Social Media Mining: An Introduction

Social Media Mining:
Fundamental Issues and Challenges
Mohammad Ali Abbasi, Huan Liu, and Reza Zafarani
Data Mining and Machine Learning Lab
Arizona State University
December 10, 2013

2Social Media Mining Measures and Metrics 2Social Media Mining ICDM 2013 Tutorial
http://dmml.asu.edu/smm/

Broadcast Media
One-to-Many
Communication Media
One-to-One Traditional Data
Traditional Media and Data

• Everyone can be a media outlet or producer
• Disappearing communication barrier
• Distinct characteristics
– User generated content: Massive, dynamic,
extensive, instant, and noisy
– Rich user interactions
– Collaborative environment, and wisdom of the
crowd
– Many small groups (the long tail phenomenon)
– Attention is expensive
Social Media: Many-to-Many

Publishing
Sharing
Gaming
Location
Marketing
Networking
Discussing
A Big Variety of Social Media

Two DMML Books of SMM
Twitter Data Analytics
Nov. 2013
Social Media Mining
Feb. 2014

Networks are Pervasive
• A network is a graph.
– Elements of the network have meanings
• Network problems can usually be represented
in terms of graph theory
Internet
CitationNetworks
TwitterNetworks

Nodes/Edges/Degree/Neighborhood
Node/
Vertex
/Actor
Edge/Tie/
Relationship
Arc For any node v, the set of
nodes it is connected to via
an edge is called its
neighborhood and is
represented as N(v)
N(v1) = {v2,v7}
The size of the neighborhood
of a node is its degree
d(v1)=2
Subgraph

Degree Distribution
When dealing with very large graphs, how nodes’
degrees are distributed is an important concept to
analyze and is called Degree Distribution
Degree distribution histogram
– The x-axis represents the degree
and the y-axis represents the
fraction of nodes having that
degree

Graph Representation: Adjacency Matrix



ijA
0, otherwise
1, if there is an edge between nodes vi and vj
Social media networks have
very sparse Adjacency matrices
• Diagonal Entries are self-links or loops

Types of Graphs
Directed
graph
A  AT
1
23
4
Undirected
Graph
A = AT
Simple graph Multigraph


 

jandibetweenedgenoisThere0,
Rww,
ijA
G(V, E, W)
Weighted graph
Signed graph

Adjacent nodes/Incident Edges/Paths/
Shortest Paths/Connectivity
Two nodes are
adjacent
if they are
connected via an
edge.
Two edges are
incident, if they
share one end-
point
• A walk where nodes and edges are distinct is called a path and
a closed path is called a cycle
• The length of a path or cycle is the number of edges visited in the path or cycle
Length of path= 4
Strongly connected Weakly connectedConnected Disconnected
Component

Shortest Path
• Shortest Path is the path between two nodes that
has the shortest length.
– We denote the length of the shortest path between nodes vi
and vj as
• The concept of the neighborhood of a node can be
generalized using shortest paths.
– An n-hop neighborhood of a node is the set of nodes that
are within n hops distance from the node.
• The diameter of a graph is the length of the longest
shortest path between any pairs of nodes in the graph

Special Graphs/Properties
A forest with
3 trees
Spanning Tree
Steiner Tree
Complete Graph
Bipartite Graph Affiliation Network Social-Affiliation
Network
Regular Graph
Bridges

Graph Algorithms
• Graph/Tree Traversal Algorithms
– Depth-First Search (DFS)
– Breadth-First Search (BFS)
• Shortest Path Algorithms
– Dijktra’s Algorithm
– Bellman-Ford Algorithm
– Floyd-Warshall Algorithm
• Minimum Spanning Tree Algorithms
– Prim’s Algorithm
– Kruskal’s Algorithm
• Maximum Flow Algorithms
– Ford-Fulkerson Algorithm
• Matching Algorithms
– Bipartite Matching
– Weighted Matching

Network Measures

Why do we need measures?
1. Who are the central figures (influential
individuals) in the network?
2. What interaction patterns are common
in friends?
3. Who are the like-minded users and how
can we find these similar individuals?
To answer these and similar questions, one
first needs to define measures for
quantifying centrality (centrality
measures), level of interactions, and
similarity (similarity measures), among
others.

