Link Prediction

in
Social Networks
 Contents

 Introduction




 Feature based Link Prediction




 Bayesian Probabilistic Models




 Probabilistic Relational Models




 Relational Markov Network
 Link Prediction

MDPI and ACS Style

Zhu, J.; Zhang, J.; Wu, Q.; Jia, Y.; Zhou, B.; Wei, X.; Yu, P.S. Constrained Active Learning for Anchor Link Prediction Across
Multiple Heterogeneous Social Networks. Sensors 2017, 17, 1786.
 Link Prediction

 Predict the likelihood of a future association between

two nodes, knowing that there is no association

between the nodes in the current state of the graph.
 G[t, t'] denotes the sub graph of G restricted to the

edges with time-stamps between t and t'.
 Training interval [t0, t0' ] and a test interval [t1, t1' ]

where t0' < t1.
 Link prediction task is to output a list of edges not
present in G[t0, t0' ], but are predicted to appear in
the network G[t1, t1' ].
 Application Areas

 Internet and www  automatic web hyper-link creation,

web site hyper-link prediction
E-commerce  build recommendation systems

 Bibliography and library science  de-duplication and

record linkage
 Bioinformatics  protein-protein interaction (PPI)

prediction
 Security related applications  identify hidden groups of
terrorists and criminals
 Methods of Link Prediction

 The traditional (non-Bayesian) models  extract a set of


features to train a binary classification model.
 The probabilistic approach  model the joint-
probability among entities in a network by Bayesian
graphical models.
 The linear algebraic approach  computes the
similarity between nodes in a network by rank-
reduced similarity matrices.
 Notation: For a node x, Γ(x) represents the set of
neighbors of x. degree(x) is the size of the Γ(x).
 Feature based Link Prediction
(Supervised Classification)


u, v ∈ V are two vertices in the graph G(V,E) and the label of the data point u, v is y(u,v)

 Assume that the interactions between u and v are symmetric







 Classification model predicts the unknown labels of pair of
vertices, u, v  E in the graph G[t1, t1].



 Popular supervised classification tools, such as naive Bayes,
neural networks, support vector machines (SVM) and k
nearest neighbors can be used.



 Major challenge is to choose a set of features for the
classification task.
 Feature Set Construction

 Each data point corresponds to a pair of vertices with

the label denoting their link status, so the chosen
features should represent some form of proximity

between the pair of vertices. 
 Graph topological features:

(1) Node neighborhood based (2) Path based
 Compute the similarity based on the node
neighborhoods or based on the ensembles of paths
between a pair of nodes.
 Node Neighborhood based Features

 Common Neighbors. For two nodes, x and y, the

size of their common neighbors is |Γ(x) ∩ Γ(y)|.
 As the number of common neighbors grows higher, the chance

that x and y will have a link between them increases.
 Jaccard Coefficient: normalizes the size of
common neighbors.
 Probability that a common neighbor of a pair of vertices x and y
would be selected if the selection is made randomly from the
union of the neighbor-sets of x and y.
 Adamic/Adar: Weighs the common neighbors
with smaller degree more heavily, works better than
the previous two metrics.
 Path based Features

 Shortest Path Distance: Path distance between two nodes can


influence the formation of a link between them.
 Katz: Directly sums over all the paths that exist between a pair
of vertices x and y. Exponentially damps the contribution of a path
by a factor of βl. Katz score between all the pairs of vertices can be
-1
 computed by finding (I − βA) − I
 Hitting time: Expected number of steps required for a random
walk starting at x to reach y. Shorter hitting time denotes that the
nodes are similar to each other, so they have a higher chance of
linking in the future.

 Rooted Pagerank: The stationary probability of y in a random
walk that returns to x with probability 1 − β in each step, moving
to a random neighbor with probability β.
 Bayesian Probabilistic Models

 Obtain a posterior probability that denotes the

chance of co-occurrence of the vertex pairs.




 Link Prediction By Local Probabilistic
Models




 Network Evolution based Probabilistic Model




 Hierarchical Probabilistic Model
 Link Prediction By Local Probabilistic Models

 Uses Markov Random Field (MRF), an undirected graphical

 model.
 Central neighborhood set consists of other nodes that appear
 in the local neighborhood of x or y.
 Compute the joint probability P({w, x, y, z}), which represents
 the probability of co-occurrence of the objects in this set.
 Step-1: Find a collection of central neighborhood sets  find a
shortest path between x and y and then all the nodes along this path
 can belong to one central neighborhood set.
 Step-2:Obtain the training data for the MRF model, which is
 taken from the event log of the social network.
 Step-3: An MRF model is trained from the training data.
This training process is translated to a maximum entropy
optimization problem. Once the model PM(Q) is built, one
can estimate the joint probability between the vertex x and y.
Network Evolution based Probabilistic Model

For graph G(V, φ), V is the set of nodes and


φ : V × V → [0, 1] is an edge label function
 φ(x, y) denotes the probability that an edge exists
between node x and y in G.
 φ(x, y) = 1 if an edge exists

φ(x, y) = 0 if an edge does not exist.

