Int. J. Applied Decision Sciences, Vol. 5, No.

1, 2012 47
Copyright 2012 Inderscience Enterprises Ltd.











A model for bundling mobile value added services
using neural networks
Arpan Kumar Kar*
Information Systems Area,
XLRI School of Business and Human Resources,
CH Area (East), Jamshedpur 831001, Jharkhand, India
E-mail: arpan.kumar.kar@gmail.com
*Corresponding author
Priyal Singh
Operations Management Area,
XLRI School of Business and Human Resources,
CH Area (East), Jamshedpur 831001, Jharkhand, India
E-mail: priyalsing@gmail.com
Abstract: Mobile value added services has emerged as one of the significant
revenue generators for the mobile telecommunication industry. With the large
array of service offerings, telecom providers engage in active advertising of
service bundles to attract new as well as retain the existing customers. Focused
advertising by personalising the service bundles based on individual customer
preferences is a key success factor of such campaigns. This paper proposes
the application of a neural network-based decision support model for pattern
association using which mixed bundles of such services may be identified, as
per customer preferences, so as to induce higher conversion in cross-sales and
up-sales. Consumer service preferences can be combined with some standard
packages of VAS bundles to personalise the service bundles and these service
bundles would be generated dynamically. Identifying and promoting such
services identified through pattern association rules will have a direct impact on
the overall profitability of the service provider.
Keywords: bundling mobile services; pattern association; neural networks;
mobile commerce; marketing mobile services; decision support system.
Reference to this paper should be made as follows: Kar, A.K. and Singh, P.
(2012) A model for bundling mobile value added services using neural
networks, Int. J. Applied Decision Sciences, Vol. 5, No. 1, pp.4763.
Biographical notes: Arpan Kumar Kar is a Research Fellow in XLRI School
of Business and Human Resources. He specialises in e-business, technology
management, management science and supply chain management. He has
published in multiple international journals and conference proceedings. He is a
member of IEEE Computer Society, Association of Computing Machinery and
Association of Information Systems.











48 A.K. Kar and P. Singh












Priyal Singh is a Research Fellow in XLRI School of Business and Human
Resources. She specialises in supply chain management and management
science. She has multiple international publications to her name.

1 Introduction
The telecommunication industry is characterised by increasing operators working on
similar tariff rates due to fierce competition. With price being a non-competing
parameter, it is imperative to focus on unique value propositions for customers in the
form of services to maintain the profitability for a particular service provider. In the
Indian context, the explosion of value added services (VAS) is a testimony to this current
trend of tight competition between all service providers. However, with the availability of
large customer data, companies can engage in selective promotion of VAS to its customer
in order to increase the revenue as well increase customer satisfaction.
Service differentiation using VAS can also be used to increase customer loyalty,
particularly if services are priced to encourage usage and if they involve some time
investment from the customer such as backing up personal data or creating personalised
content (Balakrishnan, 2010). This is of immense importance in an industry marked with
extremely high customer churn rate and a stagnant (if not dipping) average revenue per
user (ARPU). In this study, a dynamic model using artificial neural networks has been
proposed for bundling of VAS for improving the adoption of the same by the consumers.
These bundles are being generated dynamically based on past consumer purchase
behaviour. The model proposes the usage of dynamic generation of pattern association
rules for bundling of VAS obtained from a neural network configuration. The model also
evaluates every bundle, thus created for its probable profitability before suggesting the
same for the VAS marketer.
2 Literature review
2.1 Mobile value added services
Mobile value added services (MVAS) are those services that are not part of the basic
voice offer and are availed separately by the end user. Mobile services differ from
traditional services in their ability to provide service offerings regardless of temporal
and spatial constraints. Mobile services additionally differ from each other in their
value-added factor to the end-user. Some services such as e-mailing and voice are more
business oriented while on the other hand, various free-time oriented services are
provided in new smartphones, such as imaging and music playback (Verkasalo, 2008).
They are also different from traditional interpersonal services that are delivered
face-to-face, or from other types of e-services, such as wireless online services, where the
service delivery is linked to a specific fixed local area network or specific location
(Heinonen and Pura, 2006). This study proposed a classification of the mobile services
based on service management perspective as given in Figure 1.










A model for bundling mobile value added services using neural networks 49













Figure 1 Classification of mobile services

Source: Heinonen and Pura (2006)
For the purpose of service differentiation, all four levels should be taken into account
when designing mobile services, pricing services, and segmenting services into bundles
that offer similar value propositions, which thereby can be marketed to the potential users
with similar marketing communication. The four levels are related to each other and each
level contributes to offering value-in-use to the customer.
MVAS offers an opportunity to re-invent the customer and to bring the needs of the
customer into the interface with the mobile technology (Carlsson and Walden, 2002).
Due to its inherent characteristics such as ubiquity, personalisation, flexibility, and
dissemination it provides businesses with promise of unprecedented market potential,
great productivity and high profitability (Siau et al., 2001). It also offers a greater chance
of adoption and consequently repeated usage by the telecom customers, than for most
other telecom services. As demonstrated in Figure 2, the mobile service reputation and
the interest it generates in the network depend on the existing services provided (Amiri
and Kian, 2008). This further demonstrates the need for delivering effective service
packages, which when extended to MVAS, refers to suitable bundling of VAS based on
customer consumer behaviour.
Figure 2 Mobile service cause and effect loop diagram (see online version for colours)

Source: Amiri and Kian (2008)









