A Dynamic Algorithm for Local Community Detection in Graphs : NOTES

A Dynamic Algorithm for Local Community
Detection in Graphs
Anita Zakrzewska and David A. Bader
Computational Science and Engineering
Georgia Institute of Technology
Atlanta, Georgia 30332
azakrzewska3@gatech.edu, bader@cc.gatech.edu
Abstract—A variety of massive datasets, such as social net-
works and biological data, are represented as graphs that reveal
underlying connections, trends, and anomalies. Community de-
tection is the task of discovering dense groups of vertices in a
graph. Its one specific form is seed set expansion, which finds
the best local community for a given set of seed vertices. Greedy,
agglomerative algorithms, which are commonly used in seed
set expansion, have been previously designed only for a static,
unchanging graph. However, in many applications, new data is
constantly produced, and vertices and edges are inserted and
removed from a graph. We present an algorithm for dynamic
seed set expansion, which incrementally updates the community
as the underlying graph changes. We show that our dynamic
algorithm outputs high quality communities that are similar
to those found when using a standard static algorithm. The
dynamic approach also improves performance compared to re-
computation, achieving speedups of up to 600x.
I. INTRODUCTION
Graphs are used to represent relationships and communica-
tion between entities in fields such as web traffic, financial
transactions, online communications, and biology. A com-
monly studied feature of graphs is community structure. A
graph community may be broadly defined as a set of vertices
that is densely connected. In datasets representing online social
networks, multiplayer games, or on-line project management,
graph communities correspond to groups of friends on Face-
book, on-line players, or officemates who work together on
the same project. They can also be found in a variety of other
graphs, such as protein-protein interaction networks.
Global community detection methods divide the entire
graph into groups. Existing global algorithms include random
walk methods, spectral partitioning, label propagation, greedy
agglomerative and divisive algorithms, and clique percola-
tion [1]. While most methods partition the graph into mutually
disjoint groups, there is a growing body of work in detecting
overlapping communities [2]. However, there has been little
work done in detecting a local community relevant to a small
set of vertices of interest, which we call seed vertices. This
problem is called seed set expansion. Because many graphs
may now have billions of vertices, visualization is difficult
and many computationally intensive algorithms cannot be run
on commodity platforms. Seed set expansion can be used in
such cases to extract a relatively small, relevant subgraph.
A. Contributions
In contrast to static seed set expansion, which is run once
on an unchanging graph, dynamic seed set expansion is a new
area. Edges may be inserted or removed to reflect evolving
actions, communications, or relationships between entities.
When the graph changes, the community of a seed vertex must
be updated as well. Here we develop a dynamic algorithm to
incrementally update a local community when the underlying
graph changes. This algorithm improves performance com-
pared to re-computing. It can handle batch updates of various
sizes and is easily parallelized for multiple expansions from
different seeds. To the best of our knowledge, this is the first
dynamic algorithm for greedy seed set expansion.
B. Definitions and Related Work
Let G = {V, E} be a graph, where V is the set of vertices
and E the set of undirected edges. An edge (u, v, ω) ∈ E
consists of two unordered vertices u, v, and a weight ω. Let
kC
in be the sum of all edge weights interior to community C
and kC
out be the sum of all edge weights on the border of C.
kC
in =
X
(u,v,ω)∈E|u∈C∧v∈C
ω (1)
kC
out =
X
(u,v,ω)∈E|u∈C∧v /
∈C
ω (2)
The quality of a community C is often measured using a fit-
ness function. As there is no single definition of a community,
many fitness functions are commonly used. Modularity, shown
in Equation 3, compares the number of intra-community edges
to the expected number under a random null model [3].
Q(C) =
1
|E|
(kC
in −
(2kC
in + kC
out)2
4 |E|
) (3)
Conductance is another popular fitness score and measures the
community cut, or inter-community edges [4].
