Is chi-squared feature selection better than Mutual information based feature selection mechanism?
They are related, so I don't suspect there to be a big difference (hence, go for mutual information if it's easier to calculate).
I haven't seen a formal argument for this, but my logic is:
- A g-test is a derivate of mutual information ($G=2\cdot N \cdot MI(r,c)$, cfr. wiki link)
- A Chi-squared leads to the same conclusion as a g-test for reasonably sized samples
Therefore, Chi-squared and MI lead to more or less the same results for reasonably sized samples. In other cases, it will deterministically depend on the dataset properties.
Just as a follow up to @ciri answer, the same arguments have been developed in the following paper: Richter et al., 2018