Journal of Algorithms 41, 388–403 (2001)

doi:10.1006/jagm.2001.1199, available online at http://www.idealibrary.com on

On Bipartite and Multipartite Clique Problems

Milind Dawande

University of Texas, Dallas, Texas 75080; and T. J. Watson Center,

IBM, Yorktown Heights, New York 10598

Pinar Keskinocak

School of Industrial and Systems Engineering, Georgia Institute of Technology,

Atlanta, Georgia 30332

Jayashankar M. Swaminathan

The Kenan-Flagler Business School, University of North Carolina,

Chapel Hill, North Carolina 27599

and

Sridhar Tayur

GSIA, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213

Received January 24, 2000

In this paper, we introduce the maximum edge biclique problem in bipartite

graphs and the edge/node weighted multipartite clique problem in multipartite
graphs. Our motivation for studying these problems came from abstractions of real
manufacturing problems in the computer industry and from formal concept analy-
sis. We show that the weighted version and four variants of the unweighted version
of the biclique problem are NP-complete. For random bipartite graphs, we show that
the size of the maximum balanced biclique is considerably smaller than the size of
the maximum edge cardinality biclique, thus highlighting the difference between the
two problems. For multipartite graphs, we consider three versions each for the edge
and node weighted problems which differ in the structure of the multipartite clique
(MPC) required. We show that all the edge weighted versions are NP-complete in
general. We also provide a special case in which edge weighted versions are poly-
nomially solvable.  2001 Elsevier Science
Key Words: bipartite graph; multipartite graph; clique; complexity.

388
0196-6774/01 $35.00
 2001 Elsevier Science
All rights reserved.
 clique problems 389

1. INTRODUCTION

In this paper, we study biclique and multipartite clique problems. Given

a bipartite graph B = V1 ∪ V2  E, a biclique C = U1 ∪ U2 is a subset of
the node set, such that U1 ⊆ V1 , U2 ⊆ V2 , and for every u ∈ U1 , v ∈ U2 the
edge u v ∈ E. In other words, a biclique is a complete bipartite subgraph
of B. Maximum edge cardinality biclique (MBP) in B is a biclique C with a
maximum number of edges. In an edge weighted bipartite graph B, there
is a weight wuv associated with each edge u v. A maximum edge weight
(MWBP) biclique is a biclique C, where the sum of the edge weights in the
subgraph induced by C is maximum among all the bicliques in B.
A multipartite graph with n levels G = V1 ∪ V2 ∪ · · · ∪ Vn  E is defined
as a graph such that for every edge e = u v, we have u ∈ Vi and v ∈ Vi+1
for some i ∈ 1     n − 1.1 A multipartite clique M = U1 ∪ U2 ∪ · · · ∪ Um
within a multipartite graph G is defined such that Ui ⊆ Vk+i ∀ i 1 ≤ i ≤ m,
m ≥ 2 for some k ≥ 0 and for every u1 ∈ Ui and u2 ∈ Ui+1 the edge
u1  u2  ∈ E. In an edge weighted multipartite graph, the maximum edge
weighted multipartite clique is one which has the maximum sum in terms of
the weights of the edges in the multipartite clique. Similarly, a maximum
node weighted multipartite clique has the maximum sum in terms of weights
of nodes in the multipartite clique.
A well known problem related to biclique and multipartite clique prob-
lems is the maximum clique, which is one of the most widely studied NP-
complete problems in the literature. Given a graph G = V E, a clique
(or a complete subgraph) C is a subset of the node set, such that for every
pair of nodes u v ∈ C, the edge u v ∈ E. A maximum clique in G is
a clique with the maximum number of nodes. In the weighted version of
the maximum clique problem, there is a weight wv associated with each
node v and the weight W C of a clique C is the sum of the weights of the
nodes in C.
Besides their relation to the maximum clique problem, our motivation
for studying the biclique and multipartite clique problems came from a
real manufacturing problem in the computer industry. Consider a set of
components V1 = 1     n and a set of products V2 = 1     m. The
relationship between these products and components can be modeled on a
bipartite graph B with node set V1 ∪ V2 and edge set E, such that i j ∈ E
if and only if component i is part of product j. Several products share
one or more common components. One way of reducing the lead times

