Discriminate Networks Using a Few Graphlets

Reproducibility of the results from ''A Simple Embedding for Classifying Networks with a few Graphlets''

Generation of Tables and Figures

Generate Table I

run stats.m

Generate Table II

PCA :

run run_pca.m

To change the normalisation, modify global variable type_norm (line 43)

LDA :

run run_lda.m

To change the normalisation, modify global variable type_norm (line 43)

Generate Table III

run run_pca.m

To change the kind of k-node graphlets to be used, modify motif3/4/5 (lines 16-18)

Generate Table IV

run run_graph2vec.m

To change the size of the embeddings, modify the global variable t_emb (line 21)

Precomputed embedding sizes are 2, 4, 8, 16, 32, 64, 128, 256, 512.

To change the depth of the WL kernel, modify the global variable t_wl (line 21)

Precomputed WL depths are 1, 2, 3

Generate Table V

Go in the GCNs folder and create the dataset : python3 create_graphtest.py

GCNs

Train and run the neural network :python3 run_GCN.py -pooling max/mean/sum -hidd 4/8/12 -nbLay 1/2/3

  You may also change the variable ``nbItes`` (line 15), so that ``nbItes = 10``

Obtain scores : python3 analysis.py GCN_poolmax/mean/sum_nbLay1/2/3_hidd4/8/12

RGCNs

Train and run the neural network : python3 run_RGCN.py -pooling max/mean/sum -hidd 4/8/12 -nbLay 1/2/3

  You may also change the variable ``nbItes`` (line 15), so that ``nbItes = 10``

Obtain scores : python3 analysis.py GCN_poolmax/mean/sum_nbLay1/2/3_hidd4/8/12

Generate Table VI

Our method :

run run_pca.m

Gl2Vec :

run run_gl2vec.m

Graph2Vec :

run run_graph2vec.m

GCNs (takes a long time)

Train and run the neural network : python3 run_GCN.py

  Change the variable ``nbItes = 50`` (line 15)

Obtain scores : python3 analysis.py GCN_poolmean_nbLay2_hidd12

RGCNs (takes a long time)

Train and run the neural network : python3 run_RGCN.py

  Change the variable ``nbItes = 50`` (line 15)

Obtain scores : python3 analysis.py RGCN_poolmean_nbLay2_hidd8

Generate Table VII

run run_pca_wiki.m

Generate Figures VI, VII, VIII

run run_Wiki_figures.m

Generate Figure IX

Gamma-score :

run run_GammaAnalysis.m

RF-score :

run run_RFAnalysis.m

Generate Table VIII

Our method :

run run_feature_select.m

RF feature selection :

run run_RFAnalysis.m

Generate Figure X and Confusion Matrices from Table X

Figures Affinity Matrix and Threshold Sparsification / Confusion Matrix to the Left :

run afty_threshold.m

Figure Closest Neighbour Sparsification / Confusion Matrix to the Right :

run afty_knn.m

Architecture :

Folder Stock :

Matlab files, scripts and functions used by the main run_* scripts

Processing/Networks :

Networks used in the tests, on the following format :

# one or several lines
# that give indication about
# the network
!n:number_of_nodes
!m:number_of_egdes
v_src1 v_tgt1
v_src2 v_tgt2
...

Processing/Wiki_Graph_remove_40_0_500 :

Wikipedia Networks, on the following format :

# Article : Name_of_the_Wikipedia_Article
# FROM date_of_the_first_version_of_interest TO date_of_the_last_version_of_interest
number_of_nodes number_of_edges
v_src1 v_tgt1
v_src2 v_tgt2
...

MatNetworks :

Each mat file contains a struct Pbm that contains information about the network:

-> Pbm.entete : textual information (website, preprocessing, etc.)
-> Pbm.nb_nodes/nb_edges : number of nodes/edges (a bidirected edges counts for two edges)
-> Pbm.motif3 : a matrix 13x2. Pbm.motif3(k,1) : id of 3-node kth motif
                               Pbm.motif3(k,2) : occurrence number of motif k in the network
-> Pbm.motif4 : same for 4-node motif
-> Pbm.motif5 : same for 5-node motif (does not exist for all networks)
-> Pbm.edges : a matrix Pbm.nb_edges x 2 where (Pbm.edges(i,1), Pbm.edges(i,2)) = (v_srci,v_tgti)

Gl2Vec :

MotifCounts

Outputs of the java code from https://github.com/kuntu/JGraphlet-JMotif for our benchmarks.

EmbMat

The SRPs of each networks in Matlab files (generated using convert2Mat.m)

Graph2Vec :

A Python code using NetworkX and karate-club Benchmarks to generate the embeddings using graph2vec, for deep of WL-kernel from 1 to 4 and embedding size from 2 to 512.

Embedding

Outputs of the Python code

EmbMat

The corresponding Matlab files (generated using convert_csv2mat.m)

FeatureSelectionRF

A Python Code to obtain the average Gini importance of each graphlets by training a forest of 100 trees (using scikit-learn).

Clusterix :

Matlab files to run the unsupervised clustering algorithm used in Section V. (See https://pdfs.semanticscholar.org/6235/cf4b551f768fa793ed759de75f2a01475e77.pdf for the algorithm description).