & imagery from Weibo and Dianping.
- NLP × CV workflow — BERT sentiment classification with Faster R-CNN scene recognition.
- Emotion recognition — Face++ facial-emotion API quantifying space–emotion correlations.
