Overview gnn (graph neural network)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
In many ways, graphs are the main modality of data we receive from nature. This is due to the fact that most of the patterns we see, both in natural
and artificial systems, are elegantly representable using the language of graph
structures. Prominent examples include molecules (represented as graphs of
atoms and bonds), social networks and transportation networks. This potential has already been seen by key scientific and industrial groups, with alreadyimpacted application areas including traffic forecasting, drug discovery, social
network analysis and recommender systems. Further, some of the most successful domains of application for machine learning in previous years—images, text
and speech processing—can be seen as special cases of graph representation
learning, and consequently there has been significant exchange of information
between these areas. The main aim of this short survey is to enable the reader
to assimilate the key concepts in the area, and position graph representation
learning in a proper context with related fields

Given such a GNN layer, we can learn (m)any interesting tasks over a graph, by
appropriately combining hu. I exemplify the three principal such tasks, grounded in
biological examples:

+ Node classification. If the aim is to predict targets for each node u ∈ V, then
our output is equivariant, and we can learn a shared classifier directly on hu. A
canonical example of this is classifying protein functions (e.g. using gene ontology
data (Zitnik and Leskovec, 2017)) in a given protein-protein interaction network, as
first done by GraphSAGE (Hamilton et al., 2017).

+ Graph classification. If we want to predict targets for the entire graph, then
we want an invariant output, hence need to first reduce all the hu into a common representation, e.g. by performing L
u∈V hu, then learning a classifier over the resulting
flat vector. A canonical example is classifying molecules for their quantum-chemical
properties (Gilmer et al., 2017), estimating pharmacological properties like toxicity
or solubility (Duvenaud et al., 2015; Xiong et al., 2019; Jiang et al., 2021) or virtual
drug screening (Stokes et al., 2020).

+ Link prediction. Lastly, we may be interested in predicting properties of edges
(u, v), or even predicting whether an edge exists; giving rise to the name “link prediction”. In this case, a classifier can be learnt over the concatenation of features
hukhv, along with any given edge-level features. Canonical tasks include predicting links between drugs and diseases—drug repurposing (Morselli Gysi et al., 2021),
drugs and targets—binding affinity prediction (Lim et al., 2019; Jiang et al., 2020),
or drugs and drugs—predicting adverse side-effects from polypharmacy (Zitnik et al.,
2018; Deac et al., 2019).

Link tham khảo: https://arxiv.org/pdf/2301.08210.pdf

1
2
3
4
5
6
7
8
9
10
11
12
13
14
Transforming initial data into a graph for input to a Graph Neural Network (GNN) can be done using several techniques, depending on the type of data and the desired properties of the resulting graph.

+ If the data is in a structured format such as a table, it can be transformed into a graph by creating a node for each row and an edge between two nodes if their values match a certain criterion.

+ If the data is in an unstructured format such as text, it can be transformed into a graph by creating a node for each word or phrase and an edge between two nodes if they appear in close proximity to each other in the text.

+ If the data is in an image, it can be transformed into a graph by creating a node for each pixel and an edge between two nodes if they are adjacent in the image.

+ If the data is in a time series format, it can be transformed into a graph by creating a node for each time step and an edge between two nodes if they are close in time.

It's important to note that, the choice of representation and the way of creating edges will depend on the specific application and the size of the graph.

Additionally, It's also possible to use pre-trained graph embeddings such as graph2vec or node2vec to represent the graph.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
Jiang W. Graph-based Deep Learning for Communication Networks: A Survey[J]. Computer Communications, 2022, 185:40-54. Link

He S, Xiong S, Ou Y, et al. An overview on the application of graph neural networks in wireless networks[J]. IEEE Open Journal of the Communications Society, 2021. Link

Suárez-Varela J, Almasan P, Ferriol-Galmés M, et al. Graph Neural Networks for Communication Networks: Context, Use Cases and Opportunities[J]. IEEE Network, 2022. Link

