Neo4j as a data persistence layer has many, many advantages over other data persistence layers for graph-like problems (social, recommendation, etc.), principal

Neo4j as a data persistence layer has many, many advantages over other data persistence layers for graph-like problems (social, recommendation, etc.), principally from the point of view of performance. You can do it in Python but you're going to have to persist and query your data somehow, and I think at that point you'll "get" why a graph DB might be beneficial. It's not an ML-related thing though, it's a data query performance thing.