[Question] restriction of y_max

[ftmtas212]

Short Question Description

A clear single sentence question we can try to help with?

In the predict() of Class SimpleRegressionPipeline, there are restrictions on the min/max values of y.

def predict(self, X, batch_size=None):
    y = super().predict(X, batch_size=batch_size)
    y[y > (2 * self.y_max_)] = 2 * self.y_max_
    if self.y_min_ < 0:
        y[y < (2 * self.y_min_)] = 2 * self.y_min_
    elif self.y_min_ > 0:
        y[y < (0.5 * self.y_min_)] = 0.5 * self.y_min_
    return y

My question is should we also consider y_max < 0 like y_min?

def predict(self, X, batch_size=None):
    y = super().predict(X, batch_size=batch_size)
    if self.y_max > 0:
        y[y > (2 * self.y_max_)] = 2 * self.y_max_
    elif self.y_max < 0:
        y[y > (0.5 * self.y_max_)] = 0.5 * self.y_max_
    if self.y_min_ < 0:
        y[y < (2 * self.y_min_)] = 2 * self.y_min_
    elif self.y_min_ > 0:
        y[y < (0.5 * self.y_min_)] = 0.5 * self.y_min_
    return y

If I have missed anything, please let me know.
Many thanks!

[aron-bram]

Hi,
thank you for pointing this out. It's a fair question. And I do agree with you on the fact that max should be capped in the way you suggested. However, this part of the code is very old, and I did not write it so I will rather ask before changing it. Maybe there has been a good empirical reason for this.

I'm really sceptical, though, so I will mark it as a bug until all doubts are cleared.

Will let you know once I found out.

[aron-bram]

According to the author, this is indeed a bug.

Great catch! Again, thank you very much for pointing it out.

I will quickly get onto fixing it. For now, I will just use your code snippet as it's a reasonable solution and definitely the quickest one.
Later, we will likely use something more robust, like a TransformedTargetRegressor, instead of manually capping like above.