1
Note the difference between multipartite graphs and the well known class of k-partite
graphs. A graph G = V E is k-partite if V can be partitioned into k subsets V1      Vk such
that for every edge u v ∈ E, u and v belong to different vertex sets of the partition. The
class of multipartite graphs is contained in the class of k-partite graphs.
390 dawande et al.

perceived by the customers for these products is to reduce the final assem-
bly time, where such a reduction can be obtained by creating subassemblies
(or vanilla boxes) in advance (see [11] for details). A vanilla box U1 contain-
ing parts i1      ik can be used only in products which contain all of these
parts. In other words, the set of products U2 = j  il  j ∈ E, l = 1     k
can use the vanilla box U1 . Let tij be the assembly time of component i in
product j. If the total assembly time of the components in vanilla box U1
is T , then we can obtain a reduction of T in the lead times of all the prod-
ucts in U2 by having enough inventory of these vanilla boxes. On the other
hand, to obtain a large T , we have to include many parts in the vanilla
box, which will usually decrease the number of products which can use the
vanilla box (size of U2 ). Then, there is a trade-off between constructing a
large vanilla box and using it in many products. The problem of finding
a “good” vanilla box can be modeled by finding a maximum edge weight
biclique in the bipartite graph B. If all the parts have (approximately) the
same assembly time the problem reduces to the maximum edge cardinality
biclique problem (MBP). A natural generalization of the bipartite clique
problem is the multipartite clique problem. In such a case, each multipar-
tite clique in the graph represents a possible storage of vanilla boxes at
different levels in the assembly process such that a vanilla box in a later
level in assembly is itself assembled in part from another vanilla box from
the previous level (because a biclique between any two levels i and i + 1
acts as vanilla box for that level).
Bicliques have also been studied in the area of formal concept analysis
[3, 4]. Consider two sets V1 and V2 (the set of “attributes” and the set of
“objects”) and a relation R between V1 and V2 (i j ∈ R if object j has
attribute i). For subsets P ⊂ V1 and Q ⊂ V2 , let

P  = the set of all objects which have all the attributes in P, and
Q = the set of all attributes which all the objects of Q have.

TABLE 1
Variants of Biclique Problems

Abbreviation Problem

MBP Maximum edge cardinality biclique

MWBP Maximum edge weight biclique
MNWBP Maximum node weight biclique
MBBP Maximum balanced node cardinality biclique
EBNCD Exact balanced node cardinality decision problem
EECD Exact edge cardinality decision problem
MOFCP Maximum One-sided edge cardinality problem
EBPNCD Exact balanced prime node cardinality decision problem
 clique problems 391
TABLE 2
Variants of Multipartite Clique Problems

Abbreviation Problem

MPC Multipartite clique

MPCP Multipartite clique which includes nodes from all levels
MPCF Multipartite clique including the first level
MPCS Multipartite clique problem which includes nodes from
some levels

Then, a formal concept of V1  V2  R is a pair P Q such that P ⊂ V1 ,

Q ⊂ V2 , P  = Q, and Q = P. We can associate V1 ∪ V2 with the node set
of a bipartite graph B and the relation R defines the edge set E. Then the
concepts are the maximal bicliques of B. In formal concept analysis, the goal
is to cover the bipartite graph by “fat” concepts, i.e., large bicliques. Current
methods in the area do a brute force search for finding large (i.e., one with
the maximum number of edges) bicliques to cover all the edges [3, 8].
The rest of the paper is organized as follows. In Section 2, we present
complexity results related to the biclique problem and compare the size of
balanced biclique and edge cardinality bicliques in random bipartite graphs.
In Section 3, we present the alternative versions of the multipartite clique
problem and develop complexity results. We conclude in Section 4. Variants
of biclique and multipartite clique problems mentioned in the paper are
summarized in Tables 1 and 2.

2. THE BICLIQUE PROBLEM

In this section, we first present the formulation for the biclique problem
and discuss known results. Then we show that MWBP and four variants
of MBP are NP-complete. Note that since the complement of a bipartite
graph is not bipartite in general, the polynomial-time solvability of the inde-
pendent set problem on bipartite graphs does not imply a polynomial time
algorithm for MBP. Finally, we compare the sizes of maximum balanced
bicliques and maximum edge cardinality bicliques in random graphs.
In a node weighted bipartite graph B = V1 ∪ V2  E, there is a weight wv
associated with each node v. The maximum node weight biclique problem
(MNWBP) can be formulated as a 0–1 integer program as



max wu xu + w v xv
u∈V1 v∈V2

subject to xu + xv ≤ 1 u ∈ V1  v ∈ V2  u v ∈
/E (1)

xv ∈ 0 1 for all v ∈ V1 V2  (2)
392 dawande et al.

