I have a set of features where typical machine learning techniques do not work very good. All features have very different distributions, some heteroscedasticity is also present.
The distribution shapes still convey a lot of information. For example one can see that if the value of a feature is below a threshold highly likely the target variable can never be higher than a value.
I know about bayesian estimates but they do not work well with categorical features (some are) and also they require that all features independently contribute to the target variable. This does not hold true for my features.
Do you know any other machine learning technique that can learn from the same of the features (or in other words from their distribution) and then output a distribution of how the target variable might look like?
Thanks a lot Alex