To apply Bayes inference for data analysis or machine learning, we have to construct prior and likelihood, right? But if I fail to come up with a reasonable prior and likelihood, then the Bayes model will not be meaningful, right?

I wonder is there any technique could be utilized to construct reasonable Bayes models?

Let me make it concrete by an example, given a dataest with features $X$ and target $y$, where $y=\{0,1\}$ and $X$ is composed of $p$ variables. This is an ordinary binary classification problem, to do Bayes inference, what and how should I specify for the prior and likelihood?

One more question, if we could not specify a reasonable prior, then what's the point of using Bayes modelling?


1 Answer 1


You would specify a prior for parameters. You haven't mentioned any!

You need to write some model for how the distribution of $Y$ relates to $X$, so your first step is to identify a distribution for $Y$ (it's 0-1, so there's an obvious first choice there) and then write a model for its parameters in terms of $X$.

Once you have that model, you need to think about priors on the parameters. Again, a first attempt should be relatively simple, then make it more complicated if you think it needs to be.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.