Download presentation
Presentation is loading. Please wait.
Published byΝικίας Νικολάκος Modified over 6 years ago
1
NED: An Inter-Graph Node Metric Based On Edit Distance
Haohan Zhu (Facebook, Inc) Xianrui Meng (Apple, Inc) George Kollios (Boston University)
2
Node Similarity Applications
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Node Similarity Applications Transfer Learning Communication Networks Protein–Protein Interaction Networks (PPI) Graph De-anonymization Social Networks Label-Free Inter-Graph Metric From From S. Rajagopala, et al., The binary protein-protein interaction landscape of Escherichia coli, Nature Biotechnology 32, 285–290 (2014) From
3
Existing Node Similarity Measures
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Existing Node Similarity Measures
4
NED: An Inter-Graph Node Metric Based On Edit Distance
08/30/2017 Node Similarity: NED Use the neighborhood topological structure to describe a node. Use Neighborhood Trees rather than Sub-Graphs Computability - Let 𝛿(𝑢, 𝑣) be a distance function between a pair of nodes 𝑢 and 𝑣. The neighborhood subgraphs of 𝑢 and 𝑣 are isomorphic, iff 𝛿(𝑢, 𝑣) = 0. Then 𝛿 is not polynomial-computable. (As hard as graph isomorphism) Effectiveness - Non-identity is not good for top-k similar nodes ranking and k-nearest-neighbor similar nodes finding.
5
𝛿 𝑁𝐸𝐷 (𝑢,𝑣) = 𝛿 𝑇𝐸𝐷 ∗ (𝑇(𝑢,𝑘),𝑇(𝑣,𝑘))
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Tree Edit Distance The similarity between two nodes is measured by the edit distance between neighborhood adjacent-trees 𝛿 𝑁𝐸𝐷 (𝑢,𝑣) = 𝛿 𝑇𝐸𝐷 ∗ (𝑇(𝑢,𝑘),𝑇(𝑣,𝑘)) The computation of tree edit distance on unordered unlabeled trees belongs to NP-Complete, even MaxSNP-Hard.
6
Modified Tree Edit Distance: TED*
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Modified Tree Edit Distance: TED* Modified Tree Edit Distance: No operation can change the depth of any existing node in the tree Edit Operation in TED* : Insert a leaf node Delete a leaf node Move a node at the same level Advantage of TED* ? Metric: The same as TED Efficient: Polynomial time computable
7
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 Iterative Bottom-Up Algorithm
8
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L Level 3 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Padding Padding Lv 3 = 1
9
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L Level 3 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Canonization
10
Bipartite Graph Construction
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L Level 3 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Bipartite Graph Construction
11
Bipartite Graph Matching
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L Level 3 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Bipartite Graph Matching Matching Lv 3 = 0
12
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L Level 3 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Re-Canonization
13
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F Level 2 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Padding Padding Lv 2 = 0
14
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F 𝛿𝛿 2 2 𝜀 𝜁𝜁 1 G H I J K
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F Level 2 𝛿𝛿 2 2 𝜀 𝜁𝜁 1 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Canonization
15
Bipartite Graph Construction
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F Level 2 2 𝛿𝛿 𝜀 2 𝜁𝜁 1 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Bipartite Graph Construction D 𝛿𝛿 1 1 E 𝜀 1 1 1 F 2 𝜁𝜁
16
Bipartite Graph Matching
