=Paper=
{{Paper
|id=None
|storemode=property
|title=Using Bayesian Belief Networks for Modeling of Communication Service Provider Businesses
|pdfUrl=https://ceur-ws.org/Vol-818/paper8.pdf
|volume=Vol-818
}}
==Using Bayesian Belief Networks for Modeling of Communication Service Provider Businesses==
https://ceur-ws.org/Vol-818/paper8.pdf
Using Bayesian Belief Networks for Modeling of Communication
Service Provider Businesses
Pekka Kekolahti
Department of Communications and Networking
Aalto University
Espoo, Finland
Abstract factors. These include new business players like Google
and Facebook, technologies like the Internet, cloud
computing and smart-phones, as well as a growing
This paper analyses the usage of Bayesian Belief number and size of applications. It is clear that the CSP
Networks (BBNs) for Communication Service value chain structure has to be re-evaluated. To respond to
Provider (CSP) business modeling and these changes and customer requirements, and to adapt
simulation. Large and complex BBNs have been successfully to new business challenges, CSP top
created to describe the causal relationships in management needs reliable methods to model and to
CSP business domains. As a part of the study, a analyze the essential factors driving the change, and to
novel method to collect knowledge from a large understand the impact of these factors on their current and
number of independent experts living in different future business. In addition, trusted and unified
countries has been introduced. A BBN from each information is needed for strategy planning processes and
expert result was created (referred to here as a day-to-day management. Today the strategic decisions are
sub-BBN). Business model ontology was often made by a small group and they are based on
utilized to combine sub-BBNs together into a insufficient knowledge due to lack of data or expert
comprehensive model. The resulting BBN knowledge and under time constraints.
represents typical business circumstances in the
European telecommunications domain. The Bayesian Belief Networks (BBNs), also called belief
experts participating in the study represented networks, Bayes Nets and causal probabilistic networks
expertise in different business related categories are increasingly popular methods for modeling uncertain
such as technology, processes, customer
and complex domains (Uusitalo, 2007). In this paper we
experience, regulation, organization and
examine how the BBN methodology can be utilized to
products. Experts were asked to list causality
triplets for business categories including causal help CSP management in their day-to-day work, strategy
connection strengths, in order to assess the belief planning and to better control the business.
part as well. The triplets were manually
converted to a graphical causal map and A BBN is a probabilistic model which represents a set of
conditional probability tables constructed. The random variables and their conditional dependencies via a
benefit of the method is the capability to directed acyclic graph. Two basic approaches are used to
introduce rapidly a high number of variables and construct Bayes networks: data-based and knowledge-
causal relationships. A challenge is that experts based approaches. Data-based methods use conditional
use different terms with the same underlying independence semantics of Bayes networks to infer
meaning. models from data whereas the knowledge-based approach
utilizes causal knowledge from domain experts to
construct BBNs. The benefits of BBNs in data analysis
1. INTRODUCTION are, according to Nadkarni, 2004; Uusitalo, 2007; Jensen,
2001; Lee, 2009:
The Communication Service Provider (CSP) business is 1) Possibility to combine prior knowledge and data,
facing major restructuring due to several disruptive 2) Managing situations where some data is missing,
3) Modeling of causal relationships, and validated with experts through subsequent interviews
4) Structural learning possibilities, and finally the probability assessment done either
5) Support for different kind of analyses, such as manually or by using noisy-OR method or weighted sum
making inferences about probabilities of algorithm utilizing compatible parent configurations
different causes given the consequences and (DAS, 2004) to reduce the number of probability
6) Fast response to queries from the model. assessments.