where 
1 if node v is in the biclique
xv =
0 otherwise.
If we relax this integer program by replacing the integrality constraints (2)
with 0 ≤ xv ≤ 1 for all v ∈ V1 ∪ V2 , we obtain a linear program. Note that
the matrix defining the constraint set (1) is the node-edge incidence matrix
of a bipartite graph, which is totally unimodular, and hence the solution to
the linear programming relaxation will be integer [9, p. 544, Corollary 2.9].
Therefore, the maximum node weight biclique problem is polynomially solv-
able [5]. It follows that the maximum node cardinality biclique problem is
also polynomially solvable. A restricted version of these problems, where
there is an additional requirement that U1  = U2 , is called the maximum
balanced node cardinality biclique problem (MBBP), which is NP-complete
[5]. (This problem is referred to as the balanced complete bipartite sub-
graph problem in [5, p. 196].) Note that for the same bipartite graph, solu-
tions to MBBP and MBP may be quite different from each other. Hence,
node-cardinality biclique problems do not provide good approximations for
MBP in general. In Section 2.2, we quantify the difference between the
solutions to MBP and MBBP in random graphs.
Hochbaum [7] considers a related problem to MWBP and MNWBP,
where the objective is to minimize the total weight of the nodes or edges
deleted so that the remaining subgraph is a biclique. She provides a
2-approximation for the edge deletion version for general and bipartite
graphs and a 2-approximation for the node deletion version for general
graphs.

2.1. Complexity of Biclique Problems

Theorem 1. MWBP is NP-complete.
Proof. We prove this by a reduction from the maximum clique problem.
Let G = V E be a graph with node set V and edge set E. Create a
bipartite graph BG = V1 ∪ V2  E   from G, such that V1 = V2 = V and
i j ∈ E  (for i ∈ V1 and j ∈ V2 ) if and only if i = j or i j ∈ E. Let the
edges i i of BG have weight 1 and let all the other edges have weight
zero.
With the edge weights as defined, there is a maximum weight biclique
U1 ∪ U2 in BG, such that i ∈ U1 if and only if i ∈ U2 (i.e., U1  = U2  and
the biclique is “symmetric”). Such a maximum weight “symmetric” biclique
can be obtained easily by deleting the nodes i ∈ U1 , i ∈ / U2 and i ∈ U2 ,
i∈/ U1 from a maximum weight biclique. It follows that if C is a maximum
clique in G, then U1 ∪ U2 , where U1 = U2 = C, induces a maximum weight
biclique in BG. Similarly, if U1 ∪ U2 is a symmetric maximum weight
biclique in BG, then C = U1 = U2 is a maximum clique in G.
 clique problems 393

Note that the reduction in Theorem 1 does not imply the NP-hardness
of MBP, since we used a weighted bipartite graph in the reduction in which
some edge weights were zero. An NP-completeness proof for MBP has been
recently provided in [10].
Next, we consider three decision problems and an optimization problem,
which are related to MBP:
• Exact balanced node cardinality decision problem (EBNCD): Given
a bipartite graph G = V1 ∪ V2  E and a positive integer a ∈ Z+ , does there
exist a biclique C = U1 ∪ U2 with U1  = U2  = a?
• Exact node cardinality decision problem (ENCD): Given a bipartite
graph G = V1 ∪ V2  E and two positive integers a b ∈ Z+ , does there
exist a biclique C = U1 ∪ U2 with U1  = a and U2  = b?
• Exact edge cardinality decision problem (EECD): Given a bipartite
graph G = V1 ∪ V2  E and a positive integer k ∈ Z+ , does there exist a
biclique with exactly k edges?
• Maximum one-sided edge cardinality problem (MOFCP): Given a
bipartite graph G = V1 ∪ V2  E and a positive integer k ∈ Z+ , find a maxi-
mum cardinality biclique with exactly k nodes on one side of the bipartition.
Lemma 2.1. EBNCD and ENCD are NP-complete.
Proof. It is known that the maximum balanced node cardinality biclique
problem (MBBP) is NP-complete [5]. Then, it follows that EBNCD is
NP-complete, since MBBP can be solved using a polynomial number of
instances of EBNCD. Note that EBNCD is just a special case of ENCD and
hence ENCD is also NP-complete. Note that the reductions for EBNCD
and ENCD are Turing reductions rather than Karp reductions [5].
Theorem 2.2. EECD is NP-complete.
To prove this theorem, first we define the following decision problem and
show that it is NP-complete:
Exact balanced prime node cardinality decision problem (EBPNCD):
Given a bipartite graph G = V1 ∪ V2  E and a prime number p, such
that the maximum degree in G is less than p2 , does there exist a biclique
C = U1 ∪ U2 , with U1  = U2  = p?
Lemma 2.3. EBPNCD is NP-complete.
Proof. Given an instance of EBNCD, let l = max V1  V2  + 1 and
let p be any prime number such that l ≤ p ≤ 2l. Such a prime number
is guaranteed by Bertrand’s theorem [6]. Let a < p be a positive integer,
where a is the specification for EBNCD. Add p − a nodes on both sides
of the bipartition and connect each of these additional nodes to all the
394 dawande et al.