50 A.K. Kar and P. Singh












May (2001) proposed that MVAS can be classified in three primary categories,
namely-base service platforms (e-mail and messaging, web access, voice, location
finding and digital content products), services for customers (travel, ticketing, banking,
stock trading, news and sports, gambling, games and shopping), and services for business
(mobile markets and the independent professional, mobile collaboration in action and
mobile commerce closes the management loop). MVAS is fast becoming a major part of
the revenue mix for many of the telecom service providers around the world. Mobile data
is now typically between 25% and 40% of revenues in mature markets and by far the
fastest growing service.
The growth of MVAS industry also has led to the development of an entirely new
business eco-system, namely those of supporting industries focusing on content
development, application development and content aggregation. A typical value chain in
the MVAS industry encompasses content creators, providers, aggregators, technology
enablers, and operators. However, as we see in the value chain, at the content
portals/aggregators as well as the previous stages in the value chain if customer data is
used to generate patterns of usage and create customer segment profiles, it will help in
creating and promoting customer driven services and would increase the effectiveness of
the marketing programme and consequently the revenues. This has been emphasised by
reports which state that there is a significance of looking at the needs of specific
segments and to assemble a portfolio of VAS products, internal and from third parties,
which meet those needs by doing data mining analysis and using local intelligence
(Balakrishnan, 2010).
One of the most dominant strategies to provide service differentiation for the
customers is to bundle a number of VAS. Bundling can be defined as selling of two or
more products in a package for a special price. Bundling is particularly relevant in
markets where the interdependency between products is critical and where products
depend directly on each other with regard to their adoption, as is the case in many
computer, information and telecommunication services. The packaging of mobile (data)
services is less common. Many mobile data and content services are offered as separate
services on mobile operators or content providers portals, either as a subscription
service or as single downloads. With such a wide foray of VAS, service bundles are
expected to play a central role in the success of next generation mobile services.
Bundling of core products and/or services is a well-known phenomenon in the
telecommunication industry. Packages in which telephony services are combined with
wireless and internet services are common. Bundling of mobile services with an almost
for free handset is common practice. Since the objectives of telecom service providers
are very well defined, i.e., increasing revenue, lower churn-rates and improve margins
(Bouwman et al., 2005), bundling of VAS offers an exciting strategy to achieve all the
objectives.
Anckar and DIncau (2002) found that from a consumers point of view, the results
indicated a rather low willingness to use mobile services in general, but an exceptionally
high willingness to use some applications. One reason for slow adoption of VAS is a
mismatch between launched applications and the everyday needs of target users
(Carlsson and Walden, 2002). These studies indicate that there is an urgent requirement
among telecom service providers to device a model to understand and predict consumer
purchase behaviour for VAS so as to increase the cross-sale and up-sale of such VAS.
Pau and Dits (2002) proposed a formal six-level framework for personalisation features
in current and next generation mobile services, which can be used to drive the business









A model for bundling mobile value added services using neural networks 51













modelling of mobile business services from as service provider or system supplier points
of view. The study also analysed the economic, sociological, information and psychic
drivers for personalisation in mobile services and why they sometimes differ from
web-based services. Yang and Laura (2006) found that the emotional value and
functional value aspects of mobile data services motivate consumers to adopt and further
use repeatedly mobile data services. Since emotional value is derived from the consumer
operator interaction interface that is difficult to control. But based on the functional value
as known from the past consumer history, data mining can be used effectively to find
patterns of the type of functional value a consumer is looking for and hence, promote
such products to the respective consumer. Despite their potential, little is known about
what drives consumer value and ultimately the adoption of mobile service bundles
(Haaker and Voss, 2007). Verkasalo (2008) proposed that mobile services are used
differently in different contexts (e.g., office, home, on the move) and based on the kind of
context, the service would be chosen. Li et al. (2008) proposed collaborative filtering
(CF) as one of successful techniques to produce personalised recommendations for users
where the contextual information of a user and the multi-criteria ratings of an item have
been considered besides the typical information on users and items.
However, there are very few studies conducted in the context of services industry
specially telecommunications that address to this need of data mining to determine
customer preferences and thereby promote selective services to its customers. Data
mining and business intelligence proved to be useful in business where transactional data
turned out to be a mine of information about customer purchase habits. Therefore,
developing customer models (called also profiles in the literature) based on data mining
is an important step for targeted marketing (Romdhane et al., 2010). Moreover, the
telecommunications report of India, the ESI Report (2010) states that a significant
challenge is going to be how todays communications provider will evolve to become a
convergent service provider and offer service bundles to their subscribers that might
include entertainment, infotainment, content, mobile transaction, advertisements and
sponsorships on top of the data apart from voice services.
While propositions for initial services may be made through segmentation via
classification, the proposed methodology attempts to create value after an initial purchase
is made by the end customer. In this study, the technique being proposed creates
association rules for the customers of the telecom service provider, when they make a
VAS purchase, to generate a dynamic VAS bundle for cross-sell or up-sell. The
technique also evaluates each generated association rule vis--vis cost of sales
conversion, to generate only those rules which would be beneficial to the service
provider, in terms of profit generation. This would eventually provide the marketer with a
set of highly profitable VAS bundles for selling which would have a higher adoption rate.
2.2 Product/services bundling
Venkatesh and Mahajan (1993) classified the major rationale for bundling under
price segmentation, price discrimination, product range restriction, reduction in
classification/processing costs, scope economies, consumers search economies, and risk
reduction. Carlton and Waldman (2002) investigated how the tying of complementary
products can be used to preserve and create monopoly positions. Nalebuff (2004)
proposed the use of bundling as an entry deterrent strategy adopted by firms and provide