Centrality defines how important a node is within a network.
Centrality

Centrality: Degree Centrality
• The degree centrality
measure ranks nodes with
more connections higher in
terms of centrality
• di is the degree (number of
adjacent edges) for vertex vi
In this graph degree
centrality for vertex v1
is d1 = 8 and for all
others is dj = 1, j  1
In directed graphs, we can
either use the in-degree,
the out-degree, or
the combination as the
degree centrality value:

Let T

This means that Ce is an
eigenvector of adjacency
matrix A and  is the
corresponding eigenvalue
Eigenvector Centrality
• Having more friends does
not by itself guarantee that
someone is more important,
but having more
important friends
provides a stronger signal
• Eigenvector centrality tries
to generalize degree
centrality by incorporating
the importance of the
neighbors (undirected)
• For directed graphs, we can
use incoming or at times,
outgoing edges

Katz Centrality
• A major problem with
eigenvector centrality arises
when it deals with directed
graphs
• Centrality only passes over
outgoing edges and in special
cases such as when a node
is in a directed acyclic graph
centrality becomes zero even
though the node can have
many edge connected to it
• To resolve this problem, we
add bias term  to the
centrality values for all nodes
• We select α < 1/λ, where λ is
the largest eigenvalue of AT
For the matrix (I- αAT) to be
invertible
Rewriting equation in a vector form
Katzcentrality:
vector of all 1’s

PageRank
• Problem with Katz Centrality:
in directed graphs, once a
node becomes an authority
(high centrality), it passes all
its centrality along all of its
out-links
• This is less desirable since not
everyone known by a well-
known person is well-known
• To mitigate this problem we
can divide the value of passed
centrality by the number of
outgoing links, i.e., out-degree
of that node such that each
connected neighbor gets a
fraction of the source node’s
centrality

Betweenness Centrality
Another way of looking
at centrality is by
considering how
important nodes are in
connecting other nodes
In undirected graphs we can
assume s<t
the number of shortest paths
from vertex s to t – a.k.a.
information pathways
the number of shortest paths
from s to t that pass through vi

Closeness Centrality
• The intuition is that
influential and central
nodes can quickly reach
other nodes
• These nodes should
have a smaller average
shortest path length to
other nodes

Transitivity and Reciprocity

Transitivity
• Mathematic
representation:
– For a transitive relation R:
• In a social network:
– Transitivity is when a
friend of my friend is
my friend
– Transitivity in a social
network leads to a denser
graph, which in turn is
closer to a complete graph
– We can determine how
close graphs are to the
complete graph by
measuring transitivity
[Global] Clustering Coefficient
Local clustering coefficient measures
transitivity at the node level

Reciprocity
If you become my friend, I’ll be yours
• Reciprocity is a more simplified version
of transitivity as it considers closed
loops of length 2
• If node v is connected to node u,
u by connecting to v, exhibits reciprocity

• Measuring stability for an
observed network
Balance and Status

Social Balance/Social Status
Social balance theory discusses
consistency in friend/foe
relationships among individuals.
Unstable
configuration
Stable
configuration
• Status defines how prestigious an
individual is ranked within a societ
• Social status theory measures how
consistent individuals are in assigning
status to their neighbors

• How similar two nodes are
in a network?
Similarity

Structural Equivalence
• In structural equivalence, we look at the
neighborhood shared by two nodes; the size of
this neighborhood defines how similar two
nodes are.
For instance, two brothers have in
common sisters, mother, father,
grandparents, etc. This shows that they
are similar, whereas two random male or
female individuals do not have much in
common and are not similar.

Structural Equivalence: Definitions
• Vertex similarity
• In general, the definition of neighborhood N(v) excludes
the node itself v.
– Nodes that are connected and do not share a neighbor will be
assigned zero similarity
– This can be rectified by assuming nodes to be included in their
neighborhoods
Jaccard Similarity:
Cosine Similarity:

Regular Equivalence
• In regular equivalence, we do not look at
neighborhoods shared between individuals, but
how neighborhoods themselves are similar
For instance, athletes are similar not
because they know each other in person, but
since they know similar individuals, such as
coaches, trainers, other players, etc.
vi, vj are similar when
their neighbors vk and
vl are similar

Regular Equivalence
vi, vj are similar when
vj is similar to vi’s
neighbors vk
A vertex is highly similar
to itself, we guarantee this
by adding an identity
matrix to the equation

Network Models

Why should I use network models?
• In may 2011, Facebook had 721 millions users. A
Facebook user at the time had an average of 190
users -> a total of 68.5 billion friendships
– What are the principal underlying processes that help
initiate these friendships
– How can these seemingly independent friendships
form this complex friendship network?
• In social media there are many networks with
millions of nodes and billions of edges.
– They are complex and it is difficult to analyze them

So, what do we do?
• We design models that generate, on a smaller scale,
graphs similar to real-world networks.
• Hoping that these models simulate properties
observed in real-world networks well, the analysis of
real-world networks boils down to a cost-efficient
measuring of different properties of simulated
networks
– Allow for a better understanding of phenomena observed
in real-world networks by providing concrete
mathematical explanations; and
– Allow for controlled experiments on synthetic networks
when real-world networks are not available.
• These models are designed to accurately model
properties observed in real-world networks

Power-law Distribution,
High Clustering Coefficient, and
Small Average Path Length
Properties of Real-World
Networks

Power-law Degree Distribution
• Many sites are visited less than a 1,000 times a
month whereas a few are visited more than a
million times daily.
• Social media users are often active on a few sites
whereas some individuals are active on hundreds
of sites.
• There are exponentially more modestly priced
products for sale compared to expensive ones.
• There exist many individuals with a few friends
and a handful of users with thousands of friends
(Degree Distribution)

Power-Law Distribution
• When the frequency of an
event changes as a power of
an attribute:
– the frequency follows a
power-law
• Let f(k) denote the number
individuals having degree k.
b: the power-law exponent and
its value is typically in the range of [2, 3]
a: power-law intercept
Log-Log
plot
Networks with power-law degree
distribution are often called scale
free networks

High Clustering Coefficient
• In real-world networks, friendships are highly
transitive, i.e., friends of an individual are often
friends with one another
– These friendships form triads ->
high average [local] clustering coefficient
• In May 2011, Facebook had an average clustering
coefficient of 0.5 for individuals who had 2
friends.