nodes on the opposite side of the bipartition. The maximum degree of any
node in this graph is p − a + max V1  V2  ≤ 3l. Since p2 ≥ l2 it follows
that p2 > 3l and the maximum degree is less than p2 , for l > 3. Then,
EBNCD has a yes (no) answer if and only if EBPNCD has a yes (no)
answer, implying that EBPNCD is NP-complete.
Now we prove Theorem 2.2.
Proof of Theorem 2.2. Consider a bipartite graph G = V1 ∪ V2  E and
an instance of EBPNCD. A biclique of edge cardinality p2 in G can be
possible only in two ways: (1) one node on one side of the bipartition and
p2 nodes on the other side and (2) exactly p nodes on both sides of the
bipartition. Since the maximum degree in G is strictly less than p2 , the first
case is not possible. Thus, EBPNCD has a yes (no) answer if and only if an
instance of EECD with k = p2 has a yes (no) answer. Then it follows that
EECD is NP-complete since EBPNCD is.
Our next result is about the complexity of the optimization problem
MOFCP:
Theorem 2.4. MOFCP is NP-complete.
To prove Theorem 2.4, we define the following decision problem:
Maximum fixed intersection problem (MFIP): Given k ∈ Z+ , a
ground set V , and a set system
= S1      Sn , where the Si ’s are sub-
sets of V , find k subsets from
such that their intersection has maximum
cardinality.
Lemma 2.5. MFIP is NP-hard.
Proof. It is well known that the decision problem CLIQUE, “Given a
graph G = V E and a positive integer k, does there exist a clique of size
k in G?,” is NP-complete [5]. Given an instance of CLIQUE, construct the
following set system on the ground set V . For each edge e = u v ∈ E,
construct one set Se = V \ u v. Let
= Se  e ∈ E. There exists a clique
on k nodes in G if and only if there exist p = kk−1 2
subsets in
whose
intersection has cardinality at least V  − k. Thus, there exists a clique of
size k in G if and only if the cardinality of the maximum intersection in the
optimal solution to MFIP of p sets is V  − k.
Proof of Theorem 2.4. Consider an instance of MFIP. Construct a bipar-
tite graph G = V1 ∪ V2  E as follows: For each set Si is
, create a node
i in V1 ; for each element j of the base set V , create a node j in V2 . For
every element j ∈ Si , include an edge e = i j. Note that the maximum
edge cardinality biclique with exactly k nodes in V1 solves MFIP.
 clique problems 395

