PairwiseHingeLoss

• used in ranking problems where we are predicting a positive instance and negative instance.
  • basically, it means that the model outputs two scores, the positive instance is > 0 the negative instance is < 0
• an extension of HingeLoss
• the loss:

  L(max(0, margin - [score_u - score_v]))
  

  • where margin = 1 (in most cases)
  • score_u is the predicted score for the positive instance
  • score_v is the predicted score for the negative instance
• notice how the larger the gap between score_u and score_v results in a lower loss!
• since the loss depends on two scores, it’s used for learning to rank