Small Average [Shortest] Path Length
• In real-world networks, any two members of the
network are usually connected via short paths.
In other words, the average path length is small
– Six degrees of separation:
• Stanley Milgram, in the well-known small-world
experiment conducted in the 1960’s, conjectured that people
around the world are connected to one another via a path of
at most 6 individuals
– Four degrees of separation:
• Lars Backstrom et al. in May 2011, the average path
length between individuals in the Facebook graph was 4.7.
(4.3 for individuals in the US)

Random Graphs

Random Graphs
• We start with the most basic assumption on how
friendships are formed.
Random Graph’s main assumption:
Edges (i.e., friendships) between
nodes (i.e., individuals) are
formed randomly.
Formally, we can assume that for a graph with a fixed number of nodes n,
any of the edges can be formed independently, with probability p. This
graph is called a random graph and we denote it as G(n, p) model
This model was first proposed independently by Edgar Gilbert and
Solomonoff and Rapoport.

Expected Degree
The expected number of edges connected to a node
(expected degree) in G(n, p) is c=(n - 1)p
• Proof:
– A node can be connected to at most n-1 nodes (or n-1
edges)
– All edges are selected independently with probability p
– Therefore, on average, (n - 1)p edges are selected
• c=(n-1)p or equivalently,

Properties of Random Graphs
• Degree Distribution
– This is a binomial degree distribution. In the limit this
will become the Poisson degree distribution
• Global Clustering coefficient
– The global clustering coefficient of any graph defines the
probability of two neighbors of the same node that are
connected. This probability is the same for any two nodes
and is p
• Average Path length
– (sketch) When moving average path length number of
steps away from a node, almost all nodes are visited.

Small-World Model

Small-world Model
• In real-world interactions, many individuals have
a limited and often at least, a fixed number of
connections
• In graph theory terms, this assumption is
equivalent to embedding individuals in a regular
network.
• A regular (ring) lattice is a special case of regular
networks where there exists a certain pattern on
how ordered nodes are connected to one another.
• In particular, in a regular lattice of degree c,
nodes are connected to their previous c/2 and
following c/2 neighbors. Formally, for node set
V an edge exists between node i and j if and only if

Small-World Model Properties
• Degree distribution:
• In the graph generated by the small world model, most nodes
have similar degrees due to the underlying lattice.
• Clustering Coefficient
and Average Path Length:
C(0) and L(0) are the clustering
Coefficient and average path
Length of the lattice. We
change the value of p
such that C(p)/C(0) and L(p)/L(0)
Are desirable

Preferential Attachment
Model

Preferential Attachment Model
• Node v arrives
– P(1) = 1/7
– P(2) = 1/7
– P(3) = 0
– P(4) = 3/7
– P(5) = 2/7
• Networks:
– When a new user joins the network, the probability of
connecting to an existing node vi is proportional to the
degree of vi
• Distribution of wealth in the society:
– The rich get richer

Preferential Attachment Model Properties
• Degree Distribution:
• Clustering Coefficient:
• Average Path Length:

• Why connected people are
similar?
Assortativity in Social
Networks

Birds of a feather flock together

Social influences on obesity

Why connected people are similar?
• Influence
– Influence is the process by which an individual (the influential)
affects another individual such that the influenced individual
becomes more similar to the influential figure.
• If most of one’s friends switch to a mobile company, he might be
influenced by his friends and switch to the company as well.
• Homophily
– It is realized when similar individuals become friends due to
their high similarity.
• Two musicians are more likely to become friends.
• Confounding
– Confounding is environment’s effect on making individuals
similar
• Two individuals living in the same city are more likely to become
friends than two random individuals

Influence, Homophily, and Confounding

Similarity of Connected Nodes in Social Networks
• Race
• Religion
• Education
• Income level
• Job and skills
• Language
• Interests and preferences

Topics
• Measuring Assortativity
• Measuring Influence and Homophily
• Distinguishing Influence and Homophily

Measuring Assortativity

Measuring Assortativity for Nominal Attributes
• Where nominal attributes are assigned to nodes
(race), we can use edges that are between nodes of
the same type (i.e., attribute value) to measure
assortativity of the network
– Node attributes could be nationality, race, sex, etc.
Kronecker delta function
t(vi) denotes type of vertex vi

