Project A6Resource-Efficient Graph MiningDr. Nils M. Kriege [from 2017], Prof. Dr. Petra Mutzel,Prof. Dr. Christian Sohler, Prof. Dr. Kristian Kersting [until 2017]

Pro

blem

Met

hodo

logy

Res

ults

Informatik 11Algorithm Engineering

Informatik 2Efficient Algorithms and Complexity Theory

Graphs are ubiquitous and complex

Graphs occur, e.g., in chem-, bioinformatics,and social network analysis

Real-world graphs . . .I are largeI are irregularI are sparseI have a hierarchical

structureI have rich information

attached to nodes andedges

Graph classification

How similar are two graphs?

K( ), =?

I No obvious vector representation for graphsI Standard machine learning approaches cannot be

applied directly

Vertex refinementLocal k -dimensional Weisfeiler-Lehman kernel [ICDM 2017]

Local neighbourhood Global neighbourhood

I Parallel sampling algorithm withprovable approximation guarantees

I Considers local as well as globalproperties

Assignments via treesKernels from optimal assign-ments [NIPS 2016]

a

a

a

b

c

a

b

b

c

c

a b c

a

b

c

4 3 1

3 5 1

1 1 2a b c

Assignment Base-kernel Treefg

I Base kernels that induce PSDassignment kernels

I Linear running timeI Weisfeiler-Lehman Optimal

Assignment kernel

Theoretical expressivityProperty testing frameworkfor graph kernels [IJCAI 2018]

ε

Graph property: connectivity

I Current kernels cannot dis-criminate between simplegraph properties

I More expressive kernelbased on k -disks

ENZYMES NCI1 PROTEINS REDDITDatasets

0

20

40

60

80

100

Accu

racy

[%]

Graphlet3-WLLocal 3-WLSam. local 3-WL1-WL

Par. Sampl. LA Exact Par. Exact ExactVariant of local 3-WL

0

1000

2000

3000

4000

5000

6000

7000

Runn

ing

time

[s]

Dataset: PROTEINS

0

10

20

30

40

50

60

70

80

Accu

racy

[%]

aSplineCNN B2

a

Local 3-WL accuracy Speed-up by sampling and parallelism Geometric deep learning framework B2 Cuneiform classification B2

Additional selected resultsScalability

I Spectrum approximation via random walks[KDD 2018]

I Scalable kernels for graphs with continuous labels[ICDM 2016]

I Scalable kernels via propagation[Machine Learning 2016]

I Scaling lifted probabilistic inference and learningvia graph databases [SDM 2016]

ExpressivityI Counting cycles of large graphs [Algorithmica 2018]I Parameterizing the distance distribution of

undirected networks [UAI 2015]I Architecture for tractable multivariate Poisson

distributions [AAAI 2017] B4

I Machine Learning meets data-driven journalism[#Data4Good@ICML 2016] A1

Towards the third phaseI Cuneiform classification on limited

data [COST@SDM 2018] B2

I Hierarchical pooling [NIPS 2018]I Fast deep graph learning via

B-spline kernels [CVPR 2018] B2

I Protein complex similarity[Preprint] C1

I Group equivariant deep learning[NIPS 2018] B2