φ(C) =
kC
out
min(2kC
in + kC
out, 2k
V C
in + k
V C
out )
(4)
Many overlapping community detection methods use a modi-
fied ratio of intra-community edges to all edges with at least
one endpoint in the community, such as in Equation 5 [5].
f(C)MONC =
2kC
in + 1
(2kC
in + kC
out)α
(5)
2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining
559
ASONAM '15, August 25-28, 2015, Paris, France
© 2015 ACM. ISBN 978-1-4503-3854-7/15/08 $15.00
DOI: http://dx.doi.org/10.1145/2808797.2809375

Andersen et al. [6] use the Spectral PageRank-Nibble
method. Their final community minimizes conductance and
is formed by adding vertices in order of decreasing PageRank
values. In the random walk approach of Andersen and Lang [7]
some vertices in the seed set may not be placed in the final
community. Clauset gives a greedy method that starts from
a single vertex and then iteratively adds neighboring vertices
maximizing the local modularity score [8]. Riedy et al. expand
multiple vertices via maximizing modularity [9].
Several algorithms for detecting global, overlapping com-
munities use a greedy, agglomerative approach and run multi-
ple separate seed set expansions [10] [5] [2]. Lancichinetti
et al. run greedy seed set expansions, each with a single
seed vertex [10]. Overlapping communities are produced by
sequentially running expansions from a node not yet in a
community. Lee et al. use maximal cliques as seed sets [11].
Havemann et al. [5] greedily expand cliques. Each of these
greedy algorithms can be generalized to the form given by
Algorithm 1. The community is iteratively expanded by adding
the neighboring vertex that maximizes the chosen fitness
function. The algorithm terminates when there exists no vertex
whose inclusion in the community increases the fitness score.
In Algorithm 1, seed represents the initial set of seed vertices,
fit(C) the fitness score for a community C, and Nb(C) the
set of vertices not in C with at least one neighbor in C. We use
this static algorithm as the basis for our new dynamic method,
presented in Section II, using the fitness metric f(C)MONC
from Equation 5, though the approach will work for other
fitness functions as well.
Algorithm 1: Static, Greedy Seed Set Expansion
Data: graph G and seed set seed
C = seed;
while progress do
maxscore = −1;
maxvtx = −1;
for v ∈ Nb(C) do
s(v) = fit(C ∪ v) − fit(C);
if s(v) > maxscore then
maxscore = s(v);
maxvtx = v;
end
end
if maxscore > 0 then
C = C ∪ maxvtx;
end
end
II. DYNAMIC SEED SET EXPANSION ALGORITHM
A. Motivation
Our dynamic seed set expansion algorithm incrementally
updates the community when the underlying graph changes.
Since incremental updates are faster than re-computation,
our method can be used to improve performance for any
application of seed set expansion, as described in Section I.
Should be
removed
(a) Undetected community splitting.
(b) Undetected seed migration.
Fig. 1: Shortcomings of the simplistic algorithm from Sec-
tion II-A. Undesired community evolution shown left to right.
We begin with an initial graph G and perform a static
seed set expansion, as in Algorithm 1, resulting in the initial
community C. Next, a sequence of updates is applied to G
and we incrementally update C to reflect changes in graph
structure. Each graph update is of the form (u, v, ∆ω), where
u and v are edge endpoint vertices and ∆ω is an increment or
a decrement in edge weight. An edge insertion is represented
by a weight increment to a non-existent edge, while a deletion
is represented by a decrement of the edge weight to 0.
To motivate our approach, we first discuss a simple algo-
rithm for dynamic seed set expansion and the problems it
may run into. After every edge update, the simple method
can check each endpoint vertex and calculate the fitness score
of the community C both with and without that vertex. For
example, consider an edge update (u, v, ∆ω) with ∆ω > 0,
u ∈ C and v /
∈ C. We check fit(C), fit(C∪v), and fit(Cu)
and select the community update with the highest score. For an
update with ∆ω < 0, u ∈ C, and v ∈ C, we calculate fit(C),
fit(C v), fit(C u), and fit(C u, v) and choose the set
with the highest score. Next, all vertices in Nb(C) can be
rechecked for inclusion in C. This is a simple local approach
because only endpoints of updated edges are evaluated.
Unfortunately, this simple method has severe shortcomings.
For example, the community may split apart and the algorithm
may not be able to detect this because the removal of no
individual vertex increases the community score. This is shown
in Figure 1a, where the community is no longer optimal as
it is actually composed of two natural communities. In the
set of vertices that has split off and should be removed,
most neighbors of each vertex are also in the community
and therefore no vertex will be removed. Even if we evaluate
multiple vertices at once for removal, the same problem may
occur if the set that has split off is large enough.