Assortativity Significance
• Assortativity significance measures the difference
between the measured assortativity and its expected
assortativity
– The higher this value, the more significant the assortativity
observed
• Example
– Consider a school where half the population is white and
half the population is Hispanic. It is expected for 50% of the
connections to be between members of different races. If all
connections in this school were between members of
different races, then we have a significant finding

Assortativity Significance: Measuring
The expected assortativity in the whole
graph
Assortativity
This measure is called modularity

Measuring Assortativity for Ordinal Attributes
• A common measure for analyzing the
relationship between ordinal values is
covariance.
• It describes how two variables change together.
• In our case we are interested in how values of
nodes that are connected via edges are
correlated.
B
A C
18 21
20

Covariance Variables
• We construct two variables XL and XR, where for
any edge (vi, vj) we assume that xi is observed
from variable XL and xj is observed from variable
XR.
• In other words, XL represents the ordinal values
associated with the left node of the edges and XR
represents the values associated with the right
node of the edges
• Our problem is therefore reduced to computing
the covariance between variables XL and XR

Covariance Variables: Example
• XL : (18, 21, 21, 20)
• XR : (21, 18, 20, 21)

B
A C
18 21
20
List of
edges:
((A, C),
(C, A),
(C, B),
(B, C))

• Measuring Influence
• Modeling Influence
Social Influence

Social Influence: Definition
• The act or power of producing an effect without
apparent exertion of force or direct exercise of
command

Measuring Influence
• Measuring influence is assigning a number to each node
that represents the influential power of that node
• We assume that an individual’s attribute or the way she
is situated in the network predicts how influential she
will be.
• For instance, we can assume that the gregariousness
(e.g., number of friends) of an individual is correlated
with how influential she will be. Therefore, it is natural
to use any of the centrality measures to calculate
influence
• An example:
– On Twitter, in-degree (number of followers) is a benchmark for
measuring influence commonly used

Measuring Social Influence
on Twitter

Measuring Social Influence on Twitter
• In Twitter, users have an option of following
individuals, which allows users to receive tweets
from the person being followed
• Intuitively, one can think of the number of
followers as a measure of influence (in-degree
centrality)

Measuring Social Influence on Twitter:
Measures
• Number of Followers
– The number of users following a person on Twitter
– Indegree denotes the “audience size” of an individual.
• Number of Mentions
– The number of times an individual is mentioned in a tweet,
by including @username in a tweet.
– The number of mentions suggests the “ability in engaging
others in conversation”
• Number of Retweets:
– Tweeter users have the opportunity to forward tweets to a
broader audience via the retweet capability.
– The number of retweets indicates individual’s ability in
generating content that is worth being passed on.
• Number of Tweets
• PageRank (TwitterRank)

• Linear threshold model
Influence Modeling

Influence Modeling
• At time stamp t1, node v is activated and node u is not
activated
• Node u becomes activated at time stamp t2, as the effect
of the influence
• Each node is started as active or inactive;
• A node, once activated, will activate its neighboring
nodes
• Once a node is activated, this node cannot be
deactivated

Linear Threshold Model (LTM)
An actor would take an action if the number of his
friends who have taken the action exceeds (reach)
a certain threshold
• Each node i chooses a threshold ϴi randomly
from a uniform distribution in an interval
between 0 and 1.
• In each discrete step, all nodes that were active
in the previous step remain active
• The nodes satisfying the following condition will
be activated

Linear Threshold Model, an Example
0.5
0.8
0.8
0.4
0.6
0.6
0.9
0.3
0.2
0.3
0.7
0.3
0.2

“Birds of a feather
flock together”
Homophily

Homophily- Definition
• Homophily (i.e., "love of the same") is the
tendency of individuals to associate and bond
with similar others
• People interact more often with people who are
“like them” than with people who are dissimilar
• What leads to Homophily?
• Race and ethnicity, Sex and Gender, Age, Religion,
Education, Occupation and social class, Network positions,
Behavior, Attitudes, Abilities, Believes, and Aspirations

Measuring Homophily: Idea
• To measure homophily, one can measure how the
assortativity of the network changes over time
– Consider two snapshots of a network Gt(V, E) and
Gt’ (V, E’) at times t and t’, respectively, where t’ > t
– Assume that the number of nodes stay fixed and edges
connecting them are added or removed over time.

Measuring Homophily
• For nominal attributes, the homophily index is
defined as
• For ordinal attributes, the homophily index can be
defined as the change in Pearson correlation

Distinguishing influence
and Homophily

Distinguishing Influence and Homophily
• We are often interested in understanding which
social force (influence or homophily) resulted in an
assortative network.
• To distinguish between an influence-based
assortativity or homophily-based one, statistical
tests can be used
• Note that in all these tests, we assume that several
temporal snapshots of the dataset are available
where we know exactly, when each node is activated,
when edges are formed, or when attributes are
changed

Shuffle Test
IDEA:
• The basic idea behind the shuffle test comes from the
fact that influence is temporal but homophily is not!
• When u influences v, then v should have been
activated after u.
– Define a temporal assortativity measure.
– Assume that if there is no influence, then a shuffling of the
activation timestamps should not affect the temporal
assortativity measurement.