Tam P, Song I, Kang S, et al. Graph Neural Networks for Intelligent Modelling in Network Management and Orchestration: A Survey on Communications[J]. Electronics, 2022, 11(20): 3371. Link

Ivanov A, Tonchev K, Poulkov V, et al. Graph-Based Resource Allocation for Integrated Space and Terrestrial Communications[J]. Sensors, 2022, 22(15): 5778. Link

Lee M, Yu G, Dai H, et al. Graph Neural Networks Meet Wireless Communications: Motivation, Applications, and Future Directions[J]. IEEE Wireless Communications, 2022, 29(5): 12-19. Link

Li Y, Xie S, Wan Z, et al. Graph-powered learning methods in the Internet of Things: A survey[J]. Machine Learning with Applications, 2023, 11: 100441. Link

https://mlabonne.github.io/blog/gat/

https://paperswithcode.com/task/graph-classification

https://github.com/jwwthu/GNN4Traffic

https://github.com/jwwthu/GNN-Communication-Networks

https://github.com/jmhIcoding/fgnet

federated learning into graph
https://github.com/huweibo/Awesome-Federated-Learning-on-Graph-and-GNN-papers

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
https://github.com/gorgen2020/SDGCN/tree/main/SDGCN
https://github.com/YanJieWen/STGMT-Tensorflow-implementation
https://github.com/wengwenchao123/DDGCRN
https://github.com/kaist-dmlab/MG-TAR
https://github.com/Bounger2/ST-CGCN
https://github.com/kaist-dmlab/MG-TAR
https://github.com/tsinghua-fib-lab/Traffic-Benchmark
https://github.com/csyanghan/PGECRN
https://github.com/346644054/ST-3DGMR
https://github.com/MathiasNT/NRI_for_Transport
https://github.com/ZikangZhou/QCNet
https://github.com/LMissher/STWave
https://github.com/HKUDS/AutoST
https://github.com/deepkashiwa20/MegaCRN
https://github.com/Echo-Ji/ST-SSL
https://github.com/zhengdaoli/AGC-net
https://github.com/trainingl/STG4Traffic
https://github.com/liuxu77/LargeST
https://github.com/jdcaicedo251/transit_demand_prediction
https://github.com/jwwthu/GNN4Traffic

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
https://github.com/newlei/FairGo
https://github.com/joeybose/Flexible-Fairness-Constraints
https://github.com/akaxlh/KHGT
https://github.com/tsinghua-fib-lab/MBGCN
https://github.com/WHUIR/PPGN
https://github.com/twchen/SEFrame
https://github.com/xiaxin1998/COTREC
https://github.com/RUCAIBox/RecBole/blob/master/recbole/model/sequential_recommender/gcsan.py
https://github.com/userbehavioranalysis/SR-GNN_PyTorch-Geometric
https://github.com/PaddlePaddle/PaddleRec/tree/release/1.8.5/models/recall/gnn/
https://github.com/tsinghua-fib-lab/SIGIR21-SURGE
https://github.com/retagnn/RetaGNN
https://github.com/zhuty16/GES
https://github.com/Coder-Yu/QRec
https://github.com/lcwy220/Social-Recommendation
https://github.com/Wang-Shuo/GraphRec_PyTorch
https://github.com/wenqifan03/GraphRec-WWW19
https://github.com/Kanika91/diffnet
https://github.com/PeiJieSun/diffnet
https://github.com/Wenhui-Yu/LCFN
https://github.com/hanliu95/HS-GCN
https://github.com/jeongwhanchoi/HMLET
https://github.com/wujcan/SGL-TensorFlow
https://github.com/liufancs/IMP_GCN
https://github.com/Tingting2477/DGCF_torch
https://github.com/xiangwang1223/disentangled_graph_collaborative_filtering
https://github.com/gusye1234/LightGCN-PyTorch
https://github.com/gusye1234/LightGCN-PyTorch
https://github.com/tsinghua-fib-lab/GNN-Recommender-Systems