temporal graphspeople.cs.vt.edu/~badityap/classes/cs6604-fall17/student... · 2017-12-15 · graph...
TRANSCRIPT
Temporal GraphsSanket Lokegaonkar
CS6604: Data Mining Large Networks
Fall 2017Virginia Tech
Outline• Motivation
• Categories
• Yu, Wenchao, Charu C. Aggarwal, and Wei Wang. "Temporally Factorized Network Modeling for Evolutionary Network Analysis.“
• Gu, Yupeng, Yizhou Sun, and Jianxi Gao. "The Co-Evolution Model for Social Network Evolving and Opinion Migration."
• Amiri, Sorour E., Liangzhe Chen, and B. Aditya Prakash. "SnapNETS: Automatic Segmentation of Network Sequences with Node Labels."
• Take-aways
Motivation
Graph Definitions
• Temporal Graphs: • G= { N, A(t) } where N is the set of nodes in the graph and A(t) represents a
set of adjacency matrices as a function of time
• For unweighted networks, the adjacency network is binary
• For weighted networks, A(t) represents adjacency matrices with weights changing over time.
Ubiquity of Temporal Graphs [1] [2]
IRIS Co-author NetworkTwitter Follower-FolloweeNetwork
And many more,Such as disease spread network
Applications
• Link Prediction with Temporal Context• Which people would you like to be friends with, given you have been
unfriending certain people?
• Community Evolution Dynamics• Which communities grow? Which communities shrink?
• Anomaly/Event Detection• When user’s connection network changes significantly in single timestep?
• Detecting change in virus behavior with mutation
• Immunization Policies• Create changes in immunization policy as disease mutates
Categories
• Maintenance Methods: • Correct your predictions/classifications/communities over time
• Learning Dynamics: • Learn how the network is changing over time
• Hybrid methods• Work that we discuss today
Method-based Characterization[3]
Graph-based Characterization[3]
• Slowly Evolving Networks• Time-stamp based
• Classical definition
• Explored in literature
• Streaming Networks• Edges and nodes come as a stream
• Real-time streaming algorithms are expected.
• Much more challenging
Yu, Wenchao, Charu C. Aggarwal, and Wei Wang. "Temporally Factorized Network Modeling for Evolutionary Network Analysis.“ [4]
Tags: Learning Dynamics, Slowly Evolving Networks, Anomaly Detection
All figures are credited to the respective paper
Core Idea [4]
• Simple Idea
• Decompose the series of adjacency matrices into a 1. changing component w.r.t. time and 2. constant component
Definitions [4]
• Given A(t), we can factorize as follows:• A(t) = f( U V(t)T)
• where U and V(t) are n x k dimensions. V(t) is a function with time t.
• k corresponds to latent dimension size
• f(.) is an element-wise function
• For undirected graphs , A(t) is symmetric. • A(t) = f( V(t) V(t)T)
• where V(t) = n x k dimensions
• can be represented with the above formulation as well, but authors took occam’s razor approach
Optimization Problem [4]
• Least Squares Optimization
• Find me f( U V(t)T) which matches A(t) as much as possible
• D(t) = Decay function with time t, current state of the network is more important , compared to the previous
• Quite memory-intensive for sparse networks.
• Better formulation for sparser networks, , let E(t) be the set of edges for which the weights in A(t) are non-zero at time t. Let S(t) be a sample of edges (i, j) at time-stamp t such that the value of edge is 0
Model Choices [4]
• V(t) is a polynomial function:
• With added regularization:
• How to optimize:• Gradient descent:
• Compute partial derivatives with respect to U and V(t), update and repeat
Algorithm [4]
Complexity: O(m + n)
• Last T-1 training data, T as test set
• Comparative algorithms:• Common Neighbors
• Adamic Adar
• Preferential Attachment
• Nonparametric Link Prediction
• Link Prediction via Matrix Factorization
• CP Tensor Model ( Only one which takes into account temporal context)
Application: Temporal Link Prediction [4]
Application: Temporal Link Prediction Results[4]
Application: Temporal Anomalies and Events Detection [4]
• Temporal anomalies/Events = unexpected changes in the network structure over time
• A(T+1) is observed matrix and መ𝐴(T+1) is predicted matrix
• If the change is large , then possible anomaly
• How do we localize the anomaly?
