1
$\begingroup$

I'm very new to machine learning, and have a problem I'm trying to solve but don't know where to get started.

I've got a set of feature vectors that are "matched pairs," so

D={(x1,y1), (x2,y2), ..., (xn,yn)}

where each xi and yi are vectors of some length. xi and yi have the same length, but different pairs of xi and yi may have different lengths.

I want to have some algorithm learn the difference of x's and y's, such that, if I have another vector, I want to predict whether it's more like x's or more like y's.

Can someone point me in the right direction that I might try? I'm not sure what types of algorithms or methods I should be looking into, so even just the names of appropriate algorithms or methods would be great.

Improve this question
$\endgroup$
2
  • $\begingroup$ This seems a little unusual. Can you provide a simple example? $\endgroup$
    – gung - Reinstate Monica
    Commented Jan 16, 2016 at 1:28
  • $\begingroup$ An example would be a study of genetic mutations. So x's and y's are genetic sequences, where y is a mutation of x. For example: x1=ACCTGG, and y1=GGCTGG, so the "AC" in x1 mutates to become "GG" in y1. In this example, I would be trying to find if there is some pattern in the mutation so that I can determine if a new sequence is likely to be x (unmutated) or y (mutated). $\endgroup$
    – leontp587
    Commented Jan 16, 2016 at 17:54

1 Answer 1

Reset to default
0
$\begingroup$

What you are asking is not ML. In your case, if you have a lot of $x$'s and $y$'s, then simply determine the correlation matrix for all pairwise combinations of all of your $x$'s and $y$'s. Then look at the values and statistical significance of each correlation coefficient. Your question really is asking if a given $y$ is associated more strongly with a given $x$.

Improve this answer
$\endgroup$
3
  • $\begingroup$ Correlation testing is one valid approach, but it's certainly not the only one. This is a pretty open-ended problem. $\endgroup$
    – shadowtalker
    Commented Jan 16, 2016 at 7:42
  • $\begingroup$ @LEP: I'm not sure if I understand, I'm asking if I have a new vector that I don't know whether belongs to X or Y, then I'm trying to determine whether it's more likely to belong to X or belong to Y? I'm not trying to ask if a given y is associated more strongly with a given x. $\endgroup$
    – leontp587
    Commented Jan 16, 2016 at 18:01
  • $\begingroup$ Ok, then how about replacing each base with 1,2,3,4 and then determine the Euclidean distance, d(x,y), between the two vectors? The same vectors will give you a distance of zero. You can also weight the distance and reward with a 1 if the vectors are the same and near zero if very different by using: exp(-d(x,y)). This called a radial basis function, RBF. Why are you not using a pairwise alignment for two DNA sequences? $\endgroup$
    – user32398
    Commented Jan 17, 2016 at 13:28

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.