0
$\begingroup$

I'm looking for suggestions on how to proceed with predicting on separate but correlated models.

The example I will use is housing data. I have three inputs:

  1. Latitude
  2. Longitude
  3. 1-Google Street View Image

I want to create models that predict things like:

  1. Square Footage (regression)
  2. Home Price (regression)
  3. Number of Floors (categorical)
  4. For Sale or Not (binary)

The dataset is supervised, so we are given 7 features. The model in production will only have 3 inputs (lat, long, image). Ideally, we could solve this dependence by using 6 input features for every model, but production will only use 3 inputs.

So if I start out creating four independent models (sqft, price, floors, sale/no-sale) based only on 3 inputs, I may run into some potential issues. For example, the models may independently predict a small square footage and a high number of floors in the home. So as a manual reviewer, I can say something like, "Hey, one of these models is probably wrong. I wish that they could take into account each other's predictions and confidence levels.".

I'm trying to read up on how this is done in any literature, but I can't really find the correct terminology to search.

My thoughts so far:

  • Maybe create one-model on only 3 inputs that has the four multiple outputs / output types, probably a NN of some sort.

  • Another thought would be to create each model with all the other 6 inputs, and use a model that deals with missing data well (naive bayes?).

  • Some sort of post ensemble model that looks at all the four model outputs.

I'm sure this type of problem has been looked at before in statistical learning, I just am unsure where to look.

$\endgroup$

1 Answer 1

2
$\begingroup$

In general this is a prime opportunity to use Geospatial Data Science methods. These are more complex in some cases than ML models and sometimes more simplified.

In this case if you have sufficient examples to sufficiently populate your entire study region in a random spread across the landscape, you could use the variables in the actual data as points on a map and create meshes of value across the landscape and use this to interpolate the values of the four individual outcomes which you are looking to predict.

QGIS, an open source geospatial modeling application and some youtube searches are a great place to start solving this problem.

$\endgroup$
1
  • 1
    $\begingroup$ Thank you! I will check this out and see if this works. $\endgroup$
    – nfmcclure
    Oct 7, 2022 at 18:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.