CHAPTER SEVEN

Segmentation Applications in Telecommunications

MOBILE TELEPHONY

It is a jungle out there for mobile telephony network operators, with an environment of strong competition, especially in the case of mature markets. Offering high-level quality services is essential for becoming established in the market. In times of rapid change and fierce competition, focusing only on customer acquisition, which is nevertheless becoming more and more difficult, is not enough. Inevitably, organizations have to also work on customer retention and on gaining a larger “share of customers” instead of trying only to gain a bigger slice of the market. Growth from within is sometimes easier to achieve and equally as important as winning customers from competitors.

Hence, keeping customers satisfied and profitable is a one-way street to success. In order to achieve this, operators have to focus on customers and understand their needs, behaviors, and preferences. Behavioral segmentation can help in the identification of different customer typologies and in the development of targeted marketing strategies.

Nowadays, customers can choose from a huge variety of services. The days of voice-only calls are long gone. Mobile phones are communication centers and it is up to users to select the way of usage that suits their needs. People can communicate via SMS and MMS messages. They can use their phones to connect to the Internet, to send e-mails, to download games or ringtones, and to communicate with friends and family when they travel abroad. Mobile phones are perceived differently by various people. Some customers use them only in rare circumstances and mainly for receiving incoming calls. Others are addicted to their devices and cannot live without them. Some treat them as electronic gadgets, while for others they are a tool for their work.

Clearly this multitude of potential choices results in different usage patterns and typologies. Once again the good news is that usage is recorded in detail: CDRs (Call Detail Records) are stored, providing a detailed record of usage. They contain information on all types of calls. When aggregated and appropriately processed, they can provide valuable information for behavioral analysis.

All usage history should be stored in the organization’s mining data mart and MCIF, as described in the respective chapter of this book. Information on frequency and intensity of usage for each type of call (voice, SMS, MMS, Internet connection, etc.) should be taken into account when trying to identify the different patterns of behavior. In addition, information such as the day/time of calls (work days versus non-work days, peak versus off-peak hours, etc.), roaming usage, direction of calls (incoming versus outgoing), and origination/destination network type (on-net, off-net, etc.) could also contribute to the formation of a rich segmentation solution.

In this section we present a segmentation example from the mobile telephony market. The marketers of a mobile telephony network operator decided to segment their customers according to their behavior. They used all the available usage data to reveal the natural groupings in their customer base. Their goal was to fully understand their customers in order to:

• Develop tailored sales and marketing strategies for each segment.
• Identify distinct customer needs and behaviors and proceed to the development of new products and services, targeting the diverse usage profiles of their customers. This could directly lead to increased usage on behalf of existing customers but might also attract new customers from the competition.

Moreover, the operator’s marketers also decided to segment their customers according to revenue and to distinguish high-value from medium-and low-value customers. Their intention was to use this information to incorporate prioritization strategies and handle each customer accordingly.

MOBILE TELEPHONY CORE SEGMENTS – SELECTING THE SEGMENTATION POPULATION

Mobile telephony customers are typically categorized in core segments according to their rate plans and the type of relationship with the operator. The first segmentation level differentiates residential (consumer) from business customers. Residential customers are further divided into two, postpaid and prepaid:

• Postpaid – contractual customers: Customers with mobile phone contracts. Usually these customers comprise the majority of the customer base. They have a contract and a long-term billing arrangement with the network operator for the services received. They are billed on a monthly basis and according to the traffic of the past month, hence they have unlimited credit.
• Prepaid customers: These customers do not have a contract-based relationship with the operator and buy credit in advance. They do not have ongoing billing and they pay for the services before actually using them.

Additionally, business customers are further differentiated according to their size into Large Business – Corporate, SME (Small and Medium Enterprise), and SOHO (Small Office, Home Office) customers.

The typical core segments in mobile telephony are depicted in Figure 7.1.

