Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
임성수 충남대학교 컴퓨터공학과
2018년 한국정보처리학회 추계학술발표대회
충남대학교 컴퓨터공학과 임성수 0212
Research Interests Mining and modeling large-scale complex networks to study
their structural properties (eg identifying community structure)
and dynamical behaviors (eg analyzing information diffusion)
Research Experience
Assistant
Professor
Data Intelligence Lab
Dept of Computer Science and Engineering CNU
2018~now
PhD Data Mining Lab (Advisor Prof Jae-Gil Lee)
Graduate School of Knowledge Service Eng KAIST
2011~2016
Researcher Machine Intelligence Lab (Advisor Prof Kyomin Jung)
School of Computing KAIST
2010~2013
BSMS Statistical Inference Lab (Advisor Prof Sung-Ho Kim)
Dept of Mathematical Sciences KAIST
2004~2011
Brief Bio
Research Projects
Structure of Networks (2013-Present) - Advisor J-G Lee
- Detecting community structure in social networks
- Publications PhD thesis 2 top-tier conferences 3 SCI journals 1 preprint
- Awards Qualcomm Innovation Award (2016) kt Big Contest Award (2014)
Dynamics on Networks (2011-2013) - Advisor K Jung Collaborator JCS Lui (CUHK)
- Analyzing information diffusion in complex networks
- Publications 2 SCI journals 1 preprint
- Award Microsoft Research Asia Fellowship Nomination Award (2012)
Communication Networks (2010-2012) - Advisor K Jung Collaborator M Andrews (Nokia Bell Labs)
- Studying the stability of routing in wireless networks
- Publications 1 top-tier conference 1 SCI journal
- Award Samsung Humantech Paper Award (2012)
Graph Compression Privacy-Preserving Data Analysis (2018-Present) - Principal Investigator Joint work with NOTA Inc Ajou Univ
- Developing efficient learning algorithms for deep neural networks
- Designing algorithms for network data guaranteeing differential privacy
충남대학교 컴퓨터공학과 임성수 0312
1 J Kim et al ldquoLinkBlackHole Robust Overlapping Community Detection Using Link
Embeddingrdquo accepted to IEEE TKDE (SCI)
2 J Kim et al ldquoDifferential Flattening A Novel Framework for Community Detection in
Multi-Layer Graphsrdquo ACM TIST 2017 (SCIE)
3 S Lim and J-G Lee ldquoMotif-Based Embedding for Graph Clusteringrdquo JSTAT 2016 (SCIE)
4 S Lim et al ldquoBlackHole Robust Community Detection Inspired by Graph Drawingrdquo
IEEE ICDE 2016 (BK21+)
5 S Lim et al ldquoPhase Transition for Information Diffusion in Random Clustered Networksrdquo
EPJ B 2016 (SCI)
6 S Lim et al ldquoAnalysis of Information Diffusion for Threshold Models on Arbitrary
Networksrdquo EPJ B 2015 (SCI)
7 S Lim et al ldquoStability of the Max-Weight Protocol in Adversarial Wireless Networksrdquo
IEEEACM TON 2014 (SCI)
8 S Lim et al ldquoLinkSCAN Overlapping Community Detection Using the Link-Space
Transformationrdquo IEEE ICDE 2014 (BK21+)
Selected Publications
충남대학교 컴퓨터공학과 임성수 0412
Topic Analyzing the spread of information in complex networks
Previous Work Limitations on graph structure (tree-like) and states (binary)
Contributions Proposing a generalized mean-field approximation that relaxes the condition on
strong symmetry (arbitrary graphs and multiple adoption states)
Proving that the cascade sizes is highly concentrated around the expected value
with high probability
Highlight 1 Information Diffusion [EPJB 15 16 Preprint]
Irsquoll buy a smart
phone if 60 of
my friends use it
119907 has a threshold 120579119907 and a function 119891119907
119907 becomes active if 120579119907 ge 119891119907(119907primes neighbors)
Information diffusion model Probabilistic method on graphs
119907
충남대학교 컴퓨터공학과 임성수 0512
Topic Identifying overlapping communities in social networks
Previous Work Conventional algorithms usually find disjoint communities
Contributions Proposing the link-space transformation that transforms a given graph into the
link-space graph
Developing an algorithm that performs a non-overlapping clustering on the link-
space graph (easier problem) which enables us to discover overlapping clustering
Highlight 2 Overlapping Clustering [ICDE 14 TKDE 18+]