Shuffle Test
The key idea of the shuffle test is that if influence
does not play a role, the timing of activations
should be independent of users. Thus, even if we
randomly shuffle the timestamps of user activities,
we should obtain a similar assortativity
Test of Influence:
After we shuffle the timestamps of user activities, if the new estimate of social
correlation is significantly different from the estimate based on the user’s
activity log, there is evidence of influence.
User A B C
Time 1 2 3
User A B C
Time 2 3 1

The Edge-reversal Test
If influence resulted in activation, then the
direction of edges should be important (who
influenced whom).
• Reverse directions of all the edges
• Run the same logistic regression on the data
using the new graph
• If correlation is not due to influence, then α
should not change
A
B
C
A
B
C

Community Detection

Social Media Communities
• Any formation of a community requires
– 1) a set of at least two nodes sharing some interest and
– 2) interactions with respect to that interest.
• Two types of groups in social media
– Explicit Groups: formed by user subscriptions
– Implicit Groups: implicitly formed by social interactions
• (individuals calling Canada from the United States) -> phone
operator considers them one community for marketing purposes
• We may see group, cluster, cohesive subgroup, or
module in different contexts instead of “community”
• Community detection
– Discovering implicit communities

Definition
• Community Detection is the
process of finding clusters
(‘‘communities’’) of nodes
with strong internal
connections and weak
connections between
different clusters
• An ideal decomposition of a
large graph is into
completely disjoint
communities (groups of
particles) where there are no
interactions between
different communities.
• In practice, the task is to find
a partition into communities
which are maximally
decoupled.

Member-Based Community Detection
Methods that concentrate on properties of nodes
and in most cases assume that nodes with similar
characteristics represent a community
• Node Characteristics:
– Degree: node with same (or similar) degree are in
one community
• cliques
– Reachability: nodes that are close (small shortest
path) are in the same community
• k-clique, k-club, and k-clan
– Similarity: similar nodes are in the same community

Node-Degree
• Clique: a maximum complete subgraph in which all
nodes are adjacent to each other
• NP-hard to find the maximum clique in a network
• Finding cliques is computationally expensive
• The definition of clique is very strict; often, cliques are
– Relaxed: k-plex; or
– used as cores or seeds to find larger communities
• CPM is a method to find communities with overlap
Nodes 5, 6, 7 and 8 form a clique

Using cliques as seeds:
Clique Percolation Method (CPM): Algorithm
• Input
– A parameter k, and a network
• Procedure
– Find out all cliques of size k in the given network
– Construct a clique graph.
• Two cliques are adjacent if they share k-1 nodes
– Each connected components in the clique graph form
a community

Clique Percolation Method: Example
Cliques of size 3:
{1, 2, 3}, {1, 3, 4}, {4, 5,
6}, {5, 6, 7}, {5, 6, 8}, {5,
7, 8},
{6, 7, 8}
Communities:
{1, 2, 3, 4}
{4, 5, 6, 7, 8}

Node-Reachability
Any node in a group should be reachable in k hops
• k-Clique: a maximal subgraph in which the largest geodesic
distance between any nodes <= k
• k-Club: it follows the same definition as k-clique with an
additional constraint that nodes on the shortest paths should
be part of the subgraph
• k-Clans: it is a k-clique where for all shortest paths within the
subgraph the distance is equal or less than k. All k-clans are k-
cliques, but not vice versa.
Cliques: {1, 2, 3}
2-cliques: {1, 2, 3, 4, 5}, {2, 3, 4, 5, 6}
2-clubs: {1,2,3,4}, {1, 2, 3, 5}, {2, 3, 4, 5, 6}
2-clans: {2, 3, 4, 5, 6}

Node Similarity
• Apply k-means or similarity-based clustering to
nodes
• Vertex similarity is defined in terms of the similarity
of their neighborhood
• Structural equivalence: two nodes are
structurally equivalent iff they are connecting to the
same set of actors
• Structural equivalence is too restrict for practical
use.
Nodes 1 and 3 are
structurally
equivalent,
So are nodes 5 and 7.

