1
$\begingroup$

Referencing this question, I know that if $x_1$ and $x_2$ are conditionally independent given $y$ (big assumption), then

$$P(y | x_1,x_2) = \frac{P(x_1,x_2 | y)P(y)}{P(x_2 | x_1)P(x_1)}$$ $$ = \frac{P(x_1| y)P(x_2| y)P(y)}{P(x_2 | x_1)P(x_1)}$$ $$ = \frac{P(y| x_1)P(x_2| y)}{P(x_2 | x_1)}$$

How do I generalize to $n$ variables and compute $P(y | x_1,...,x_n)$? I don't know any of the priors, but I have all the single conditional probabilities (complete matrix)!


Summary

  1. Known: $P(y|x_i), P(x_i|y)$, and $P(x_i|x_j), \forall i,j$
  2. Assumption: $x_1,...,x_n$ are conditionally independent given $y$
  3. Problem: Compute $P(y|x_1,...,x_n)$.

Any help would be appreciated!


UPDATE (reply to Xian):

So to further clarify my problem: I have a disease set $D=\{d_1,...,d_m\}$ and a symptom set $S=\{s_1,...,s_n\}$.

For a given disease, $d_i$, I know the probabilities of the symptoms, $p(s_1| d_i),...,p(s_n|d_i)$ (sparse). For a given symptom $s_j$, I have probabilities $p(d_1 | s_j),...,p(d_m | s_j)$ (also, sparse).

Now, I want to compute $p(d_i | s_{\alpha_1},...,s_{\alpha_k}), \forall i\in[1:m]$, for $k\leq n$ (probability of each disease given a subset of symptoms).

If I understand your answer correctly, you're saying that for a given disease $d$, I can sample a synthetic patient with some symptoms based on the distribution of conditionals $p(s_j | d), \forall j$. But how would I incorporate $p(y|x_1)=p(d_i|s_1),\forall i$ into the sampling procedure so that I can account for the fact that, say, the common cold occurs more frequently than tuberculosis given cold-like symptoms?

Sorry for the confusion!

$\endgroup$
  • $\begingroup$ It seems to me that you need the full set of conditional probabilities $p(x_i | x_{-i})$, not just the pairwise conditional probabilities. $\endgroup$ – jbowman Jun 14 '18 at 22:43
1
$\begingroup$

Since the $X_i$'s are independent given $Y$, the joint density of $(Y,X_1,\ldots,X_n)$ writes down as$$p(y)p(x_1|y)\cdots p(x_n|y)$$and hence the conditional of $Y$ given $(X_1,\ldots,X_n)$ is $$\dfrac{p(y)p(x_1|y)\cdots p(x_n|y)}{\int p(y)p(x_1|y)\cdots p(x_n|y)\text{d}y}$$It simplifies into $$\dfrac{p(x_1)p(y|x_1)p(x_2|y)\cdots p(x_n|y)}{\int p(x_1)p(y|x_1)p(x_2|y)\cdots p(x_n|y)\text{d}y}=\dfrac{p(y|x_1)p(x_2|y)\cdots p(x_n|y)}{\int p(y|x_1)p(x_2|y)\cdots p(x_n|y)\text{d}y}$$but I see no further simplification.

$\endgroup$
  • $\begingroup$ Thanks for the response @Xi'an. I don't actually have access to a dataset for this problem; however, I have the numerical values for the elements of the joint density $p(y|x_1),p(x_2|y),\cdots, p(x_n|y)$. To numerically evaluate the integral in the denominator, would I just compute the joint density in the numerator for each possible value of $y$ and sum? Wouldn't I still be ignoring the prior in $dy$? $\endgroup$ – D. Rad Jun 23 '18 at 17:20
  • $\begingroup$ If you can evaluate the numerator, you can simulate from this distribution without knowing the value of the normalising integral in the denominator. This formula is correct for the conditional in $y$ : hence no you do not ignore the prior. $\endgroup$ – Xi'an Jun 24 '18 at 5:43
  • $\begingroup$ My reply is long, so I updated the post! $\endgroup$ – D. Rad Jun 26 '18 at 22:16
  • $\begingroup$ If $Y$ has a discrete support, the integral in the denominator becomes a summation over all possible values $y$. $\endgroup$ – Xi'an Jun 27 '18 at 5:42

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.