Received August 29, 2020, accepted September 18, 2020, date of publication October 1, 2020, date of current version

October 13, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3028112

NK-MaxClique and MMCQ: Tow New Exact

Branch and Bound Algorithms for
the Maximum Clique Problem
NEDA MOHAMMADI AND MEHDI KADIVAR
Department of Computer Sciences, Shahrekord University, Shahrekord 8818634141, Iran
Corresponding author: Mehdi Kadivar (m_kadivar@aut.ac.ir)

ABSTRACT The maximum clique problem (MCP) is a fundamental problem in combinatorial optimization
which finds important applications in real-word. This paper describes two new efficient branch-and-bound
maximum clique algorithms NK-MaxClique and MMCQ, designed for solving MCP. We define some prun-
ing conditions based on core numbers and vertex ordering to efficiently remove many of the search space.
With respect to this ordering, the algorithms consider the vertices respectively to find the corresponding
maximum clique in subproblems. Simulation results demonstrate that the algorithms outperform the previous
well-known algorithms for many instances when applied to DIMACS benchmark and random graphs.

INDEX TERMS Branch and bound algorithms, complexity, core number, maximum clique.

I. INTRODUCTION
Let G = (V , E) be a undirected graph, where V =
{v1 , v2 , · · · , vn } is the set of vertices and E ⊆ V × V is
the set of its edges. Two vertices are said to be adjacent
if they are connected by an edge. A clique of G is a set
of vertices, any two of which are adjacent. The maximum
clique problem (MCP) is the problem of finding the clique
with the largest number of vertices in G. The MCP has a
wide range of practical applications in numerous fields such
as, in computer vision [1], robotics [2], bioinformatics [3],
computational biology [4], coding theory [5], social network
analysis [6] and scheduling [7].
The MCP is an NP-hard problem [11] and also is hard to
approximate [12]. So the optimal solutions cannot be reached
within a reasonable time and therefore various heuristic
and metaheuristic algorithms have been devised to provide
approximate solutions as good as possible to large problems
within an acceptable time. In [8] a survey of these algorithms
for MCP is presented. In [13] a polynomial-time algorithm
with an approximation guarantee of O(n(log log n)2 /(log n)3 )
is presented.
The exact methods to solve MCP are designed based on
the general branch-and-bound (B&B) technique. These meth-
ods differ from each other by their specific techniques to
The associate editor coordinating the review of this manuscript and
approving it for publication was Dongxiao Yu . work, lead us to an improved bit parallel exact algorithm
determine the lower and upper bounds and their branching
strategies. To estimate the upper bound of the maximum
clique, graph coloring techniques are used. The reason is that
if a graph G can be colored with k colors, then the number
of vertices in a maximum clique in G is smaller or equal
to k. However, coloring is an NP-hard problem and may
be also time consuming, and therefore fast greedy coloring
heuristic algorithms must be used (e.g. DSATUR algorithm in
[16]). An early classic B&B algorithm (denoted by BT) which
employs DSATUR algorithm for finding maximum cliques
can be find in [15]. Tomita et al. in [17] presented the MCQ
algorithm that used the coloring of the subgraph induced by
the candidate set to provide a bound on the number of vertices
in a maximum clique and serve a branching strategy. The
MCR [18] improved MCQ with another sorting of vertices
of the input graph G. The coloring and sorting procedure of
MCQ is further improved in the MaxCliqueDyn algorithm
presented by Konc et al. in [19]. In MaxCliqueDyn, before
coloring of the vertices, these vertices may be re-ordered non-
increasingly with respect to their degrees. This process may
reduce the number of used colors. The ColorSort algorithm
was first described in [19] and used a pruning color threshold
that vertices with color number less than that will be cut in the
pruning step. The MCS [21] algorithm improved the greedy
coloring procedure of MCQ algorithm, by using a recolor-
ing strategy. Integrating this strategy in a bit string frame-