52 A.K. Kar and P. Singh












evidence for this in an oligopolistic environment. Thus, a VAS bundling model can
provide the telecom service provider with immense benefits in terms of increased
adoption of VAS and higher sales returns in the short run, and a sustainable competitive
advantage in the long run.
Offering two goods both separately and as a bundle (at a price other than the sum of
the components prices) is known as bundling (Salinger, 1995; Venkatesh and Mahajan,
1993). Stremersch and Tellis (2002) classify various bundling strategies based on the
focus of bundling, whether on price or product, and based on the form of bundling,
whether pure or mixed. Price bundling is the sale of two or more separate products at a
discount, without any integration of the products. Product bundling is the integration and
sale of two or more separate products at any price. Pure bundling is the practice of
offering two (or more) goods only in bundled form. Chung and Rao (2003) stated that the
two distinct approaches adopted in the marketing literature to bundle choices
can be classified according to the type of units for analysis: component-based and
attribute-based approaches (as given in Figure 3). Component-based approaches refer to
bundle choice models, which employ products (components) of a bundle as the ultimate
unit of analysis in describing the utility of a bundle. In this stream, complementarity
among the components is described directly in terms of the products included in the
bundle. In contrast, the attribute-based approaches develop a model for bundle utility to
specify the complementarity in terms of the attributes of the products (components).
Figure 3 Approaches to bundle services

Source: Chung and Rao (2003)









A model for bundling mobile value added services using neural networks 53













Bundling articles commonly assume that a consumers reservation price for the bundle is
equal to the sum of his or her separate reservation prices for the component products
(Adams and Yellen, 1976). This is called the assumption of strict additivity and products
having these assumptions are independently valued products. When products are
complements, a consumers reservation price for the bundle is super-additive in those for
the component products. Alternatively when the products are substitutes, a consumers
reservation price for the bundle will be sub-additive in those for the components
(Venkatesh and Kamakura, 2003).
In this study, the proposed model focuses predominantly on providing a product
bundling decision support model using pattern association where the association rules are
generated dynamically using artificial neural networks. A probabilistic approach of VAS
adoption has been used to evaluate the probable profitability of VAS bundles, by the
model, to generate only profitable association rules. The profitability of the proposed
VAS bundles have been calculated based on the product of VAS price vis--vis
possibility of conversion.
2.3 Models for bundling choices
Prior marketing literature on bundling has examined the optimality of bundling,
consumer evaluation of bundles, and firms pricing and promoting of bundles.
Economists have focused mainly on the optimality of bundling for monopolists,
equilibrium theories of bundling and the welfare implications of bundling (Stremersch
and Tellis, 2002).
Most often, component prices are observed and researchers ask under what conditions
might bundling be profitable (e.g., Adams and Yellen, 1976; Salinger, 1995). In pure
bundling case, Salinger (1995) found that when bundling lowers costs, it tends to be more
profitable when demands for the components are highly positively correlated and
component costs are high and established that as marginal costs rise, bundling becomes
less attractive. Bundling creates inefficiency in that some consumers are forced to buy
the bundle even though they value one of its components at below production cost
(Adams and Yellen, 1976). Thus, it is not surprising that software is especially conducive
to bundling since this countervailing force does not arise. The potential for very large
scale bundling of information goods is demonstrated in Bakos and Brynjolfsson (1999)
who consider the limiting case of bundling together an infinite number of information
goods. They found that bundling can create economies of aggregation for information
goods if their marginal costs are very low, even in the absence of network externalities or
economies of scale or scope. However, their results do not extend to physical goods as
the marginal cost for production of goods not used by the buyer typically negate any
value from the predictive value of large-scale bundling. However, bundling may also
contain situations wherein the bundle price is higher than the sum of components.
DeGraba and Mohammed (1999) studied the inter-temporal mixed bundling in the case of
rock concert tickets. In this case the high-valuation customers buy a bundle because they
expect quantity rationing when units are sold individually even at an excess cost. This
can be of paramount importance in case of services bundling, like that of VAS, where the
consumers perception of value of the service can be escalated drastically by offering
suitable bundles.