The simple updating method may fail even when it outputs
a valid community in the graph. This is because seed set
560

position 0 1 2 ... n
members m0 m1 m2 . . . mn
inner edges k0,in k1,in k2,in . . . kn,in
border edges k0,out k1,out k2,out . . . kn,out
fitness score s0 s1 s2 . . . sn
TABLE I: Community evolution sequence Ψ
expansion differs from global community detection in an
important way: the local community is chosen for a particular
seed set. The task is not simply to find any good community
in the graph, but rather the appropriate community for the
seed. Changes to the graph may shift the community C to one
not centered around the original seed, as shown in Figure 1b.
While C may still have a good fitness score, it may not be a
local community of the seed and would not be produced by a
complete re-computation using static seed set expansion.
Given these considerations, quality evaluation for an up-
dated community of a seed is more difficult than for general
communities. We must consider not only the degree to which
the chosen set of vertices resembles a community, but also
whether it is a good community for the seed. A static seed set
expansion algorithm detects the best community for the seed
set using full information. Thus, one method of determining
quality is to use the community found using static seed set
expansion as a baseline and consider an incremental updating
algorithm to be successful if it produces similar results.
B. Algorithm Overview
The dynamic seed set expansion algorithm begins with
the computation of a community using a static expansion
on the initial graph as in Algorithm 1. When the algorithm
begins, the community initially contains only the seed, and
new vertices are then iteratively added. In each iteration, the
neighboring vertices of the current community are potential
new members and the vertex producing the greatest increase in
the fitness score is chosen. The initial computation thus results
in an ordered sequence of vertices added to the community
and a corresponding sequence of nested sets, each with an
increasingly greater fitness score. As the goal is to maintain a
community centered around the seed, it is necessary to keep
track of the order in which vertices were added.
Let mi denote the ith
vertex added as a member of the
community in Algorithm 1 and Mi = {mj | j ≤ i}. Mi has
an interior edge weight sum of ki,in, a border edge weight
sum of ki,out, and a fitness score of si. Note that ki,in is
equal to kMi
in and ki,out is equal to kMi
out as in Equations 1
and 2. If mi is vertex v, then we say that v has position
i or ρ(v) = i. We refer to this collection of sequences by
Ψ = {mi, ki,in, ki,out, si | 0 ≤ i ≤ end}, as shown in Table I.
Here by end we represent the last position in the sequence Ψ,
and Mend is the current community, which we also call C.
The dynamic algorithm works as follows. In phase A, we
start with the initial graph and perform static seed set expan-
sion (Algorithm 1) to produce Ψ. In phase B, a stream of graph
updates is applied. With each graph update, the algorithm
updates the community by modifying Ψ while ensuring that
the sequence si remains monotonically increasing. That is,
we require the updated community to contain vertices that,
if added one by one as in the static algorithm, result in
an increasing sequence si. This guarantees that the resulting
community remains relevant to the source seed.
For each update, we modify the sequence of community
members mi to ensure that the corresponding fitness scores
are increasing. After a batch of updates, the algorithm detects
any decreases in the sequence of fitness scores and removes
vertices from the community to eliminate any such decrease.
Next, it checks if any new vertices should be added and
updates Ψ if needed.
C. Algorithm Details
The dynamic algorithm updates Ψ after each graph update
and ensures both that the sequence of fitness scores si remains
monotonically increasing and that there are no additional ver-
tices in G whose inclusion in the community would increase
the fitness score further. For each batch of edge updates, the
following four steps are performed (some may be omitted
depending on the case). Further explanations are given later.
1) Values ki,in, ki,out, and si in Ψ are updated to reflect
new internal and border edges.
2) Vertices that are endpoints of an updated edge are
checked for removal. If a vertex v is removed, the com-
munity members are further pruned and Ψ is updated to
reflect the pruning.
3) Ψ is scanned to check that the sequence of fitness
scores si is still monotonically increasing. If a dip
exists at position i, Ψ is truncated after position i (set
Ψ = Ψ0,i−1).
4) The static seed set expansion algorithm is restarted to
check whether neighboring vertices in Nb(C) should be
added to the community.
