0

I have a data in which each sample has feature vector consisting of x and about 9000 other features and also corresponding y(target value). in which x and y are both continuous values(between 0 to 20). x a noisy data but we can not recognize the source of the noise. The goal is to predict y from x and other features(features are not noisy). number of samples are about 900,000. what are the machine learning approaches I can use in this problem. also famous networks in neural network or deep learning.

sandra
  • 949
  • 3
  • 12
  • 25
  • I'm a little confused what you are looking for? Would you like suggestions on algorithms, configurations of algorithms, both? – Andnp Jun 17 '16 at 18:12
  • I'm looking for algorithms or common machine learning approaches can be use to solve this problem. For example this problem can be seen as a regression problem but input data which is x is noisy that we don't have this property in regression. I want to know if there is a common method or algorithm in machine learning is like to solve this kind of problem or problems near this problem. if there is any I would appreciate to just name it. – sandra Jun 17 '16 at 18:23
  • 1
    In this case, I think you answered the question yourself. The first approach that comes to mind for me would be a neural network. The interactions between features of your dataset would determine the architecture of that network (so we can't comment there), and you would have a single output node that would handle the regression. – Andnp Jun 17 '16 at 18:32
  • Do I read this correctly? You're trying to predict Y from X alone, without using the other 9000 features? – Prune Jun 17 '16 at 18:57
  • 1
    Take a look at "How to approach machine learning problems with high dimensional input space?" http://stackoverflow.com/questions/2255833/how-to-approach-machine-learning-problems-with-high-dimensional-input-space – WestCoastProjects Jun 17 '16 at 19:11
  • @Prune No I will consider all features in this problem, but I separate x from the other in my explanation because just this feature is noisy – sandra Jun 17 '16 at 19:54
  • Excellent! In that case, my answer below should be of some use to you. – Prune Jun 17 '16 at 20:56

1 Answers1

1

This sounds to me like a standard regression problem, although your prediction correlation is going to suck (technical term :-) ) in direct proportion to the noisiness of x. Look up all the educational examples for predicting housing prices (often used to illustrate gradient descent). You have 9000 features instead of 3 or 4, but that's just a matter of training time.

You might also consider some "factor analysis", so that you can eliminate the features that don't contribute enough to y (correlation coefficient near 0.0). This is called "dimensionality reduction"; look for PCA (Principal Component Analysis).

Prune
  • 76,765
  • 14
  • 60
  • 81