A Sommelier's Guide to Recommendation Algorithms: Classical and Graph-Based Recommender Systems
summary
Compares classical embeddings (feature-scaling, one‑hot, target encoding, sentence‑transformers) and matrix‑factorization with SGD against a graph pipeline built in Neo4j that models wines, attributes and users as nodes, stores ratings on edges, projects the graph and runs Neo4j GDS fastRP to get wine embeddings then averages them into user embeddings. Uses the X‑Wines: 100k wines, 21M ratings, 1M users dataset, scores recommendations by 0.5×similarity + 0.5×rating, and evaluates with precision@K, recall@K, MRR and business metrics; demonstrates working implementations and tradeoffs but no single approach is claimed universally superior.
AI-written summary · about
chapters (14)
- 0:00Introduction and welcome
- 1:25What are recommendation systems
- 4:57Challenges in recommendation systems
- 5:18Wine dataset overview
- 7:02Content-based filtering explained
- 9:06Classical embedding approaches
- 12:58Graph-based embeddings
- 16:11How graph algorithms work
- 18:17Content-based filtering with graphs
- 19:41Collaborative filtering basics
- 22:09Matrix factorization in collaborative filtering
- 24:43Graph-based collaborative filtering
- 27:13Evaluating recommendation systems
- 27:54Q&A session
related talks
30mGraph Machine Learning in All Its Glory!
21mUnleashing the Power of Data Science
38mRecommendation Engine with Neo4j, Piece of Cake
58mRecommender Systems – A Walkthrough
27mUsing Graph Theory to Build a Recommendation Engine in Node js
32mLet's talk about Recommendation Systems!
01h 14mBuilding a Recommendation System with a Hybrid Search From Scratch
01h 31mHands on - Build a Recommender system