54 A.K. Kar and P. Singh












Adams and Yellen (1976) demonstrated that the profitability of commodity bundling
can also stem from its ability to sort customers into groups with different reservation
price characteristics, and hence to extract consumer surplus. Mackay and Weaver (1983)
applied bundling to public sector and developed an analytical model of a political market
in which a multi-activity governmental unit practices commodity bundling in order to
advance the interests of the setter. Based on two scenarios of voting process based on
bundling, they explored the impact of form of monopoly power on policy outcomes as
the governmental units total budget, its budgetary mix, and the distribution of net
benefits from collective action. Guiltinan (1987) developed a framework to provide
some guidance for the potential use of price bundling by a firm with a broad line of
complementary products. Venkatesh and Mahajan (1993) proposed a probabilistic
approach to optimally price a bundle of products or services that maximises sellers
profits. Their focus is on situations in which consumer decision making is on the basis of
multiple criteria. After comparing the pure and mixed strategy, their results indicate that a
mixed bundling strategy is more profitable than pure components or pure bundling
strategies provided the relative prices of the bundle and components are carefully chosen.
Chen (1997) demonstrated how bundling is an equilibrium strategy of competing firms
for its role as a product-differentiation device. They also found that mixed bundling,
while dominating pure bundling under monopoly, becomes a dominated strategy in the
presence of competing firms. When it is possible for firms to commit to (pure) bundling,
the profits of all firms in the industry are higher, but social surplus is unambiguously
reduced. Koschat and Putsis (2002) conducted a hedonic analysis of magazine advertising
rates under all the three bundling strategies (pure component, pure and mixed). They
found that the magnitude of the price and revenue premiums earned by unbundling a
magazines readers can be substantial. Thatcher and Clemons (2000) proposed pure
bundling strategy in insurance market to provide better access to maximum people at
equitable premiums. Chung and Rao (2003) developed a general random utility model for
bundle choices regardless of the type of bundles. Their model can be used for used to find
markets segments for bundles with heterogeneous products in multiple product
categories, to estimate individual reservation prices or bundles, and to determine the
optimal bundle prices for different market segments.
Apart from the economic perspective, bundling has also been studied from buyers
perspective. Yadav and Monroe (1993) found that additional savings offered directly on
the bundle have a greater relative impact on buyers perceptions of transaction value than
savings offered on the bundles individual items. The effect of each saving is also
influenced by the magnitude of the other saving. Jedidi et al. (2003) studied that the
product-line pricing policy depends on the degree of heterogeneity in the reservation
prices of the individual products and the bundle. A uniformly high-price strategy for all
products and bundles is optimal when heterogeneity is high. Otherwise, a hybrid strategy
is optimal. Venkatesh and Kamakura (2003) found that bundles with complements and
substitutes should typically be priced higher than independently valued products. Dewan
and Freimer (2003) applied bundling for software applications and demonstrated that the
price of the bundle will be less than the sum of the prices of the base and add-in software
when they are sold separately. They also argued that the total consumer surplus and the
social welfare increase if the base software producers profit increases with bundling.
Goh et al. (2003) did a simulation study and found that agents have the highest preference
for complementary product bundles with a good, certain quality focal item than for other
types of product bundles. Bhargava and Feng (2004) compared the two strategies, pure









A model for bundling mobile value added services using neural networks 55













(proprietary connection software) and mixed (access only and connection software for
different consumer segments). Based on economic modelling they propose that pure
strategy is optimal as power users (one segment of mixed strategy) impose exceptionally
high cost on system and may outweigh the benefits. Also with learning curve effects
novice users that use connection software might shift to access only segment and thus
increase costs for the service provider.
In this study, dynamically generated pattern association rules have been proposed for
the usage of bundling of VAS for the telecom service users.
2.4 Pattern association
From the early 1970s, pattern association was primarily studied in the context of market
research model development. Perhaps the most common precursors to the exploration of
the associative relationship between two variables in marketing involve the use of
bivariate cross-tabulations or multivariate analysis. Several alternative measures are
available for assessing the extent of association in a contingency table.
Current pattern association studies in data mining started developing from 1993.
Agrawal et al. (1993), Holsheimer et al. (1995), and Houtsma and Swami (1995) studied
association in mining approaches and are highly cited works. The Apriori algorithm
(Mannila et al., 1994; Agrawal and Srikant, 1994) was shown to have superior
performance to previous methodologies and is one of the most referred (cited) algorithm.
All these previous studies focused on static data only. There had been few studies in
the field of neural continuum models in a 1-d or a 2-d cortical domain in associative
memory which were the precursor to studies in associative memory research of more
recent times and focused more on pattern generation, pattern retrieval and visual
invariance and not on the generation of association rules. Hopfield (1982) and Hopfield
and Tank (1985) proposed methodologies for neural networks to express collective
computational capabilities.
Cohen and Grossberg (1983) studied pattern association and focused their efforts on
the process whereby input patterns are transformed and stored by competitive cellular
networks, besides focusing on the stability of the pattern rules and their parallel memory
storage in competitive neural networks. Carpenter and Grossberg (1985) proposed the
adaptive resonance architectures (ART) for neural networks that self-organise stable
pattern recognition codes in real-time in response to arbitrary sequences of input patterns,
for adaptive pattern recognition. Carpenter and Grossberg (1987a) proposed a parallel
architecture with neural networks for pattern association where they presented a
methodology to form invariant pattern recognition rules with ART architecture for a
dynamic data. Subsequently, Carpenter and Grossberg (1987b) proposed the ART 2
architecture, a class of adaptive resonance architectures which rapidly self-organise
pattern recognition categories in response to arbitrary sequences of either analog or
binary input patterns. Aumann et al. (1999) presented a new algorithm BORDERS, with
three variants, for generating associations in dynamic databases. Very little work has
been done in the field of incremental and dynamic databases using neural networks.
While in current times, mining of databases needs to be done not only real time, but also
such mining needs to analyse the latest trends in data, the need for techniques to analyse
dynamic databases is strongly felt.