Group-Based Community Detection
• In group-based community detection, the global
network information and topology is considered
to determine communities
• We search for communities that are:
– Balanced -> spectral clustering
– Modular -> modularity maximization
– Hierarchical -> hierarchical clustering / Girvan-
Newman algorithm
– Dense -> Quasi-cliques
– Robust -> k-connected graphs

Balanced Communities: Spectral Clustering: Cut
• Most interactions are within group whereas
interactions between groups are few
• Cut: A partition of vertices of a graph into two
disjoint sets
• Minimum cut problem: find a graph partition
such that the number of edges between the two
sets is minimized
• Community detection  minimum cut problem

Ratio Cut & Normalized Cut
• Minimum cut often returns an imbalanced
partition, with one set being a singleton
• Change the objective function to consider
community size

Spectral Clustering
• Both ratio cut and normalized cut can be
reformulated as
• Spectral relaxation:
Represnts diagonal degree matrix
Optimal solution:
Top eigenvectors with the smallest eigenvalues

Spectral Clustering: Example
D = diag(2, 2, 4, 4, 4, 4, 4, 3, 1)
Eigenvectors

Modular Communities: Modularity Maxmization
• Consider a graph G(V, E), |E| = m where the degrees
are known beforehand however edges are not
– Consider two vertices vi and vj with degrees di and dj.
• Now what is an expected number of edges between
these two nodes?
• For any edge going out of vi randomly the probability
of this edge getting connected to vertex vj is

Modularity Maximization: Main Idea
• Given a degree distribution, we know the
expected number of edges between any pairs of
vertices
• We assume that real-world networks should be
far from random. Therefore, the more distant
they are from this randomly generated network,
the more structural they are.
• Modularity defines this distance and modularity
maximization tries to maximize this distance

Normalized Modularity
Consider a partitioning of the data, P = (P1, P2, P3, …, Pk)
For partition Px, this distance
can be defined as
This distance can be generalized
for a partitioning P
The normalized version of this
distance is defined as Modularity

Modularity Maximization
Modularity matrix
d  Rn*1 is the degree vector for all nodes
Reformulation of the modularity

Modularity Maximization: Example
Modularity Matrix
k-means
Two Communities:
{1, 2, 3, 4} and {5, 6, 7, 8, 9}

Information Diffusion

Information Diffusion
• Sender(s). A sender or a small set of senders that
initiate the information diffusion process;
• Receiver(s). A receiver or a set of receivers that
receive diffused information. Commonly, the set of
receivers is much larger than the set of senders and
can overlap with the set of senders; and
• Medium. This is the medium through which the
diffusion takes place. For example, when a rumor is
spreading, the medium can be the personal
communication between individuals

Information Diffusion Types
We define the process of interfering with information diffusion
by expediting, delaying, or even stopping diffusion as
Intervention

• Network is observable
• Only public information is
available
Herd Behavior

Herd Behavior Example
• Consider people participating in an online auction.
• In this case, individuals can observe the behavior of others by
monitoring the bids that are being placed on different items.
• Individuals are connected via the auction’s site where they can
not only observe the bidding behaviors of others, but can also
often view profiles of others to get a feel for their reputation
and expertise.
• In these online auctions, it is common to observe individuals
participating actively in auctions, where the item being sold
might otherwise be considered unpopular.
• This is due to individuals trusting others and assuming that
the high number of bids that the item has received is a strong
signal of its value. In this case, Herd Behavior has taken place.

Herding: Elevator Example
http://www.youtube.com/watch?v=zNNz0yzHcw
g

Herding: Urn Experiment
• There is an urn in a large class with
three marbles in it
• During the experiment, each student
comes to the urn, picks one marble,
and checks its color in private.
• The student predicts majority blue or
red, writes her prediction on the
blackboard, and puts the marble back
in the urn.
• Students can’t see the color of the
marble taken out and can only see the
predictions made by different students
regarding the majority color on the
board
B B R R R B
50% 50%

Designing a Herd Behavior Experiment
1. There needs to be a decision made.
- In our example, it is going to a restaurant
2. Decisions need to be in sequential order;
3. Decisions are not mindless and people have
private information that helps them decide; and
4. No message passing is possible. Individuals
don’t know the private information of others,
but can infer what others know from what is
observed from their behavior.

Herding Intervention
In herding, the society only has access to public
information.
Herding may be intervened by releasing private
information which was not accessible before
The little boy in “The
Emperor’s New Clothes” story
intervenes the herd by
shouting “There is no clothe”

• In the presence of a
network
• Only local information is
available
Information Cascade

Information Cascade
• In social media, individuals commonly repost content posted by
others in the network. This content is often received via immediate
neighbors (friends).
• An Information Cascade occurs as information propagates through
friends
• An information cascade is defined as a piece of information or
decision being cascaded among a set of individuals, where
– 1) individuals are connected by a network and
– 2) individuals are only observing decisions of their immediate neighbors
(friends).
• Therefore, cascade users have less information available to them
compared to herding users, where almost all information about
decisions are available.
In cascading, local information is available to the
users, but in herding the information about the
population is available.

Underlying Assumptions for Cascade Models
• The network is represented using a directed graph.
Nodes are actors and edges depict the communication
channels between them. A node can only influence nodes
that it is connected to;
• Decisions are binary - nodes can be either active or
inactive. An active nodes means that the node decided to
adopt the behavior, innovation, or decision;
• A node, once activated, can activate its neighboring
nodes; and
• Activation is a progressive process, where nodes change
from inactive to active, but not vice versa 1.

