I have a binary target variable that I try to predict with using a Bernoulli likelihood and a Sigmoid function for the p of the likelihood like logistic regression. I am only interested in the probability of success.
I have to say I am also quite new to Bayesian modelling and PyMC.
I would like to somehow weight the observed data points. I googled a lot and found the following not really satisfying results
- If I had a Normal likelihood incorporating weights into the model is rather easy by modifying the sd of the likelihood. I have a Bernoulli likelihood. Is it maybe somehow possible to “convert” the problem into one that uses a likelihood with a variance component that I could modify with the weights?
- Following this idea https://groups.google.com/d/msg/stan-users/v4CoBWUehwU/f6ynwWKxCQAJ I could weight the logp values. How could I do this with PyMC3? I’m using Minibatches if this is important. I thought about wrapping the Bernoulli likelihood and modifying the logp value. Could this work? How would I get the correct index for the weights then?
- Upsampling the data by weights. I would prefer to not do this.
- Are there any other ways to weight binary outcomes, maybe some papers?