The objective of the marketers was to enrich the core segmentation scheme with refined sub-segments. Therefore they decided to focus their initial segmentation attempts exclusively on residential postpaid customers. Prepaid customers need a special approach in which attributes like the intensity and frequency of topups (recharging of their credits) should be taken into account. Business customers also need a different handling since they comprise a completely different market. It is much safer to analyze these customers separately, with segmentation approaches like the ones (value based, size based, industry based, etc.) presented in Chapter 5.

Moreover, only MSISDNs (telephone numbers) with current status active or in suspension (due to payment delays) were included in the analysis. Churned (voluntary and involuntary churners) MSISDNs have been excluded from the analysis. Their “contribution” is crucial in the building of a churn model but they have nothing to offer in a segmentation scheme mainly involving phone usage. In addition, the segmentation population was narrowed down even more by excluding users with no incoming or outgoing usage within the past three months. These users have been flagged as inactive and selected for further examination and profiling. They could also form a target list for an upcoming reactivation campaign, but they do not have much to contribute to a behavioral analysis since, unfortunately, inactivity is their only behavior at the moment.

BEHAVIORAL AND VALUE-BASED SEGMENTATION – SETTING UP THE PROJECT

The methodological approach followed was analogous to the general framework presented in detail in the relevant chapter. In this section we just present some crucial points concerning the project’s implementation plan, which obviously affected the whole application.

Customers may own more than one MSISDN, which may be used in a different manner to cover different needs. In order to capture all the potentially different usage behaviors of each customer, it was decided to implement the behavioral segmentation at MSISDN level. Therefore, relevant input data have been aggregated accordingly and the derived cluster model assigned each MSISDN to a distinct behavioral segment.

A two-way approach was decided for the value-based segmentation. The value segments were identified at both an MSISDN and a customer level, providing a complete view of profitability. Since the methodological approach does not depend on the level of the analysis, only the MSISDN value segmentation will be described here.

The behavioral segmentation implementation included the application of a data reduction technique (PCA in particular) to reveal the distinct dimensions of information, followed by a clustering technique to identify the segments. Once again the mining data mart tables, presented in detail in the corresponding chapter, comprised the main sources of input data. Table 7.1 outlines the main usage aspects that were covered.

On the other hand, value-based segmentation relies only on a single field. It does not need the application of a data mining algorithm either. It only involves a simple sorting of records (MSISDNs or customers) according to a profitability index and an assignment to corresponding groups. In the case presented here, it is assumed that the respective value index is already calculated and stored in the organization’s MCIF. The marketers of the telephone operator used the already calculated marginal average revenue per user (MARPU), a marginal profitability index that denotes the revenue minus the costs of products/services, for building the value segments.

Both implementations included a phase of extensive profiling of the revealed segments. All available information was used in this phase, including demographic data and contract information details.

The analysis was based on data from the last six months, a time frame that generally ensures the capturing of stable, non-volatile behavioral patterns instead of random or outdated ones. Summarizing fields (sums, counts, percentages, averages, etc.) covering the six-month period under study were used as clustering fields for the behavioral segmentation. For the same reasons, the MARPU six-month average was used for the value-based segmentation, instead of one month’s profitability data.

Table 7.2 summarizes the setup for the two segmentation implementations. The list of input fields is presented in the next section.

SEGMENTATION FIELDS

The list of candidate inputs for behavioral segmentation initially included all usage fields contained in the mining data mart and the MCIF. In a later stage and according to the organization’s specific segmentation priorities, the marketers arrived at a limited list of clustering fields, which is presented in Table 7.3. Obviously this list cannot be considered a “magic” approach that covers all the needs of all organizations. It represents the approach adopted in the specific implementation, but it also outlines a general framework of potential types of fields that could prove useful in similar applications.

Table 7.2 Mobile telephony segmentation: summary of implementation methodology.