 φ(t)  the edge label function at time t, which changes over
 time;
 The model is Markovian, i.e., φ(t+1) depends only
on φ(t).
 Model Evolution

 An edge label is copied from node l to node m randomly

 with probability wlm.
 First, the model decides on l and m, then chooses an edge
label uniformly from one of l’s |V | − 1 edge labels
 (excluding φ(l,m)) to copy as m’s edge label.
 The model satisfies the following probability constraints.



 Through the edge label copying process, l can become
 friend of one of m’s friend.
 The learning task in the model is to compute the weights
wij and the edge labels φ(t+1) given the edge label φ(t)
from training dataset.
 Hierarchical Probabilistic Model

 Let G be a graph with n vertices.


 A dendogram D is a binary tree with n leaves
corresponding to the vertices of G. Each of the n−1
internal nodes of D corresponds to the group of vertices
 that are descended from it.
 A probability pr is associated with each internal node r.
Then, given two vertices i,j of G, the probability pij that
they are connected by an edge is pij = pr where r is the
 lowest common ancestor in D.
 The combination, (D, {pr}) of the dendogram and the set
of probabilities then defines a hierarchical random
graph.
 Cont..

 One of the dendograms are chosen by Markov chain

Monte Carlo(MCMC) sampling method with probabilities
 proportional to their likelihood.
 To create the Markov chain, the method creates a set of
transitions between possible dendograms through
 rearrangement.
 For link prediction, a set of sample dendograms are
obtained at regular intervals once the MCMC random walk
 reaches an equilibrium.
 For the pair of vertices x and y for which no connection
exists, the model computes a mean probability pxy that
they are connected by averaging over the corresponding
probability pxy in each of the sampled dendograms.
 Probabilistic Relational Models

 Concrete modeling tool that provides a systematic

way to incorporate both vertex and edge attributes to
model the joint probability distribution of a set of
entities and the links that associate them.
 Considers the object-relational nature of structured
data by capturing probabilistic interactions
between entities
 Type 1: Based on Bayesian networks, which
consider the relation links to be directed
 Type 2: Based on relational Markov networks,
which consider the relation links to be undirected.
 Example : Co-authorship Network

 PRM can have heterogeneous entities in the model. Ex:

 article, author, conferenceVenue, and institution.
 Author may have attributes like name, affiliationInstitute,
 status
 An article may have publicationYear, conferenceVenue;

 An institute may have location, Conference venue may have
 attributes like ResearchKeywords.
 Then there can be relational links between these entities.

 Two persons can be related by an advisor/advisee
 relationship.
 A person can be related to a paper by an author relationship.
A paper can be related to a conference venue by publish
relationship.
 Relational Markov Network

 Relational counterpart of undirected graphical models or

 Markov Networks.
 A Markov network for V defines a joint distribution over
V through an undirected dependency network and a set
 of parameters.
 For a graph G, if C(G) is the set of cliques , the Markov
network defines the distribution p(v)

Where Z is the standard normalizing factor

vc is the vertex set of the clique c φc
is a clique potential function.
 Cont..

 Given a particular instantiation I of the schema, the

RMN M produces an unrolled Markov network

over the attributes of entities in I
 There exists one clique for each c ∈ C(I), and all of these cliques are

associated with the same clique potential φC.
 In a large network with a lot of relational attributes,

the network is typically large
 RMN also uses belief propagation for inference.
 References

 https://pdfs.semanticscholar.org/e7d3/0fefe1b99c21813
 873f976e46d03dc82b4fc.pdf
 https://www.slideshare.net/SinaSajadmanesh/probabili
 stic-relational-models-for-link-prediction-problem
 http://www.robotics.stanford.edu/~koller/Papers/Fried
 man+al:IJCAI99.pdf
 https://networkx.github.io/documentation/networkx-
 1.10/reference/algorithms.link_prediction.html
 https://courses.cs.washington.edu/courses/cse574/05sp
 /slides/rmn-danny.pdf
 https://pdfs.semanticscholar.org/presentation/fad1/fb7
85125ba8db0e8bf417f8a68729327269e.pdf