[Task] Target generation for Youtube-DNN retrieval model

[sararb]

Description

Implement a mapping function to apply on top of the data loader in order to extract targets from the input sequence. The goal is to define different strategies for setting the target of the Item Prediction task.

We particularly want to support 4 different strategies:

• Last: extract the last non-padded item from the input sequence and set it as the target.

• Sliding window: Generate multiple targets at different positions based on a sliding-window and extend the input batch with the related truncated input sequences.

• Random: Randomly select a cutting point and set the last item as the target. The input sequence is all interactions from the left context.

• Random held-out: Randomly select a cutting point as the label. The input sequence is all interactions from both contexts, left and right.

Additional context In the Youtube-DNN paper, they tested the following label selection strategies: