Learning to Rank Using Privileged Information

Learning from the Mistakes of Others: Matching Errors in Cross-Dataset Learning

“…If we assume that X ⋆ is of much better quality than X as in for example [35,28], the distribution of the slack variables on privileged space d ⋆ will have a mean value in the negative region (for a correct prediction with a high confidence, the functional margin y ⋆ i w ⋆ , x ⋆ i will be large and therefore the slack variable will be negative). Whereas, the distribution of the slack variables in the original space, p d , will have a mean value around zero and tails that accounts for correctly (left tail) classified samples at negative region and incorrectly (right tail) classified samples at positive region.…”

“…The idea is that if it is difficult to classify the instance in the privileged domain X ⋆ , which is often assumed to have better quality instances [35,28], then it is going to be even more difficult in the original X domain. Non-slack based methods.…”

“…For this, a common approach found in the literature is to consider that the privileged information can be used to distinguish between easy and difficult instances [35,28,14]. This extra knowledge can be used to bias the learning process towards finding a latent function f with better generalization properties.…”

“…Most LUPI methods (e.g. [35,28,20,14]) follow the assumption that the extra information is useful to discriminate between easy and difficult examples. This knowledge is then used to determine the influence of each instance in the training process.…”

“…LUPI has been shown useful in a variety of learning scenarios such as ranking [28], categorization [37], structured prediction [9], data clustering [8], metric learning [10], face/gesture recognition [38], glaucoma detection [7], and recently learning with annotation disagreements [27]. Most LUPI methods (e.g.…”