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

VOLUME 8, 2020 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 180045
 N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms
presented by Segundo et al. in [22]. Ostergard in [14] pro-
posed an iterative algorithm, known as Cliquer algorithm,
which uses a basic idea similar to dynamic programming.
Li et al. in [29] computed tighter bounds for the number
of vertices in a maximum clique by reducing each colored
subproblem to maximum satisfiability problem(MaxSAT).
Based on this idea Segundo et al. in [28] presented an effi-
cient way to compute related infra-chromatic upper bounds
at each step of the search with no MaxSAT encoding.
Segundo et al. in [20], proposed an exact bit-parallel algo-
rithm, called BB-MaxClique, for the MCP. BB-MaxClique
uses an improved approximate coloring procedure which
relies on a new implicit global degree branching rule to
obtain tighter upper bounds to the candidate set. Moreover,
BBMCL [23] improves BB-MaxClique using a selective
coloring scheme. PMC [24], [25] and FMC [26] use an
adjacency list representation of the network and unroll the
first level of the search tree to enforce early pruning. FMC
employs several pruning filters for branching and uses the
degree of the vertices to compute the bounds. PMC employs
core numbers to estimate the bounds. Segundo et al. in
[27] designed BBMCSP for large and massive sparse graphs
using a novel sparse encoding for the adjacency matrix. Like
PMC, in BBMCSP core numbers and coloring techniques are
used to estimate the bounds. Verma et al. introduced scale
reduction algorithms for large and low-density graphs which
are based on clique relaxations such as k-community and
k-core to find the maximum cliques and vertex colorings [34].
Li et al. in [35] presented a branch and bound algorithm,
called IncMC2, which uses graph coloring and MaxSAT rea-
soning. Kanazawa et al. proposed an FPGA-based accelerator
to solve partial MaxSAT-encoded maximum clique problems
[32]. Belachew et al. introduced a continuous formulation
of the MCP based on the symmetric rank-one non-negative
approximation of a given matrix and showed that stationary
points of the formulation are correspond to cliques of a given
graph and vice versa [33]. Recently, Hungerford et al. devel-
oped a general regularization-based continuous optimization
framework for the MCP [31].
In this paper, we define a new order on the vertices of a
given graph and employ bounds, determined based on vertex
core number, during search which leads a faster pruning.
Moreover, instead of neighbor set of u ∈ V , a subset of this
set is searched to find a clique including u.
The remaining part of the paper is structured as fol-
lows. Section II covers notations and preliminary con-
cepts. Section III describes the new algorithms and finally
Section IV shows the simulation results. Section V concludes
the paper.
II. BASIC CONCEPTS
We now introduce the notations used in the remainder of this
paper. We refer to N (u) as the neighbor set of u ∈ V , and
N [u] = N (u) ∪ {u}. The degree of a vertex u is denoted
by degG (v). We refer to G[W ] as the subgraph of G that is
induced by W ⊆ V .
180046
The size of the maximum clique ω(G) of a given graph G is
known as the clique number of the graph. Let 1(G) and χ(G)
denote the maximum degree and the chromatic number of G,
respectively. Obviously, ω(G) ≤ χ (G) ≤ 1(G) + 1.
Definition 1: Let G = (V , E) be a graph and W ⊆ V .
A subgraph Hk = G[W ] is a k-core or a core of order k if and
only if Hk is the maximal subgraph of G that all its vertices
have degree at least k, i.e. ∀u ∈ W , degHk (u) ≥ k. The core
number of vertex u in G, denoted by core(u), is the highest
order of a core that contains u [10].
An O(|E|) time algorithm, which we call Core-algorithm,
is presented in [9] to determine the core numbers of vertices
of G = (V , E). In this algorithm, the core number of a vertex
u is degG (u) if it is a vertex with minimum degree in G.
Then all edges (u, v) which degG (u) < degG (v) are removed.
This process is iteratively done on G until the core number
of all the vertices are computed. We refer to K|V |×|V | as the
core-matrix of G in which Kuv = 1 for u, v ∈ V , if v ∈
N (u) and the core number of u is computed before the core
number of v by Core-algorithm. The Core-matrix algorithm
compute the core number of vertices and core-matrix K (see
Algorithm 3 in the Appendix).
For each u ∈ V , we define NK (u) = {u} ∪ {v | Kuv = 1}.
Clearly, for each v ∈ NK (u) we have core(v) ≥ core(u).
Computation of NK (u) takes only O(1) time. Since K is
computed at the same time with the core numbers in the
Core-matrix algorithm, computation of it takes no additional
times. Therefore NK (u) as a row of K is provided in a constant
time.
Note: let u1 , u2 , · · · , un be any vertex-ordering of these
vertices and the set S(ui ) = N [ui ] ∩ {ui , ui+1 , · · · , un } where
1 ≤ i ≤ n. The set S(ui ) are used as search area to
find the maximum clique including ui in [36]–[38]. If the
vertices are sorted in the degeneracy ordering computed by
the Core-matrix algorithm, then NK (ui ) = S(ui ). Since S(ui )
and NK (ui ) are computed in O(n + m) [36] and O(1) time
respectively, NK (ui ) is preferred to use as search area.
Proposition 1: The core number of elements in NK (u) is
at least core(u). Therefore the members in NK (u) have the
necessary condition (i.e. the core number of each member
must be at least core(u)) to construct a clique of size core(u)+
1 (if it exist(s)).
Definition 2: We refer to the core number of u in G[N [u]]
as the local core number of u which is denoted by LCN (u).
Similarly, we refer to the core number of u in G[NK (u)] as the
local-NK core number of u which is denoted by LCNNK (u).
Note that since NK (u) ⊆ N (u) we have LCNNK (u) ≤
LCN (u).
Lemma 1: Let K be core-matrix for a given graph G and
Q = {v1 , · · · , vt } be an arbitrary clique in G. There exist
vi ∈ Q such that Q ⊆ NK (vi ).
Proof: Let vi be the first vertex in Q that the core number
of which is computed by the Core-matrix algorithm. Since vi
is the first vertex in Q considered by the algorithm, no vertices
of Q have been deleted previously and therefore Kvi vj = 1 for
every vj ∈ Q. Thus {v1 · · · , vt } ⊆ NK (vi ).
VOLUME 8, 2020
N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms

