Balancing learning to rank efficacy and efficiency

Search engines are used to apply dynamic ranking methods on the flight. Therefore efficiency of learning to rank methodologies is as important as efficacy (accuracy). Anyway, very little attention has been paid so far to efficacy.

Learning to efficiently rank introduces a principled and uni ed framework for learning ranking functions that jointly optimize both retrieval e ectiveness and efficiency. The key intuition is that those features having a weight under a threshold can be discarded, after the learning phase. An optimal parameter threshold is learned by direct optimization by using a non-analytical optimization procedure and performing a linear search in the optimization space. Each function evaluation requires measuring both efficiency and eff ectiveness of the current parameter setting . This can be costly for large training sets and large feature sets.