For classification, I guess the recommended value is 1/N for multi-class (N represents the number of classes) classification. For regression, I read somewhere that mean of the target could be a better initial value.
I tried tuning the other parameters for a regression problem with the default base_score, initialized to the mean of the target and initialized to 0.5. All experiments ended up with more or less same MAPE (difference was only after the 5th decimal places). However, the runtime was much higher for optuna HPO when base_score was set to the mean of the target.
So, my question is what is the recommended value of base_score for a regression problem? Does it also needs to be tuned? If yes, how do people choose the range?