In the e-commerce, the labels are usually impression, clicks, add to cart, and order. The ranking model is typically trained on these labeled data. It is possible to design multiple objective function using these labels.
One key observation is that these labels are strongly depended on one another. For instance, a merchandise item would not be purchased if it was never clicked. This work attempts to model these label relationship through the sequential models.
The model is trained to predict the clicks, impression, atc, etc. For each label, the hidden representation is created through the feed forward networks. Due to the sequential dependent between clicks, atc, impression. The hidden representation for clicks is a function of the impression; and same for ATC as well as the number of orders. The simple recurrent neural nets is used to learn these hidden representation.
The multi-objective functions are weighted by the learnable weight parameters. One modification to the objective function is to enforce the predicted impression to be greater the predicted clicks and so on.
In sum, for some instances of the multi-task learning problem, we can try to make each task to be correlated. This idea seems to be useful in the e-commerce application.
Reference:
[Sage]Blog 