2.2. Comparing Maximum Balanced Bicliques and

Maximum Edge Cardinality Bicliques
In this section, we show that the size of a maximum balanced biclique
may be considerably smaller than the size of a maximum edge cardinal-
ity biclique in random bipartite graphs, thus highlighting the difference
between these seemingly similar problems.
We denote a random bipartite graph by B = V1 ∪ V2  p, where 0 ≤ p ≤
1 is the probability that a particular edge exists in B. We denote the size
of a biclique by a × b, if it has a nodes in V1 and b nodes in V2 . For V1  =
V2  = n and for sufficiently large n, we show that the maximum balanced
biclique will be of size a × a with high probability, where an ≤ a < 2an
and an = log n/ log p1 . Note that the size of a maximum edge cardinality
biclique in a random bipartite graph will be at least np (consider a single
node and all its neighbors) with high probability, which is much larger than
a × a for constant p.
Theorem 2.6. Consider a random bipartite graph B = V1 ∪ V2  p, where
0 < p < 1 is a constant, V1  = V2  = n, and an = log n/ log p1 . If the
maximum balanced biclique in this graph has size a × a, then an ≤ a ≤
2an with high probability ( for sufficiently large n).
Proof. The proof consists of two main steps. First, we show that the
probability of having a balanced biclique of size 2an × 2an is very small
(i.e., the probability approaches zero as n approaches ∞). Second, we show
that the probability of having a balanced biclique of size at least an × an
approaches 1 as n approaches ∞.
Let Za = number of a × a bicliques in G. First, we need to show that
the probability of having a balanced biclique of size a × a is very small if
a ≥ 2an. We use the fact that ProbZa ≥ 1 ≤ EZa .
 2
n 2
Prob Za ≥ 1 ≤ EZa  = pa
a
 a 2
n 2
≤ pa 
a!
The computation of EZa  follows from the following argument. A subset
of nodes A ∪ Q, A ⊆ V1 , Q ⊆ V2 form a biclique, if there is an edge between
every pair of nodes u ∈ A, v ∈ Q. Suppose both A and Q have size a. Since
the probability of an edge is p, the probability
  that a given node set A ∪ Q
forms a biclique is pa . There are na different ways of choosing a node
2

subset
  A ⊆ V1 or Q ⊆ V2 of size a. Hence, the number of a × a subgraphs
is na na and the expected number of a × a bicliques is
  
n n a2
EZa  = p 
a a
396 dawande et al.
−2
Note that for a ≥ 2 log n/ log p1 , pa = pa a ≤ plogp n a = n−2a and
2

hence ProbZa ≥ 1 ≤  a!1 2 . Thus, for a ≥ 2 log n/ log p1 ,

ProbZa ≥ 1 → 0 as n → ∞ (3)

Now, we need to show that there is a balanced biclique of size an × an
in B with high probability; i.e., ProbZa = 0  a = an is very small. From
the second moment method [2], we have ProbZa = 0 ≤ VarZa /EZa 2 .
Let XA Q be an indicator variable which assumes value 1, if the nodes in
A ⊆ V1 and Q ⊆ V2 form a biclique, and zero otherwise:
 


E Za2 = ProbXA Q = 1 XA  Q = 1
A Q A  Q



= ProbXA  Q = 1  XA Q = 1 ProbXA Q = 1
A Q A  Q

 Q:
Since all the A Q look alike, fix A Q as A 
 


E Za2 = ProbXA  Q = 1  XA
Q  = 1 ProbXA Q = 1
A Q A  Q



= ProbXA Q = 1 ProbXA  Q = 1  XA
Q  = 1
A Q  
A Q

a
a


= ProbXA Q = 1 ProbXA  Q = 1  XA
Q  = 1
A Q i=0 j=0  
A ∩A=i
 
Q ∩Q=j

a
a     
a n − a a n − a a2 −ij
= ProbXA Q = 1 p 
A Q i=0 j=0
i a−i j a−j

Letting A Q
  ProbXA  Q = 1  XA
Q  = 1 = %, we get

VarZa  %
ProbZa = 0 ≤ 2
= − 1
EZa  EZa 
We can write
a


a
%
= T 
EZa  i=0 j=0 ij

where
an−aan−a
i a−i j a−j
Tij = nn p−ij 
a a
 clique problems 397

We want to show that %/EZa  = 1 + on−3/2 . First, we look at the first

few terms of the sequence Tij :
n−a2
T00 = a2
n
a

    2
a a a
= 1− 1− ··· 1 −
n n−1 n − a − 1

2 2
a
= 1− + on−3/2  
n
an−a an−a
0
T10 = 1
na−1
 na
a a

a2
= T 
n − 2a + 1 00
n−a a

n−a

The second equality in T10 follows, since a−1
= n−2a+1 a
. Similarly,
2
a
T01 =
T 
n − 2a + 1 00
Adding up the first three terms, we obtain
 
2a2
T00 + T01 + T10 = T00 1 +
n − 2a + 1

2  
a2 2a2
= 1− + on−3/2  1+
n n − 2a + 1
log n
= 1 + on−3/2  for a = 
log p1
Now, we want to show that the remaining part of the summation is also
small. To be able to do that, first we will bound the terms Tij i j ≥ 1 in
terms of T11 :
an−a an−a
Tij j a−j
= ai n−a
a−i
 an−a p−ij+1 
T11 1 a−1 1 a−1
Since
an−a
n − 2a + 1! a − 1!2
ai n−a
a−i
=
1 a−1
n − 2a + i! i! a − i!2
we obtain
 i−1  j−1
Tij a2 a2
≤ p−ij+1 
T11 n − 2a n − 2a
398 dawande et al.

