2
$\begingroup$

I have a dataset for which every object has multiple features, and some of those features exist in tuples, like as in list of $(x, y)$ coordinates for each object. I know I can create an input layer of number of neurons equal to number of features of each object, but how do I pass these 'tuple' sets to the input. Do I break them and append to create a longer list of features, but will that not lead to loss in some correlation between these 2 sets of points from tuples?

$\endgroup$
2
$\begingroup$

This sounds like an ideal situation for a complex valued neural net. A readable introduction is https://makeyourownneuralnetwork.blogspot.com/2016/05/complex-valued-neural-networks.html

$\endgroup$
2
$\begingroup$

You can try dimensionality reduction on those tuples.

Draw each tuple-valued feature separately on a 2-dimensional plane. If the (x,y) points follow a line (maybe a curved line), then the tuple can be replaced with a single value. One point on the line is taken as an origin, and all other (x, y)-points are replaced with the distances to the origin along the line. If the line is curved, the distances are also curved, not the Euclidean distances.

If (x,y) points in a feature don't follow a line, but spread inside a two dimensional region then it is safe to treat them as two separate features.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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