Critical point	Decision
Segmentation level	• Behavioral segmentation: MSISDNs – subsequent aggregation of results at a customer level
	• VBS: two-way approach, MSISDN and customer level
Segmentation population	• Residential postpaid MSISDNs, currently active (or suspended) with usage activity during the last three months
Data sources	• Mining data mart and MCIF
Time frame of used data	• Six months
Segmentation fields	• Behavioral segmentation: usage KPIs contained in the MCIF (percentages, averages, etc.) summarizing usage of the last six months
	• VBS: value index (MARPU) summarizing MSISDN’s profitability
Profiling fields	• All fields of interest: usage KPIs for the last six months, demographics, contract details, etc.
Technique used	• Behavioral segmentation: clustering technique to reveal segments
	• VBS: binning

The selected fields indicate the marketers’ orientation toward a segmentation scheme that would capture usage differences in terms of preferred type of calls (voice, SMS, MMS, Internet, etc.), roaming usage (calls made in a foreign country), frequency of international calls (calls made in home country to international numbers), call day (peak vs. off-peak), and hour (work vs. non-work day) of usage.

For the needs of value-based segmentation, the only field used was the value index. The mobile telephony operator had already implemented a relevant project for the calculation of ARPU (Average Revenue Per User) and MARPU for each MSISDN. ARPU is a revenue-only index. It does not take into account costs, but covers all types of revenues (e.g., monthly fees, activation fees, outgoing usage revenue, incoming usage revenue, i.e., interconnection costs paid by other operators to recipient’s operator) and possible discounts. In MARPU, however, all costs of providing products and services (e.g., interconnection costs, roaming costs, acquisition/renewal costs such as handset costs, distribution/dealers’ costs, retention costs) are taken into account and subtracted from ARPU.

The full list of used fields is presented in Table 7.3.

Table 7.3 Mobile telephony segmentation fields.

