This project analyzes approximately 14 million geotagged tweets from Helwig et al. (2015) to examine spatial and temporal patterns of Twitter activity across the United States. The workflow includes large-scale data cleaning, aggregation, and visualization to quantify tweet intensity across longitude, latitude, and daily time windows.
A Random Forest regression model is used to estimate tweet counts per spatial–temporal grid cell, revealing strong geographic clustering (major urban corridors) and relatively stable day-to-day dynamics. Although limited to a seven-day sample and lacking textual context for topic-level analysis, the model serves as a baseline estimator for spatiotemporal tweet intensity and illustrates the challenges of short-horizon social signal prediction.
- Processing of 14M+ records using
pandasandscikit-learn - Geo-temporal aggregation into 1° latitude–longitude grid cells
- Daily tweet count modeling and spatial intensity visualization
- Out-of-sample performance: RMSE ≈ 20% of mean counts, R² ≈ 0.99
- Discussion of modeling limitations (event-driven surges, limited time window, lack of textual context)
Helwig, N., Gao, Y., Wang, S., & Ma, P. (2015).
Twitter Geospatial Data. UCI Machine Learning Repository.
https://doi.org/10.24432/C5RS5J