Graph Mining with Deep Learning: challenges and pitfalls
summary
Survey of graph embedding methods contrasts matrix factorization and shallow methods with random-walk approaches and highlights practical tradeoffs: DeepWalk and node2vec use biased BFS/DFS walks to capture local vs. global context, optimize low‑dim node vectors with negative sampling, and avoid full adjacency factorization. Claims: random walks scale better than per‑node factorization but require careful walk length, walk count and sampling bias; open problems include incremental updates for evolving graphs, combining node features with structure, and memory cost of one embedding per node.
AI-written summary · about
chapters (28)
- 0:00Start
- 3:29Many Data are Networks
- 3:44Why Networks? Why Now?
- 4:18Machine Learning with Networks
- 4:57Example: Node Classification
- 5:18Example: Link Prediction
- 7:19Why Is It Hard?
- 8:18Embedding Nodes
- 8:28Setup
- 9:22Learning Node Embeddings
- 9:46Two Key Components
- 10:02"Shallow" Encoding
- 10:43How to Define Node Similarity?
- 11:58Main approaches
- 12:33Adjacency-based Similarity
- 14:15Multi-hop Similarity
- 15:44Summary so far
- 16:26Random-walk Embeddings
- 17:07Why Random Walks?
- 17:58Random Walk Optimization
- 19:08Random Walks: Stepping Back
- 19:23How should we randomly walk?
- 19:44node2vec: Biased Walks
- 20:40Experiments: Micro vs. Macro
- 21:12Training the Model
- 22:51Other random walk ideas
- 24:21Conclusion
- 25:47Future challenges
related talks
21mGraph Representation Learning with Deep Embedding Approach
01h 41mGraph Intelligence: Enhance Reasoning and Retrieval Using Graph Analytics
39mUsing Graph Neural Networks to Embrace the Dependency Within Your Data
23mGraph ML – The Next Level Of Machine Learning; Learn How To Use It And When Not To
01h 02mGraph Powering
01h 39mBoot Camp on Graph Learning
31mLet's do graph machine learning on Cloud
31mGraph-Based Machine Learning with Python