Link-Space
Transformation
Non-Overlapping Clustering
(with edge sampling)
Membership
Translation
12
34 13
23 35
1
2
3
4
5
0 03
45
1
2
3
4
5
0
12
34 13
23 35
03
45
Original
Graph
Overlapping
Communities
Link
Communities
Link-Space
Graph
충남대학교 컴퓨터공학과 임성수 0612
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
충남대학교 컴퓨터공학과 임성수 0212
Research Interests Mining and modeling large-scale complex networks to study
their structural properties (eg identifying community structure)
and dynamical behaviors (eg analyzing information diffusion)
Research Experience
Assistant
Professor
Data Intelligence Lab
Dept of Computer Science and Engineering CNU
2018~now
PhD Data Mining Lab (Advisor Prof Jae-Gil Lee)
Graduate School of Knowledge Service Eng KAIST
2011~2016
Researcher Machine Intelligence Lab (Advisor Prof Kyomin Jung)
School of Computing KAIST
2010~2013
BSMS Statistical Inference Lab (Advisor Prof Sung-Ho Kim)
Dept of Mathematical Sciences KAIST
2004~2011
Brief Bio
Research Projects
Structure of Networks (2013-Present) - Advisor J-G Lee
- Detecting community structure in social networks
- Publications PhD thesis 2 top-tier conferences 3 SCI journals 1 preprint
- Awards Qualcomm Innovation Award (2016) kt Big Contest Award (2014)
Dynamics on Networks (2011-2013) - Advisor K Jung Collaborator JCS Lui (CUHK)
- Analyzing information diffusion in complex networks
- Publications 2 SCI journals 1 preprint
- Award Microsoft Research Asia Fellowship Nomination Award (2012)
Communication Networks (2010-2012) - Advisor K Jung Collaborator M Andrews (Nokia Bell Labs)
- Studying the stability of routing in wireless networks
- Publications 1 top-tier conference 1 SCI journal
- Award Samsung Humantech Paper Award (2012)
Graph Compression Privacy-Preserving Data Analysis (2018-Present) - Principal Investigator Joint work with NOTA Inc Ajou Univ
- Developing efficient learning algorithms for deep neural networks
- Designing algorithms for network data guaranteeing differential privacy
충남대학교 컴퓨터공학과 임성수 0312
1 J Kim et al ldquoLinkBlackHole Robust Overlapping Community Detection Using Link
Embeddingrdquo accepted to IEEE TKDE (SCI)
2 J Kim et al ldquoDifferential Flattening A Novel Framework for Community Detection in
Multi-Layer Graphsrdquo ACM TIST 2017 (SCIE)
3 S Lim and J-G Lee ldquoMotif-Based Embedding for Graph Clusteringrdquo JSTAT 2016 (SCIE)
4 S Lim et al ldquoBlackHole Robust Community Detection Inspired by Graph Drawingrdquo
IEEE ICDE 2016 (BK21+)
5 S Lim et al ldquoPhase Transition for Information Diffusion in Random Clustered Networksrdquo
EPJ B 2016 (SCI)
6 S Lim et al ldquoAnalysis of Information Diffusion for Threshold Models on Arbitrary
Networksrdquo EPJ B 2015 (SCI)
7 S Lim et al ldquoStability of the Max-Weight Protocol in Adversarial Wireless Networksrdquo
IEEEACM TON 2014 (SCI)
8 S Lim et al ldquoLinkSCAN Overlapping Community Detection Using the Link-Space
Transformationrdquo IEEE ICDE 2014 (BK21+)
Selected Publications
충남대학교 컴퓨터공학과 임성수 0412
Topic Analyzing the spread of information in complex networks
Previous Work Limitations on graph structure (tree-like) and states (binary)
Contributions Proposing a generalized mean-field approximation that relaxes the condition on
strong symmetry (arbitrary graphs and multiple adoption states)
Proving that the cascade sizes is highly concentrated around the expected value
with high probability
Highlight 1 Information Diffusion [EPJB 15 16 Preprint]
Irsquoll buy a smart
phone if 60 of
my friends use it
119907 has a threshold 120579119907 and a function 119891119907
119907 becomes active if 120579119907 ge 119891119907(119907primes neighbors)
Information diffusion model Probabilistic method on graphs
119907
충남대학교 컴퓨터공학과 임성수 0512
Topic Identifying overlapping communities in social networks
Previous Work Conventional algorithms usually find disjoint communities
Contributions Proposing the link-space transformation that transforms a given graph into the
