Online Set Cover with Rated Subsets
Christine Markarian
Department of Engineering and Information Technology,
University of Dubai, Dubai, U.A.E.
Keywords:
Online Set Cover, Rating, Optimization, Online Algorithms, Competitive Analysis, Randomized Rounding.
Abstract:
In this paper, we introduce the Online Set Cover With Rated Subsets problem (OSC-RS), a generalization of
the well-known Online Set Cover problem, in which we are given a universe of elements and a collection of
subsets of the universe, each associated with a subset cost and a rating cost. In each step, the algorithm is given
a request containing elements from the universe. The algorithm serves a request by assigning it to a number
of purchased subsets that jointly cover the requested elements. The algorithm pays the subset costs associated
with the subsets purchased and for each request, it pays the sum of the rating costs associated with the subsets
assigned to the request. The aim is to serve all requests as soon as revealed, while minimizing the total subset
and rating costs paid. OSC-RS is motivated by intrinsic client-service-providing scenarios in which service
providers are rated and their ratings are included in the decision-making process, so as higher-rated service
providers are associated with lower rating costs. That is, the decisions about serving clients take into account
the quality of the services provided. We propose the first online algorithm for OSC-RS and evaluate it using
the standard notion of competitive analysis. The latter compares the performance of the online algorithm to
that of an optimal offline algorithm that is assumed to know all the input sequence in advance.
1 INTRODUCTION
In our digital world today, nearly every service or
product offered is accompanied with a rating (Landy
and Farr, 1980). We rely on customer reviews to make
most of our day-to-day decisions, including which ho-
tel to stay in, which restaurant to go to, and which
course to register for. The customer review system
is applied in almost every industry, urging companies
to keep good reputation in order to succeed or even
survive (Trenz and Berger, 2013; Zhou et al., 2014).
Despite the vital role played by customer review sys-
tems in ensuring quality control for companies, most
real-world optimization problems studied thus far in
algorithmic theory do not include the rating factor in
their optimization model.
In this paper, aiming to better reflect today’s real-
world companies, that highly rely on customer review
systems, we introduce the concept of rating in the op-
timization objective. In particular, we address a vari-
ant of the well-known Set Cover (SC), a fundamen-
tal problem in operations research, computer science,
and combinatorics (Feige, 1998; Slavık, 1997). SC
has numerous applications in scheduling, budgeting,
networks among others (Vemuganti, 1998). Given a
universe of elements and a collection of subsets of the
universe, each associated with a positive cost, SC asks
to purchase subsets of minimum total costs that can
cover the elements of the universe.
In the online setting (Borodin and El-Yaniv,
2005), elements of the universe are revealed to the
online algorithm over time and the problem is known
as the Online Set Cover problem (OSC) (Alon et al.,
2009). OSC appears, among many other applica-
tions, in client-service-providing scenarios in which
elements represent clients and subsets represent ser-
vice providers. Covering elements is translated to
serving clients and subset costs are equivalent to ser-
vice providing costs, where each service provider can
serve a subset of the clients.
2 ONLINE SET COVER WITH
RATED SUBSETS (OSC-RS)
In this paper, we study the Online Set Cover With
Rated Subsets problem (OSC-RS), a generalization
of OSC, in which we are given a universe of n ele-
ments and a collection of m subsets of the universe,
each associated with a positive subset cost and a pos-
itive rating cost. In each step, the algorithm is given
a request containing at most k elements from the uni-
verse. Each element may belong to at most d subsets.
The algorithm serves a request by assigning it to a
number of purchased subsets that jointly cover the re-
Markarian, C.
Online Set Cover With Rated Subsets.
DOI: 10.5220/0010976900003179
In Proceedings of the 24th International Conference on Enterprise Information Systems (ICEIS 2022) - Volume 1, pages 455-461
ISBN: 978-989-758-569-2; ISSN: 2184-4992
Copyright
c
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
455
quested elements. The algorithm pays the subset costs
associated with the subsets purchased and for each re-
quest, it pays the sum of the rating costs associated
with the subsets assigned to the request. The aim is to
serve all requests as soon as revealed, while minimiz-
ing the total subset and rating costs paid. Note that in
OSC (Alon et al., 2009), a single element arrives in
each time step. Hence, OSC-RS generalizes OSC by
setting the number of elements per request to 1 and all
rating costs to 0.