Corollary 1: Let Qmax be a maximum clique in G. Then Algorithm 1 The MMCQ(V 0 , LCN , C) Algorithm
Qmax ⊆ NK (u) in which u is the first vertex among elements 1: Input: vertex set V 0 ⊆ V , Local Core number set LCN ,
in Qmax that its core number is computed by the Core-matrix Color set C.
algorithm. 2: Output: A maximum clique of G[V 0 ].
The existing algorithms for the maximum clique problem 3: while V 0 6 = ∅ do
are branch-and-bound solvers which search in G[N (u)] when 4: Choose a vertex v with maximum color C(v) from V 0 .
vertex u is selected for branching. Corollary 1 shows that 5: V 0 := V 0 \ {v};
there exist(s) a vertex u such that Qmax ⊆ NK (u). Therefore, 6: if LCN (v) + 1 > |Qmax | and |Q| + C(v) > |Qmax |
instead of N (u), our algorithm searches in NK (u) to find then
Qmax . Since NK (u) ⊆ N (u), the search area of our algorithm 7: Q := Q ∪ {v};
is smaller than the search area of the others. Moreover, instead 8: if V 0 ∩ N (v) 6 = ∅ and |V 0 ∩ N (v)| + |Q| > |Qmax |
of 1(G), the local core numbers are used as a smaller bound then
for pruning. 9: C 0 = ColorSort(G[V 0 ∩ N (v)]);
We define V̄ = {v1 , v2 , · · · , vn } as the ordered vertex 10: MMCQ(V 0 ∩ N (v), LCN , C 0 );
set of a given graph G in which the vertices G are sorted 11: else
non-increasingly with respect to their LCNNK . Our algorithm 12: if (|Q| > |Qmax |) then
searches in NK (vi ) where vi is selected from V̄ in the order to 13: Qmax := Q;
find maximum clique(s). 14: end if
Proposition 2: Let u ∈ V̄ and Qumax and QuNK −max are 15: end if
maximum cliques in G[N [u]] and G[NK (u)], respectively. 16: Q := Q \ {v};
By the definition of core number we can see that |Qumax | ≤ 17: else
LCN (u) + 1 ≤ core(u) + 1 and |QuNK −max | ≤ LCNNK (u) + 18: return;
1 ≤ LCN (u) + 1 ≤ core(u) + 1. Let LCNmax = 19: end if
max{LCNNK (u) | u ∈ V }. We have ω(G) ≤ LCNmax + 1. 20: end while

III. TOW NEW MAXIMUM CLIQUE ALGORITHMS

In this section we present MMCQ and NK-MaxClique
algorithms to find a maximum clique in a given graph
G = (V , E).
First we describe the MMCQ algorithm which is presented
in Algorithm 1 and is a modified version of the MCQ +
CS algorithm [19]. Let Q be the set of vertices of current
maximum clique. Initially, the algorithm puts Q = {u} and
V 0 = NK (u). Vertices are added to Q respectively until it can
be verified that there are no larger clique. At each step, the
vertex v ∈ NK (u) is selected with the maximum color C(v).
|Q| + C(v) is an upper bound for the number of vertices in a
maximum clique including u and v. Moreover, proposition 2
indicates that LCN (v) + 1 is an other upper bound for the
maximum clique including v. Therefore, if |Qmax | < |Q| +
C(v) and |Qmax | < LCN (v) + 1, Q is extended to Q ∪ {v}.
Then V 0 is replaced by V 0 ∩ N (v) and this process is done
recursively until Q could not be extended (for more details
see [19]). The difference between the MCQ + CS algorithm
and our modified version is that inequalities
LCN (v) + 1 > |Qmax | (III-.1)
and
|V 0 ∩ N (v)| + |Q| > |Qmax | (III-.2)
are added to the condition of extension of Q as pruning
conditions.
We now illstrate the NK-MaxClique algorithm. The exist-
ing algorithms for the maximum clique problem are branch-
and-bound solvers which search in G[N (u)] when vertex u
is selected for branching. In Corollary 1 we have shown
that there exist(s) a vertex u such that Qmax ⊆ NK (u).
Therefore, instead of N (u), we can search in NK (u) to find
maximum clique Qmax . Since NK (u) ⊆ N (u), the search
area of our algorithm is smaller than the search area of the
others. Moreover, instead of 1(G), the local core numbers
are used as a smaller bound for pruning. The algorithm
computes NK (u) for vertex u and finds a maximum clique
in G[NK (u)], and then replaces it by the current founded
maximum clique if its size is larger than the size of current
maximum clique. Let Qmax be the set of vertices of the
largest clique found until now. Initially, the NK-MaxClique
algorithm puts Qmax = ∅. Then it finds a maximum clique in
G[NK (u)] by using the MMCQ algorithm. By proposition 2
we can see that LCNNK (u) + 1 is an upper bound for the
number of vertices in a maximum clique in G[NK (u)]. Hence
if the inequality |Qmax | ≥ LCNNK (u) + 1 hols, it is implied
that there is no clique larger than Qmax in G[NK (u)]). Oth-
erwise, the algorithm finds a maximum clique in G[NK (u)].
In the NK-MaxClique algorithm, for every u ∈ V , the
vertices of G[NK (u)] are colored and sorted by the ColorSort
algorithm presented in [19]. The number of colors used to
color a graph is an upper bound for the clique number of
the graph. It is known from [17] and [30] that if the ver-
tices of a graph be sorted non-increasingly with respect to
their degree, then a tighter (smaller) bound for the clique
number is obtained. Hence, in Algorithm 2, before vertex
coloring, the vertices in G[NK (u)] are sorted. The above
descriptions lead us to the algorithm which is introduced
in Algorithm 2.