56 A.K. Kar and P. Singh












A particularly well-studied problem in data mining is the search for association rules
in market basket data and the studies go as early as the early 1990s (Agrawal and Srikant,
1994). Association rules are primarily intended to identify rules of the type, customer
purchasing service A is likely to also purchase service B. This behaviour has been used
in this paper to propose real time VAS bundles. Most of this research in data mining from
large databases has focused on the discovery of Boolean association rules (a => b).
What has not been extensively studied is the formation of association rules in market
basket data obtained from telecom settings, during the purchase of VAS by the telecom
customers. In this paper, previously proposed algorithms have been suitably modified and
adapted for the generation of pattern association rules which can be further developed
into economically profitable VAS bundles so as to enable the marketer invest suitably to
obtain better results for his cross-selling efforts.
3 Research gap and contribution
While the need of bundling VAS has been established in the prior survey of existing
literature, no previous significant study has attempted to address the gap of building a
decision support model, which can dynamically generate bundles of VAS based on past
consumer purchase behaviour. This paper presents a model which not only generates
bundles of VAS based on pattern association, the neural network based model also
addresses the latest trends in consumer purchase behaviour. The model also has the
capability to evaluate which bundles of VAS would be profitable for the firm based on a
probabilistic approach, and screens association rules so as to provide the most profitable
VAS bundles which also would have a higher selling possibility. This would ensure that
the telecom service provider would benefit substantially from the cross-sell and up-sell of
VAS.
4 Mathematical proposition
So based on previous research done in the field of pattern association, a methodology is
being suggested for the generation of pattern association rules for the purchase of VAS
from a telecom service provider by the customer. For the generation of association rules,
each already purchased VAS or to be purchased VAS, be it a ring-tone or a call-alert, for
example, would be considered to be a unit service. Each service purchase or transaction is
to be treated as an individual vector in this system using the neural network mesh and
would have a unique serial number.
Thus say an individual purchases a VAS be of a serial number ABC-123. Say,
purchaser of that particular VAS often purchases another VAS with a serial number
XYZ-123. So, the purchase of the item ABC-123 can be associated with the purchase of
the accessory XYZ-123. In this paper, both purchases would be represented as separate
transactions and each transaction would be encoded as a vector. Let x be the vector for
the purchase of the original item, ABC-123, in this case. Then t would be the vector for
the purchase of the associative product, XYZ-123, in this case.
Thus, each associative purchase pair can be represented as an input-output vector
pair, say x-t. If each vector t is the same as the vector x with which it is associated, then
the net is called an auto-associative memory. If the ts are different from the xs, the net









A model for bundling mobile value added services using neural networks 57













is called a hetero-associative memory. In both types, the neural network configuration not
only learns from the specific pattern pairs that were used for training, but also is able to
recall the desired response pattern when given an input stimulus that is similar, but not
identical, to the training input. This is the part that gives it the pattern recognition
capability.
For training the system for pattern association, we use the extended form of the Delta
rule, as proposed by Widrow and Hoff (1960), for pattern association since this rule may
be used for input patterns that are linearly independent but not orthogonal. We denote our
training vector pairs as x-t and then denote our testing input vector as x. Let be the
learning rate, x be the training input vector and t be the target output for input vector x.
The original Delta rule assumes that the activation function for the output units is the
identity function; or, equivalently, it minimises the square of the difference between the
net input to the output units and the target values. An analogy can be drawn from this as
the minimisation of RMS error. That the delta rule will produce the least squares solution
when input patterns are not linearly independent was established by Rumelhart et al.
(1986). Using this process, we are able to compute the initial output node values, being
referred here as y.
Using y as the computed output for the input vector x, y is calculated as follows:
( ) j i ij
y x *w for all i =

The weight updates are: w
ij
(new) = w
ij
(old) + (t
j
y
j
)x
i
for all (i = I,,n; j = I,,m)
Or we can denote the same in terms of weight change as w
ij
= (t
j
y
j
)x
i
.

Now, we use the extended Delta rule, which allows for an arbitrary, differentiable
activation function to be applied to the output units. The update for the weight from the
I
th
input unit to the J
th
output unit is as follows:
( ) ( )
lJ J J I J
W t y x f y_in . =
Now, a hetero-associative memory neural network is being used for the next step.
Associative memory neural networks are nets in which the weights are determined in
such a way that the net can store a set of P pattern associations. Each association is a pair
of vectors (x(p), t(p)), with p = 1, 2,,P. Each vector x(p) is an n-tuple (has n
components), and each t(p) is an m-tuple. The weights of the network are determined by
the extended Delta rule as mentioned earlier in the paper.
The algorithm to find out the association between the vectors is as follows:
Step 1 For each input vector X
i
, do Steps 2 to 4.
Step 2 Set activations for input layer units equal to the current input vector X
i
.
Step 3 Compute net input to the output units:
i ij
y_inj X W . =

Step 4 Determine the activation of the output units:
j j
Y I if y_in 0 = >
j j
Y 0 if y_in 0. =









58 A.K. Kar and P. Singh












Here, X
i
denotes the inputs to the neural network mesh. Y
i
denotes the output of the
neural network mesh, and would provide the final association rules, that would indicate
whether association exists between the products. W
ij
denotes the weights of the
interconnecting network branches of the neural network mesh, and would be initialised in
the beginning with the extended Delta rule, as indicated earlier. The network is denoted
along with its symbols in the following diagram.
Figure 4 The artificial neural network configuration

A threshold function has been used here to generate the final association rules by the
proposed methodology. The utility of the threshold function would be to minimise the
implications of cross-talk, unless the same is not too high.
The output unit Y
j
denotes a positive association if Y
j
= 1, which would indicate, for
the concerned pair of vectors, X
i
and Y
j
. This means that if transaction X
i
happens, Y
i
can
be inferred to happen by the rules of association. Thus in such a scenario, a purchase of
the VAS corresponding to X
i
would infer a strong probability of purchase of the VAS
corresponding to Y
i
, following the rules of association. The output unit Y
j
denotes a
negative association if Y
j
= 0, which would indicate, for the concerned pair of vectors, X
i

and Y
j
. This means that if transaction X
i
happens, the probability of the transaction Y
j

happening can be inferred to be low, or if VAS corresponding to X
i
is purchased, the
probability that the buyer would also be interested to purchase the VAS corresponding to
Y
j
is rather low, following similar rules of association.
Now, the rules generated needs to be further tested before actual implementation to
see whether they create economic profit for the firm where economic profit is defined as
the increase in wealth that the seller has from making a transaction, taking into
consideration the revenue generated from the transaction and all the costs associated with
that transaction. For the same, the strength of each generated rule would need to be
tested on a continual basis. For doing so, the ratio of the transactions that lead to a
success of the association rule generated, to the total number of transactions involving the
first product which may or may not have led to the purchase of the secondary product (as
predicted by the association rule) needs to be calculated. This would be used to determine
which rules would finally be used on a regular basis, and which would need to be
rejected, based on the profit that may be generated by the association rule to the business.