Known challenges in BBNs are This study focuses on BBNs as a methodology for
1) Difficulty to obtain prior knowledge in a form modeling and analysis of CSP business. As part of the
that can be converted into probability study, both multiple sub-BBNs (one per expert) and a
distributions. However, for example a weighted comprehensive CSP BBN combining sub-BBNs have
sum algorithm utilizing compatible parent been created. The experts were asked to list and
configurations has been developed to ease the categorize the variables they considered to have an effect
calculation of conditional probability tables in on CSP business and also how strong this effect would be.
complex environments (Das 2004). The used seven categories are the same as in typical
2) Handling of continuous variables only in a Balanced Scorecards and business models (Kasperskaya,
limited manner (Uusitalo 2007) and 2006; Osterwalder, 2002 and 2005; Faber, 2003) namely
3) Lack of support of feedback loops due to acyclic financial variables, customer-related variables, product
nature of a BBN. Feedback loops are useful and service innovations, staff and internal processes,
when analyzing phenomena like new disruptive technology and architecture, strategy and competition,
CSP technologies as a function of time (Casey et local and global economy and legislation.
al. 2010).
The following types of information can be derived from
According to our knowledge, BBNs have in the past not the comprehensive model and sub models:
been used to model the CSP industry in a large scale. The Financial variables: Effect of variables like
utilization of causality itself is wide spread in business customer experience on revenue, OPEX
management due to widely used performance measuring (operating expense) and CAPEX (capital
and management tools such as the Balanced Scorecard expenditure).
(BSC) and Tableau de Board. 66% of enterprises used Customers: The causes and consequences related
BSC in 2007 (Rigby, 2007). Both the BSC and the to customer satisfaction.
Tableau de Board rely on causal assumptions R&D organization: How do organization agility,
(Kasperskaya, 2006). Causal mapping tools like fishbone managerial structures, salary and incentives
diagrams, cause-and effect diagrams, impact wheels, issue affect on efficiency, productivity, OPEX and
trees, strategy maps, and risk-assessment mapping are customer experience.
tools to help managers to understand and improve Technologies: How do new technologies like
complex systems in the areas of quality, strategy, and rapid growth of smart-phones affect on CAPEX,
information systems. (Scavarda et al., 2006). The revenue and data traffic.
causalities in the performance measuring and strategy
creation have been normally deduced by using human BBNs that include all the seven categories are very
interaction techniques such as brainstorming or complex. The number of variables and arcs, and
interviews. These methods rely on person-to-person or especially the size of conditional probability tables play
group interaction in eliciting the knowledge and are great effect on the practical usability of the BBN for CSB
fraught with biases associated with inter-person business analysis purposes. Optimization between
dynamics. Methods to elicit a non-biased knowledge in practical usability and model granularity and accuracy is
large scale have been developed (Nadkarni et al., 2004; examined through creating the comprehensive BBN from
Scavarda et al., 2006). Scavarda introduces a formal sub-BBNs.
Collective Causal Mapping Methodology (CCMM),
which collects information asynchronously from an expert The remainder of this article is structured as follows:
group which is dispersed and diverse. Person-to-person
interaction possibility is eliminated and a large amount of Chapter 2 introduces a novel method for the collection of
experts can be utilized in a controlled way. Nadkarni the expert knowledge, and describes how the expert
introduced a procedure for constructing BBNs from knowledge is converted into BBNs.
domain knowledge experts, where through four steps of a
text analysis process the first round interview results can Chapter 3 describes the constructed sub-BBNs and
be converted into causal relationships. Once the causal comprehensive BBN and elaborates on key variables and
map is available, the states of the variables can be defined their analysis states. Also some result examples are given.
Chapter 4 discusses challenges in eliciting and conversion Table 1: Part of given example triplets.
of prior knowledge into BBN and how well these models
truly represent different aspects of CSP businesses. Also
future research topics for this line of study are identified. Causing - List of variables Effected
variable(s) variable(s)
2. METHODS
Number of staff 2 Marketing effort 2 Market share, 1 aver.