VOLUME 8, 2020 180047

 N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms

TABLE 1. CPU times and number of steps (recursive procedure calls) for DIMACS benchmark graphs. ω is the clique number, 1G is the maximum degree
and LCNmax is maximum local-NK core number in G. The best times and minimum number of steps for each row are written in bold and underline style.
Bold values in columns 6, 7 and 8 indicate the best performance between MMCQ, MCQ + CS and NK-MaxClique and bold values in the last two columns
of Table 1 indicate the priority of the MaxCliqueDyn or NK- MaxCliqueDyn algorithms.

IV. SIMULATION RESULTS vs MaxCliqueDyn are compared. The parameter Tlimit in

This section reports experimental tests carried out to eval-
uate the performance of the new algorithms. The MMCQ,
MCQ + CS, MaxCliqueDyn [19], NK-MaxClique and
NK-MaxCliqueDyn (see Algorithm 5 in the Appendix)
algorithms are considered. In the previous section,
NK-MaxClique algorithm is presented in which maximum
cliques are found in NK (u), u ∈ V̄ by MMCQ algorithm
(see line 10 in Algorithm 2). In NK-MaxCliqueDyn algorithm
MMCQ is replaced by Modified-MaxCliqueDyn algorithm
(see Algorithm 4 in the Appendix). Similar to the MMCQ
algorithm, in the Modified-MaxCliqueDyn algorithm prun-
ing conditions III-.1 and III-.2 are considered. In our sim-
ulations, the performance of MMCQ and NK-MaxClique
vs MCQ + CS and the performance of NK-MaxCliqueDyn
MaxCliqueDyn algorithm presented in [19] is set to 0.025.
All algorithms are coded in the same programmatic style,
using the same graph representation and data structures,
so that there was no performance variation due to differences
in implementation details. We coded the algorithms in C++
and compiled it with gcc-5.4. We conducted the computa-
tional tests on a personal computer with a 2.67 GHz Intels
Core i5 processor with 4.0 GB of memory.
Table 1 reports the output of the algorithms on DIMACS
benchmark graphs and Table 3 includes the comparison of
these algorithms on random graphs.
Table 1 compares the performance of the algorithms for
DIMACS graphs. With respect to the CPU time and num-
ber of steps (i.e. recursive procedure calls), in most of

180048 VOLUME 8, 2020

N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms

TABLE 2. Comparing the effect of pruning condition, vertex ordering and search area reduction: S1 depicts the effect of vertex ordering and search area
reduction and S2 shows the effect of pruning condition.

these graphs (44 out of 56) NK-MaxClique outperforms
MCQ + CS and in 38 of 56 instances NK-MaxCliqueDyn
outperforms MaxCliqueDyn. The results show that aver-
agely NK-MaxClique is 40.26% faster than MCQ +
CS and it is about 36% faster than MMCQ. This
value for NK-MaxCliqueDyn and MaxCliqueDyn is about
28.36%. The number of steps in NK-MaxClique and in
NK-MaxCliqueDyn algorithms is about 36.61%, 40.59%
and 33.76% smaller than MMCQ, MCQ + CS and Max-
CliqueDyn algorithms, respectively. The reason is that a
stronger pruning parameters is used by the NK-MaxClique
and NK-MaxCliqueDyn algorithms. In the MCQ + CS and
MaxcliqueDyn algorithms 1(G) + 1 colors are used for
coloring the input graph. Therefore, 1(G) + 1 is the upper