A model for bundling mobile value added services using neural networks 59













Now, in this study, the economical value of each generated association rule is
evaluated using the mentioned algorithm.
Step 1 Do Step 2 to Step 4 until all association rules (X
i
Y
j
) have been evaluated.
Step 2 Calculate the profit that may be generated on successful completion of the
transaction, i.e., profit = Y
j-Val
* (X
i
Y
j
) C(X
i
Y
j
).
Step 3 If (Y
j-Val
* (X
i
Y
j
) C
a
) > 0, keep the association rule.
Step 4 If (Y
j-Val
* (X
i
Y
j
) C
a
) 0, reject the association rule.
Here, in this algorithm, every association rule is referred as (X
i
Y
j
). Y
j-Val
is the revenue
generated on completion of the transaction Y
j
. The strength of the association rule
X
i
Y
j
) is depicted as (X
i
Y
j
). C
a
is the cost of converting a customer through
targeted advertising into buying incurring transaction Y
j
and in this scenario, would
involve the cost of advertising. Finally only those rules are finally kept which propose
economical value to the business.
Now, on the actual purchase of product X
i
, a profit maximisation algorithm is being
proposed where the objective function represents the goal of a profit maximisation
problem and therefore reflects the microeconomic framework of the marketer.
Step 1 For transaction Y
j
, do steps 2 to 3 for all values of j.
Step 2 For each Y
j
, calculate CY
j
.
Step 3 For all Y
j
, find maximum (Y
j-val
Cy
j
) * (X
i
Y
j
).
Step 4 Choose Y
j
for which (Y
j-val
Cy
j
) * (X
i
Y
j
) has the maximum value.
Here, CY
j
represents the cost of sales of Y
j
which would include the sum of the cost of
the VAS as is incurred by the telecom company for the transaction. Y
j-val
represents the
monetary value of the transaction Y
j
. The strength of the association rule (X
i
Y
j
) is
again depicted as (X
i
Y
j
).
With this algorithm, the associated rule which will generate the maximum expected
profit for the marketer for the transaction Y
j
after X
i
has taken place, will be generated
eventually, and thus will address the profit maximisation objective of the telecom firm.
5 Model testing methodology
The model was tested with hypothetical data generated after a series of interviews with
eight telemarketing executives of a telecom service provider. The interview identified
five main VAS which was classified internally as follows:
1 P2P: person to person SMS are the most common form of mobile communication
apart from voice-based.
2 Ringtones: this is inclusive of mono-tunes, poly-tunes, true-tunes and caller ring
back tones. From the interview, it was understood that the purchase of caller back
ring tones have grown in popularity the maximum in the last two years.
3 P2A and A2P: person to application (P2A) SMS includes messages sent by end users
for contests and for seeking other information like news and updates; while









60 A.K. Kar and P. Singh












application to person (A2P) SMS are inclusive of service push by enterprise service
providers. These also include calls on IVRS for all other services like astrology,
jokes and business news.
4 Games and data: games include download of one play games offered by the telecom
service provider. Data includes downloads of wallpapers and screen savers.
5 Others: include multi media messages (MMS) and subscription charges for WAP and
3G services.
Based on the interviews, the VAS class of ringtones consisting of mono-tunes,
poly-tunes, true-tunes and caller ring back tones was chosen to test out the model since
they had one of the highest frequencies of adoption across customer segments. For all the
four product types, data was simulated based on the interview with the telemarketing
executives to understand the trends in consumer behaviour. The dataset had five types of
products in each of the mono-tunes, poly-tunes, true-tunes and caller ring back tones
VAS category. The dataset also included an assumed cost of the VAS delivery, and
probability of cross-sale for each product based on the interviews. After the dataset was
completely developed, the same was reviewed with the tele-marketing executives for
their representativeness. This dataset with hypothetical 500 transactions was used to test
the model.
6 Results
The model was coded and tested with the dataset consisting of the 500 transactions as has
been described in the earlier section. Only 200 data points were used to train the model.
A total of 24 association rules were generated during training. After training the
programme on the hypothetical dataset, in the testing phase it was found that the
algorithm successfully identified nine out of the 24 association rules (bundles) which
were having a positive monetary benefit to the service provider, and thus revealed that
not only 37.5% the possible VAS bundles are profitable, but also marketing the other
rules, even if some conversion was done in the sales, would not create more value for the
company in the long run based on probability of conversion of such sale of bundles. The
bundles identified as being profitable are also those which have a higher adoption factor
based on past consumer purchase behaviour. Thus, there was a definite identification of
viable VAS bundles which can be sold more profitably to the end consumer. Considering
such a limited dataset was used for training the model, the results of the model seemed
interesting and would add a lot of value to the telecom service provider. However, the
current model needs to be tested on a larger scale on actual data of customer buying
behaviour.
7 Limitations and concluding discussion
A major limitation of the proposed model is that it needs to be tested on real life telecom
data. At its current state, it has been tested only on simulated representative data with
satisfactory results. In the present context where MVAS have been proposed as the new
strategy for maintaining growth, service bundles have been proposed as a viable strategy