2.1 THE KNOWLEDGE COLLECTION METHOD service usage, 2 OPEX
Five targets were set for the developed method: 1) to Network equip. need 1, Number of staff 3 OPEX, 2 marketing
combine different expert knowledge from various current network effort
business categories with the help of a broad expert team equipment capability 1
and 2) give the experts freedom to focus on those
causalities they feel, by their expertise, to be important in
order to make sure that new innovative cause-
consequence –relationships would arise, 3) to discover as
much as possible variable candidates from CSP business
domains, 4) to ensure that the experts acted as individuals
and no group –thinking possibility existed and 5) to
facilitate also disruptive proposals. Thus a pre-defined
variable list was not introduced but instead experts had
freedom to also name the variables. The financial
category was seen more a deterministic than a
probabilistic cause- consequences structure and thus it
was decided that only a few experts need to be dedicated
to financial topics.
An email was sent to 100 expert candidates working in
12, mostly European countries, for CSPs, universities,
CSP infrastructure vendors and software and consulting
companies which offer services to CSPs. The email
included extensive background information about the
study targets, introductions of BBN and causality,
example variables and an excel template based on the Figure 1: A causal map of two triplets from Table 1
seven CSP business categories. With the help of the including strengths.
template, experts were asked to list variables they
considered to have effect on CSP businesses and to
categorize the variables to the correct category. Basically 2.2 SUCCESS OF KNOWLEDGE COLLECTION
experts were asked to list causality triplets of “variable X Out of 100 expert candidates, 48 answered with survey
has some cause on variable Y, which has some effect on results. The resulting causal models were reviewed with
variable Z”, see table 1. It was supposed, that with this 60% of these 48 experts. The distribution of expertise
method, an expert can easily just start to write the triplets was:
without need to first have a big picture in mind. In
addition, experts were asked to estimate the strength of Product and service innovations 21%
effect by using numbers: Technology and architecture 20%
Staff and internal processes 20%
Strong effect=3, Strategy and competition 19%
Moderate effect = 2, Customers-related 11%
Weak effect = 1 Local & global economy and legislation 5%
These values were used for measuring the expert’s Financial 4%
degree-of-belief value for causal connections. The plan Experts used between 1 and 5 hours for the survey, with
was to use a simplistic method, where both weight and the average being 2,5 hours. More than 2200 variables
belief parts originate from this strength of effect. and 3400 arcs and 40 sub- BBNs were created from the
Triplets are in fact mini causal maps (see Figure 1) and survey results. Text analysis (www.textanalyser.net) was
constructing of one full BBN required combining these used in order to understand word frequencies used in
triplets together. This was done with a BBN tool called variable names. Out from about 5000 used words, 40%
BayesiaLab (www.bayesia.com) by hand. The plan was to were unique. The top 12 used words for variable names
review the achieved model with each expert. were “product and service” 80 times, “customers” 60
times, “costs” 56 times, “market” 50 times, “product” 36
times, “brand 28” times, “new” 22 times, “revenue” 20
times, “price”, “marketing”, “personnel”, “network” 16
times.
From the text analysis it was clear that:
The process to create a comprehensive BBN is
challenging because of the high number of
different variable names that have closely the
same meaning. The plan was to give full freedom
to experts in order to make sure that there were
innovative approaches, but this study
demonstrated clearly the need of business
dictionary if Bayes Belief Networks are to be
widely used in CSP business modeling and
simulation.
The competition for customers and tight cost
control in European CSP markets might explain Figure 2: Sub-BBNs derived from expert surveys covered
the top 12 used words, as the majority of experts in 70% of cases all seven categories of CSP businesses
were from European countries. but granularity varied greatly depending on expert.
2.3 CONSTRUCTION OF A SUB-BBN converted them in the following way: 3=> 90%, 2=>
75%, 1=> 60%, -3=>10%, -2=>25% and 1=>40%.