Algorithm 2 summarizes our method. To save space, two
simplifying assumptions are made. First, as edges are undi-
rected, the order of vertices in an edge update (u, v, ∆ω) is
arbitrary. Therefore, we only consider ρ(u) < ρ(v). Second,
we only consider a single edge update, even though the
algorithm can handle batch updates. In the case of a batch, step
1 is first performed for each edge, then step 2 is performed
for each edge, and then steps 3 and 4 are only executed once.
The complexity of the static seed set expansion is O(n2
d),
where n is the final community size and d is the average
degree, though this is an overestimate for graphs whose
vertices share many neighbors. In the worst case, the dynamic
algorithm must recompute a large portion of the community,
in which case the complexity is O(n2
d) as well. In practice,
many of the updates result in no decrease in the fitness score
sequence so that only a scan of Ψ is needed and the complexity
becomes O(n). Additional details of each step are given next.
Step 1: First, ki,in, ki,out, and si in Ψ are updated to reflect
new edges internal to and on the border of the community. The
input is (u, v, ∆ω), where u and v are vertices and ∆ω the
corresponding change in weight.
561

Algorithm 2: Dynamic Seed Set Expansion Algorithm
Data: edge (u, v, ∆ω)
//Step 1
if u ∈ C and v ∈ C then
for p = ρ(u) to ρ(v) − 1 do
kp,out+ = ∆ω;
update sp;
end
for p = ρ(v) to end do
kp,in+ = ∆ω;
update sp;
end
else if u ∈ C and v /
∈ C then
for p = ρ(u) to end do
kp,out+ = ∆ω;
update sp;
end
//Step 2
if ∆ω < 0 and u ∈ C and v ∈ C then
Queue ← v
else if ∆ω > 0 and u ∈ C and v /
∈ C and u 6= seed then
Queue ← u
else if ∆ω > 0 and u ∈ C and v ∈ C and u 6= seed then
Queue ← u
while Queue not empty do
u ← Queue;
if u ∈ C and sρ(u)−1 ≥ sρ(u) then
remove u from C;
update Ψ to reflect removal;
for neighbors w of u do
if w ∈ C and ρ(w) > ρ(u) then
Queue ← w;
end
end
end
end
//Step 3
for i = max(ρ(u), 1) to end do
if si−1 ≥ si then
end ← i − 1;
C = Mi−1;
break;
end
end
//Step 4
Check for new members using static algorithm;
Step 2: Once Ψ has been updated in step 1, the fitness score
sequence si may no longer be monotonically increasing. For
some edge updates, keeping one of the edge endpoints in C
may cause a decrease in fitness score. For an edge update
(u, v, ∆ω) with v ∈ C, we check whether sρ(v)−1 ≥ sρ(v). If
so, then keeping v as the ρ(v)th
member of C causes a non-
increase in the fitness score. Accordingly, v is removed from
C and Ψ must be updated: ki,in, ki,out, and si for ρ(v) ≤
i ≤ end must be recalculated to reflect the fact that edges of
v are no longer inside C. For each edge (v, u), if u ∈ C, the
edge changes from an internal community edge (contributing
to kin), to a border edge (contributing to kout). If u /
∈ C, the
edge changes from a border edge to an edge with no influence
on the fitness score. Only entries in Ψ after position ρ(v)
must be updated because previous entries were added to the
community before v.
The removal of a vertex v from C in step 2 may cause
other vertices in C to be removed as well. Candidate vertices
are neighbors of v that were added to C after v. Let u be
such a neighbor. At the time of u’s inclusion in C, adding u
increased the fitness score by increasing kin, which was due
to u having neighbors already in the community. However, at
least one such neighbor was v, which is now no longer in C.
Thus, it is possible that without v in C, u would not have
enough neighbors in C to be added. We can check this by
testing if sρ(u)−1 ≥ sρ(u). If v is removed from C, all such
neighbors u of v in C are also checked. Neighbors of v added
to C before v (ρ(u) < ρ(v)) need not be checked because they
were added to C without the assistance of v. If any neighbor
u of v is removed, then we must in turn check neighbors of
u that were added to C after u. In order to perform the entire
pruning process, a selective breadth first search beginning from
v is performed, as in step 2 of Algorithm 2.