link-space graph
Developing an algorithm that performs a non-overlapping clustering on the link-
space graph (easier problem) which enables us to discover overlapping clustering
Highlight 2 Overlapping Clustering [ICDE 14 TKDE 18+]
Link-Space
Transformation
Non-Overlapping Clustering
(with edge sampling)
Membership
Translation
12
34 13
23 35
1
2
3
4
5
0 03
45
1
2
3
4
5
0
12
34 13
23 35
03
45
Original
Graph
Overlapping
Communities
Link
Communities
Link-Space
Graph
충남대학교 컴퓨터공학과 임성수 0612
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Research Projects
Structure of Networks (2013-Present) - Advisor J-G Lee
- Detecting community structure in social networks
- Publications PhD thesis 2 top-tier conferences 3 SCI journals 1 preprint
- Awards Qualcomm Innovation Award (2016) kt Big Contest Award (2014)
Dynamics on Networks (2011-2013) - Advisor K Jung Collaborator JCS Lui (CUHK)
- Analyzing information diffusion in complex networks
- Publications 2 SCI journals 1 preprint
- Award Microsoft Research Asia Fellowship Nomination Award (2012)
Communication Networks (2010-2012) - Advisor K Jung Collaborator M Andrews (Nokia Bell Labs)
- Studying the stability of routing in wireless networks
- Publications 1 top-tier conference 1 SCI journal
- Award Samsung Humantech Paper Award (2012)
Graph Compression Privacy-Preserving Data Analysis (2018-Present) - Principal Investigator Joint work with NOTA Inc Ajou Univ
- Developing efficient learning algorithms for deep neural networks
- Designing algorithms for network data guaranteeing differential privacy
충남대학교 컴퓨터공학과 임성수 0312
1 J Kim et al ldquoLinkBlackHole Robust Overlapping Community Detection Using Link
Embeddingrdquo accepted to IEEE TKDE (SCI)
2 J Kim et al ldquoDifferential Flattening A Novel Framework for Community Detection in
Multi-Layer Graphsrdquo ACM TIST 2017 (SCIE)
3 S Lim and J-G Lee ldquoMotif-Based Embedding for Graph Clusteringrdquo JSTAT 2016 (SCIE)
4 S Lim et al ldquoBlackHole Robust Community Detection Inspired by Graph Drawingrdquo
IEEE ICDE 2016 (BK21+)
5 S Lim et al ldquoPhase Transition for Information Diffusion in Random Clustered Networksrdquo
EPJ B 2016 (SCI)
6 S Lim et al ldquoAnalysis of Information Diffusion for Threshold Models on Arbitrary
Networksrdquo EPJ B 2015 (SCI)
7 S Lim et al ldquoStability of the Max-Weight Protocol in Adversarial Wireless Networksrdquo
IEEEACM TON 2014 (SCI)
8 S Lim et al ldquoLinkSCAN Overlapping Community Detection Using the Link-Space
Transformationrdquo IEEE ICDE 2014 (BK21+)
Selected Publications
충남대학교 컴퓨터공학과 임성수 0412
Topic Analyzing the spread of information in complex networks
Previous Work Limitations on graph structure (tree-like) and states (binary)
Contributions Proposing a generalized mean-field approximation that relaxes the condition on
strong symmetry (arbitrary graphs and multiple adoption states)
Proving that the cascade sizes is highly concentrated around the expected value
with high probability
Highlight 1 Information Diffusion [EPJB 15 16 Preprint]
Irsquoll buy a smart
phone if 60 of
my friends use it
119907 has a threshold 120579119907 and a function 119891119907
119907 becomes active if 120579119907 ge 119891119907(119907primes neighbors)
Information diffusion model Probabilistic method on graphs
119907
충남대학교 컴퓨터공학과 임성수 0512
Topic Identifying overlapping communities in social networks
Previous Work Conventional algorithms usually find disjoint communities
Contributions Proposing the link-space transformation that transforms a given graph into the
link-space graph
Developing an algorithm that performs a non-overlapping clustering on the link-
space graph (easier problem) which enables us to discover overlapping clustering
Highlight 2 Overlapping Clustering [ICDE 14 TKDE 18+]
Link-Space
Transformation
Non-Overlapping Clustering
(with edge sampling)
Membership
Translation
12
34 13
23 35
1
2
3
4
5
0 03
45
1
2
3
4
5
0
12
34 13
23 35
03
45
