Home

TFM Wiki

Summary tables
Graph Neural Network models
- Graph Neural Network: scenario, arch, training, non neural alternatives
- Graph Convolutional Network
- Gated Graph
- MPNN
- GraphSAGE
- GIN
Datasets
Theoretical background
- Linear Algebra
- Graph Theory (isomorphisms vs homomorphisms)
- Network metrics & theory
- Embeddings
- Kernels
- Random Walks
- Matrix Factorization
- Spectral Matrix analysis
Alternatives to GNN
- Matrix factorization
- Random Walks
- Graph Kernels
Applications
- Biochemistry
- Recommendation Systems
- Code analysis
- Relational learning
- Reinforcement learning
Open Research topics on Graph Neural Networks
- combinatorial generalization
- size invariance
- ...
Research & Experiments
References
Tasks

Summary tables

Algorithm types

Algorithm	Description	Type	ML Tasks	Paper
GNN	Forward iterative convergence pass + backprop iterative conv. pass	Spectral based?	node/sub-graph/graph classification	The Graph Neural Network Model
GCN	Aggregation of node neighborhood	Spatial based method	node classification, embedding	Semi-Supervised Classification with Graph Convolutional Networks

Implementations

Algorithm	Source code	Datasets	Tests	Results
GNN				Slow
GCN				Faster but still not scalable

Applications

Algorithm	Applications	Research topics
GNN
GCN

Datasets

Category	Dataset	dimensions	Source
Citation networks
	Cora	1 graph, 2708 nodes, 5429 edges, 1433 features, 7 labels
	Citeseer	1 graph, 3327 nodes, 4732 edges, 3703 features, 6 labels
	Pubmed	1 graph, 19717 nodes, 44338 edges, 500 features, 3 labels
	DBLP
Social networks
	DBBlogCatalogLP	1 grph 10312 nodes, 333983 edges, - , 39 labels
	Reddit	1 graph, 232965 nodes, 11606919 edges, 602 features, 41 labels
	Epinions		www.epinions.com
Chemical/Biological
	PPI	24 graphs, 56944 nodes, 818716 edges, 50 features, 121 labels
	NCI-1	4100 graphs, 37 features, 2 labels
	NCI-109	4127 graphs, 38 features, 2 labels
	MUTAG	188 graphs, 7 features, 2 labels,
	D&D	1178 graphs 2 labels
	QM9	133885 graphs 13 labels
	tox21	12707 graphs 12 labels	tripod.nih.gov/
tox21/challenge/
	PROTEINS	3 labels
	PTC	19 labels
Unstructured graphs
	MNIST	70000 graphs 10 labels	yann.lecun.com/exdb/mnist/
	Wikipedia	1 graph 4777 nodes, 184812 edges, 40 labels	www.mattmahoney.net/dc/textdata ? dbpedia link?
	20NEWS	1 graph, 18846 nodes 20 labels
Others
	METR-LA
	Movie-Lens1M	1 graph 10000 ndoes 1 Million edges	grouplens.org/datasets/movielens/1m/
	Nell	1 graph 65755 nodes, 266144 edges, 61278 features 210 labels

PENDING: http://web.stanford.edu/class/cs224w/data.html

Graph/ML Tasks

Task	Description	Algorithms
Node classification	semi-superv. + transductive	convolutional, non-convolutional (=spectral?)
Node classification	semi-superv. + inductive	convolutional
Node classification	supervised + inductive	convolutional
(Sub)Graph classification		convolutional, non-convolutional
Link prediction
Graph generation
Node/network embedding
Spatial-temporal forecasting

Research topics

Research Topic	Description	Links
Scalability	not for large graphs: multiple layers of gcn =node's final state involves large num of it's neighbors = complex for backpropagation. Also embedding of each node fitting in memory is not realistic	A Comprehensive Survey on Graph Neural Networks Representation Learning on Graphs: Methods and Applications
Deep or not?	CNN go deep succesfully, GNN don't. It is still not understood why	A Comprehensive Survey on Graph Neural Networks
Receptive Field	how to select a representative receptivel field(neigborhood)	A Comprehensive Survey on Graph Neural Networks
Dynamics and heterogeneity	static structures and from a single source: not realistic, ex social networks sign ups, recommender systems product types	A Comprehensive Survey on Graph Neural Networks
Dynamic, temporal graphs	incorporate timing info about edges in embeddings for example. Exciting topic	Representation Learning on Graphs: Methods and Applications
Decoding higher-order motifs	pairwise decoders for generated node embeddings assuming pairwise node relationships. Higher order graph structures involving more than two nodes are needed	Representation Learning on Graphs: Methods and Applications
Reasoning about large sets of candidate subgraphs	subgraph embedding limitation = require the traget subgraphs to be pre-specified before learning process. Needs: discover subgraphs, reason over the combinatorially large space of possible candidate subgraphs	Representation Learning on Graphs: Methods and Applications
Interpretabilty	interpreations and underlying biases of embeddings are unknown. Ensure models are learning to represent relevant graph info and not just exploiting statistical tendencies of benchmarks	Representation Learning on Graphs: Methods and Applications Relational inductive biases, deep learning, and graph networks
how to extract graphs from any kind of date	Example: gnn for raw sensory data? Unless DNN that succeeds, we don't know how to force the sparsity of a graph and yet compute a good model	Relational inductive biases, deep learning, and graph networks
GraphSAGE - incorporate directed or multi-modal graphs		Hamilton et al., Inductive Representation Learning on Large Graphs . NIPS , 2017
GraphSAGE - explore non-uniform neighborhood sampling functions		Hamilton et al., Inductive Representation Learning on Large Graphs . NIPS , 2017
GraphSAGE - Optimize neighborhood sampling functions during training		Hamilton et al., Inductive Representation Learning on Large Graphs . NIPS , 2017