First, note that T12 /T11 = a − 12 /2n − 2a + 2p ≤ 1, for sufficiently large
n. Similarly, T21 /T11 ≤ 1, for sufficiently large n.
For i ≥ 2,

−ij + 1
−i − 1j ≤ 
2

Similarly, for j ≥ 2,

−ij + 1
−j − 1i ≤ 
2

Thus,
 i−1  j−1
Tij a2 −ij+1 a2 −ij+1
≤ p 2 p 2
T11 n − 2a n − 2a
 i−1  j−1
a2 a2
≤ p−j p−i 
n − 2a n − 2a

For the choice of a = a∗ n = 1 − ' log n/ log p1 , we get Tij /T11 ≤ 1 for
sufficiently large n.
Noting that

a4
n−a2
n−2a+12 a
T11 = n2
a
p
a a a a
and i=1 j=1 Tij ≤ i=1 j=1 T11 → 0 as n → ∞ for a = a∗ n, we get

%
= 1 + on−3/2 
EZa 

Hence,

ProbZa = 0 ≤ on−3/2  (4)

From (3) and (4), we get the claimed result.

Our use of the probabilistic method in Theorem 2.6 was inspired by the
work presented in [2]. The fundamentals of the method and similar results,
on random graphs, for combinatorial quantities such as the clique number
and the chromatic number are presented in [2].
 clique problems 399

3. MULTIPARTITE CLIQUE PROBLEM

In this section we introduce the following three versions of the multipar-

tite clique problem: (1) maximum edge-weighted multipartite clique which
includes nodes from all levels (MPCP), (2) maximum edge-weighted multi-
partite clique which starts from the first level (product level) and includes
nodes from a contiguous subset of remaining levels (MPCF), and (3) maxi-
mum edge-weighted multipartite clique problem which includes nodes from
a subset of levels (MPCS).
Figure 1(a) shows a multipartite clique, namely, 1 2 5 6 7 10 11
13 15, which includes nodes from all the levels of the graph. The graph
given in Fig. 1(b) (where, for simplicity, we assumed all the edge weights to
be equal to 1) illustrates the difference between MPCP, MPCF, and MPCS.
Here, the optimum solution to MPCP is 1 3 5 9 10 13, whereas
the optimum solution to MPCF is 1 3 5 7 8 9 10 and the optimum
solution to MPCS is 3 4 5 7 8 9 10.
Note that, in general, problems MPCP, MPCP, and MPCS may not be
solved by solving a sequence of biliclique problems on successive bipartite
subgraphs of the multipartite graph.

3.1. Formulations
i i+1 i i+1
Let G = GVi  Vi+1  E  be the bipartite graph induced by node
i i+1
sets Vi and Vi+1 . Define the variable xe to be 1 if edge e of E i i+1
is not in the multipartite clique; 0, otherwise. For an edge e in E i i+1 , let
Ae be the edges in E i+1 i+2 which are adjacent to e and let Be be
i i+1
the edges in E i−1 i which are adjacent to e. we is the weight of edge

First Level (Final Products)

1 5 9 13 3 7
11
2 6 10 14 1 4 8
12
3 15 5 9
7 11 2
13
8 16 6 10
4 12
(a) (b)
A multi-partite clique (MPC) Difference between MPCP, MPCF
and MPCS.
(spanning all levels of the graph)

FIG. 1. MPCP and its variants: MPCF and MPCS.

400 dawande et al.

e in E i i+1 . We assume that there are n levels in the graph and they are
numbered 1     n where 1 represents the first level.