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F Level 2 𝛿𝛿 2 𝜀 2 1 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Bipartite Graph Matching Matching Lv 2 = 0 D 𝛿𝛿 1 1 E 𝜀 1 1 1 F 2 𝜁𝜁
17
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F 𝛿𝛿 2 2 𝜀 3 𝜁𝜁 1 G H I J
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F Level 2 𝛿𝛿 2 2 𝜀 3 𝜁𝜁 1 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Re-Canonization
18
Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F 𝛿𝛿 2 𝜀 3 𝜁𝜁 1 G H I J K
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C Level 1 𝛽𝛽 𝛾𝛾 D 2 3 E 1 F 𝛿𝛿 2 𝜀 3 𝜁𝜁 1 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Padding Padding Lv 1 = 0
19
Computation of TED* A 𝛼𝛼 B 5 6 C 𝛽𝛽 7 𝛾𝛾 4 D 2 3 E 1 F 𝛿𝛿 2 3 𝜀 1 𝜁𝜁 G
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B 5 6 C Level 1 𝛽𝛽 7 𝛾𝛾 4 D 2 3 E 1 F 𝛿𝛿 2 3 𝜀 1 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Canonization
20
Bipartite Graph Construction
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 5 B 6 C Level 1 7 𝛽𝛽 𝛾𝛾 4 D 2 3 E 1 F 𝛿𝛿 2 3 𝜀 1 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Bipartite Graph Construction B 𝛽𝛽 1 2 2 C 𝛾𝛾 1
21
Bipartite Graph Matching
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 5 B 6 C Level 1 7 𝛽𝛽 𝛾𝛾 4 D 2 3 E 1 F 𝛿𝛿 2 3 𝜀 1 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Bipartite Graph Matching Matching Lv 1 = 1 B 𝛽𝛽 1 2 2 C 𝛾𝛾 1
22
Computation of TED* A 𝛼𝛼 7 B 5 6 4 C 7 𝛽𝛽 𝛾𝛾 4 D 2 3 E 1 F 2 𝛿𝛿 3 𝜀 1
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 7 B 5 6 4 C Level 1 7 𝛽𝛽 𝛾𝛾 4 D 2 3 E 1 F 2 𝛿𝛿 3 𝜀 1 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Re-Canonization
23
Computation of TED* A 𝛼𝛼 B 7 4 C 7 𝛽𝛽 𝛾𝛾 4 D 2 3 E 1 F 𝛿𝛿 2 3 𝜀 𝜁𝜁 1 G
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* Level 0 A 𝛼𝛼 B 7 4 C 7 𝛽𝛽 𝛾𝛾 4 D 2 3 E 1 F 𝛿𝛿 2 3 𝜀 𝜁𝜁 1 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝜇𝜇 Padding Padding Lv 0 = 0
24
𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) = 𝑖=0 𝑘 𝑃 𝑖 + 𝑀 𝑖
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* A 𝛼𝛼 B C 𝛽𝛽 𝛾𝛾 D E F 𝛿𝛿 𝜀 𝜁𝜁 G H I J K L 𝜂𝜂 𝜃𝜃 𝜄 𝜅𝜅 𝜆𝜆 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) = 𝑖=0 𝑘 𝑃 𝑖 + 𝑀 𝑖 𝛿 𝑇𝐸𝐷 ∗ 𝑇 𝐴,3 , 𝑇(𝛼, 3) =2
25
𝑚( 𝐺 𝑖 2 )= 𝑀𝑖𝑛 ∀𝑥𝜖 𝐿 𝑖 (𝑢) 𝑤(𝑥, 𝑓 𝑖 (𝑥))
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 Computation of TED* Canonization Padding 𝑃 𝑖 = 𝐿 𝑖 (𝑢) − 𝐿 𝑖 (𝑣) Matching 𝑀 𝑖 =(𝑚( 𝐺 𝑖 2 )− 𝑃 𝑖+1 )/2 Bipartite Graph Bipartite Matching 𝑚( 𝐺 𝑖 2 )= 𝑀𝑖𝑛 ∀𝑥𝜖 𝐿 𝑖 (𝑢) 𝑤(𝑥, 𝑓 𝑖 (𝑥))
26
NED: An Inter-Graph Node Metric Based On Edit Distance
08/30/2017 TED* vs TED & GED
27
TED* is a metric, so is NED
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 TED* is a metric, so is NED TED* is a metric for trees 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) = 𝑖=0 𝑘 𝑃 𝑖 + 𝑀 𝑖 = 𝑖=1 𝑘 𝑃 𝑖 + 𝑖=0 𝑘−1 𝑚( 𝐺 𝑖 2 )/2 Metric properties: 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) ≥0 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) = 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑣,𝑘),𝑇(𝑢,𝑘) 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) =0, iff 𝑇 𝑢,𝑘 ⋍𝑇(𝑣,𝑘) 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑣,𝑘) ≤ 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑢,𝑘),𝑇(𝑤,𝑘) + 𝛿 𝑇𝐸𝐷 ∗ 𝑇(𝑤,𝑘),𝑇(𝑣,𝑘) NED is a metric for inter-graph nodes
28
TED* vs TED on Efficiency
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 TED* vs TED on Efficiency
29
NED Case Study: De-anonymization
NED: An Inter-Graph Node Metric Based On Edit Distance 08/30/2017 NED Case Study: De-anonymization
30
Thank You !
Similar presentations
© 2025 SlidePlayer.com Inc.
All rights reserved.