Graph Classification: A Comparison Study 02/04/19 Presented by: - PowerPoint PPT Presentation

Graph Classification: A Comparison Study 02/04/19 Presented by: Camilo Muñoz Juan Carrillo

Graph Classification Graph Classification: A Comparison Study PAGE 2

Graph Classification Induce a mapping f(x) : X → {±1} given a set of training samples Input graph Classifier Label Action Comedy Romance Sci-fi Graph Classification: A Comparison Study PAGE 3

Graph Classification Kernel methods Embedding Sequential methods methods Graph Classification: A Comparison Study PAGE 4

The Comparison Study ● Kernel CNN (KCNN) ● Deep Graph Kernels (DGK) ● graph2vec >> Embedding ● Multi-hop Assortativity (MHA) Graph Classification: A Comparison Study PAGE 5

Motivation ● Lack of evaluation of graph similarity techniques across categories ● Lack of experimental evaluation regarding multiclass classification https://ls11-www.cs.tu-dortmund.de/staff/morris/graph kerneldatasets Graph Classification: A Comparison Study PAGE 6

Kernel CNN (KCNN) Graph Classification: A Comparison Study PAGE 7

Kernel CNN (KCNN) Graph Classification: A Comparison Study PAGE 8

Deep Graph Kernels Graph Classification: A Comparison Study PAGE 9

Deep Graph Kernels Graph Classification: A Comparison Study PAGE 10

graph2vec Graph Classification: A Comparison Study PAGE 11

graph2vec Graph Classification: A Comparison Study PAGE 12

Multi-hop Assortativity (MHA) Graph Classification: A Comparison Study PAGE 13

Multi-hop Assortativity (MHA) Graph Classification: A Comparison Study PAGE 14

Experimental Evaluation ● RQ1: How can the nature of the graph data (e.g. number of nodes, average number of edges per node) impact the performance of the techniques? ● RQ2: Is there a clear difference in performance when using binary classification datasets versus using multiclass graph data? ● RQ3: Is there a technique that clearly outperforms the others in terms of performance? Graph Classification: A Comparison Study PAGE 15

Microsoft datasets ● Semantic image processing ● The class for each graph corresponds to its semantic meaning. For example building, grass, tree, face, car, bicycle Graph Classification: A Comparison Study PAGE 16

First-MM dataset ● First-MM stands for Flexible Skill Acquisition and Intuitive Robot Tasking for Mobile Manipulation in the Real World ● The graphs represent 3d point clouds of household objects Graph Classification: A Comparison Study PAGE 17

IMDb datasets Reddit datasets ● Movie collaboration graphs ● User discussion datasets ● The class correspond to the genre of ● Binary dataset contains posts from 4 the movie such as Action, Romance, popular subreddits: IAmA, AskReddit, Comedy , and Sci-Fi TrollXChromosomes , and atheism ● Multiclass dataset contains posts from 5 subreddits: worldnews, videos, AdviceAnimals, aww , and mildlyinteresting Graph Classification: A Comparison Study PAGE 18

Experimental Evaluation Graph Classification: A Comparison Study PAGE 19

Experimental Setup ● Dataset preprocessing ● Code customization ● Selection of initialization parameters ● Graph transformation technique ● Graph Classification Graph Classification: A Comparison Study PAGE 20

Evaluation Metrics ● Mean prediction accuracies and standard deviations ● Graph transformation runtime ● Additional storage required for transformed data Graph Classification: A Comparison Study PAGE 21

Evaluation Results ● Mean prediction accuracies and standard deviations Graph Classification: A Comparison Study PAGE 22

Evaluation Results Graph Classification: A Comparison Study PAGE 23

Evaluation Results ● Graph transformation runtime Graph Classification: A Comparison Study PAGE 24

Evaluation Results ● Additional storage required for transformed data Graph Classification: A Comparison Study PAGE 25

Potential Extensions Graph Classification: A Comparison Study PAGE 26

Discussion Graph Classification: A Comparison Study PAGE 27

Appendix PAGE 28

PAGE 29

PAGE 30