Field name	Description
Fields for value-based segmentation:
MARPU	Monthly average of MARPU value index
Clustering fields for behavioral segmentation:
Community:
OUT_COMMUNITY_TOTAL	Total outgoing community: monthly average of distinct phone numbers that the holder called (includes all call types)
OUT_COMMUNITY_VOICE	Outgoing voice community
OUT_COMMUNITY_SMS	Outgoing SMS community
PRC_OUT_COMMUNITY_VOICE	Percentage of outgoing voice community: outgoing voice community as a percentage of total outgoing community
PRC_OUT_COMMUNITY_SMS	Percentage of outgoing SMS community
IN_COMMUNITY_VOICE	Incoming voice community
IN_COMMUNITY_SMS	Incoming SMS community
IN_COMMUNITY_TOTAL	Total incoming community
PRC_IN_COMMUNITY_VOICE	Percentage of incoming voice community
PRC_IN_COMMUNITY_SMS	Percentage of incoming SMS community
Number of calls by call type:
VOICE_OUT_CALLS	Monthly average of outgoing voice calls
VOICE_IN_CALLS	Monthly average of incoming voice calls
SMS_OUT_CALLS	Monthly average of outgoing SMS calls
SMS_IN_CALLS	Monthly average of incoming SMS calls
MMS_OUT_CALLS	Monthly average of outgoing MMS calls
EVENTS_CALLS	Monthly average of event calls
INTERNET_CALLS	Monthly average of Internet calls
TOTAL_OUT_CALLS	Monthly average of outgoing calls (includes all call types)
TOTAL_IN_CALLS	Monthly average of incoming calls (includes all call types)
PRC_VOICE_OUT_CALLS	Percentage of outgoing voice calls: outgoing voice calls as a percentage of total outgoing calls
PRC_SMS_OUT_CALLS	Percentage of SMS calls
PRC_MMS_OUT_CALLS	Percentage of MMS calls
PRC_EVENTS_CALLS	Percentage of event calls
PRC_INTERNET_CALLS	Percentage of Internet calls
Minutes/traffic by call type:
VOICE_OUT_MINS	Monthly average number of minutes of outgoing voice calls
VOICE_IN_MINS	Monthly average number of minutes of incoming voice calls
EVENTS_TRAFFIC	Monthly average of events traffic
GPRS_TRAFFIC	Monthly average of GPRS traffic
International calls/roaming usage:
OUT_CALLS_ROAMING	Monthly average of outgoing roaming calls (calls made in a foreign country)
OUT_MINS_ROAMING	Monthly average number of minutes of outgoing voice roaming calls
PRC_OUT_CALLS_ROAMING	Percentage of outgoing roaming calls: roaming calls as a percentage of total outgoing calls
OUT_CALLS_INTERNATIONAL	Monthly average of outgoing calls to international numbers (calls made in home country to international numbers)
OUT_MINS_INTERNATIONAL	Monthly average number of minutes of outgoing voice calls to international numbers
PRC_OUT_CALLS_INTERNATIONAL	Percentage of outgoing international calls: outgoing international calls as a percentage of total outgoing calls
Usage by day/hour:
OUT_CALLS_PEAK	Monthly average of outgoing calls in peak hours
OUT_CALLS_OFFPEAK	Monthly average of outgoing calls in off-peak hours
OUT_CALLS_WORK	Monthly average of outgoing calls in work days
OUT_CALLS_NONWORK	Monthly average of outgoing calls in non-work days
PRC_OUT_CALLS_PEAK	Percentage of outgoing calls in peak hours
PRC_OUT_CALLS_OFFPEAK	Percentage of outgoing calls in non-peak hours
PRC_OUT_CALLS_WORK	Percentage of outgoing calls in work days
PRC_OUT_CALLS_NONWORK	Percentage of outgoing calls in non-work days
IN_CALLS_PEAK	Monthly average of incoming calls in peak hours
IN_CALLS_OFFPEAK	Monthly average of incoming calls in off-peak hours
IN_CALLS_WORK	Monthly average of incoming calls in work days
IN_CALLS_NONWORK	Monthly average of incoming calls in non-work days
PRC_IN_CALLS_PEAK	Percentage of incoming calls in peak hours
PRC_IN_CALLS_OFFPEAK	Percentage of incoming calls in non-peak hours
PRC_IN_CALLS_WORK	Percentage of incoming calls in work days
PRC_IN_CALLS_NONWORK	Percentage of incoming calls in non-work days
Days with usage:
DAYS_OUT	Monthly average number of days with any outgoing usage
DAYS_IN	Monthly average number of days with any incoming usage
Average call duration:
ACD_OUT	Average duration of outgoing voice calls (in minutes)
ACD_IN	Average duration of incoming voice calls (in minutes)
Other usage fields – profiling fields
Contract information fields – profiling fields (for instance, tenure, rate plan, acquisition channel, payment method, handset category, etc.)
Customer information fields – profiling fields (customer demographics)

VALUE-BASED SEGMENTATION

All customers are not equal. Some are more valuable than others. Identifying those valuable customers and understanding their importance should be considered a top priority for any organization.

Value-based segmentation is one of the key tools for developing customer prioritization strategies. It can enable service-level differentiation and prioritization of churn prevention activities. The marketers at the mobile telephony network operator decided to develop a value-based segmentation scheme in order to assign each customer to a segment according to his or her value. They believe that this segmentation will enable them to separate valuable customers from the rest and provide insight into their differentiating characteristics.

The value-based segmentation was based on the previously calculated marginal revenue for each MSISDN. Thus, all MSISDNs were ranked according to their MARPU. The ordered records were divided (binned) into equal-sized subsets, or quantiles, and these tiles were then used to construct the value segments.

In value-based segmentation the number of binning tiles to be constructed depends on the specific needs of the organization. The derived segments are usually of the form highest n%, medium high n%, medium low n%, low n%. In general, it is recommended to select a sufficiently rich and detailed segmentation level, especially in the top groups, in order to discern the most valuable customers. A detailed segregation level may not be required for the bottom of the value pyramid.

