1
$\begingroup$

I model my problem as a simple V-structured Bayesian network. There is an $outcome$ variable, the binary $switch$ variable, and some environment features $X$. All the variables are observed during training but $outcome$ and $switch$ are not known during inference. I am interested in estimating the marginal $P(outcome \mid X)$. From the domain knowledge,

$P(outcome \mid switch = 0, X) = 0,$

so, the model can be decomposed as

$P(outcome \mid X) = P(switch = 1 \mid X) \cdot P(outcome \mid switch = 1, X).$

Fitting $P(switch = 1 \mid X)$ is straightforward: I just use logistic regression. But how to fit $P(outcome \mid switch = 1, X)$? My current approach is to filter the training data to retain only examples for which $[switch=1]$ as dictated by the form of this probability. However, it does not seem sensible when it comes to inference. There will be some $[switch=0]$ examples that have non-zero $P(switch = 1 \mid X)$ prediction, which can be amplified by $P(outcome \mid switch = 1, X)$ as we did not train that model on similar examples.

Should I instead train $P(outcome \mid switch = 1, X)$ on the $[switch=0]$ examples as well, just weighting them proportionally to the $P(switch = 1 \mid X)$ prediction? Or should I reformulate my model completely?

Will appreciate any references.

$\endgroup$
0
$\begingroup$

You know that your Data may be distributed in an absolute different way than you latent space. You are sampling your latent space given/from your data.

The Bernoulli Dstribution for your latent space can either only be obtained through backprop. or through a parameterization given your Latent distribution´like a reverse generative model.

Though you have:

--data distro-- --Bernoulli distro--

X1   -     ...   -    Y1
X2   -     ...   
X3   -     :W:   -    Y2
X4   -     ...
X5   -     ...   -    Y3

A generative model would learn a Gassian/Beta/Gamma/Poisson data noise with Bernoulli prior, so for the encoding case, you need a Bernoulli data noise with any data prior.

The parameters can be obtained with the KL-Divergence or EM... Be sure not to learn too much values. E.g. $diag( \sigma_.^2)$ and the bias $\mu_x$ for simple space invariance. But you could even use just a recurrent / convolutional layer for the spacial invariance.

Note that usually your $f(x,W)$ shall already depict your mean value of Y! (And $f^{-1}(W^{-1},y)$ mathematically you data mean... if I didn't overcomplicate something.)

Bayesian it would be:

$P(X|Y,\Theta) = \frac{P(Y|X,\Theta)P(x)}{P(Y)}$

which you can approximate with $\frac{P(Y|X,\Theta)P(X)}{\sum_{x' \in ROI(X)}P(Y|x',\Theta)P(x')}$ where $ROI(X)$ would be the region of interest from X given the $M'$ highest marginal probabilities over your Bernoulli distribution.

There are papers about binary to binary generative models which you should read considering the math you are going to implement!

$\endgroup$

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.