Model Exploration

[VasLem]

• Investigate models availability for Geographic Data Clustering and Analysis
• Create a backbone of a plausible model

[adriana-madi]

Research Stage Update (with @OlympiaG): Discussion points for 3/05/21

• Document with propositions
• Several sites/papers on Geographic Data Clustering
• How is the data pre-processing going? what factors are you looking at?

Action for next stand-up:

• [x] Compile research in one document accessible to everyone
• [ ] Create a flowchart for possible model backbone

[OlympiaG]

Research Findings Document - Sprint 1: Research Summary -Sprint 1 (2).pdf

It needs further investigation.

[bajo1207]

Ok, so, from my understanding this is how the prediction is gonna work:

1. We select x amount of cities from all countries. The cities that we select are based on some metric e.g. biggest city.
2. With the help of the labeled dataset and unlabeled datasets (Aquastat, education, economics, development) we predict the water security for those cities.
   • This predictor/classifier also considers geographical data
3. We create a geographical perimeter around these cities for which we can make predictions
4. For a given lat lng within these perimeters we predict the risk based on clustering to nearby cities
5. The prediction can optionally be given with a confidence interval

[VasLem]

Regarding my findings, it seems that the proposed way from @adriana-madi and @OlympiaG is fairly straightforward and clean. Another direction to take would be to consider the problem as a graph, with edges weights produced by the distance matrix of the cities and with the features being nodes attributes. To this end, if we have time or we see that the model is not efficient enough, we could try to get into Graph Convolutional Networks in a simliar way as described here and more specifically to our task there

[bajo1207]

We have agreed on the general structure - lets continue the more detailed model discussion in #29 and #31.