Recommender Systems - Using Neo4j vs traditional NMF techniques


I see a bunch of articles about how Neo4j could be used to develop recommender systems.
How do recommender systems developed using Neo4j fare in comparison to those developed using NMF (Non-negative Matrix Factorization) techniques?
From my understanding, we can write targeted Cypher queries to get recommendations. But these would be hard-coded queries that work with predefined relations. How would hidden features be discovered and used for giving better recommendations?
I'd really appreciate some discussion regarding the two approaches.



Please look into Graph Data Science library for more details on how system works.

This is a great question. I would love to see a legitimate answer.