Heterogeneous Subgraph Features for Information Networks Andreas - - PowerPoint PPT Presentation
Heterogeneous Subgraph Features for Information Networks Andreas Spitz , Diego Costa, Kai Chen, Jan Greulich, Johanna Gei, Stefan Wiesberg, and Michael Gertz June 10, 2018 GRADES-NDA, Houston, Texas, USA Heidelberg University, Germany
Heidelberg University, Germany Database Systems Research Group
◮ Scientific publication networks ◮ Knowledge bases ◮ Metabolic networks ◮ · · · 1
◮ Scientific publication networks ◮ Knowledge bases ◮ Metabolic networks ◮ · · ·
1
◮ Scientific publication networks ◮ Knowledge bases ◮ Metabolic networks ◮ · · ·
◮ With features, of course 1
◮ Scientific publication networks ◮ Knowledge bases ◮ Metabolic networks ◮ · · ·
◮ With features, of course ◮ But how do you get the features? 1
◮ Require domain knowledge ◮ Are time-consuming to engineer ◮ Require metadata that may not be available 2
◮ Require domain knowledge ◮ Are time-consuming to engineer ◮ Require metadata that may not be available
◮ Sample neighbourhoods through random walks ◮ Require extensive parameter tuning 2
◮ Require domain knowledge ◮ Are time-consuming to engineer ◮ Require metadata that may not be available
◮ Sample neighbourhoods through random walks ◮ Require extensive parameter tuning
2
◮ Encode neighbourhood information ◮ Are extremely diverse in heterogeneous networks 3
◮ Encode neighbourhood information ◮ Are extremely diverse in heterogeneous networks
3
◮ Encode neighbourhood information ◮ Are extremely diverse in heterogeneous networks
3
◮ Encode neighbourhood information ◮ Are extremely diverse in heterogeneous networks
3
◮ Encode neighbourhood information ◮ Are extremely diverse in heterogeneous networks
3
4
◮ Explore the local neighbourhood around each node ◮ Represent subgraphs by their characteristic string ◮ Count subgraphs by hashing the characteristic string ◮ Use the counts of subgraphs as node features 5
◮ Explore the local neighbourhood around each node ◮ Represent subgraphs by their characteristic string ◮ Count subgraphs by hashing the characteristic string ◮ Use the counts of subgraphs as node features
◮ Encode each node as a block ◮ Blocks start with the node label ◮ Subsequent entries denote neighbours of all given labels ◮ Blocks are sorted lexicographically 5
◮ Are a pseudo-canonical encoding ◮ May result in colliding encodings 6
◮ Are a pseudo-canonical encoding ◮ May result in colliding encodings
◮ Can only be enumerated (no closed formula) ◮ Depend on the network structure and the labels ◮ Have negligible frequency in practice 6
◮ Skewed degree distributions ◮ Highly connected nodes (hubs) 7
◮ Skewed degree distributions ◮ Highly connected nodes (hubs)
◮ Feature extraction time is strongly increased ◮ Random walks retrieve non-local information 7
◮ Skewed degree distributions ◮ Highly connected nodes (hubs)
◮ Feature extraction time is strongly increased ◮ Random walks retrieve non-local information
7
◮ Heterogeneous network ◮ Some nodes with missing labels
◮ Missing node labels 8
◮ Heterogeneous network ◮ Some nodes with missing labels
◮ Missing node labels
◮ Model as a classification task using logistic regression ◮ Evaluate with F1-score 8
◮ Star-shaped structure around movies ◮ Low edge density
◮ Intermediate structure ◮ Papers form the core component
◮ Cooccurrences of named entities in text ◮ Strongly connected structure ◮ High edge density 9
◮ Maximum number of edges: 5 ◮ No exploration beyond 10% of highest degree nodes ◮ Masked starting node label
◮ DeepWalk ◮ LINE ◮ node2vec 10
11
0.2 0.3 0.4 0.5 0.6 10% 30% 50% 70% 90%
Subgraph node2vec DeepWalk LINE
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 10% 30% 50% 70% 90%
Subgraph node2vec DeepWalk LINE
0.2 0.3 0.4 0.5 0.6 0.7 0.8 10% 30% 50% 70% 90%
Subgraph node2vec DeepWalk LINE
12
0.2 0.3 0.4 0.5 0.6 0% 15% 30% 45% 60% 75%
Subgraph node2vec DeepWalk LINE
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0% 15% 30% 45% 60% 75%
Subgraph node2vec DeepWalk LINE
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0% 15% 30% 45% 60% 75%
Subgraph node2vec DeepWalk LINE
13
◮ Scientific publication network ◮ A range of years ◮ A set of conferences
◮ For upcoming conferences ◮ By accepted papers ◮ For the next conference
14
◮ Scientific publication network ◮ A range of years ◮ A set of conferences
◮ For upcoming conferences ◮ By accepted papers ◮ For the next conference
◮ Model as a regression task for the institution relevance score ◮ Evaluate with normalized discounted cumulative gain (NDCG20)
14
◮ Institutions I ◮ Authors A ◮ Papers P ◮ Publication data from 2011 - 2016
◮ Focus on 5 conferences
◮ Use citations to a depth of 3 15
◮ Previous relevance scores, publication counts, etc. (8) ◮ Linguistic features (32)
◮ Maximum number of edges: 5 ◮ No maximum degree exploration limit
◮ DeepWalk ◮ LINE ◮ node2vec 16
0.00 0.20 0.40 0.60 0.80 1.00
Random Forest
0.00 0.20 0.40 0.60 0.80 1.00
Linear Regression Classic Subgraphs Combined node2vec DeepWalk LINE
0.00 0.20 0.40 0.60 0.80 1.00
Decision Tree
0.00 0.20 0.40 0.60 0.80 1.00
KDD FSE ICML MM MOBICOM Bayesian Ridge
17
18
19
◮ Extracted by local exploration and enumeration ◮ Avoid isomorphism test by encoding degree sequences 20
◮ Extracted by local exploration and enumeration ◮ Avoid isomorphism test by encoding degree sequences
◮ Similar performance ◮ Require no domain knowledge for extraction ◮ No engineering process necessary 20
◮ Extracted by local exploration and enumeration ◮ Avoid isomorphism test by encoding degree sequences
◮ Similar performance ◮ Require no domain knowledge for extraction ◮ No engineering process necessary
◮ Beter predictive performance ◮ Longer extraction time 20
◮ C++ (core extraction routines) ◮ Python (wrapper)
21
◮ C++ (core extraction routines) ◮ Python (wrapper)
21