• Edges where the change is above threshold
Application: Temporal Anomalies and Events Detection Results [4]
Application: Community Analysis [4]
Summary (tl;dr) [4]
• Decompose the series of adjacency matrices into a 1. changing component w.r.t. time and 2. constant component
• Use least squares optimization for learning the two components
• Link prediction: Corresponds to predicting based on previous history
• Event detection: Corresponds to detecting large changes in expected vs observed
• Can be used for detecting expanding communities
• Complexity manageable
• Cons:• Does not consider the individual characteristics of the nodes
Gu, Yupeng, Yizhou Sun, and Jianxi Gao. "The Co-Evolution Model for Social Network Evolving and Opinion Migration." [5]
Tags: Learning Dynamics, Slowly Evolving Networks, Anomaly Detection
All figures are credited to the respective paper
Core Idea [5]
• Consider a dynamic graph, • Step 1 : Actors(nodes) change their opinions or behaviors, based on their
social influence
• Step 2: The network structure gets affected by these changes
• Go back to Step 1
• This paper proposes a co-evolution model which jointly models• Network changes
• Actors changes
• Allows to do lots of interesting stuff:• Opinion Convergence, Intra-community divergence
• Social reach of actors as opinions
Definitions & Background [5]
• Derives from “Influence” and “Homophily”
• Influence:• Individuals are affected by their friends, or the company they keep
• Homophily• Individuals interact with people that are similar
• Each individual can be represented by a feature vector
• Network is a reflection of “latent” variables
Definitions & Background [5]
• Collective motion:• Each particle moves at a constant rate v, while the direction of motion is
determined by the average direction of all others within its neighborhood of radius r, plus some random perturbation
Co-evolution Model [5]
• Project the nodes of a network to latent variable space ( Network generation)
• Opinions change depending on the interactions (Opinion Migration)
Co-evolution Model : Social Network Generation [5]
• Each link independent of other links
• Links specified by scoring function: RK x RK -> R, that assigns a score to a pair of node features indicating the likelihood of link between two nodes
• Dot Product-based Score Function:
• Cosine Similarity is the scoring function
• In other words, people with extreme stances (i.e. large norms of latent feature vector) will become the opinion leader.
• Not true in real-world case
• Gravity-based Score Function:
• Each node has a popularity hyper-parameter b.
• The scoring function is similar to gravity equation:
• Earlier approaches in opinion migration• is the average position of neighbors of xn at time t• Leads to a model where opinions do not change after the communities are
formed
• Does not capture the propagation model
• The paper chooses to regularize the direction instead of the position
• = average direction of neighbors• A scholar tends to raise interest in a research topic that is trending among her
collaborators.
Co-evolution Model : Opinion Migration [5]
Co-evolution Model: Unified Model Algorithm[5]
• Sparsity parameter d -> the sparsity of the graph (i.e. the average number of friends)
• Noise level σ -> the deviation of one’s direction from the expected value
Applying to Simulated Dataset [5]
• Simulated Dataset:• Node randomly initialized with initial position in a lattice [−L/2,L/2]× [−L/2,L/2] where L = 5
• Each node has popularity b ∼ Uniform([1,2]). b fixed throughout the process
• Average Normalized Velocity:
• Two hyper-parameters:.• Noise level σ
• Sparsity Parameter d
Applying to Simulated Dataset Results [5]
Applying to Simulated Dataset Results [5]
Applying to Co-sponsorship dataset [5]
• Co-sponsorship dataset:• .A sponsor of a bill is a legislator (usually a member from the congress) who introduces a bill or resolution for consideration. A
cosponsor is another congress member who adds his or her name as a supporter to the sponsor’s bill. Co-sponsorship contains important information about the social support network between legislators: the closer the relationship between a sponsor and a cosponsor, the more likely it is that the sponsor has directly petitioned the cosponsor for support
• Optimization objective:
• Method:• Step 1 : Initialize the latent variables by maximizing likelihood on the first observed graph and directions are initialized randomly• \\ Co-ordinate descent –like optimization• Step 2 :Update b• Step 3 : Update X, Θ• Go back to Step 2 \\ Stop when the likelihood does not change - converges
• Baselines:• CoNNdot : Probability of link = dot-product based similarity functions• Latent feature propagation model (LFP): binary latent feature propagation model• Dynamic social network with latent space models : Only latent vectors matter• Phase transition model :
• Large-scale information network embedding (LINE) :baseline for static methods
Applying to Co-sponsorship Prediction Results[5]
• Calculate the pairwise probability of a link from un to all other users, rank them and evaluate the AUC score
• Two metrics:• Normalized discounted cumulative gain
(NDCG)• AUC Score
Applying to Co-sponsorship Prediction Results[5]
• Calculate the pairwise probability of a link from un to all other users, rank them and evaluate the AUC score
• Two metrics:• Normalized discounted cumulative gain
(NDCG)• AUC Score
Summary (tl;dr)
• Models changes in the network structure and opinions
• Network Structure:• Embed to latent vector using HMM
• Opinion Migration:• Flocks of bird fly together.