Step 2 is only performed if there is a specific candidate
vertex for removal, which will always be an endpoint of
an updated edge. An edge update (u, v, ∆ω) can cause the
removal of an endpoint v only when sρ(v)−1 ≥ sρ(v) due to
either a decrease in kρ(v),in or an increase in kρ(v),out. This
occurs in three cases. The first case is an edge decrement
with v ∈ C, u ∈ C, and ρ(u) < ρ(v). The second is an
edge increment with v ∈ C and u /
∈ C. The third is an edge
increment with v ∈ C, u ∈ C, and ρ(v) < ρ(u). This third
case may seem counter-intuitive because an intra-community
edge is incremented, densifying the community. However, we
must maintain an increasing sequence of fitness scores si in
Ψ. As v was added to C before u, the edge between v and u
is a border edge at position ρ(v), and becomes internal only
starting at position ρ(u). Thus, by incrementing it, the sum
of border edges kρ(v),out increases. If, due to this increase,
sρ(v) ≤ sρ(v)−1, then v must be removed from C. In a later
step, v may be re-added to C, but it must be removed from
position ρ(v) of Ψ because it causes a non-increase of si.
Step 3: Next we scan all of Ψ to check if the si are still
monotonically increasing. If si−1 ≥ si, we truncate Ψ at
position i − 1 and set end = i − 1. The community is now
C = Mi−1 with fitness score si−1, and the sequence of fitness
scores in Ψ is monotonically increasing.
Step 3 differs from step 2 because instead of a selective
pruning, all of Ψ after the chosen position is deleted. It also
serves a different purpose than step 2. We perform step 2
only when there is a specific candidate vertex to check for
removal from C. Step 3 can check all vertices. For example,
let the update be (u, v, ∆ω), with v ∈ C, u ∈ C, ρ(v) <
ρ(u), and ∆ω > 0. By incrementing an intra-community edge,
562

kρ(u),in increases and the set Mρ(u) becomes denser. Thus, any
vertex added after position ρ(u) may no longer increase the
fitness score, and all Ψ after position ρ(u) must be scanned
for such vertices. In addition, after step 2 the entire sequence
of fitness scores may still no longer be increasing. Step 3
is more computationally expensive than step 2 because after
detecting a score drop at position i, step 3 truncates all of Ψ
after i − 1, while step 2 selectively prunes. However, unlike
step 2, it guarantees a monotonically increasing sequence of
fitness scores. Of course, step 3 could replace step 2 entirely,
but this increases running time. In section III we show results
for a modified dynamic algorithm that skips step 2.
Step 4: Finally, new vertices can be added to the community.
Vertices neighboring C are checked for inclusion by running
the loop in Algorithm 1. For every vertex added to C, Ψ is
updated by appending a new entry that includes that vertex
and the corresponding sum of interior edges kend,in, sum of
border edges kend,out, and fitness score send.
III. RESULTS
We test the dynamic seed set expansion algorithm on three
social network graphs, listed in Table II. The first graph is
built from Slashdot data [12] [13], where vertices represent
website users and edge weights correspond to numbers of
replies between users. The second graph represents user replies
on Digg [14], and the third is a computer science co-authorship
network DBLP [15]. As these graphs represent social interac-
tions, they are likely to display group structure. The Slashdot
and Digg datasets contain timestamps, which define the order
for edge insertion or incrementation. There are no timestamps
in the DBLP dataset, so the edge order is defined by random
permutation. In phase A we run static seed set expansion on
the initial graph, which is formed from the first half of edges.
In phase B, we sequentially insert the remaining edges and
dynamically update the community. For edge decrements and
removals, old data is removed from the graph. The fitness
function fMONC is used with α = 1.0 and α = 0.5. For
seeds, we chose the top 1000 vertices in the graph by degree
centrality, and then used each of them as a seed vertex in a
separate seed set expansion. The code was implemented in C
and run on an 8 core Intel i7-2600K CPU at 3.40GHz.