OSC-RS is motivated by intrinsic client-service-
providing scenarios in which service providers are
rated and their ratings are included in the decision-
making process, so as higher-rated service providers
are associated with lower rating costs. That is, the
higher the rating of a service provider, the lower is its
rating cost. Consider a company that assigns service
providers to serve its clients. Unlike in the Online Set
Cover problem, in OSC-RS, the decisions about serv-
ing clients take into account the quality of the ser-
vices provided. Hence, the company does not only
aim to minimize its classical costs, but considers low-
rated services to be additional costs that represent its
reputation-maintainance costs. In OSC-RS, this is re-
flected in the rating costs associated with the subsets
in addition to the subset costs.
2.1 Our Contribution
We propose the first online algorithm for OSC-RS and
evaluate it using the standard notion of competitive
analysis, which is defined as follows. An online al-
gorithm has competitive ratio c (or is c-competitive)
if for all input sequences, the cost incurred by the on-
line algorithm is at most c times the cost of an optimal
offline algorithm that knows the entire input sequence
in advance and is optimal. The input to an online al-
gorithm is revealed in portions over time, and the al-
gorithm is expected to react to each portion while tar-
geting the given optimization goal against the whole
input sequence. Unlike offline algorithms that react
once to a given input sequence, the challenge of an
online algorithm is to react to each of the input por-
tions as soon as revealed, without knowing about the
remaining input sequence in advance.
2.2 Relationships with Other Problems
The special case of OSC-RS in which each request
is composed of one element, can be solved by trans-
forming a given instance of the problem into an in-
stance of the Non-metric Online Facility Location
problem (non-metric OFL) (Alon et al., 2006), de-
fined as follows. Given a collection of potential facil-
ity locations. Each facility can be opened by paying
its so-called opening cost. Clients arrive at different
locations over time. Non-metric OFL asks to connect
each arriving client to an open facility. Connecting
a client to an open facility incurs the so-called con-
necting cost between the client and the facility, which
is the distance between the client location and the fa-
cility location. The transfomation works as follows.
Each subset is formed as a facility location and fa-
cility opening costs correspond to subset costs. Each
element is formed as a client and the connecting cost
between a client and a facility is set to the rating cost
of the subset corresponding to the facility. If an ele-
ment does not belong to a subset, the connecting cost
between the corresponding client and the correspond-
ing facility is set to infinity. Every time a client is
connected to a facility, the corresponding subset’s rat-
ing cost is paid.
Our algorithm for OSC-RS which we describe in
the coming sections will handle requests composed of
any bounded number of elements.
As mentioned earlier, OCS-RC generalizes the
Online Set Cover problem (OSC) due to (Alon et al.,
2009), in which the number k of elements in each re-
quest is 1 and all rating costs are 0.
There is (
lognlogm
loglogn+loglogm
) lower bound on the
competitive ratio of any deterministic algorithm
for OSC, where n is the number of elements and
m is the number of subsets, due to (Alon et al.,
2009).
There is (log nlog m) lower bound on the com-
petitive ratio of any randomized polynomial-time
algorithm for OSC, under the assumption that
BPP 6= NP, due to (Korman, 2005).
2.3 Competitive Ratio
Our algorithm for OSC-RS has an O(logd log n) com-
petitive ratio, where:
d is the maximum number of subsets any element
can belong to.
n is the total number of elements.
2.4 Algorithmic Techniques
Our online algorithm is based on randomized round-
ing, a technique used to solve many online optimiza-
tion problems (Markarian, 2021; Markarian et al.,
2021; Hamann et al., 2018; Markarian, 2018; Alon
et al., 2006). Moreover, the algorithm is based on
a transformation of a given OSC-RS instance into a
graph-theoretic problem instance, as input portions
appear over time.
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
456
Figure 1: OSC-RS graph instance of three elements, three
subsets, and a request of two elements.
3 TRANSFORMATION OF
OSC-RS INSTANCE
Given an instance I of the Online Set Cover With
Rated Subsets problem (OSC-RS). We transform I
into a directed edge-weighted graph problem in-
stance, as follows. We refer the reader to Figure 1
for an illustration.
The Nodes of the Graph Are Formed as Follows:
We associate each element in the universe with an
element node.
We associate each subset with two nodes: a subset
node and a subset-copy node.
We associate each request with a request node.
The Edges of the Graph Are Formed as Follows:
We add a directed edge from each subset-copy
node to each element node if the corresponding
subset contains the corresponding element. These
edges have weight 0.