A model for bundling mobile value added services using neural networks 61













to increase the revenue. However, there is scant research to propose methods to
effectively design such bundles. Another issue of scrutiny would be the scalability of the
mentioned model to evaluate consumer data when millions of transactions would be
evaluated in a proper business setting. However, given the self learning capability of the
model, adaptation in such a situation should not be problematic.
By utilising this model, telecom service providers would be able to provide VAS
bundles such that there can be better cross-selling and up-selling of VAS. This paper
proposes a model for identifying the unique service preferences of customers based on
past consumer behaviour. With this neural network approach using association rules, the
telecom company will be able to predict after a customer makes a transaction, which
purchase would he be interested in, next, and whether attempting to convert such a
purchase would create economic value for the business, and thus create unique bundled
VAS which may provide greater value proposition for both the customer and the telecom
service provider.
The current study also proposes a very effective way for service providers to club the
standard services with the individually preferred services and thus provide unique
personalised services for each customer. This kind of mixed bundling strategy can be
useful in creating specific consumer segments and thereby utilise the price differentiation
strategy for increasing profits. Also, when the churn rate is so high in the telecom sector,
this would help to increase customer loyalty. Also, this would be increasing revenue
generation from the existing customer-base and thus create cost savings from the high
costs of acquisition of new customers. However, since association rules are used, this
proposed method has the limitation of predictability only if historical consumer
behaviour patterns continue in future although the model can dynamically update the
changing consumer preferences. Further, we have not dealt with the pricing issue for
service bundles. This can be further incorporated in the proposed model to develop a
complete bundling strategy in future research.
Acknowledgements
We gratefully acknowledge the guidance and the feedback of the editor and the reviewers
for the constructive and helpful comments which has enabled us to bring the study to its
current state.
References
Adams, W. and Yellen, J. (1976) Commodity bundling and the burden of monopoly, Quarterly
Journal of Economics, Vol. 90, No. 3, p.475.
Agrawal, R. and Srikant, R. (1994) Fast algorithms for mining association rules, 20th Intl. Conf.
on Very Large Data Bases, Santiago, Chile.
Agrawal, R., Imielinski, T. and Swami, A. (1993) Mining association rules between sets of items
in large databases, ACM, SIGMOD Intl. Conf. Management of Data, Washington.
Amiri, A.N. and Kian, M.T. (2008) Increasing mobile operator revenue out of VAS, a system
dynamics approach, Proceedings of the 26th International Conference of the System
Dynamics Society, Greece.
Anckar, B. and DIncau, D. (2002) Value creation in mobile commerce: findings from a consumer
survey, Journal of Information Technology Theory and Application, Vol. 4, No. 1, pp.4364.









62 A.K. Kar and P. Singh












Aumann, Y., Feldman, R. and Lipshtat, O. (1999) Borders: an efficient algorithm for association
generation in dynamic databases, Journal of Intelligent information Systems, Vol. 12, No. 1,
pp.6173.
Bakos, Y. and Brynjolfsson, E. (1999) Bundling information goods: pricing, profits, and
efficiency, Management Science, Vol. 45, No. 12, pp.16131630.
Balakrishnan, V. (2010) Mobile value added services in India: filling the VAS vacuum to drive
high performance, Accenture Report.
Bhargava, H.K. and Feng, J. (2004) Pure bundling better than mixed? Or, why doesnt AOL offer
standardized dial-up service, Proceedings of the 37th Hawaii International Conference on
System Sciences, Hawaii, USA.
Bouwman, H., Faber, E. and Haaker, T. (2005) Designing business models for mobile service
bundles, Paper presented to the Hong Kong Mobility Roundtable, Hong Kong, China.
Carlsson, C. and Walden, P. (2002) Mobile commerce: a summary of quests for value-added
products and services, Proceedings of BLED Electronic Commerce Conference, pp.463476.
Carlton, D.W. and Waldman, M. (2002) The strategic use of tying to preserve and create market
power in evolving industries, RAND Journal of Economics, Vol. 33, No. 2, pp.194220.
Carpenter, G.A. and Grossberg, S. (1985) Category learning and adaptive pattern recognition: a
neural network model, Proceedings, 3rd Army Conference on Applied Mathematics and
Computing, ARO Report, Vol. 86, No. 1, pp.3756.
Carpenter, G.A. and Grossberg, S. (1987a) A massively parallel architecture for a self-organizing
neural pattern recognition machine, Computing and Vision Graphics Image Processing,
Vol. 37, No. 1, pp.54115.
Carpenter, G.A. and Grossberg, S. (1987b) ART 2: self organization of stable category recognition
codes for analog input patterns, Applied Optics, Vol. 26, No. 23, pp.49194930.
Chen, Y. (1997) Equilibrium product bundling, Journal of Business, Vol. 70, No. 1, pp.85103.
Chung, J. and Rao, V.R. (2003) A general choice model for bundles with multiple-category
products: application to market segmentation and optimal pricing for bundles, Journal of
Marketing Research, Vol. 40, No. 2, p.115.
Cohen, M. and Grossberg, S. (1983) Absolute stability of global pattern formation and parallel
memory storage by competitive neural networks, IEEE Transactions on Systems and
Cybernetics, SMC, Vol. 13, No. 15, pp.815826.
DeGraba, P. and Mohammed, R. (1999) Intertemporal mixed bundling and buying frenzies,
RAND Journal of Economics, Vol. 30, No. 4, pp.694718.
Dewan, R.M. and Freimer, M.L. (2003) Consumers prefer bundled add-ins, Journal of
Management Information Systems, Vol. 20, No. 2, pp.99111.
ESI Report (2010) The Indian Telecom Industry.
Goh, K.Y., Lee, C.S. and Lee, C.H. (2003) Information technology product bundling in the
presence of complementarities, quality uncertainty, and network effects: an agent-based
approach, Proceedings of the 24th International Conference on Information Systems, Seattle,
WA.
Guiltinan, J.P. (1987) The price bundling of services: a normative framework, Journal of
Marketing, Vol. 51, No. 2, pp.7485.
Haaker, T. and Voss, D. (2007) Customer preferences for bundled content services, Proceedings
of the 18th European Regional ITS Conference, Turkey.
Heinonen, K. and Pura, M. (2006) Classifying mobile services, Presented at Helsinki Mobility
Roundtable, Helsinki.
Holsheimer, M., Kersten, M., Mannila, H. and Toivonen, H. (1995) A perspective on databases
and data mining, 1st Intl. Conf. Knowledge Discovery and Data Mining.
Hopfield, J.J. (1982) Neural networks and physical systems with emergent collective
computational abilities, Proceedings, National Academy of Science, USA, Vol. 79, No. 8,
pp.25542558.









