[RMP] GTC Session-based Web Recommendations

[karlhigley]

Problem:

The current GTC recommender produces recommendations for users who have registered for GTC or are in the marketing email dataset. There are also GTC attendees that we don't have any information about outside the current session, and we'd like to be able to provide recommendations for them too.

Goal:

• Build a session-based recommender that serves real-time recommendations
• Build on top of NVTabular+Transformer4Rec+Triton
• Use Bert Embeddings (or the text embeddings from the Prospero team) as model features

New Functionality

• NVTabular
  • [ ] Improve NVTabular performance on Triton: https://github.com/NVIDIA-Merlin/NVTabular/issues/1251

Constraints:

• Acceptable latency
• Custom Code for preprocessing and post processing (filtering) as not available in Merlin Systems
• Using legacy NVTabular export with PyTorch

Starting Point:

• [x] POC Training model
• [ ] POC Deploying model on Triton
• [ ] Iterate on Model by adding more features/improve model

[viswa-nvidia]

@angmc @bschifferer , please update this ticket based on the latest discussions.

[viswa-nvidia]

@bschifferer please link the ticket that tracks NVT and Triton perf check

[bschifferer]

@viswa-nvidia I updated the ticket about NVTabular performance - I think this one will be relevant for GTC Spring recommender: https://github.com/NVIDIA-Merlin/NVTabular/issues/1251 -> 200ms for NVTabular will be too long

[viswa-nvidia]

@angmc @bschifferer , please update this ticket based on the latest discussions.

[bschifferer]

@viswa-nvidia @EvenOldridge should we mark this as done?

There is a new RMP ticket to update GTC recommender: https://github.com/NVIDIA-Merlin/Merlin/issues/887

[karlhigley]

Closing since the GTC recommender has already been built and deployed