For an image pixel classification task (a pixel can be either class 1 or class 0), I am fitting a logistic regression model using six features related to a specific pixel's color space properties (red channel, green channel, blue channel, hue, etc). I am interested in learning how to incorporate prior information (in the sense of a bayesian prior distribution for each model coefficient) about the behavior of these predictive features into a logistic regression model.

As an example, I know that in the images related to my problem, for each image, most pixels have a certain distribution for each color channel and for each class. However, I am not seeing very clearly how to encode/relate this information in terms of the model coefficients. Most of the literature that I've seen is on how to set a weakly informative prior for logistic regression coefficients; but given that I know how these variables behave for each one of the classes, is there a way I can use this information? In summary, my questions are:

  1. How can I translate prior knowledge of the distribution of each predictive feature in terms of the model coefficients?
  2. If I fit a "frequentist" logistic regression just to see how the coefficients behave in order to derive my prior distributions, would this be "cheating" since I am already taking a look at the data?

As a side note, I'm aware that a bayesian LR may not be the optimal solution to perform this kind of pixel classification, but I am doing this more as a learning experiment to understand how to do modelling under the bayesian perspective.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.