Original
Graph
Overlapping
Communities
Link
Communities
Link-Space
Graph
충남대학교 컴퓨터공학과 임성수 0612
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
1 J Kim et al ldquoLinkBlackHole Robust Overlapping Community Detection Using Link
Embeddingrdquo accepted to IEEE TKDE (SCI)
2 J Kim et al ldquoDifferential Flattening A Novel Framework for Community Detection in
Multi-Layer Graphsrdquo ACM TIST 2017 (SCIE)
3 S Lim and J-G Lee ldquoMotif-Based Embedding for Graph Clusteringrdquo JSTAT 2016 (SCIE)
4 S Lim et al ldquoBlackHole Robust Community Detection Inspired by Graph Drawingrdquo
IEEE ICDE 2016 (BK21+)
5 S Lim et al ldquoPhase Transition for Information Diffusion in Random Clustered Networksrdquo
EPJ B 2016 (SCI)
6 S Lim et al ldquoAnalysis of Information Diffusion for Threshold Models on Arbitrary
Networksrdquo EPJ B 2015 (SCI)
7 S Lim et al ldquoStability of the Max-Weight Protocol in Adversarial Wireless Networksrdquo
IEEEACM TON 2014 (SCI)
8 S Lim et al ldquoLinkSCAN Overlapping Community Detection Using the Link-Space
Transformationrdquo IEEE ICDE 2014 (BK21+)
Selected Publications
충남대학교 컴퓨터공학과 임성수 0412
Topic Analyzing the spread of information in complex networks
Previous Work Limitations on graph structure (tree-like) and states (binary)
Contributions Proposing a generalized mean-field approximation that relaxes the condition on
strong symmetry (arbitrary graphs and multiple adoption states)
Proving that the cascade sizes is highly concentrated around the expected value
with high probability
Highlight 1 Information Diffusion [EPJB 15 16 Preprint]
Irsquoll buy a smart
phone if 60 of
my friends use it
119907 has a threshold 120579119907 and a function 119891119907
119907 becomes active if 120579119907 ge 119891119907(119907primes neighbors)
Information diffusion model Probabilistic method on graphs
119907
충남대학교 컴퓨터공학과 임성수 0512
Topic Identifying overlapping communities in social networks
Previous Work Conventional algorithms usually find disjoint communities
Contributions Proposing the link-space transformation that transforms a given graph into the
link-space graph
Developing an algorithm that performs a non-overlapping clustering on the link-
space graph (easier problem) which enables us to discover overlapping clustering
Highlight 2 Overlapping Clustering [ICDE 14 TKDE 18+]
Link-Space
Transformation
Non-Overlapping Clustering
(with edge sampling)
Membership
Translation
12
34 13
23 35
1
2
3
4
5
0 03
45
1
2
3
4
5
0
12
34 13
23 35
03
45
Original
Graph
Overlapping
Communities
Link
Communities
Link-Space
Graph
충남대학교 컴퓨터공학과 임성수 0612
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Topic Analyzing the spread of information in complex networks
Previous Work Limitations on graph structure (tree-like) and states (binary)
Contributions Proposing a generalized mean-field approximation that relaxes the condition on
strong symmetry (arbitrary graphs and multiple adoption states)
Proving that the cascade sizes is highly concentrated around the expected value
with high probability
Highlight 1 Information Diffusion [EPJB 15 16 Preprint]
Irsquoll buy a smart
phone if 60 of
my friends use it
119907 has a threshold 120579119907 and a function 119891119907
119907 becomes active if 120579119907 ge 119891119907(119907primes neighbors)
Information diffusion model Probabilistic method on graphs
119907
충남대학교 컴퓨터공학과 임성수 0512
Topic Identifying overlapping communities in social networks
Previous Work Conventional algorithms usually find disjoint communities
Contributions Proposing the link-space transformation that transforms a given graph into the
link-space graph
Developing an algorithm that performs a non-overlapping clustering on the link-
space graph (easier problem) which enables us to discover overlapping clustering
Highlight 2 Overlapping Clustering [ICDE 14 TKDE 18+]
Link-Space
Transformation
Non-Overlapping Clustering
(with edge sampling)
Membership
Translation
12
34 13
23 35
1
2
3
4
5
