Predicting lat/long from binary features

Question

I have a number of observations that occur around my city (a small area), and several of them have latitude and longitude. I have been looking into predicting the latitude/longitude of the observations that are not geographically tagged, but I am unsure of the best approach.

One approach I have seen divides an area smaller sections and predicts the likelihood that a point will fall into any specific section.

Wing and Baldridge, 2011: https://scholar.google.com/scholar?cluster=7781354766705840422&hl=en&as_sdt=0,11

Han, Cook, and Baldwin, 2014: https://www.jair.org/papers/paper4200.html

Do you guys know of any alternative approaches to solve this problem? It seems that estimating latitude and longitude separately would be a mistake, since they are related. The grid approach would address that, but it might lack precision.

Answer 1

I just found one paper that might be at least somewhat relevant to your question. It doesn't use binary features; it uses text features instead, but you could probably modify their approach to accommodate your inputs.

The paper is Estimating User Location in Social Media with Stacked Denoising Auto-encoders. It's a deep learning approach using the text of tweets to create one model that estimates region within the U.S. and also state, and then another model that estimates latitude and longitude. The only flaw that I see with their approach is that they don't seem to be treating the earth as a sphere or an oblate spheroid. If your city happens to be bisected by the international date line, this might be a serious problem, but otherwise you may be able to ignore this and still get decent accuracy.

Answer 2

I have another approach that you may find useful. In this paper, the author attempts to predict a lat and lon based on a text string. The way that this is done is by breaking up the text string into individual "n-grams" (basically just words) and then creating a Gaussian Mixture Model (GMM) for each n-gram. To predict a lat/lon for a new text string that's never been seen before, you can find all the n-grams in the new text string, and combine the GMMs for each of those n-grams into a new GMM. This new GMM is a map of the most probable locations for where that text string originated from. Ideally, you can then pick one obvious high-priority point as the most likely location for the origin of the text string.

Similarly, you could create GMMs for each of the binary features that you have available. Whenever one of the binary features is true, just treat that as though it is one of the n-grams that has been found in a text string. If you have lots of binary features, then you should probably be able to get decent results without a lot of work.