VOLUME 8, 2020 180049

 N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms

TABLE 3. CPU times and number of steps (recursive procedure calls) for random graphs. The best times and minimum number of steps for each row are
written in bold and underline style. Bold values in columns 4 and 5 indicate the best performance between MCQ + CS and NK-MaxClique and bold values
in the last two columns of Table 3 indicate the priority of the MaxCliqueDyn or NK- MaxCliqueDyn algorithms.

bound for the maximum clique (pruning condition). Instead Algorithm 2 The NK-MaxClique(G) Algorithm
of 1(G) + 1 we use LCNNK (u) + 1, u ∈ V as the upper 1: Input: undirected graph G = (V , E).
bound for the maximum clique. Since LCNNK (u) ≤ 1(G) 2: Output: A maximum clique Qmax of G.
our pruning condition is stronger. In the c-fat family we 3: K := Core − matrix(G);
can see ω(G) = LCNmax + 1. Therefore, comparing to the 4: Qmax := ∅; Q := ∅; for every u ∈ V compute LCN (u)
other instances, the performance of the NK-MaxClique and and LCNNK (u) and order vertex set V̄ .
NK-MaxCliqueDyn algorithms is more valuable. 5: for every u ∈ V̄ in the order do
The results in Table 2 compare the effective of pruning 6: if |Qmax | < LCNNK (u) + 1 then
condition, vertex ordering and search area restriction process 7: Q := {u};
which used in the algorithms. Flag-vertex is the first vertex 8: calculate the degrees of vertices in G[NK (u)] and
that is proceed by a MCP algorithm and its corresponding sort vertices in NK (u) in a non-increasing order with
subproblem leads to a maximum clique. win − ite denotes respect to their degrees;
that a maximum clique is founded in win − iteth iteration of 9: C := ColorSort(G[NK (u)]);
the algorithms. We define S1 = |N (u)|−|NR (u)| 1
× 100 × win−ite 10: MMCQ(NK (u), LCN , C);
|N (u)|
11: Q := ∅;
in which the first term denotes the percent of normalized
12: end if
search area reduction and the second term denotes the inverse 13: end for
of win − ite. For the algorithms larger values of these two 14: retun Qmax ;
terms are preferred. NR (u) is the list of remaining neigh-
bors of vertex u in the current subproblem. Note that in
our algorithm NR (u) = NK (u). S2 denotes that how many size of maximum clique found so far. As earliest as Qmax
percent of the vertices (at the root node of the search tree) are is founded, more number of subproblems can be pruned
pruned by pruning condition used in line 6 of Algorithm 2. by the algorithms (see pruning conditions in line 6 Algo-
We consider the color number assigned by MCQ + CS and rithm 2, lines 6 and 8 in Algorithm 1). Hence smaller value of
MaxCliqueDyn algorithms to the vertices as initial coloring win − ite is preferred. Note that the value of the parameters
of the input graph (see [19]), and refer to S20 as the percent in the Table 2 are equal for MCQ + CS and MaxCliqueDyn
of the vertices that their color numbers are smaller than the and also for NK- MaxCliqueDyn and NK-MaxClique. The

180050 VOLUME 8, 2020

N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms

Algorithm 3 The Core-Matrix(G) Algorithm Algorithm 4 The Modified-MaxCliqueDyn (V 0 , LCN ,