It was quickly concluded that the creation of a
comprehensive CSP business model directly from triplets In further studies, when the model(s) will be tested in
was a too complicated task. It was decided that individual CSP environment, the dual review method with experts
BBNs, called sub-BBNs would be first created. One sub- will be used, namely first a causal model review with
BNN was created per expert and then the comprehensive states alone, and after it second review with weights and
BBN was merged from these sub-BBNs. This approach confidence values.
has two benefits: 1) it filters out excess of variables with
the same meaning in the sub-BBN review –process with 2.5 CONSTRUCTING OF COMPREHENSIVE BBN
the expert and 2) innovative sub-BBNs will be
documented individually. The 40 sub-BBNs varied in granularity and coverage
(Figure 2) because experts were not asked to focus solely
The creation of a sub-BBN is straightforward: Variables on their own expertise topic. Merging the sub-BBNs to a
and their causal connection were created manually from comprehensive BBN became challenging without a
triplets by using BayesiaLab-tool (www.bayesia.com). A standard “kernel”. The Osterwalder business model
model review was organized whenever possible with the ontology (Osterwalder, 2002) is used as a standardized
expert including the states. Each variable has typically causal kernel (Figure 3) to which sub-BBNs was merged.
only two states which describe best the variable in
question like true/false, big/small, high/low,
positive/negative, fast/slow.
2.4 PROBABILITY CALCULATIONS FOR A SUB-
BBN
The conditional probability tables were calculated with
weighted sum –algorithm utilizing compatible parent
configurations defined by Das (Das, 2004). This
algorithm allows for simplification of the calculation
through the utilization of compatible parent Figure 3: Osterwalder business model blocks, which
configurations for the evaluations performed by the are used as “a kernel” for comprehensive BBN.
expert, limiting the need of individual probability state
combinations needed to be evaluated.
The comprehensive BBN can be seen as an onion-like
For this study a simplistic method was used in structure, where the kernel is from the business model
calculation: The weights 3, 2, 1, -1, -2, -3 were used as ontology and surrounding layers represent experts’ sub-
relative weights and the same weight as probability after BBNs (Figure 4).
Figure 4: The comprehensive BBN constructs onion type of layers (black and blue circles) around the kernel model
(yellow circles)
Comprehensive BBN with different granularities (number merged from individual BBNs. Chapter 3.1 gives three
of variables and arcs) were created to test the tool and examples of innovative sub-BBNs, which can be used, not
computer environment constraints. When the number of only as an input to the comprehensive BBN but also
variables exceeds 100, and at the same time the independently. Chapter 3.2 presents results on the
relationship between number of arcs divided by number comprehensive BBN.
of variables is on the average greater than three and if a
few of variables have five to ten common effects, the
practical utilization of the comprehensive BBN for
different kind analysis decreases due to slowness of the 3.1 SUB-BBN EXAMPLES
PC-environment. The objective of this study is not to Example 1: A generic purpose financial causal map with
focus on the tool usability nor model complexity topics 32 variables and their relationships (Figure 5). Many of
but to discover a Bayes Belief Network which can be the variables and causalities are more deterministic than
utilized in practice, contains all the seven business probabilistic and values are results of mathematical
categories and which reflects the expert’s common view equations like calculation of EBITDA (Earnings before
about CSP variables effecting on business. Interest, Taxes, Depreciation and Amortization). This
The merge process was performed manually, with map can be used to analyze the effect of non financial
variables and arcs being combined from each sub-BBN to variables analyzed in other sub-BBNs connected to a
the comprehensive network around it’s kernel. If certain comprehensive set of financial variables in this model.
variable and causal connection existed in many sub- Example 2: The variable “new business opportunities” is
BBNs, the weights (used in sub-BBNs) were summed a parent variable for many new business opportunities for
together. Thus, if 10 sub-BBNs have a variable “customer CSPs in a electric-car ecosystem (Figure 6), The business
satisfaction” affecting with weight 3 “customer loyalty”, opportunities vary from traditional bit-pipe services to
then the combined weight is 30. The conditional content service opportunities. The model contains
probability tables have been calculated with the same variables such as the effect of regulator actions,
method as described in sub-BBN-case. However, a dual environmental circumstances, renewal energy portion,
review method is planned to be used when the model will new technology, price of electricity, price of a electric car,
be tested in real life. number of electric cars and emergence of new business
opportunities. The model offers ways to analyze the
3. RESULTS effect of different ecosystem variables on potential new
services. The states and probabilities of key variables in
This chapter presents both sub-BBNs, created based on the model are shown in Figure 6.
