3
$\begingroup$

I am not sure if this question makes sense or is even ideal, but here is my problem: Many proposed distance formulas only give a distance in terms of 1-dimensions. A common distance formula is the Euclidean distance (the most obvious one). Euclidean distance $d_{ecdn}$ of two points located at $(x_1, x_2, x_3, ... , x_n)$ and $(y_1, y_2, y_3, ... , y_n)$, respectively, is given by

$d_{ecdn} = \sqrt{\displaystyle\sum_{i=1}^{n} (x_i-y_i)^2}$.

Geometrically, in any number of dimensions (but imagine 3 dimensions as an example), if you were to draw a line between the two points, whose length expresses the Euclidean distance, the line itself exists in 1 dimension. Why is this a problem? I am doing a project involve data mining and information retrieval. Basically, each document in a corpus is mapped to a location in an $n$-dimensional space where $n$ is the total number of terms within the whole set of documents. They say that if document A is geometrically closer to document B than document C, then document A is more similar to B than C. Euclidean distance gives the wrong impression of similarity since each dimension matters.

I am wondering if anyone has any good algorithms that will calculate the "distance" between two points, but with a weight on each dimensions. An example would be the Pearson Product-Moment Correlation Coefficient. This formula is given by

pearson

|cite|improve this question
$\endgroup$

migrated from math.stackexchange.com Aug 22 '12 at 5:53

This question came from our site for people studying math at any level and professionals in related fields.

  • $\begingroup$ Note, lines are in 1 dimension. $\endgroup$ – Karolis Juodelė Aug 22 '12 at 5:52
  • 5
    $\begingroup$ I think you are looking for Mahalanobis Distance. $\endgroup$ – TenaliRaman Aug 22 '12 at 6:27
  • 1
    $\begingroup$ 1) Weighted version of euclidean distance is straightforward, see the formula. 2) Why do you think Pearson r incorporates weights? I don't see any weighting. $\endgroup$ – ttnphns Aug 22 '12 at 7:34
  • $\begingroup$ @TenaliRaman +1. That probably would have been appropriate enough for an answer, but regardless, thank you for that! $\endgroup$ – Sidd Singal Aug 22 '12 at 13:30
  • $\begingroup$ @ttnphns Wow, I didn't know there was a weighted Euclidean distance formula. I figured Euclidean distance was just a single formula, so thank you for that too! I wasn't really too positive about the Pearson, but it seemed like even though it wasn't directly weighting each dimension, there was still accountability for each dimension. $\endgroup$ – Sidd Singal Aug 22 '12 at 13:43
1
$\begingroup$

Just to create an official answer for this topic (based on @TenaliRaman's comment), you want the Mahalanobis distance.

This is computed similarly to the Euclidian distance, (in fact, the Euclidian distance formula is a special case of the Mahalanobis distance formula) except each squared difference term is 'corrected' based on an inverted covariance matrix.

From the Wikipedia page: 

 Mahalanobis distance is thus unitless and scale-invariant, 
 and takes into account the correlations of the data set.
|cite|improve this answer
$\endgroup$

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.