We add a directed edge from each subset node
to its corresponding subset-copy node, of weight
equal to the corresponding subset cost.
For each request, we add a directed edge from
the request node to each subset node which corre-
sponds to a subset containing at least one element
comprising the request. The weight of each out-
going edge to a subset node will be equal to the
rating cost of the corresponding subset.
Recall that the goal of the algorithm is to assign
each request to a number of subsets that jointly cover
the elements of the request. On a given time step, the
algorithm may purchase new subsets by paying their
subset costs. This goal is tansformed in our graph
problem as follows.
Initially, the subset nodes, the subset-copy nodes,
and the element nodes are formed, including the
weighted edges between them, as described above.
In each step, as soon as a new request arrives:
The algorithm forms a request node associated
with the request and adds all the outgoing edges
from it, as described above.
The algorithm finds a directed path from the re-
quest node to each of the element nodes corre-
sponding to the elements of the request. To do
this, the algorithm purchases edges of the graph
that form the desired paths.
The algorithm pays the subset costs and the rat-
ing costs associated with the purchased edges, as
follows. Whenever the algorithm purchases an
edge that connects a subset node to its subset-
copy node, the corresponding subset cost is paid.
Whenever the algorithm purchases an edge con-
necting a request node to a subset node, the corre-
sponding rating cost of the subset is paid.
It is easy to see that a feasible solution for the
graph instance formed from a given instance I of
OSC-RS according to the transformation above, im-
plies a feasible solution for I, of the same cost.
4 ONLINE ALGORITHM
In this section, we present an online algorithm that
finds a feasible solution for the graph problem de-
scribed in the previous section. This implies a feasible
solution for any given instance of OSC-RS.
The algorithm is randomized and has two phases.
In the first phase, the algorithm uses a randomized
approach to purchase edges from the graph, that may
form an infeasible solution. In the second phase, the
algorithm may buy further edges to guarantee a feasi-
ble solution.
This idea has been implemented in other algo-
rithms for similar online problems (Markarian, 2021;
Markarian et al., 2021; Hamann et al., 2018; Markar-
ian, 2018; Alon et al., 2006).
A fractional value v
e
is assigned to each edge e
in the graph, all set to 0 initially and non-decreasing
throughout the execution of the algorithm. The con-
cepts of maximum flow and minimum cut will be
used. The maximum flow from node i to node j is the
smallest total fractional values of edges which would
break the directed path from i to j if removed. These
edges are called minimum cut edges.
The algorithm chooses a random value q, indepen-
dently among 2
d
logn
e
random variables that are uni-
formly distributed in the interval [0, 1], where n is the
Online Set Cover With Rated Subsets
457
total number of elements. This value q will be used
by the algorithm in its first phase when choosing the
edges.
Recall that the edges of the graph are directed and
weighted, as per the transformation described in the
previous section. We refer to the weight of an edge e
by w
e
.
The online algorithm reacts to each new element s
in a given request as follows:
If the algorithm’s current outputted solution does
not contain a directed path from the request node to
the element node associated with s, the algorithm per-
forms the following:
Phase 1. The algorithm calculates the maximum
flow from the request node to the element node as-
sociated with s. As long as this value does not exceed
1, the algorithm constructs a minimum cut Q and in-
creases the fractional value of each edge in Q , based
on its weight and the cardinality of Q , using the equa-
tion below:
v
e
v
e
(1 +
1
w
e
) +
1
|Q | · w
e
When the maximum flow is at least 1, the algo-
rithm adds each edge whose fractional value exceeds
q (the random value described earlier) to the solution.
Phase 2. At this point, the algorithm might have
constructed an infeasible solution - that depends on
the randomization process. To ensure a feasible solu-
tion, the algorithm observes the edges in the solution
so far outputted. If the solution does not contain a di-
rected path from the request node to the element node
associated with s, the algorithm calculates a shortest
weighted directed path from the request node to the
element node and adds the edges of this path to the
solution.
The two phases of the algorithm are depicted in
Algorithm 1.
5 COMPETITIVE ANALYSIS
In this section, we prove an upper bound on the com-
petitive ratio of the algorithm presented in the previ-
ous section.