individual expert’s survey and the comprehensive BBN,
Figure 5: A generic purpose financial –related causal map. The red circles represent financial, yellow customers, blue
staff and processes, greed product, black technical and pink competition-strategy –related variables. This model in its
many parts is deterministic in nature and don’t contain probabilities in this study. Sub-BBN’s end up often to Sales,
ARPU, R&D etc (blue disks) and with this map the financial analysis towards EBITDA , Earnings per share,
expected cash flow and investment decisions (orange disks) can be extended.
Figure 6: Key causal structure (upper part), states and conditional probabilities (lower part) of some variables in electric
car ecosystem model. The variable “New business opportunities” represents potential new business for CSPs.
Figure 7: High level sub-BBN for typical European CSP Operations Support System (OSS) based on one expert’s views.
OPEX and CAPEX targets have been set to 100% in order to test the consequences: Automation rate needs to be
enhanced, similarly more investment, head count reduction and activities to enhance perceived user quality are needed
(red arrows).
Example 3: The task of Operations Support Systems Belief Network as a methodology for business reasoning
(OSS) is to take care of day-to-day infrastructure and what-if analysis.
management so that the network and related services
Also other innovative sub-BBNs were created, such as
work properly with high quality and in an optimized way.
IPTV model, customer experience & satisfaction model,
OSS BBN parent variables are the number of today’s regulator causalities model.
management platforms (rather low), investment capability
of the company (often restricted), current OSS 3.2 THE COMPREHENSIVE BBN
architecture (often complex), network performance (often
not enough) and harmonization need (typically high). The comprehensive Bayesian Belief Network was created
The target variables in the model are OPEX and from sub-BBNs as described in chapter 2. The BBN
revenue/profitability. The model covers variables like contains the kernel shown in Figure 3. The model
training needs, technology, head count, perceived quality (Figures 8 and 9) contains the 32 most used variables and
seen by customer, customer experience and automation their 93 causal connections. It is remarkable that three
need (Figure 7). The model, even though it is on a rather variables are very central in the model: 12 variables have
high level, demonstrates the great potential of Bayes variable called “Customer experience & satisfaction”, 9
Figure 8: The comprehensive Bayesian Belief Network representing the overall feedback of the survey with granularity
of 32 variables and 93 causal connections. The blue disks are the six variables, which have highest node forces as a sum
of entering and outing arcs forces. Three from them, namely Customer experience & satisfaction, Product portfolio and
Activity & efficiency are the central variables in the model.
Figure 9: The states and probabilities of comprehensive BBN. The probability of the first state of variable “Product
portfolio”, “Customer experience and satisfaction” has been set to 100% (green bars).
variables “Product portfolio” and 5 variables “Efficiency” environment where customer experience, efficiency and
as a common variable. On the other hand there is only one portfolio play important role. The 32 variables and 93 arcs
purely technical variable even though 20% of experts had in the model were selected as a compromise between
technical expertise. The reason for the lack of technical model granularity and usability and based on response
variables might be the fact that most of the experts were times in analysis.
from Europe and the model represents thus mostly a
A light validation has been done for the model to verify
mature European mobile and convergent operator’s
whether it gives logical results especially because a
simplistic conditional probability calculation method, 50 experts, who spent hours by documenting their
where both weights for arcs and probabilities originate opinions, deserve a big thank. Special thanks to Gareth
from same strength of effect value given by experts, has Smith and Matti Aksela, who created the excel and R –
been used. Figure 9 gives an example of the tests: It based macros for calculation of conditional probability
shows that when the probability of “Product portfolio” tables and finally big thanks to Antti Myllykangas, who
state “competent” is set to 100% and “Customer has altruistic supported me in many aspects of this study.
