PHH_Graph_Representation_Learning_11092020

Transcript

1. Graph Representation Learning An introduction Phan Hoang - AI Research Team
2. Do we need deep graph neural network?
3. Graph Representation Learning?!
4. Graph Representation Learning?! - Viblo recommender engine - Graph-based recommender system for paper citation network - https://api.semanticscholar.org/v1/paper/arXiv:2006.07739 - https://jsoneditoronline.org/#left=cloud.f3946e860ca54e98934899f6f30eb475 - https://www.connectedpapers.com/
5. Graph Representation Learning?!
6. Graph Representation Learning?!
7. Prerequisite knowledge - https://viblo.asia/p/6J3ZgP0qlmB#_1-so-ly-thuyet-do-thi-co-ban-1
8. Tasks - https://viblo.asia/p/6J3ZgP0qlmB#_1-so-bai-toan-dien-hinh-2 - Node Classification - Link Prediction - Graph Clustering & community detection - ...
9. Node Embedding - Random Walk / DeepWalk - Word2Vec
10. Node Embedding - Random Walk
11. Node Embedding - Skip-gram model
12. DeepWalk - karate dataset
13. Node Embedding - Cons - Unseen node? - Node feature? - Tranductive learning vs inductive learning
14. Graph Neural Network - an introduction
15. Example of GCN model
16. Example of GCN model
17. Graph Convolution Network - Kipf (2016)
18. GraphSage - L.Hamilton (2017)
19. GraphSage - L.Hamilton (2017)
20. GraphSage - inductive learning
21. The number of layers
22. GraphSage - L.Hamilton (2017) - Loss function, un-supervised learning
23. How to apply?
24. Real use-case applications - PinSage / UberEat
25. Real use-case applications - PinSage / UberEat - Based on GraphSage paper (Haminton / 2017) - Heterogeneous graph / bipartite graph - Node Embedding: image + text feature, user info - Edge: Pin-Board, User-Dish, User-Restaurant relation
26. Real use-case applications - Decagon
27. Real use-case applications - Decagon
28. Real use-case applications - Goal-directed generation
29. How to model the GNN for a specific dataset?
30. Task: Text Classification
31. Task: Relation Extraction / NLP
32. Task: Key-Information Extraction
33. Task: Recommender System
34. Other tasks - 3D Object Detection - Action Recognition / Pose Estimation - GAN / VAE - GAT / Graph Transformer - Feature Matching - Key Information Extraction - Scene Graph Generation - Recommender System - ...
35. Papers - Cluster-GCN / GraphSAINT - graph/sub-graph/node sampling - GAT (graph attention network) - GIN (graph isomorphism network) - Deep-GCN / Deeper-GCN - going deeper with GCN - ...
36. Cons - non-injective aggregate function
37. Cons - adversarial attack
38. Thank you! https://viblo.asia/p/deep-learning-graph-neural-network-a-literature-review-and-applications-6J3ZgP0qlmB