Graph Machine Learning in All Its Glory!
summary
Builds a movie recommender as a link‑prediction GNN using Deep Graph Library and GraphSAGE convolutions, training on ~940 users, ~1,700 movies and 100k ratings to regress 1–5 scores with MSE/RMSE. Features enter either as input vectors or as extra categorical nodes; training uses transductive node sets with an 80/20 edge split and negative sampling. Practical pitfalls highlighted: convolution depth equals hop range (oversmoothing on small graphs) and data leakage causing 100% accuracy if splits or implementation details are wrong.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 2:10Why explore graph ML
- 4:36Graphs as real-world structures
- 6:45Neighborhoods and message passing
- 10:37Movie recommendations setup
- 13:27Building GNN with DGL
- 16:21Adding features to models
- 20:11Splitting train/test data
- 21:36Creating graphs from tables
- 23:04Training and evaluating
- 24:10Homogeneous vs heterogeneous graphs
- 25:14Link prediction examples
- 26:57Transductive vs inductive training
- 27:40Choosing convolutional layers
- 28:45Blackbox frameworks in GNNs
- 29:26Summary and Q&A
related talks
01h 31mHands on - Build a Recommender system
33mA Sommelier's Guide to Recommendation Algorithms: Classical and Graph-Based Recommender Systems
23mGraph ML – The Next Level Of Machine Learning; Learn How To Use It And When Not To
50mMachine Learning + Graph Databases for Better Recommendations
29mFurther enhancements of recommender systems using deep learning
21mUnleashing the Power of Data Science
31mGraph-Based Machine Learning with Python
57mHands-On with LLM-Powered Recommenders: Hybrid Architectures