MPCP: Multipartite Clique which Includes Nodes from All Levels

n−1


i i+1 i i+1
W ∗ = min wk xk
i=1 k∈E i i+1

subject to
i i+1 i i+1
xk + xl ≥1 if edges k and l in E i i+1 cannot be in
the same biclique, ∀ pair k l ∈ E ii+1

i i+1 i−1 i
xe ≤ Ap + xp −1 ∀ p ∈ E i−1 i  2 ≤i≤n−1
e∈Ap

i−1 i i i+1
xe ≤ Bp + xp −1 ∀ p ∈ E i i+1  2 ≤i≤n−1
e∈Bp


x1
e
2
≤ E 1 2  − 1
e∈E 12
i i+1
xe ∈ 0 1 ∀ e ∈ E i i+1  1 ≤ i ≤ n − 1
For each bipartite subgraph Gi i+1 , due to the first set of constraints, the
i i+1
variables xe having value 0 form a biclique. The second set of con-
i−1 i
straints “links” these bicliques together. That is, if variable xe is 0 (i.e.,
edge e is in the biclique of Gi−1 i ), then at least one edge adjacent to e
in Gi i+1 should be in the MPC. Note that the second set of constraints
is required only for levels 2 through n − 1. Similarly, the third constraint
i i+1
makes sure that if a variable xe is 0 then at least one edge adjacent to
e in Gi−1 i should be in the MPC. The fourth constraint makes sure that at
least one edge from the first level is included in the MPC.

MPCF: Multipartite Clique Including the First Level

n−1


i i+1 i i+1
W ∗ = min wk xk
i=1 k∈E i i+1

subject to
i i+1 i i+1
xk + xl ≥1 if edges k and l in E i i+1 cannot be in
the same biclique, ∀ pair k l ∈ E i i+1
 clique problems 401

i i+1 i−1 i
xe ≤ Ap + xp − zi ∀ p ∈ E i−1 i  2 ≤i≤n−1
e∈Ap

i−1 i i i+1
xe ≤ Bp + xp − zi ∀ p ∈ E i i+1  2 ≤i≤n−1
e∈Bp

zi+1 ≤ zi 2 ≤i≤n−2

i i+1
xe ≥ E i i+1 1 − zi  2 ≤i≤n−1
i i+1
e∈E


x1
e
2
≤ E 1 2  − 1 (I)
1 2
e∈E
i i+1
xe ∈ 0 1 ∀ e ∈ E i i+1  2 ≤i≤n−1
zi ∈ 0 1 1 ≤ i ≤ n − 1

In this case, zi = 0 indicates that no edges from levels i and above can
be in the MPC. Notice that if zi = 0, then none of the edges in E i i+1
can be in the MPC due to the fifth set of constraints and zj = 0 ∀ j =
i + 1     n − 1 due to the fourth set of constraints. When zi = 1, the
second set of constraints “links” level i with level i + 1 in the MPC (similar
to MPCP) and when zi = 0, they are redundant.

MPCS: Multipartite Clique Problem which Includes Nodes from Some Levels
Problem MPCS can be formulated in a way similar to that of MPCF by
removing constraint (I) and using variables δi in addition to variables zi .
In the formulation for MPCF, zi = 0 indicates that no edges from levels
i and above can be in the MPC. Problem MPCS will have an additional
set of similar constraints involving variables δi where δi = 0 indicates that
no edges from levels i and below can be in the MPC. We avoid giving the
entire formulation for MPCS since the basic idea of the formulation is the
same as that of MPCF.
Since the biclique problem is a special case of MPCP, the complexity of
several optimization and decision problems regarding the MPCP follows
directly from the results for the biclique problem proved in Section 2. We
list these results in Lemma 3.1.

Lemma 3.1. Given a multipartite graph M = V1 ∪ V2 ∪ · · · ∪ Vn  E, the

following optimization/decision problems regarding MPCP, MPCF, and MPCS
are NP-complete.

• Maximum edge weight multipartite clique: Find a multipartite clique C,

where the sum of the edge weights in the subgraph induced by C is maximum.
402 dawande et al.

• Exact balanced node cardinality decision problem: Given M and a pos-

itive integer a ∈ Z+ , does there exist a multipartite clique C = U1 ∪ U2 ∪ · · · ∪
Un  E with U1  = U2  = · · · = Un  = a?
• Exact edge cardinality decision problem: Given M and a positive integer
k ∈ Z+ , does there exist a multipartite clique with exactly k edges?
• Maximum one-sided edge cardinality problem: Given M and a positive
integer k ∈ Z+ , find a maximum cardinality multipartite clique with exactly k
nodes on any level.
However, an interesting special case of MPCP can be solved in polyno-
mial time.
Theorem 3.2. Given a multipartite graph GV E, if an optimum multi-
partite clique M ∗ is such that for every level i (i = 1 2     n − 1), M ∗ has
a node (say vi ) such that all neighbors of vi in Gi i+1 are also in M ∗ , then
MPCP is polynomially solvable.
Proof. For a node u in level i, let N r u denote the neighbors of u in
i i+1
G and let N l u denote the neighbors of u is Gi−1 i . For a set S ⊆ Vi ,
N S = ∩i∈S N r i denotes the common neighborhood, in Gi i+1 , of nodes
r