1: Input: the graph G = (V , E). C, level) Algorithm
2: Output: the core-matrix K and the ordered vertices set V̄ . 1: Input: vertex set V 0 ⊆ V ,Local Core number set LCN ,
Order the set of vertices V in increasing order of Color set C.
their degrees; 2: Output: A maximum clique of G[V 0 ].
3: i := 1; 3: S[level] := S[level] + S[level − 1] − Sold [level];
4: V̄ = ∅; 4: Sold [level] := S[level − 1];
5: for u ∈ V in the order do 5: while V 0 6 = ∅ do
6: core(u) := deg(u); 6: Choose a vertex v with maximum color C(v) from V 0 .
7: order(u) := i; 7: V 0 := V 0 \ {v};
8: i := i + 1; 8: if LCN (v) + 1 > |Qmax | and |Q| + C(v) > |Qmax |
9: V̄ := V̄ ∪ {u}; then
10: for v ∈ N (u), where v ∈ / V̄ do 9: Q := Q ∪ {v};
11: Kuv := 1; 10: if V 0 ∩ N (v) 6 = ∅ and |V 0 ∩ N (v)| + |Q| > |Qmax |
12: if deg(v) > deg(u) then then
13: delete edge (u, v) from G; 11: if S[level]/ALL_STEPS < Tlimit then
14: reorder V accordingly; 12: calculate the degrees of vertices in G[V 0 ∩
15: end if N (v)];
16: end for 13: sort vertices in V 0 ∩ N (v) in a non-increasing
17: end for order with respect to their degrees;
14: end if
last column depicts the values of the parameters corresponded 15: C 0 = ColorSort(G[V 0 ∩ N (v)]);
to MCQ + CS and MaxCliqueDyn and the column before the 16: S[level] := S[level] + 1;
last column shows these values for NK- MaxCliqueDyn and 17: ALL_STEPS := ALL_STEPS + 1;
NK-MaxClique. 18: Modified-MaxCliqueDyn(V 0 ∩
The results show that in most cases (48 of 56 instances), the N (v), LCN , C 0 , level + 1);
algorithm with larger value of S1 has lower elapsed time. The 19: else
reason is that an algorithm with larger value of S1 whether can 20: if (|Q| > |Qmax |) then
reduce valuably the search area, or can determine the maxi- 21: Qmax := Q;
mum clique earlier than the other algorithms or both of them. 22: end if
In all instances reported in table 2, S2 > S20 which means that 23: end if
our pruning condition is stronger than this conditions used in 24: Q := Q \ {v};
the two other algorithms. 25: else
Table 3 reports average CPU times and number of steps of 26: return;
algorithms over 15 random graphs with fixed n and p, where 27: end if
28: end while
n and p denote the number of vertices and the probability
that an edge between a pair of vertices exists, respectively.
The results show that NK-MaxClique and NK- MaxClique- the new pruning parameter and search area restriction on the
Dyn are about 31.71% and 9.5% faster than MCQ + CS well-known existing algorithms.
and MaxCliqueDyn, respectively. The reason is that smaller
upper bounds for pruning is used in the NK-MaxClique and APPENDIX
NK- MaxCliqueDyn algorithms. THE CORE-MATRIX, NK-MaxCliqueDyn AND
MODIFIED-MaxCliqueDyn ALGORITHMS
V. CONCLUSION AND FUTURE WORKS In what follows we briefly describe the Core-matrix,
In this paper, we study the Maximum Clique Problem (MCP), NK-MaxCliqueDyn and Modified-MaxCliqueDyn algo-
an important problem in graph theory and real-world appli- rithms. The Core-matrix algorithm provides matrix K
cations. For MCP, we present two new Branch-and-Bound and orders the vertices of a given graph. The Modified-
exact algorithms, which employ a novel pruning bound (i.e. MaxCliqueDyn algorithm is a new version of the Max-
LCNNK ) to ignore the subproblems that can not contain larger CliqueDyn algorithm which is modified with respect to the
clique(s) than the current solution. The other advantage of the new pruning conditions III-.1 and III-.2. In the Modified-
algorithms is that they are designed to find a clique including MaxCliqueDyn algorithm, level denotes the number of
vertex u in NK (u)(⊆ N (u)), instead of N (u). The computa- branches from the root of the recursion tree to the current
tional results provided in the manuscript reveal that the new leaf. S[level] and Sold [level] denote the sum of steps which
bound and smaller search area help to improve the MCQ and the Modified-MaxCliqueDyn algorithm performs from the
MaxCliqueDyn algorithms. As a future works we will study root node up to and including the current level, and the sum

