Possible enhancement at feature selection

[caiogimenes]

SelectKBest transformer currently ranks features based on raw Pearson correlation values [-1,1]. So features that are negatively correlated with the target might be excluded, even if they are highly predictive.
e.g:
0.8 → ranked high
−0.9 → ranked low or excluded, despite being stronger

Possible fix is to add a new param use_abs to allow user to control using absolute value and adding it to learn_one method.

def learn_one(self, x, y):
    for i, xi in x.items():
        self.similarities[i].update(xi, y)
        if self.use_abs:
            similarity_value = abs(self.similarities[i].get())
        else:
            similarity_value = self.similarities[i].get()
        self.leaderboard[i] = similarity_value

[MaxHalford]

Good idea!