Based upon the natural distribution of the features and the outliers and the nature of the outliers, it is important to choose the feature scaling technique accordingly. In the first place, if it would have made no difference, whether to go for min max scaling and robuts scaling or standard scaling, the different techniques would have not resulted.
Since I deal with a lot of sensor data and scaling features built on it, I use robust scaling. This allows me to make features less susceptible to the extreme outliers. At the same time, the distribution of the values between the q1 and q3 is well preserved.
Can you think of a scenario where you’d apply a powertransformer (hint: its used to reduce the skewness and stabilize variance)?
See sklearn’s documentation for more details.