Independent Cascade Model: An Example

• The network is not observable
• Only public information is
observable
Diffusion of Innovations

Diffusion of Innovation
• an innovation is “an idea, practice, or object that
is perceived as new by an individual or other unit
of adoption”
• The theory of diffusion of innovations aims to
answer why and how these innovations spread.
It also describes the reasons behind the diffusion
process, individuals involved, as well as the rate
at which ideas spread.

Diffusion of innovation: Adopter Categories
Cumulative Adoption Rate
Time

Rogers: Diffusion of Innovations: The Process
• Awareness
– The individual becomes aware of the innovation, but her
information regarding the product is limited
• Interest
– The individual shows interest in the product and seeks
more information
• Evaluation
– The individual tries the product in his mind and decides
whether or not to adopt it
• Trial
– The individual performs a trial use of the product
• Adoption
– The individual decides to continue the trial and adopts the
product for full use

Social Media Mining
Challenges

Some New Challenges in Mining Social Media
• Evaluation Dilemma
– Evaluation without conventional test data, but
how?
• Big-Data Paradox
– Often we get a small sample of (still big) data. How
can we ensure if the data can offer credible
findings?
• Noise-Removal Fallacy
– How do we remove noise without losing too much?

Challenge 1: Evaluation Dilemma
• In conventional data mining, training and test
datasets are used to validate findings and
compare performance.
• Without training-test data and with the need
to evaluate, how can we do it?
– User study, Amazon Mechanical Turk, …
– Are they scalable, reproducible, or applicable?
• We need to explore new ways of evaluation.

Understanding User
Migration Patterns in Social
Media
Joint work with Shamanth Kumar and
Reza Zafarani
AAAI 2011, San Francisco, CA

Migration in Social Media
• What is migration?
– Migration can be described as the movement of
users away from one location toward another,
either due to necessity, or attraction to the new
environment.
• Migration in social media can be of two
types
– Site migration
– Attention migration

Why is Migration Important?
• Users are a primary source of revenue
– Ads, Recommendations, Brand loyalty
• New social media sites need to attract new
users to expand their user base
• Existing sites need to retain their users by
migration prevention
• Competition for attention entails the
understanding of migration patterns

Obtaining User Migration Patterns
• Goal: Identifying trends of attention
migration of users across the two phases of
the collected data.
• Process

Patterns from Observation

Facing an Evaluation Dilemma
• Important to know if they are valid or not
– If yes, we investigate further how we use the
patterns to: prevention or promotion.
– If not, why not? And what can we do?
• We would like to evaluate migration
patterns, but without ground truth
• How?
– User study or AMT?

Evaluating Patterns’ Validity
• One way is to verify if these patterns are
fortuitous
• Null Hypothesis: Migration of individuals is a
random process
– Generating another similar dataset for comparison
• Potential migrating population includes overlapping
users from Phase 1 and Phase 2
• Shuffled datasets are generated by picking random
active users from the potential migrating population
• The number of random users selected for each dataset
is the same as the real migrating population

A Significance Test
Shuffled
dataset
Observed
migration
dataset
Coefficients
of user
attributes
Comparing
and
Sig. Test
Logistic
Regression
Chi Square
Statistic
Coefficients
of user
attributes
Logistic
Regression

Evaluation Results
• Significant differences observed in StumbleUpon,
Twitter, and YouTube
• Patterns from other sites are not statistically
significant. Potential cause:
– Insufficient Data?

Summary
• Mitigating or promoting migration by
targeting high net-worth individuals
– Identifying users with high value to the network,
e.g., high network activity, user activity, and
external exposure
• Social media migration is first studied in this
work
• Migration patterns can be evaluated without
test data

Challenge 2: Big-Data Paradox
• What is Big Data?
– 3Vs, 4Vs, or 5Vs …
• Is Social Media Data big?
– Yes, it is obviously so (e.g., FB and Twitter)
– But, we are often limited by source APIs (e.g.,
Twitter Streaming API offers 1% data)
• What can we do facing the cold reality?
– It depends on if you’re rich, famous, lazy, or curious

Is the Sample Good Enough?
Comparing Data from Twitter’s
Streaming API and Data from
Twitter’s Firehose
Joint Work with Fred Morstatter,
Jürgen Pfeffer, and Kathleen Carley
AAAI ICWSM2013, Boston, MA

Big-Data Problems
• Twitter provides two main outlets for
researchers to access tweets in real time:
– Streaming API (~1% of all public tweets, free)
– Firehose (100% of all public tweets, costly)
• Streaming API data is often used to by
researchers to validate hypotheses.
• How well does the sampled Streaming API
data measure the true activity on Twitter?

