Non-linear regression with wide spread of datapoints

1

Hi Guys,
I have a non linear dataset with around 25 features and 2 target variables to be predicted. The data points have a wide spread. It is a regression problem. The attached image shows the non-linear distribution of the data which shows how the data is dispersed (Target1 vs each feature).

I have 2 questions:

  1. I want to know if CNN networks with non-linear activation layers can be used to predict multiple target variables in such non-linear patterns?

  2. Do I need to use any specific CNN architecture (encoder-decoder may be) to handle non-linear regression with dispersed pattern as in the pairplot image?

distribution - features vs target
distribution - features vs target1387×91 42.3 KB

2

Hi Shilpa! I am running into a problem with non-linear regression and I would like to know if you get a solution for this problem you posted some months ago. Could you manage to solve this problem?

3

Hi Santiago,
I tried both a sequential network and a CNN network. The sequential network worked better in my case. CNN did not reduce loss beyond a point even after multiple hyper-parameter adjustments. What is the error/issue that you are facing?

4

Hi Shilpa, thanks for the answer!
Okay, is good to know that the sequential network worked better than the CNN. My problem is different from yours, I think. I explained it in the following link https://discuss.pytorch.org/t/getting-a-zero-grad-in-non-linear-regression/140005

5

Hi Santiago,

Check this blog, if it helps.

In your other post, what is happening in the “class” function? Can you try with a sequential network with multiple hidden layers with non-linear activations?

6

Hi Shilpa,
Thank you for the post you shared! I answered your questions in the other post.