Weakly relevant feature

For a learning problem a feature $x_i$ of the input space is weakly relevant if

• it is not a Strongly relevant feature, and
• there exists a subset of features $S$ such that adding $x_i$ to $S$ improves the Bayeses optimal classifier.