in S. Similarly, N l S denotes the common neighborhood in Gi−1 i of

nodes in S.
For a node u in level 1, it is easy to see that the induced subgraph

Su = N l N r u × N r u × N r N r u × · · · × N r N r · · · N r u · · ·

is a MPC (provided that all the sets in the above product are nonempty).
Consider an optimal MPC (say M ∗ ) which satisfies the hypothesis. Thus, for
every level i (i = 1 2     n − 1), M ∗ has a node (say vi ) in the optimum
solution such that all neighbors of vi in Gi i+1 are also in the optimum solu-
tion. Then, it can easily be verified that Sv1 = M ∗ . Hence, the polynomial
time procedure which considers every node u from level 1 and constructs
the set Su will find M ∗ .
The above conditions on the multipartite clique may be true in certain
real environments. Many manufacturers in the computer industry offer a
base model (a complete product) as a shell and offer several options on
the base model to define other products in the product line. They store
inventory of the shell and use it as a vanilla box while customizing other
products with options. If the supplier (or supplying plant) of at least one key
component to the shell also follows a similar strategy, then the multipartite
cliques of interest are such that they require the conditions in Theorem 3.2
to be satisfied.
As with the biclique problem, the node cardinality and node weighted
counterparts of multipartite clique problems can be considered. To the best
 clique problems 403

of our knowledge, the complexity of these problems is open. To the best of

our knowledge, the complexity of the unweighted versions of MPCP, MPCF,
and MPCS is open.

4. CONCLUSIONS

In this paper, we studied biclique and multipartite clique problems.

Among biclique problems, we considered the maximum (edge) biclique
problem (MBP) and its weighted version, the maximum (edge) weighted
biclique problem (MWBP) in bipartite graphs. MBP and MWBP are inter-
esting problems from a theoretical point of view and have applications in
manufacturing and formal concept analysis. We showed that MWBP and
four variants of MBP are NP-complete. For random bipartite graphs, we
presented a result about the size of a maximum balanced biclique. This
result and an observation suggest that the number of edges in a maximum
edge cardinality biclique may be considerably larger than the number of
edges in a maximum balanced biclique, and it highlights the difference
between the well known maximum balanced node cardinality biclique prob-
lem and MBP. We also presented three versions of the multipartite clique
problem.

REFERENCES

1. R. K. Ahuja, T. L. Magnanti, and J. B. Orlin, “Network Flows,” Prentice Hall, Englewood

Cliffs, NJ, 1993.
2. N. Alon, J. H. Spencer, and P. Erdos, “The Probabilistic Method,” Wiley, New York, 1992.
3. B. Ganter, personal communication.
4. B. Ganter and R. Wille, “Formale Begriffsanalyse—Mathematische Grundlagen,”
Springer-Verlag, Berlin/New York, 1996.
5. M. S. Garey and D. S. Johnson, “Computers and Intractibility: A Guide to NP-
Completeness,” Freeman, New York, 1979.
6. G. H. Hardy and E. M. Wright, “An Introduction to the Theory of Numbers,” Clarendon,
Oxford, 1954.
7. D. S. Hochbaum, Approximating clique and biclique problems, J. Algorithms 29 (1997),
174–200.
8. S. Krolak-Schwerdt and P. Orlik, Ein Verfahren zur Klassifikation zweimodaler bin” arer
Daten (A method for classifying twomodal binary data), presented at the Annual Confer-
ence of the German Classification Society, 1996.
9. G. L. Nemhauser and L. A. Wolsey, “Integer Programming and Combinatorial Optimiza-
tion,” Wiley, New York, 1988.
10. R. Peeters, “The Maximum Edge Biclique Problem Is NP-Complete,” Tilburg University
Department of Econometrics research memorandum, 2000.
11. J. M. Swaminathan and S. Tayur, Management of broader product lines through delayed
product differentiation using vanilla boxes, Management Sci. 44 (1998), 161–172.