As mentioned in Section 3, the expected cost of
the solution for a given instance I of OSC-RS is less
or equal to the expected solution cost of the corre-
sponding graph instance. Hence, to analyze the com-
petitive ratio of the algorithm, we observe the total
expected cost of edges purchased by the algorithm in
comparison to that of an optimal offline solution.
Algorithm 1: Online Algorithm for OSC-RC.
Phase 1.
While the maximum flow from the request node to
the element node associated with s is less than 1:
Construct a minimum cut Q from the request
node to the element node associated with s and in-
crease the fractional value v
e
of each edge e Q ,
using the equation below:
v
e
v
e
(1 +
1
w
e
) +
1
|Q | · w
e
Add each edge e to the solution if its value v
e
ex-
ceeds q.
Phase 2.
If there is no directed path from the request node
to the element node associated with s in the current
solution, add to the solution the edges of a smallest
weighted path from the request node to the element
node associated with s.
Notice that the algorithm may purchase edges in
the first and second phases. Let S
1
and S
2
denote the
the edges purchased by the algorithm in Phase 1 and
Phase 2, respectively. We will analyze the expected
cost of each collection separately. We denote by Opt
the cost of an optimal offline solution.
Phase 1 - Analysis. In the first phase, the algorithm
purchases an edge whenever the fractional value as-
sociated with it becomes at least q (recall the num-
ber generated by a random process before the ar-
rival of any request). Let us observe an edge e and
i : 1 i 2
d
logn
e
. Let X
e,i
be the variable indicating
whether e has been purchased by the algorithm or not.
Recall that w
e
and v
e
denote the weight and fractional
value of an edge e, respectively. We can now express
the expected cost C
S
1
of the collection S
1
as follows:
C
S
1
=
eS
1
2
d
logn
e
i=1
w
e
· Exp [X
e,i
] (1)
= 2
d
logn
e
eS
1
w
e
v
e
(2)
To make a comparison to the cost of the optimal
offline solution, we need to observe the minimum cuts
constructed, since the algorithm purchases edges of
minimum cuts constructed in the first phase. The first
observation which allows the comparison is that each
minimum cut generated by the algorithm contains at
least one edge that is in the optimal offline solution.
Observation 1. Each minimum cut constructed by
the algorithm contains at least one optimal edge.
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
458
This observation is true because of the definition
of minimum cut and since any optimal algorithm is
supposed to output a directed feasible path for the
given element. The second observation which allows
the comparison is the number of times the algorithm
constructs a minimum cut.
Observation 2. The algorithm constructs O(Opt ·
log|Q|) minimum cuts, where |Q| is the size of the
biggest minimum cut constructed.
Proof. Observe the optimal edges. We show that each
such edge would show up in a bounded number of
minimum cuts, since its fractional value will eventu-
ally become 1 and so it won’t show up in any later
minimum cut, according to the algorithm’s design.
Let us fix any optimal edge and look at the equa-
tion of the algorithm to check how many times we
need to perform a fractional increase until the frac-
tional value of this edge becomes 1. Notice that
O(w
e
log|Q|) fractional increases are needed to make
the value of the edge become at least 1. On the other
hand, as per the first observation, each minimum cut
contains at least one optimal edge. Therefore, the
algorithm constructs O(Opt · log |Q|) minimum cuts,
where |Q| is the size of the biggest minimum cut con-
structed.
To achieve the desired upper bound, we will need
to give an upper bound on |Q|. Since each element
may belong to at most d subsets, then there are at
most d directed paths from the request node to each
element node. Therefore, |Q| d and so the num-
ber of minimum cuts constructed is upper bounded
by O(Opt · log |d|).
The next thing to look at is the total fractional in-
crease associated with each minimum cut. Since we
have an upper bound on the number of minimum cuts
constructed, we can conclude an upper bound on the
expected cost C
S
1
of the algorithm in Phase 1.
Lemma 1. Each minimum cut constructed is associ-
ated with a maximum of 2 fractional increase.
Proof. We observe any minimum cut Q. Each edge e
in Q causes a fractional increase of w
e
·
v
e
w
e
+
1
|
Q
|
·w
e
,
based on the algorithm’s equation. Notice that before
an increase is made, the maximum flow is less than 1.
Hence,
eQ
v
e
< 1. Hence, we conclude the follow-
ing.
eQ
w
e
·
v
e
w
e
+
1
|
Q
|
· w
e
< 2
Therefore, we have that
eS
w
e
v
e
O(Opt ·
logd). Thus,
C
S
1
O(Opt · log n · log d) (3)
Phase 2 - Analysis. We now bound the expected
cost C
S
2
of the collection S
2
of edges purchased by
the algorithm in Phase 2.
