[RMP] Filtering - Githubissues

NVIDIA-Merlin / Merlin

NVIDIA Merlin is an open source library providing end-to-end GPU-accelerated recommender systems, from feature engineering and preprocessing to training deep learning models and running inference in production.

Apache License 2.0

715 stars 111 forks source link

[RMP] Filtering #703

Open EvenOldridge opened 1 year ago

EvenOldridge commented 1 year ago

Problem:

In a production system customers need to filter out candidate recommendations for a variety of reasons.

Goal:

Create a set of common filtering ops and a mechanism for executing them in systems.
Determine if this can exist within NVTabular or whether it should be it's own unique library.

New Functionality

Systems

Wrap as ops:
[ ] Item attribute filtering op
[ ] User attribute based selection of filter for item attribute filtering op
[ ] Filter previous interactions op (bloom filter?)
[ ] "Do not recommend from this list" filtering op (bloom filter? / item attribute flag)

Docs/Examples

Add filtering to existing end to end examples

Constraints:

Access to item/user attributes, passed into this from the feature store or directly into the graph

Starting Point:

interaction filters are commonly implemented as bloom filters
item attribute filters are commonly implemented as queries with constraints

karlhigley commented 1 year ago

@EvenOldridge These ops would be purely for Systems, so wouldn't involve NVT or new libraries.