[Task] Improve negative sampling for retrieval

[marcromeyn]

Problem:

There is a considerable amount of tech-debt in the code that’s responsible for negative sampling in retrieval. The biggest issue is that it relies heavily on the model-context, which we would like to remove in order to simplify things.

Goal:

• Simplify the base-class PredictionTask
• Remove the usage of model-context in the prediction-tasks
• Generalize sampling-queues to allow for ranking

Constraints:

• Keep user-facing API the same.

Starting Point:

• [x] Implement PredictionBlock, BinaryPrediction & RegressionPrediction: Done
• [x] Implement a new negative-sampler base class Done
• [x] Implement DotProduct Done
• [x] Implement ContrastivePredictionBlock + tests Done
• [x] Implement DotProductPrediction Done
• [ ] Tests for two-tower/mf with BinaryPrediction & RegressionPrediction
• [x] Implement CategoricalPrediction (without contrastive capabilities) Done
• [x] Make CategoricalPrediction allow for negative-sampling Done

For later release:

• [ ] Update retrieval models with new DotProductCategoricalPrediction block Note: This is pushed to the next release as it depends on another task not captured in this ticket, which is the refactory of the top-k recommender model: https://github.com/NVIDIA-Merlin/models/issues/622

• Impelement different types of samplers one by one

[viswa-nvidia]

@EvenOldridge , is this getting added to the 22.08 POR. ?

[viswa-nvidia]

This RMP ticket has been prioritized over merlin dependencies used in T4Rec

[viswa-nvidia]

@sararb , are all the unchecked boxes on track for 22.08 ? let me know which ones are moving to later release

[sararb]

@viswa-nvidia I have just updated the ticket with the tasks that are moving to later release.