0 03
45
1
2
3
4
5
0
12
34 13
23 35
03
45
Original
Graph
Overlapping
Communities
Link
Communities
Link-Space
Graph
충남대학교 컴퓨터공학과 임성수 0612
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Topic Identifying overlapping communities in social networks
Previous Work Conventional algorithms usually find disjoint communities
Contributions Proposing the link-space transformation that transforms a given graph into the
link-space graph
Developing an algorithm that performs a non-overlapping clustering on the link-
space graph (easier problem) which enables us to discover overlapping clustering
Highlight 2 Overlapping Clustering [ICDE 14 TKDE 18+]
Link-Space
Transformation
Non-Overlapping Clustering
(with edge sampling)
Membership
Translation
12
34 13
23 35
1
2
3
4
5
0 03
45
1
2
3
4
5
0
12
34 13
23 35
03
45
Original
Graph
Overlapping
Communities
Link
Communities
Link-Space
Graph
충남대학교 컴퓨터공학과 임성수 0612
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Topic Identifying highly interconnected communities in social networks
Previous Work Conventional algorithms are often not robust to high mixing
Contributions Proposing the BlackHole transformation that transforms a given graph into the
points in a low-dimensional space
Developing an algorithm that performs clustering on the embedded space which
enables us to discover highly mixed communities
Highlight 3 Graph Embedding [ICDE 16 JSTAT 16 TKDE 18+]
BlackHole
Transformation
Point
Clustering
Membership
Translation
Communities
of Vertices
Communities
of Positions
Positions in
a Space
Original
Graph
충남대학교 컴퓨터공학과 임성수 0712
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Research Topics
Understanding complex systems with big data analytics
Developing algorithms for solving intelligence and real-world data problems
Recent Topics
Big Data Analytics Privacy-Preserving Methods
Network Science Graph Compression
Artificial Intelligence Statistical Inference
Research Plans
충남대학교 컴퓨터공학과 임성수 0812
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Goal Developing network analysis using higher-order relationships
Methods Considering higher-order graph substructures called the network motifs or graphlets
is important to capture the structural dependencies in networks
Using the motif-based weighting and graph embedding method we are working on
interesting problems for big graph data management and analysis
A
B C
119860 119861 friends
119860 119862 friends
119861 119862 likely to
become friends
Triadic closure (a concept from sociology)
A
B C
Motifs in biological networks
On-Going Higher-Order Network Analysis
충남대학교 컴퓨터공학과 임성수 0912
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Goal Developing effective privacy-preserving techniques for network data
Methods Increasing data complexity (volume variety velocity) due to data sources and sizes
causes the privacy concerns about big data
Using the differential privacy a mathematical framework for protecting data privacy
we develop data mining models that achieve good utility and privacy protection
On-Going Privacy-Preserving Network Analysis
1198631 1198632 differ in the data of a person (element)
P 119860 1198631 isin 119878 le 119890120598P 119860 1198632 isin 119878 + 120575 for all 119878
119860 Differential private algorithm Designing privacy-preserving algorithm
충남대학교 컴퓨터공학과 임성수 1012
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Goal Developing efficient learning algorithms for deep neural networks
Methods Designing and analyzing iterative pruning algorithm for deep neural networks
that is fast with quite good accuracy
Using the proposed algorithm we develop an on-device deep learning platform
for the mobile and embedded applications
On-Going Graph Compression
Iterative pruning
Remove negligibly important
connections with low weights to
produce an approximation
충남대학교 컴퓨터공학과 임성수 1112
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212
Thank You Very Much
Any Questions
충남대학교 컴퓨터공학과 임성수 1212