Facets of Twitter Data
• Compare the data along different facets
• Selected facets commonly used in social
media mining:
– Top Hashtags
– Topic Extraction
– Network Measures
– Geographic Distributions

Preliminary Results
Top Hashtags Topic Extraction
• No clear correlation
between Streaming and
Firehose data.
• Topics are close to those
found in the Firehose.
Network Measures Geographic Distributions
• Found ~50% of the top
tweeters by different
centrality measures.
• Graph-level measures give
similar results between the
two datasets.
• Streaming data gets >90%
of the geotagged tweets.
• Consequently, the
distribution of tweets by
continent is very similar.

How Good are These Results?
• Accuracy of Streaming API varies with
analysis the researcher wants to perform.
• These results are about single cases of
streaming API.
• Are these findings significant, or just an
artifact of random sampling?
• How do we verify that our results indicate
sampling bias or not?

Probing Further
• Aggregate data by day
• Select 5 days with different levels
of coverage
• Create random
samples from the
Firehose data to
compare against
the Streaming API

Comparative Results
Top Hashtags Topic Extraction
• No correlation between
Streaming and Firehose
data.
• Not so in random
samples
• Topics are close to those
found in the Firehose.
• Topics extracted with the
random data are
significantly better.

Summary
• Streaming API data can be biased in some
facets.
• Our results were obtained with the help of
Firehose.
• Without Firehose data, challenges are to
figure out which facets have biases, and
how to compensate them in search of
credible mining results

Challenge 3: Noise-Removal Fallacy
• A common complaint: “99% Twitter data is useless”.
– “Had eggs, sunny-side-up, this morning”
– Can we remove the noise as we usually do in DM?
• What is left after noise removal?
– Twitter data can be rendered useless after
conventional noise removal
• As we are certain there is noise in data,
how can we remove it?

Feature Selection with
Linked Data in Social Media
Joint Work with Jiliang Tang
SDM2012 and KDD2012

Social Media Data
• Massive and high-dimensional social media
data poses unique challenges to data mining
tasks
– Scalability
– Curse of dimensionality
• Social media data is inherently linked
– A key difference between social media data and
attribute-value data

Feature Selection of Social Data
• Feature selection has been widely used to
prepare large-scale, high-dimensional data
for effective data mining
• Traditional feature selection algorithms deal
with only “flat" data (attribute-value data).
– Independent and Identically Distributed (i.i.d.)
• We need to take advantage of linked data for
feature selection

An Example of Social Media Data
𝑢1
𝑢2
𝑢3
𝑢4
𝑝1
𝑝2
p3 𝑝5
𝑝6
𝑝4
𝑝7
𝑝8
User-post
relations

Representation for Attribute-Value Data
𝑝1
𝑝2
𝑝5
𝑝6
𝑝7
𝑝8
𝑓𝑚
…. …. ….
Posts

Representation for Social Media Data
1
1 1 1
1
1 1
𝑢1
𝑢2
𝑢3
𝑢4
𝑢1 𝑢2 𝑢3 𝑢4
𝑝1
𝑝2
𝑝5
𝑝6
𝑝4
𝑝7
𝑝8
𝑓𝑚
…. …. …. 𝑐 𝑘….
Social
Context

How to Use Link Information
• The new question is how to proceed with
additional information for feature selection
• Two basic technical problems
– Relation extraction: What are distinctive
relations that can be extracted from linked data
– Mathematical representation: How to use these
relations in feature selection formulation
• Do we have theories to guide us?

𝑢1
𝑢2
𝑢3
𝑢4
𝑝1
𝑝2
p3 𝑝5
𝑝6
𝑝4
𝑝7
𝑝8
Relation Extraction
1.CoPost
2.CoFollowing
3.CoFollowed
4.Following

Relations, Social Theories,
Hypotheses
• Social correlation theories suggest that the
four relations may affect the relationships
between posts
• Social correlation theories
– Homophily: People with similar interests are
more likely to be linked
– Influence: People who are linked are more likely
to have similar interests
• Thus, four relations lead to four hypotheses

Modeling CoFollowing
Relation
• Two co-following users have similar interested topics
||||
)(
^
k
Ff
i
T
k
Ff
i
k
F
fW
F
fT
uT kiki
 
）（
• Users' topic interests
 

u Nuu
jiF
T
uji
uTuT
,
2
2
^^
1,2
2
W
||)()(||||W||||YWX||min 

Evaluation Results on Digg

Summary
• LinkedFS is evaluated under varied circumstances
to understand how it works.
– Link information can help feature selection for
social media data.
• Unlabeled data is more often in social media,
unsupervised learning is more sensible, but also
more challenging.
• An unsupervised method is showcased in our
KDD12 paper following social correlation theories

Thank You All
Acknowledgments: Projects are, in part,
sponsored by ARO, NSF, and ONR;
thanks to passionate and creative DMML
members and our collaborators.
Thanks for Your Interests and
Participation

Two DMML Books of SMM
Twitter Data Analytics
Nov. 2013
Social Media Mining
Feb. 2014

Social Media Mining: An Introduction

More Related Content

What's hot

Viewers also liked

Similar to Social Media Mining: An Introduction

Recently uploaded

Social Media Mining: An Introduction