In order to compare the communities output by the dynamic
algorithm to those from static re-computation, we repeatedly
re-run the static algorithm as the graphs are updated. The
algorithm performance is measured by three metrics. The
first is the average ratio of the fitness scores in the dynamic
algorithm vs. those obtained by re-computation. The average
is taken across all timestamps and all expansions run. In
Table II these ratios are above 1 for each graph, showing that
the dynamic algorithm produces higher fitness scores. Two
other metrics, precision and recall, compare the members of
communities output by the dynamic and static algorithms. For
a given graph update, let CU be the of community produced
by the dynamic algorithm and CR be the community output
by the static method. Then Equations 6 and 7 give precision
and recall. In Table II precision and recall are in most cases
Graph Vertices Edges Score Ratio Recall Precision
Slashdot, α=1.0 51,083 140,778 1.0 0.92 0.82
Slashdot, α=0.5 51,083 140,778 1.2 0.84 0.70
Digg, α=1.0 30,398 187,627 2.4 0.94 0.81
Digg, α=0.5 30,398 187,627 2.8 0.88 0.57
DBLP, α=1.0 317,080 1,049,866 1.1 0.91 0.87
DBLP, α=0.5 317,080 1,049,866 3.6 0.85 0.79
TABLE II: Quality of communities found with the dynamic
algorithm compared to recomputing with the static algorithm
0.0
0.2
0.4
0.6
0.8
1.0
Recall
dynamic alg
dynamic alg, no step 2
simple alg
initial community
0 10000 20000 30000 40000 50000
Update Count
0.0
0.2
0.4
0.6
0.8
1.0
Precision
Fig. 2: The average precision and recall for the full dynamic
algorithm (solid blue line), a modified version without step 2
(dashed green line), and the simple approach from Section II-A
(red dotted line). The evolution of baseline communities vs.
the initial communities is shown in purple dots and dashes.
above 0.8, indicating good quality. Also, better performance
is obtained with α is set to 1.0, which gives smaller, denser
communities.
precision =
|CU ∩ CR|
|CU |
(6)
recall =
|CU ∩ CR|
|CR|
(7)
Figure 2 shows details of the dynamic algorithm’s perfor-
mance on the Slashdot graph. The average recall and precision
are plotted against the number of updates. Averages are taken
across multiple independent expansions, each with its own
seed. The solid blue line represents the dynamic algorithm.
Despite occasional dips, there is no downward trend in either
precision or recall, showing that the community quality does
not deteriorate over time. In addition to the full dynamic
algorithm, we also evaluate a modified version, in which step
2, selective pruning, is not performed. As expected, without
pruning, the quality of results, shown with the dashed green
line, is somewhat higher, but with a slower running time. Next
we examine if the high precision and recall of our algorithm
could be due to the fact that the communities in the dataset
simply do not change much as the graph is updated over time.
We test this possibility using two approaches. First, for each
563

10
0
10
1
10
2
10
3
10
4
batch size
10
-2
10
-1
10
0
10
1
10
2
10
3
speedup
over
recomputation
minimum
median
maximum
Fig. 3: The speedup of our dynamic algorithm over recomput-
ing with the standard, static algorithm. Higher is better.
seed vertex, we take the initial community calculated using
static seed set expansion. For each graph update over time,
we use precision and recall to compare that initial community
to the one produced by static re-computation. The results,
plotted in Figure 2 with the purple dots and dashes, show the
recall and precision of initial communities severely decrease
as the graph evolves. This means that the communities in
the Slashdot graph do change over time and our algorithm
is able to update appropriately. The second test compares the
precision and recall produced by the simple, local updating
approach described in Section II-A. These are plotted with
the red, dotted line in Figure 2. Although recall remains high,
the precision deteriorates. This occurres because the simple
approach expands communities to contain a very large portion
of the graph.
Figure 3 plots the speedup of the dynamic algorithm over re-
computation. For this experiment we used batches of updates
of varying sizes. The x-axis shows the batch size used and
the y-axis shows the ratio of re-computation running time to
dynamic running time, both on a log scale. It is clear that
the advantage of the dynamic algorithm is greatest for small
batch sizes. This is expected because the as the batch size is
increased by a factor of x, re-computation occurs x times less
frequently and its running time also decreases by a factor of
x. The dynamic algorithm also performs less work with larger
batch sizes, as steps 3 and 4 are only run once per batch,
but the decrease is not by a factor of x as some steps must
occur the same number of times regardless of the batch size.
The dynamic algorithm produces speedups of over 600x for
a batch size of 1 and over 200x for a batch size of 10. The
median speedup is over 60x for a batch size of 1 and remains
greater than 1x for batch sizes up to 1000. This suggests that
the dynamic algorithm is faster than re-computation even for
large batches of 1000 updates. On graphs that contain larger
communities, we would expect the dynamic algorithm to be
even more advantageous, as re-computing a large community
from scratch takes longer.
