2
$\begingroup$

I am studying some event for a set of objects that can be plotted on a square $[0, 100] ^ 2$. I have used logistic regression to calculate probabilities that event occur for different objects and the output can be plotted as below:

We can observe a high density around point $(100, 50)$ and that basically the farther the less probable the event is. A distance to this point is one of the predictors, but there are also few others. What I am interested in is to capture a potential movement, i.e. how valuable is to move from point $A$ to point $B$. But I would like to reward also those movements which are made far from the point $(100, 50)$, not only those near to the point. And for the former the absolute increase in probability will be small even if a movement is pretty significant (long distance), while for the latter we can observe a large increase in probabilities even with a tiny move. So I think what I would need is to somehow nullify/smooth effect of distance from $(100, 50)$, i.e. make it less significant for calculating our movement gain. And I do not really know how to accomplish that. I cannot calculate the percentage gain, because this results in a really large gains. I think I need some monotonic transformation that would make a large probabilities lower and low probabilities higher. Any ideas what I could apply here?

Edit

@AlexFrancisco, I decided to edit a post to include an image that should help me in explanations. My question is: is there a way to restrict the parameters $a, b, c, M$ to such intervals that changing values within those intervals would make $1 - \mathcal{F}(f(x))$ different concave functions ranging from identity (blue curve), by slightly concave function (purple curve) to maximally concave function (red curve)?

enter image description here

In other words, I would like to exclude those values of $a, b, c, M$ that makes $1 - \mathcal{F}(f(x))$ not concave.

$\endgroup$
  • $\begingroup$ If $f$ is known beforehand, the range of all parameters can be presumably determined as long as $f$ really depends on a certain distance function. But since there seems to be only empirical information about $f$ based on sampling, it's unlikely to determine the range. $\endgroup$ – Saad Oct 21 '18 at 9:43
  • $\begingroup$ Sad news, but thanks anyway. $\endgroup$ – jakes Oct 21 '18 at 9:48
2
+50
$\begingroup$

Assume that the density $f$ is positive on $[0, 100]^2$ and there exists a constant $M > 0$ such that $f(x) \leqslant M$ for any $x \in [0, 100]^2$. Two families of prototypical transformations are given first:$$ \mathscr{F}_1(f)(x) = \left( \ln\frac{M}{f(x)} \right)^a, \quad \mathscr{F}_2(f)(x) = \left( \left( \frac{M}{f(x)} \right)^b - 1 \right)^a, $$ where $a, b > 0$ are parameters. To see the rationale, consider two families of density functions:$$ f_1(x) = M \exp(-(d(x, x_0))^{\tfrac{1}{a}}), \quad f_2(x) = \frac{M}{((d(x, x_0))^{\tfrac{1}{a}} + 1)^{\tfrac{1}{b}}}, $$ where $x_0 = (100, 50)$ and $d$ is some distance function. Note that for $d(x, y) = \sqrt{\smash[b]{(x_1 - y_1)^2 + (x_2 - y_2)^2}}$, $a = \dfrac{1}{2}$, $b = \dfrac{2}{3}$, $f_1$ and $f_2$ are the density functions of bivariate normal distribution and that of bivariate Cauchy distribution, respectively. Solving $d(x, x_0)$ from $f_1$ and $f_2$ yields $d(x, x_0) = \mathscr{F}_1(f_1)(x)$ and $d(x, x_0) = \mathscr{F}_2(f_2)(x)$, respectively.

Now, since it may be unpractical to assume that $f$ is positive, or $f$ may be almost $0$ at points far from $x_0$, then it is useful to introduce another parameter $c > 0$, i.e.$$ \mathscr{F}_1(f)(x) = \left( \ln\frac{M + c}{f(x) + c} \right)^a, \quad \mathscr{F}_2(f)(x) = \left( \left( \frac{M + c}{f(x) + c} \right)^b - 1 \right)^a. $$ After normalization,$$ \mathscr{F}_1(f)(x) = \left( \frac{\ln\dfrac{M + c}{f(x) + c}}{\ln\dfrac{M + c}{c}} \right)^a, \quad \mathscr{F}_2(f)(x) = \left( \frac{\left( \dfrac{M + c}{f(x) + c} \right)^b - 1}{\left( \dfrac{M + c}{c} \right)^b - 1} \right)^a. $$

$\endgroup$
  • $\begingroup$ thank you. great answer, deserved bounty reward indeed. $\endgroup$ – jakes Oct 8 '18 at 11:05
  • $\begingroup$ @jakes Since $\mathscr F(f)$ are transformations to recover the “distance,” it would be inevitably decreasing to some sense. But considering $1-\mathscr F(f)$ with suitable $a,b,c$ should give what you need. $\endgroup$ – Saad Oct 19 '18 at 7:04
  • $\begingroup$ @jakes Assume that $f$ depends only on the distance $d$. Since $d$ itself changes linearly with respect to $d$, transforming $f(d)$ to $d$ would offset the variation of $f$. $\endgroup$ – Saad Oct 21 '18 at 1:35

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.