VOLUME 8, 2020 180051

 N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms
Algorithm 5 The NK-MaxCliqueDyn(G) Algorithm
1: Input: undirected graph G = (V , E).
2: Output: A maximum clique Qmax of G.
3: (K , V̄ ) := core − matrix(G);
4: Qmax := ∅; for every u ∈ V̄ compute LCN (u) and
LCNNK (u).
5: for every u ∈ V̄ in the order do
6: if |Qmax | < LCNNK (u) + 1 then
7: calculate the degrees of vertices in G[NK (u)] and
sort vertices in NK (u) in a non-increasing order with
respect to their degrees;
8: C := ColorSort(G[NK (u)]);
9: Modified − MaxCliqueDyn(NK (u), LCN , C, 1);
10: end if
11: end for
12: retun Qmax ;
of steps up to and including the previous level, respectively.
T [level] = S[level]/ALL_STEPS where ALL_STEPS is a
global counter of steps which is increased by 1 at each step
of the Modified-MaxCliqueDyn algorithm. While T [level] <
Tlimit , the algorithm performs the calculations of the degrees
and sorting. For more details see [19]. The formal description
of the Modified-MaxCliqueDyn algorithm is depicted in
Algorithm 4. The NK-MaxCliqueDyn algorithm is presented
in Algorithm 5 which uses Modified-MaxCliqueDyn as the
solver of its subproblems.
REFERENCES
[1] P. San Segundo and J. Artieda, ‘‘A novel clique formulation for the
visual feature matching problem,’’ Int. J. Speech Technol., vol. 43, no. 2,
pp. 325–342, Sep. 2015.
[2] P. San Segundo, D. Rodríguez-Losada, F. Matía, and R. Galán, ‘‘Fast exact
feature based data correspondence search with an efficient bit-parallel
MCP solver,’’ Int. J. Speech Technol., vol. 32, no. 3, pp. 311–329,
Jun. 2010.
[3] S. Butenko and W. E. Wilhelm, ‘‘Clique-detection models in computational
biochemistry and genomics,’’ Eur. J. Oper. Res., vol. 173, no. 1, pp. 1–17,
Aug. 2006.
[4] C. W. Art, and B. Sergiy, Clustering Challenges in Biological Networks.
Singapore: World Scientific, 2009, doi: 10.1142/6602.
[5] T. Etzion and P. R. J. Ostergard, ‘‘Greedy and heuristic algorithms for
codes and colorings,’’ IEEE Trans. Inf. Theory, vol. 44, no. 1, pp. 382–388,
Jan. 1998.
[6] T. Gschwind, S. Irnich, F. Furini, and R. W. Calvo, ‘‘Social net-
work analysis and community detection by decomposing a graph into
relaxed cliques,’’ Univ. Mainz, Mainz, Germany, Tech. Rep. LM-2015-07,
Dec. 2015.
[7] O. Weide, D. Ryan, and M. Ehrgott, ‘‘An iterative approach to robust
and integrated aircraft routing and crew scheduling,’’ Comput. Oper. Res.,
vol. 37, no. 5, pp. 833–844, May 2010.
[8] Q. Wu and J.-K. Hao, ‘‘A review on algorithms for maximum clique
problems,’’ Eur. J. Oper. Res., vol. 242, no. 3, pp. 693–709, 2015.
[9] V. Batagelj and M. Zaversnik, ‘‘An O(m) algorithm for cores decompo-
sition of networks,’’ Oct. 2003, arXiv:cs/0310049. [Online]. Available:
https://arxiv.org/abs/cs/0310049
[10] S. B. Seidman, ‘‘Network structure and minimum degree,’’ Social Netw.,
vol. 5, no. 3, pp. 269–287, Sep. 1983.
[11] M. R. Garey and D. S. Johnson, Computers and Intractability: A Guide
to the Theory of NP-Completeness. New York, NY, USA: W. H. Freeman,
1979.
[12] J. Hastad, ‘‘Clique is hard to approximate within,’’ Acta Math., vol. 182,
pp. 105–142, Jan. 1999.
[13] U. Feige, ‘‘Approximating maximum clique by removing subgraphs,’’
SIAM J. Discrete Math., vol. 18, no. 2, pp. 219–225, Jan. 2004.
180052
[14] P. R. J. Östergárd, ‘‘A fast algorithm for the maximum clique
problem,’’ Discrete Appl. Math., vol. 120, nos. 1–3, pp. 197–207,
Aug. 2002.
[15] L. Babel and G. Tinhofer, ‘‘A branch and bound algorithm for the
maximum clique problem,’’ Methods Models Oper. Res., vol. 34, no. 3,
pp. 207–217, May 1990.
[16] D. Brélaz, ‘‘New methods to color the vertices of a graph,’’ Commun. ACM,
vol. 22, no. 4, pp. 251–256, Apr. 1979.
[17] E. Tomita and T. Seki, ‘‘An efficient branch-and-bound algorithm for
finding a maximum clique,’’ in Proc. Discrete Math. Theor. Comput. Sci.,
vol. 2731, Jan. 2003, pp. 278–289.
[18] E. Tomita and T. Kameda, ‘‘An efficient Branch-and-bound algorithm for
finding a maximum clique with computational experiments,’’ J. Global
Optim., vol. 37, no. 1, pp. 95–111, Dec. 2006.
[19] J. Konc and D. Janežič, ‘‘An improved branch and bound algorithm for
the maximum clique problem,’’ MATCH Commun. Math. Comput. Chem.,
vol. 58, pp. 569–590, Jun. 2007.
[20] P. San Segundo, D. Rodríguez-Losada, and A. Jiménez, ‘‘An exact
bit-parallel algorithm for the maximum clique problem,’’ Comput. Oper.
Res., vol. 38, no. 2, pp. 571–581, Feb. 2011.
[21] E. Tomita, Y. Sutani, T. Higashi, S. Takahashi, and M. Wakatsuki,