The flow of a path will be the minimum value
among the edge values of a given path. We fix an
element node u and an i: 1 i 2
d
logn
e
. We calcu-
late the probability that there is no feasible path pur-
chased by the algorithm in Phase 1 for u, for a single
i. In terms of flow, the latter is the probability that q
is more than the flow of each directed path to u. Let
Q be a minimum cut constructed at the end of Phase
1. Before executing Phase 2, we have that the sum of
flow associated with all paths to u is at least 1 accord-
ing to the algorithm. Therefore, the probability that
there is no feasible path purchased by the algorithm
in Phase 1 for u, considering a single i, is:
eQ
(1 v
e
) e
eQ
v
e
1
e
Now we calculate the probability considering all
i: 1 i 2
d
logn
e
and conclude the following. The
probability that there is no feasible path purchased by
the algorithm in Phase 1 for u is at most
1
n
2
.
Hence, the algorithm purchases a smallest
weighted path from the request node to u with prob-
ability at most
1
n
2
. Obviously, this path is upper
bounded by Opt, since it is a smallest weighted path.
Since we have a total of n elements, each arriving
exactly once, we can conclude the following.
C
S
2
n ·
Opt
n
2
Therefore, we following theorem holds.
Theorem 1. There is an O(log d log n)-competitive
randomized algorithm for the Online Set Cover With
Rated Subsets problem, where d is the maximum num-
ber of subsets any element can belong to, and n is the
number of elements.
6 CONCLUDING THOUGHTS
We have initiated in this paper the study of the im-
pact of rated resources on decision-making. We have
generalized the well known Online Set Cover problem
(OSC) by associating each subset with a price that de-
pends on the rating of the subset or resource it repre-
sents. In other words, upon purchasing a resource, the
Online Set Cover With Rated Subsets
459
decision-maker pays in addition to the actual cost of
the resource, the price for good quality service.
Our simple yet effective model is one way to rep-
resent resource ratings and relate them to the opti-
mization objective. There are different ways this can
be done, for example, rather than paying the sum of
the rating costs for each subset serving a request, one
could pay the average rating cost of these subsets in-
stead.
Also, each request could be associated with an
amount of money one is willing to pay to get a higher
rated service for the request. In many real-world busi-
ness scenarios, customers are often given options to
choose from when it comes to quality of services.
Hence, the decision maker would take into account
the quality of service requested when assigning re-
sources to clients. For instance, a rating cost would
not have to be paid if the request is not made for a
high-rated service.
Furthermore, the ratings in our model are assumed
to be fixed throughout time, which is not the case in
actual rating systems. One may want to include dy-
namic pricing for these ratings. From an algorithmic
perspective, it could be that extensions of the existing
algorithm would solve the variants that arise from dy-
namic pricing - a similar study as the one in (Feldkord
et al., 2017) for leasing problems.
We have initiated this study by targeting OSC.
There are a lot of other well-studied online resource
allocation problems to explore. Many of these would
serve as real-world decision-making problems in the
context of rating, such as variants of the Online Fa-
cility Location problem (Alon et al., 2006; Meyer-
son, 2001; Markarian et al., 2021; Markarian and
Khallouf, 2021; Markarian and auf der Heide, 2019)
and variants of the Online Connected Dominating Set
problem (Hamann et al., 2018; Markarian and Kassar,
2020).
REFERENCES
Alon, N., Awerbuch, B., Azar, Y., Buchbinder, N., and
Naor, J. (2006). A general approach to online net-
work optimization problems. ACM Transactions on
Algorithms (TALG), 2(4):640–660.
Alon, N., Awerbuch, B., Azar, Y., Buchbinder, N., and
Naor, J. (2009). The online set cover problem. SIAM
Journal on Computing, 39(2):361–370.
Borodin, A. and El-Yaniv, R. (2005). Online computa-
tion and competitive analysis. Cambridge University
Press.
Feige, U. (1998). A threshold of ln n for approximating set
cover. Journal of the ACM (JACM), 45(4):634–652.
Feldkord, B., Markarian, C., and auf der Heide, F. M.
