This short paper extends the NADE-CF for implicit feedback dataset. The main difference between implicit and explicit dataset is that the implicit feedback data does not have a negative sample. For example, the lack of a number of clicks or view counts does not imply that a particular user dislikes that item. He/she may not aware the existing of that item.
The author extent CF-NADE to solve the implicit feedback problem by predicting the likeness probability:
where M is the number of items, indicates user u likes item i, when user u dislikes item i. However, we do not know whether user u will like item i. So we need to convert implicit feedback to by setting when .
Since this binarizing an implicit feedback is noisy, the confidence that user u like or dislike the given item should be included in the model. One way to estimate the confidence is:
The higher implicit feedback score user u gives the higher confidence.
Thus, the final model becomes:
This conditional probability is modeled as an autoregressive model with a hidden variable. The confidence will basically act as a weight on the loss function:
Overall, this short paper provides a simple extension of CF-NADE by binarizing an implicit feedback and adding confident random variable to the original CF-NADE model.