IV. CONCLUSION
We have presented a new algorithm that incrementally
updates the local community of a seed set when the underlying
graph changes. Our approach produces communities with high
fitness scores and similar to the communities produced by
a standard greedy algorithm that must be re-run whenever
the graph is updated. The dynamic method is faster than re-
computation, and the performance improvement is greatest
when low latency updates are needed. The speedup achieved
varies based on the size of a local community, with the dy-
namic algorithm performing relatively better on large commu-
nities. The algorithm is easily parallelized across independent
expansions, which may be addressed in future work.
ACKNOWLEDGMENT
The work depicted in this paper was sponsored by De-
fense Advanced Research Projects Agency (DARPA) under
agreement #HR0011-13-2-0001. The content, views and con-
clusions presented in this document do not necessarily reflect
the position or the policy of DARPA or the U.S. Government,
no official endorsement should be inferred.
REFERENCES
[1] S. Fortunato, “Community detection in graphs,” Physics Reports, vol.
486, no. 3, pp. 75–174, 2010.
[2] J. Xie, S. Kelley, and B. K. Szymanski, “Overlapping community
detection in networks: The state-of-the-art and comparative study,” ACM
Computing Surveys (CSUR), vol. 45, no. 4, p. 43, 2013.
[3] M. E. Newman and M. Girvan, “Finding and evaluating community
structure in networks,” Physical review E, vol. 69, no. 2, p. 026113,
2004.
[4] F. R. Chung, Spectral graph theory. American Mathematical Soc.,
1997, vol. 92.
[5] F. Havemann, M. Heinz, A. Struck, and J. Gläser, “Identification of
overlapping communities and their hierarchy by locally calculating
community-changing resolution levels,” Journal of Statistical Mechan-
ics: Theory and Experiment, vol. 2011, no. 01, p. P01023, 2011.
[6] R. Andersen, F. Chung, and K. Lang, “Local graph partitioning using
pagerank vectors,” in Foundations of Computer Science, 2006. FOCS’06.
47th Annual IEEE Symposium on. IEEE, 2006, pp. 475–486.
[7] R. Andersen and K. J. Lang, “Communities from seed sets,” in Proceed-
ings of the 15th international conference on World Wide Web. ACM,
2006, pp. 223–232.
[8] A. Clauset, “Finding local community structure in networks,” Physical
review E, vol. 72, no. 2, p. 026132, 2005.
[9] J. Riedy, D. A. Bader, K. Jiang, P. Pande, and R. Sharma, “Detecting
communities from given seeds in social networks,” 2011.
[10] A. Lancichinetti, S. Fortunato, and J. Kertész, “Detecting the overlap-
ping and hierarchical community structure in complex networks,” New
Journal of Physics, vol. 11, no. 3, p. 033015, 2009.
[11] C. Lee, F. Reid, A. McDaid, and N. Hurley, “Detecting highly overlap-
ping community structure by greedy clique expansion,” in 4th SNA-KDD
Workshop, 2010, p. 3342.
[12] V. Gómez, A. Kaltenbrunner, and V. López, “Statistical analysis of the
social network and discussion threads in slashdot,” in Proceedings of
the 17th international conference on World Wide Web. ACM, 2008,
pp. 645–654.
[13] “Slashdot threads network dataset – KONECT,” Aug. 2014. [Online].
Available: http://konect.uni-koblenz.de/networks/slashdot-threads
[14] M. D. Choudhury, H. Sundaram, A. John, and D. D. Seligmann, “Social
synchrony: Predicting mimicry of user actions in online social media,”
in Proc. Int. Conf. on Computational Science and Engineering, 2009,
pp. 151–158.
[15] J. Yang and J. Leskovec, “Defining and evaluating network communities
based on ground-truth,” in Proc. ACM SIGKDD Workshop on Mining
Data Semantics. ACM, 2012, p. 3.
564

A Dynamic Algorithm for Local Community Detection in Graphs : NOTES

More Related Content

Similar to A Dynamic Algorithm for Local Community Detection in Graphs : NOTES (20)

More from Subhajit Sahu (20)

Recently uploaded (20)

A Dynamic Algorithm for Local Community Detection in Graphs : NOTES