‘‘A simple and faster branch-and bound algorithm for finding a maximum
clique,’’ Lect. Notes Comput. Sci., vol. 5942, pp. 191–203, Feb. 2010.
[22] P. San Segundo, F. Matia, D. Rodriguez-Losada, and M. Hernando, ‘‘An
improved bit parallel exact maximum clique algorithm,’’ Optim. Lett.,
vol. 7, no. 3, pp. 467–479, Mar. 2013.
[23] P. San Segundo and C. Tapia, ‘‘Relaxed approximate coloring in exact max-
imum clique search,’’ Comput. Oper. Res., vol. 44, pp. 185–192, Apr. 2014.
[24] R. A. Rossi, D. F. Gleich, A. H. Gebremedhin, and M. M. A. Patwary,
‘‘Fast maximum clique algorithms for large graphs,’’ in Proc. 23rd Int.
Conf. World Wide Web, 2014, pp. 365–366.
[25] R. A. Rossi, D. F. Gleich, and A. H. Gebremedhin, ‘‘Parallel maximum
clique algorithms with applications to network analysis,’’ SIAM J. Sci.
Comput., vol. 37, no. 5, pp. C589–C616, Jan. 2015.
[26] B. Pattabiraman, M. M. A. Patwary, A. H. Gebremedhin, W. K. Liao,
and A. Choudhary, ‘‘Fast algorithms for the maximum clique problem on
massive sparse graphs,’’ in Proc. Int. Workshop Algorithms Models Web-
Graph, vol. 8305, Dec. 2013, pp. 156–169.
[27] P. San Segundo, A. Lopez, and P. M. Pardalos, ‘‘A new exact maximum
clique algorithm for large and massive sparse graphs,’’ Comput. Oper. Res.,
vol. 66, pp. 81–94, Feb. 2016.
[28] P. San Segundo, A. Nikolaev, and M. Batsyn, ‘‘Infra-chromatic bound for
exact maximum clique search,’’ Comput. Oper. Res., vol. 64, pp. 293–303,
Dec. 2015.
[29] C. M. Li and Z. Quan, ‘‘An efficient branch-and-bound algorithm based
on MaxSAT for the maximum clique problem,’’ in Proc. 24th AAAI Conf.
Artif. Intell., Jul. 2010, pp. 128–133.
[30] D. J. A. Welsh, ‘‘An upper bound for the chromatic number of a graph
and its application to timetabling problems,’’ Comput. J., vol. 10, no. 1,
pp. 85–86, Jan. 1967.
[31] J. T. Hungerford and F. Rinaldi, ‘‘A General Regularized Continuous
Formulation for the Maximum Clique Problem,’’ Math. Oper. Res., vol. 44,
no. 4, pp. 1–13, Nov. 2019.
[32] K. Kanazawa and S. Cai, ‘‘FPGA acceleration to solve maximum
clique problems encoded into partial MaxSAT,’’ in Proc. IEEE 12th Int.
Symp. Embedded Multicore/Many-Core Syst. Chip (MCSoC), Sep. 2018,
pp. 217–224.
[33] M. T. Belachew and N. Gillis, ‘‘Solving the maximum clique problem
with symmetric rank-one non-negative matrix approximation,’’ J. Optim.
Theory Appl., vol. 173, no. 1, pp. 279–296, Apr. 2017.
[34] A. Verma, A. Buchanan, and S. Butenko, ‘‘Solving the maximum clique
and vertex coloring problems on very large sparse networks,’’ Informs
J. Comput., vol. 27, no. 1, pp. 164–177, Feb. 2015.
[35] C. M. Li, Z. Fang, H. Jiang, and K. Xu, ‘‘Incremental upper bound for
the maximum clique problem,’’ INFORMS J. Comput., vol. 30, no. 1,
pp. 137–153, Nov. 2017.
[36] A. Buchanan, J. L. Walteros, S. Butenko, and P. M. Pardalos, ‘‘Solv-
ing maximum clique in sparse graphs: An O(nm + n2d/4 ) algorithm
for d-degenerate graphs,’’ Optim. Lett., vol. 8, no. 5, pp. 1611–1617,
Dec. 2013.
[37] G. Manoussakis, ‘‘A new decomposition technique for maximal clique
enumeration for sparse graphs,’’ Theor. Comput. Sci., vol. 770, pp. 25–33,
May 2019.
[38] L. Chang, ‘‘Efficient maximum clique computation and enumeration over
large sparse graphs,’’ VLDB J., vol. 29, no. 5, pp. 999–1022, Feb. 2020,
doi: 10.1007/s00778-020-00602-z.
VOLUME 8, 2020
N. Mohammadi, M. Kadivar: NK-MaxClique and MMCQ: Tow New Exact Branch and Bound Algorithms

NEDA MOHAMMADI received the B.S. and MEHDI KADIVAR received the B.S. degree in
M.S. degrees in applied mathematics and oper- computer sciences from Yazd University, Iran,
ation research from Shahrekord University, in 2001, and the M.S. and Ph.D. degrees in com-
Shahrekord, Iran, in 2009 and 2012, respectively, puter sciences from the Amirkabir University of
where she is currently pursuing the Ph.D. degree Technology, Tehran, in 2008 and 2012, respec-
in computer science. Her research interests include tively.
modeling, algorithm design, and analysis of prob- Since 2009, he has been an Assistant Pro-
lems related to computationally hard problems fessor with the Computer Sciences Department,
in combinatorics and graph theory and wireless Shahrekord University, Shahrekord, Iran. He has
networks, such as wireless scheduling. published several conference and journal articles.
His research interests include computer networks, wireless networks, algo-
rithms, scheduling, and distributed algorithms.

VOLUME 8, 2020 180053