1.4 Potential projects
The ideas described below are just that – nascent ideas for potential research projects in spatial or spatio-temporal data analysis.
Using long short term memory models to model spatio-temporal data.
In recent years, Long Short Term Memory Models (LSTM models) have been used in a number of different application areas to build predictive models of spatio-temporal data. These include:
- Earthquake prediction
- Analysis of human actions based on 3D skeleton data
- Weather forecasting
- Flood forecasting
- Traffic flow prediction
- Motion prediction of beating heart
- Next Point-of-Interest (POI) recommendation
- Facial expression recognition
- Traffic accident prediction
- Stock market movement
- Tumor growth prediction
- For what other spatio-temporal data can LSTM models provide improved to good prediction? For example, can LSTM models reasonably predict acts of political violence in a country, or region, in space and time? Can LSTM models reasonably predict police shootings in space and time?
- What is it about the styles of spatio-temporal data for which LSTM models can provide improved to good prediction – what are the essential characteristics of such spatio-temporal data that permit LSTM models to provide improved to good predictions ?
Non-separable analysis of political violence data.
Our current analysis of political violence in India and the United States essentially deals with location data and time series data separately. A more useful, and realistic, analysis would involve treating the location and time data together, as recommended in, for example, Harvill, J. L. (2010). Spatio‐temporal processes. Wiley interdisciplinary reviews: computational statistics, 2(3), 375-382.
Geometric graph clustering generally works very well for spatial data (although it requires data structuring through, for example, quadtrees if the data is large). Using, for example, time as a third dimension, how does geometric clustering work for spatio-temporal data?
Simulation of spatio-temporal data.
Simulation of univariate or multivariate data is generally quite straightforward, in that we can construct an empirical distribution function that allows us to choose simulated data “at random” from that probability distribution. However, to simulate spatial or spatio-temporal data we need more than just the probability distribution of locations or times: we need to take account of the actual spatial or temporal arrangement of the data (see, for example, Luc Anselin’s video between 5:00 and 6:40 minutes).