experience and satisfaction” state “high” is also set to
100%, the consequence will be that the revenue will References
increase clearly, when it is assumed that product pricing
can be higher, efficiency of internal processes will be T. Casey (2009). Analysis of Radio Spectrum Market
better, innovation capacity will increase, technical assets Evolution Possibilities. Communications and Strategies,
are modern and competent staff will be in place. These No. 75, 109-130, September 2009
validations demonstrated that the model yield logical B. Das (2004). Generating Conditional Probabilities for
results. Bayesian Networks: Easing the Knowledge Acquisition
Problem. Command and Control Division, DSTO,
4. SUMMARY AND DISCUSSIONS Edinburgh, SA 5111, Australia
An extensive study to model communication service E. Faber, P. Ballon, H. Bouwman, T. Haaker, O. Rietkerk,
provider businesses was performed. A novel method was M. Steen (2003). Design business models for mobile ICT
used to elicit this prior information. Especially the way to services. 16th Bled Electronic Commerce Conference
create the so called causal triplets in the survey and to Bled, Slovenia June 9-11, 2003
construct the initial BBN based on the triplets was found F.V. Jensen (2001). Bayesian Networks and Decision
to be a fast, innovative and effective method that can be Graphs.Springer-Verlag, New York, ISBN 0-387-95259-4.
used to create different kind of causal maps. All together
40 Bayes Belief Networks were created, each representing Y. Kasperskaya (2006). Essays on Causal Performance
an individual expert’s view. These networks were then Measurements Models. Universitat Pompeu Fabra,
merged to one comprehensive Bayes Belief Network. Department of Economics and Business, PhD
Dissertation December 2006
Models such as an IPTV model, OSS management model,
customer experience & satisfaction model, regulator E. Lee, Y.Park and J.G. Shin (2009). Large engineering
causalities model, the electric car ecosystem, financial project risk management using a Bayesian belief network.
model are examples of innovative sub-models. The Expert Systems with Applications 36 (2009) 5880-5887
biggest challenge in the knowledge collection was the S. Nadkarni and P. Shenoy (2004). A causal mapping
excessive freedom in variable naming. Creation and usage approach to constructing Bayesian networks. Decision
of a dictionary would be, from work amount and quality Support Systems 38 (2004) 259-281
points of view, a clear improvement for the eliciting of
prior knowledge and this is highly recommended for A. Osterwalder (2003). An e-Business Model Ontology
future studies. for Modeling e-Business. 15th Bled Electronic Commerce
Conference.Bled, Slovenia, June 17-19, 2002
Calculations of conditional probability tables were based
on a weighted sum algorithm utilizing compatible parent A. Osterwalder (2005). Clarifying Business Models:
configurations for the states. The process used in this Origins, Present, and Future of the Concept.
study was simplified but yielded promising results which Communications of the Association for Information
motivate the further testing of the models in a real life Systems Volume 15, May 2005
environment. However, a dual review method with J. Peppard, A. Rylander (2006). From Value Chain to
experts is needed in order to achieve more reliable results. Value Networks: Insights for Mobile Operators.
The results showed that the BBN is a potential method for European Management Journal Vol. 24, Nos. 2-3, pp.
what-if analysis and predictions in the strategy creation 128-141, 2006
process but also in daily management and decision tasks. D. Rigby and B. Bilodeau (2007). Management tools and
Future research topics are to enhance and benchmark trends 2007. Bain & Company, Management Tools 2007
some of models in mobile operator environment in A.J. Scavarda, T. Bouzdine-Shameeva, S. Meyer
focused use cases as well as synthesis of expert Goldstein, J.M. Hayas, A.V. Hill (2006). A Methodology
knowledge with data in order to enhance the dynamicity for Constructing Collective Causal Maps. Decision
of the model. Science Volume 37, Number 2, May 2006
Acknowledgements L. Uusitalo (2007). Advantages and challenges of
Bayesian networks in environmental modelling,
Colleagues in Department of Communications and Ecological modelling 203 (2007) 312-38
Networking of Aalto University as well as all the nearly