A model for bundling mobile value added services using neural networks 63













Hopfield, J.J. and Tank, D.W. (1985) Neural computation of decisions in optimization problems,
Biol. Cybernetics, Vol. 52, No. 3, pp.141152.
Houtsma, M. and Swami, A. (1995) Set-oriented mining of association rules in relational
databases, Proceedings, 11th Intl. Conf. Data Engineering.
Jedidi, K., Jagpal, S. and Manchanda, P. (2003) Measuring heterogeneous reservation prices for
product bundles, Marketing Science, Vol. 22, No. 1, p.107.
Koschat, M.A. and Putsis, W.P. (2002) Audience characteristics and bundling: a hedonic analysis
of magazine advertising rates, Journal of Marketing Research, Vol. 39, No. 2, p.262.
Li, Q., Wang, C. and Geng, G. (2008) Improving personalized services in mobile commerce by a
novel multicriteria rating approach, Proceeding of the 17th international conference on World
Wide Web, pp.2125.
Mackay, R.J. and Weaver, C.L. (1983) Commodity bundling and agenda control in the public
sector, Quarterly Journal of Economics, Vol. 98, No. 4, p.611.
Mannila, H., Toivonen, H. and Verkamo, I. (1994) Efficient algorithms for discovering association
rules, AAAI Workshop, Knowledge Discovery in Databases, pp.181192.
May, P. (2001) Mobile Commerce: Opportunities, Applications, and Technologies of Wireless
Business, Cambridge University Press, UK.
Nalebuff, B. (2004) Bundling as an entry barrier, Quarterly Journal of Economics, Vol. 119,
No. 1, pp.159187.
Pau, L. and Dits, J. (2002) Business Modeling Framework for Personalisation in Mobile Business
Services, Springer, Berlin.
Romdhane, L.B., Fadhel, N. and Ayeb, B. (2010) An efficient approach for building customer
profiles from business data, Expert Systems with Applications, Vol. 37, No. 2, pp.15731585.
Rumelhart, D.E., Mcclelland, J.L. and the PDP Research Group (1986) Parallel Distributed
Processing, Explorations in the Microstructure of Cognition, MIT Press, Cambrige, Mass.
Salinger, M.A. (1995) A graphical analysis of bundling, Journal of Business, Vol. 68, No. 1,
pp.8598.
Siau, K., Lim, Ee-P. and Shen, Z. (2001) Mobile commerce: promises, challenges, and research
agenda, Journal of Database Management, Vol. 12, No. 3, pp.413.
Stremersch, S. and Tellis, G.J. (2002) Strategic bundling of products and prices: a new synthesis
for marketing, Journal of Marketing, Vol. 66, No. 1, pp.5572.
Thatcher, M.E. and Clemons, E.K. (2000) Managing the costs of informational privacy: pure
bundling as a strategy in the individual health insurance market, Journal of Management
Information Systems, Vol. 17, No. 2, pp.2957.
Venkatesh, R. and Mahajan, V. (1993) A probabilistic approach to pricing a bundle of products or
services, Journal of Marketing Research, Vol. 30, No. 4, pp.494508.
Venkatesh, R. and Kamakura, W. (2003) Optimal bundling and pricing under a monopoly:
contrasting complements and substitutes from independently valued products, Journal of
Business, Vol. 76, No. 2, pp.211232.
Verkasalo, K. (2008) Contextual patterns in mobile service usage, Personal and Ubiquitous
Computing, Vol. 13, No. 5, pp.331342.
Widrow, B. and Hoff, M.E. (1960) Adaptive switching circuits, IRE WESCON Convention
Record, pp.96104.
Yadav, M.S. and Monroe, K.B. (1993) How buyers perceive savings in a bundle price: an
examination of a bundles transaction value, Journal of Marketing Research, Vol. 30, No. 3,
pp.350358.
Yang, K. and Laura, J. (2006) Value-added mobile data services: the antecedent effects of
consumer value on using mobile data services, International Journal of Mobile Marketing,
Vol. 1, No. 2, pp.1117.