|
I want to learn a better representation of a example based on some preference pairs. I have some pair wise preferences of data. Basically says that one example is better than other. I want to learn a better representation of the example which can be used to learn another classification task. One thing I was thinking is using a neural network to classify given two pairs which is better than other and use the hidden layer to get a better representation of data. Should we impose a sparse representation of the hidden layer for better features? Has any one some experience in this? Can anyone point to appropriate literature? |
Maybe also consider metric learning?