It is often the case that the predictors are variables that take continuous or categorical values in machine learning. However, if the predictors or some of them are vectors, how can we deal with it?

Take interpolation in traditional practices for an example, we use the neighbors to predict the values at the center location $x_0$, denoted briefly here as $f:(x_1,x_2,...,x_n) \rightarrow x_0$. We know that the relative location/distance of $x_i,i=1,2,...$ with respect to $x_0$ is important information for an effective interpolation. Therefore, I want to introduce the spatial coordinates into the predictors for modeling, i.e., $f:[(x_1,xcoord_1,ycoord_1),(x_2,xcoord_2,ycoord_2),...,(x_n,xcoord_n,ycoord_n)] \rightarrow x_0$. I did not see this type of models or their invariants in machine learning before. Is there any one who can give some ideas to deal with the vectors or take advantage of supportive information about each predictor?

  • $\begingroup$ What if you just treated each feature vector as multiple features? You could have one feature vector with three coordinates, another feature vector with three coordinates, and another feature vector with three coordinates, or you could have nine features. Yes, there are relationships between the features, but that’s often the case. $\endgroup$
    – Dave
    Feb 19 at 4:24
  • $\begingroup$ Transformers, GCNNs and DeepSets will all work with vector features $\endgroup$
    – Firebug
    Feb 23 at 14:35


Your Answer

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