Weighted Approximately Ranked Pairwise (WARP) Ranking Loss

From GM-RKB
Jump to navigation Jump to search

A Weighted Approximately Ranked Pairwise (WARP) Ranking Loss is a ranking loss function for ranking with pairwise classification.



References

2013

2012

  • http://www.hongliangjie.com/2012/08/24/weighted-approximately-ranked-pairwise-loss-warp/
    • QUOTE: To focus more on the top of the ranked list, where the top k positions are those we care about using the precision at k measure, one can weigh the pairwise violations depending on their position in the ranked list. For pair-wise learning procedure, we construct a set of all positive labelled instances, denoted as C+u and a set of negative labelled instances as C−u. The loss is defined as:
      errWARP(xi,yi) = L[rank(f(yi|xi))](1)
      where [math]\displaystyle{ \operatorname{rank}(f(y_i|x_i)) }[/math] is a function to measure how many negative labelled instances are “wrongly” ranked higher than this positive example xi:
      [math]\displaystyle{ \operatorname{rank}(f(y_i|x_i)) = ∑(x′,y′) \in C−uI[f(y′|x′) \ge f(y|x_i)] }[/math],
      where I(x) is the indicator function, and L(⋅) transforms this rank into a loss: [math]\displaystyle{ L(r) = r∑j = 1τj, }[/math] with [math]\displaystyle{ τ1≥τ2≥⋯≥0 }[/math].

2010

2009