Beyond Dashboards: Are Transformers the Future of Urban Analytics?
I've just launched the NYC Urban Analytics Hub on Hugging Face, a project I'm excited to share. It includes an interactive dashboard and a rich geospatial time-series dataset covering crime, 311 requests, and building permits across New York City. It’s a powerful tool for exploring the city's pulse, but I believe we can go even further.
The current application uses traditional machine learning and statistical models for its predictions. They work well, but being on the Hugging Face ecosystem, I can’t help but think about a more powerful tool: Transformers.
This leads me to a question I'd love to explore with the community: How can we leverage Transformer-based models to unlock deeper insights from urban data?
Ideas to Explore
1. Time-Series Forecasting as a Language Problem
Could we treat the sequence of monthly data for a single census tract as a “sentence”? Models like TimeGPT or PatchTST have shown impressive results in time-series forecasting. By reframing urban data this way, we might capture complex, long-range dependencies that traditional models miss, leading to more accurate predictions.
2. Multi-Modal Urban Understanding
The dataset is primarily tabular, but what if we combined it with other modalities?
- Text from 311 service requests
- Satellite imagery of census tracts
- Construction permit details
A Transformer model could potentially learn relationships such as:
- a rise in noise complaint texts,
- overlapping with new building permits,
- and how both relate to future crime levels.
Opening It Up
This is where things get exciting. I’ve provided the foundational data and a baseline application inside the NYC Urban Analytics Hub.
👉 What are your thoughts? Could Transformers revolutionize how we understand and plan our cities?
Let’s discuss in the comments. I’m looking forward to seeing what we can build together.