(2017). Price fluctuation in online leasing. In Gao,
X., Du, H., and Han, M., editors, Combinatorial Op-
timization and Applications - 11th International Con-
ference, COCOA 2017, Shanghai, China, December
16-18, 2017, Proceedings, Part II, volume 10628 of
Lecture Notes in Computer Science, pages 17–31.
Springer.
Hamann, H., Markarian, C., auf der Heide, F. M., and
Wahby, M. (2018). Pick, pack, & survive: Charg-
ing robots in a modern warehouse based on online
connected dominating sets. In Ito, H., Leonardi,
S., Pagli, L., and Prencipe, G., editors, 9th Inter-
national Conference on Fun with Algorithms, FUN
2018, June 13-15, 2018, La Maddalena, Italy, volume
100 of LIPIcs, pages 22:1–22:13. Schloss Dagstuhl -
Leibniz-Zentrum f
¨
ur Informatik.
Korman, S. (2005). On the use of randomization in the on-
line set cover problem. Master’s thesis, Weizmann In-
stitute of Science, Israel.
Landy, F. J. and Farr, J. L. (1980). Performance rating. Psy-
chological bulletin, 87(1):72.
Markarian, C. (2018). An optimal algorithm for online
prize-collecting node-weighted steiner forest. In Il-
iopoulos, C. S., Leong, H. W., and Sung, W., editors,
Combinatorial Algorithms - 29th International Work-
shop, IWOCA 2018, Singapore, July 16-19, 2018, Pro-
ceedings, volume 10979 of Lecture Notes in Computer
Science, pages 214–223. Springer.
Markarian, C. (2021). Online non-metric facility location
with service installation costs. In Filipe, J., Smi-
alek, M., Brodsky, A., and Hammoudi, S., editors,
Proceedings of the 23rd International Conference on
Enterprise Information Systems, ICEIS 2021, Online
Streaming, April 26-28, 2021, Volume 1, pages 737–
743. SCITEPRESS.
Markarian, C. and auf der Heide, F. M. (2019). Online
algorithms for leasing vertex cover and leasing non-
metric facility location. In Parlier, G. H., Liberatore,
F., and Demange, M., editors, Proceedings of the 8th
International Conference on Operations Research and
Enterprise Systems, ICORES 2019, Prague, Czech
Republic, February 19-21, 2019, pages 315–321.
SciTePress.
Markarian, C. and Kassar, A. (2020). Online deterministic
algorithms for connected dominating set & set cover
leasing problems. In Parlier, G. H., Liberatore, F., and
Demange, M., editors, Proceedings of the 9th Interna-
tional Conference on Operations Research and Enter-
prise Systems, ICORES 2020, Valletta, Malta, Febru-
ary 22-24, 2020, pages 121–128. SCITEPRESS.
Markarian, C., Kassar, A., and Yunis, M. M. (2021). An
algorithmic approach to online multi-facility location
problems. In Parlier, G. H., Liberatore, F., and De-
mange, M., editors, Proceedings of the 10th Interna-
tional Conference on Operations Research and En-
terprise Systems, ICORES 2021, Online Streaming,
February 4-6, 2021, pages 29–35. SCITEPRESS.
Markarian, C. and Khallouf, P. (2021). Online metric fa-
cility service leasing with duration-specific dormant
fees. In Panetto, H., Macchi, M., and Madani, K., edi-
tors, Proceedings of the 2nd International Conference
ICEIS 2022 - 24th International Conference on Enterprise Information Systems
460
on Innovative Intelligent Industrial Production and
Logistics, IN4PL 2021, October 25-27, 2021, pages
25–31. SCITEPRESS.
Meyerson, A. (2001). Online facility location. In Proceed-
ings 42nd IEEE Symposium on Foundations of Com-
puter Science, pages 426–431.
Slavık, P. (1997). A tight analysis of the greedy algorithm
for set cover. Journal of Algorithms, 25(2):237–254.
Trenz, M. and Berger, B. (2013). Analyzing online cus-
tomer reviews-an interdisciplinary literature review
and research agenda.
Vemuganti, R. R. (1998). Applications of Set Covering, Set
Packing and Set Partitioning Models: A Survey, pages
573–746. Springer US, Boston, MA.
Zhou, L., Ye, S., Pearce, P. L., and Wu, M.-Y. (2014).
Refreshing hotel satisfaction studies by reconfiguring
customer review data. International journal of hospi-
tality management, 38:1–10.
Online Set Cover With Rated Subsets
461