6
$\begingroup$

I have a (binary) classification problem where after merging single training data points (that can be tracked back to the same source) into aggregates, test accuracy (on single data points again) increases significantly. By merging I mean adding and averaging feature vectors. A SVM classifier/linear kernel is used. (The training data is noisy since it is semi-automatically generated).

So there is the weird situation that making training and test data more 'dissimilar' (aggregates vs. single data points) increases performance. (Moreover, for plain linear regression, adding together feature vectors shouldn't make a difference, I guess).

I'm trying to find out what might be the reason for that. One hypothesis is that for large margin classifiers like SVM's separability is crucial, and that by aggregating, there is less overlap between classes caused by noise.

It would be nice to have a measure indicating the amount of 'overlap' or 'separability'. What are standard measures here? I plan to do some plot of the first two PCA dimensions and see how things look, but some 'credible number' would be good in addition.

$\endgroup$
10
$\begingroup$

The most common measures of separability are based on how much the intra-class distributions overlap (probabilistic measures). There are a couple of these, Jeffries-Matusita distance, Bhattacharya distance and the transformed divergence. You can easily google up some descriptions. They are quite straightforward to implement.

There also some based on the behavior of nearest neighbors. The separability index, which basically looks at the proportion of neighbors that overlap. And the Hypothesis margin which looks at the distance from an object’s nearest neighbour of the same class (near-hit) and a nearest neighbour of the opposing class (near-miss). Then creates a measure by summing over this.

And then you also have things like class scatter matrices and collective entropy.

EDIT

Probabilistic separability measures in R

separability.measures <- function ( Vector.1 , Vector.2 ) {
# convert vectors to matrices in case they are not
  Matrix.1 <- as.matrix (Vector.1)
  Matrix.2 <- as.matrix (Vector.2)
# define means
mean.Matrix.1 <- mean ( Matrix.1 )
mean.Matrix.2 <- mean ( Matrix.2 )
# define difference of means
mean.difference <- mean.Matrix.1 - mean.Matrix.2
# define covariances for supplied matrices
cv.Matrix.1 <- cov ( Matrix.1 )
cv.Matrix.2 <- cov ( Matrix.2 )
# define the halfsum of cv's as "p"
p <- ( cv.Matrix.1 + cv.Matrix.2 ) / 2
# --%<------------------------------------------------------------------------
# calculate the Bhattacharryya index
bh.distance <- 0.125 *t ( mean.difference ) * p^ ( -1 ) * mean.difference +
0.5 * log (det ( p ) / sqrt (det ( cv.Matrix.1 ) * det ( cv.Matrix.2 )
)
)
# --%<------------------------------------------------------------------------
# calculate Jeffries-Matusita
# following formula is bound between 0 and 2.0
jm.distance <- 2 * ( 1 - exp ( -bh.distance ) )
# also found in the bibliography:
# jm.distance <- 1000 * sqrt (   2 * ( 1 - exp ( -bh.distance ) )   )
# the latter formula is bound between 0 and 1414.0
# --%<------------------------------------------------------------------------
# calculate the divergence
# trace (is the sum of the diagonal elements) of a square matrix
trace.of.matrix <- function ( SquareMatrix ) {
sum ( diag ( SquareMatrix ) ) }
# term 1
divergence.term.1 <- 1/2 * trace.of.matrix (( cv.Matrix.1 - cv.Matrix.2 ) * 
( cv.Matrix.2^ (-1) - cv.Matrix.1^ (-1) )
)
# term 2
divergence.term.2 <- 1/2 * trace.of.matrix (( cv.Matrix.1^ (-1) + cv.Matrix.2^ (-1) ) *
( mean.Matrix.1 - mean.Matrix.2 ) *
t ( mean.Matrix.1 - mean.Matrix.2 )
)
# divergence
divergence <- divergence.term.1 + divergence.term.2
# --%<------------------------------------------------------------------------
# and the transformed divergence
transformed.divergence  <- 2 * ( 1 - exp ( - ( divergence / 8 ) ) )
indices <- data.frame(
jm=jm.distance,bh=bh.distance,div=divergence,tdiv=transformed.divergence)
return(indices)
}

And some silly reproducible examples:

##### EXAMPLE 1
# two samples
sample.1 <- c (1362, 1411, 1457, 1735, 1621, 1621, 1791, 1863, 1863, 1838)
sample.2 <- c (1362, 1411, 1457, 10030, 1621, 1621, 1791, 1863, 1863, 1838)

# separability between these two samples
separability.measures ( sample.1 , sample.2 )

##### EXAMPLE 2
# parameters for a normal distibution
meen <- 0.2
sdevn <- 2
x <- seq(-20,20,length=5000)
# two samples from two normal distibutions
normal1 <- dnorm(x,mean=0,sd=1) # standard normal
normal2 <- dnorm(x,mean=meen, sd=sdevn) # normal with the parameters selected above

# separability between these two normal distibutions
separability.measures ( normal1 , normal2 )

Note that these measures only work for two classes and 1 variable at a time, and sometimes have some assumptions (like the classes following a normal distibution) so you should read about them before using them thoroughly. But they still might suit your needs.

| cite | improve this answer | |
$\endgroup$
  • 2
    $\begingroup$ I added an R function based on some code I came across a long time ago, I don't remember who wrote it but I hope it helps. $\endgroup$ – JEquihua Dec 9 '13 at 17:03
  • $\begingroup$ I believe the form of the Bhattacharyya distance used in the above script is only valid for multivariate normal distributions. At least, that's what I gather from perusing the Wikipedia page on it. $\endgroup$ – psychometriko Nov 24 '15 at 19:48

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.