• Results show interesting applications:• Detecting changes in opinion• Opinion Leaders
• Biologically-inspired
• Comparison might be nuanced:• Comparing against other state-of the-art approaches will be interesting
Amiri, Sorour E., Liangzhe Chen, and B. Aditya Prakash. "SnapNETS: Automatic Segmentation of Network Sequences with Node Labels." [6]
Tag: Network summarizations
Network Sequences [6]
Epidemiology: disease spreads over contact networks
Flu
Meme
G1 G2 G3 G4
Uninfected
Infected
G1 G2 G3 G4
Social Media: Information spreads over friendship networks
Inactive
2
Active
Making sense of network sequences [6]
Flu
when do the infection patterns change?
Star Bridge Near Clique
Reason:
• Virus mutation
• Vaccination
• …
G1 G2 G3 G4Uninfected
3
Infected
Making sense of network sequences [6]
Meme Reason:
• Event
• …
Star Clique
when do the activation patterns change?
G1 G2 G3 G4
Inactive
4
Active
Problem 1: Network sequence segmentation [6]
Given a sequence of networks with labeled nodes,
Find the best segmentation which captures:
Different distribution of node labels.
Near Clique
5
G1
Star
G2
Bridge
G3 G4
In this work:
Binary labels {0, 1}
Overview of SnapNETS [6]
Goal 1. Summarize each graph:Keep structural and label dependent properties
Goal 2. Construct Segmentation graph:Define nodes and edgesDefining edges weights
o extract the features of summarized graphs
Goal 3. Find the best segmentation:Define the best segmentation (path)Compute the best segmentation
15
Goal 1: Summarizing graph snapshots [6]
We want to preserve
Structural properties
Nodes labels
Role of Eigenvalue:
Epidemic threshold in most diffusion
models [Prakash et al. ICDM 2011]
Same Same diffusive properties
Leading eigenvalue of Adjacency matrix
Amiri, Chen, Prakash 18
Nodes: For each segment there is a node + {Source (‘s’), Target (‘t’)} Source (‘s’) = start time Target (‘t’) = end time
Edges: There is a directed edge between adjacent nodes
Goal 2: Segmentation graph [6]
Amiri, Chen, Prakash 23
Edge Weights [6]
How can we measure the distance between two segments?
w ?
Amiri, Chen, Prakash 24
Goal 3: Finding the best segmentation [6]
Observation:For each segmentation there is a path from ‘s’ to ‘t’
For each path from ‘s’ to ‘t’ there is a segmentation
Therefore,• Best segmentation problem ≡ Path optimization problem
Amiri, Chen, Prakash 28
Problem 3: Finding the best segmentation [6]
Our idea: Average longest path
Advantages: Parameter free
Naturally balances weight of the path with the number of segments.
Given a segmentation graph
Find the average longest path from ‘s’ to ‘t’
Amiri, Chen, Prakash 30
Summary (tl;dr)
• Presents a novel approach for summarizing temporal changes in graph
• It finds patterns where adjacent segments have different characteristics of nodes with the same label i.e. the ‘placement’ and ‘connection’ of active/inactive nodes are different
• Looks at global changes
• Can be extended to streaming or partially observed graphs
Conclusion
• Temporal networks come with massive applications • Link Prediction with Temporal Context
• Community Evolution Dynamics
• Anomaly/Event Detection
• Research Areas:• Link Prediction
• Learning Graph Dynamics
• Graph Summarization
• Opinion Migration
• And more.
References
• [1] : http://cmxp.bplaced.net/Commetrix/iris/2co-author_network_min3pap.png
• [2] : : http://cmxp.bplaced.net/Commetrix/iris/2co-author_network_min3pap.png
• [3] : Aggarwal, Charu, and Karthik Subbian. "Evolutionary network analysis: A survey." ACM Computing Surveys (CSUR) 47.1 (2014): 10.
• [4] Yu, Wenchao, Charu C. Aggarwal, and Wei Wang. "Temporally Factorized Network Modeling for Evolutionary Network Analysis.“
• [5] : Gu, Yupeng, Yizhou Sun, and Jianxi Gao. "The Co-Evolution Model for Social Network Evolving and Opinion Migration."
• [6] : Amiri, Sorour E., Liangzhe Chen, and B. Aditya Prakash. "SnapNETS: Automatic Segmentation of Network Sequences with Node Labels."