The segmentation bands selected by the marketers were the following:

1. Platinum: The top 1% MSISDNs with the highest MARPU

2. Gold: The top 4% MSISDNs with the second highest MARPU

3. Silver: 15% of MSISDNs with high MARPU

4. Bronze: 40% of MSISDNs with medium MARPU

5. Mass: 40% of MSISDNs with the lowest MARPU.

Since customers with no usage activity have been excluded from the procedure, a sixth segment, not listed but implied, is the one composed of “Inactive” customers.

The whole value-based segmentation procedure is graphically depicted in Figure 7.3.

IBM SPSS Modeler offers a very useful tool (node) called Binning that can automate the above procedure. It performs ranking and creates the quantiles in one step. The respective menu is shown in Figure 7.4.

In order to implement the value-based segmentation, the marketers selected MARPU as the “Bin field,” “Tiles (equal count)” as the binning method, and “Percentile (100)” for the initial grouping of the MSISDNs. These options created bands of 1% according to the ranked MARPU.

The derived bands were then been processed with a “Reclassify” IBM SPSS Modeler node, in order to refine the solution and regroup the bands into the desired segments, as indicated in Figure 7.5.

The “MARPU_TILE100” field is the one created by the “Binning” node which denotes the assignment into 100 bands of 1%. These bands were recoded into broader groups through the “Reclassify” node which mapped the original quantiles into the value-based segments. The derived “VBS” field is the final result and indicates the value-based assignment for each user.

VALUE-BASED SEGMENTS: EXPLORATION AND MARKETING USAGE

The next step before making use of the derived segments was the investigation of their main characteristics.

The initial goal of this segmentation was to capture the assumed large-scale differences between customers, in terms of value. Thus, the marketers began to investigate this hypothesis by examining what each segment represents in terms of revenue. Table 7.4 presents the number of MSISDNs and the percentage of total MARPU for each value segment.

This table shows that a substantial percentage of revenue comes from a disproportionately small number of high-value users. Almost half of the operator’s revenue arises from the top three value segments. Thus, 20% of the most valuable users account for about 50% of the total MARPU. On the other hand, low-value users which comprise the mass (bottom 40%) segment provide a disappointing 13% of the total revenue. The comparison of mean MARPU values also underlines the large-scale differences among segments. The mean MARPU value in the top 1% segment is more than 17 times higher than that for the bottom 40%. These findings, although impressive, are not far from reality. On the contrary, in other markets, in banking for example, things are even more polarized, with an even larger part of the revenue originating from the top segments of the value pyramid.

These vast differences in revenues, as expected, are also reflected in the intensity of usage of the different services. Table 7.5 summarizes some of the main usage KPIs by segment.

Once again, the interpretation is quite clear since the contrast in usage between the top and bottom segments is very intense. In almost all of these KPIs (for instance, total outgoing calls, total outgoing voice calls and minutes) users at the top of the value pyramid present values up to 10 times higher than those at the bottom. For example, the average number of total outgoing calls reaches 700 calls per month for top users, while this value does not exceed 70 calls per month among mass users.

The two-fold value segmentation project concluded by developing an analogous segmentation scheme at a customer level, at the end of which each customer was allocated to a corresponding value group.

The business benefits from the identification of value segments are prominent. The implemented segmentation can provide valuable help to the marketers in setting appropriate objectives for their marketing actions according to customer value. High-value customers are the heart of the organization. Their importance should be recognized and rewarded. They should perceive their importance every time they interact with the operator. Preventing the defection of such customers is vital for any organization. Identifying valuable customers at risk of attrition should trigger an enormous effort to avoid losing them to competitors. For medium-and especially low-value customers, marketing strategies should focus on driving up revenue through targeted cross-and up-selling campaigns to make it approach that of high-value customers.

PREPARING DATA FOR CLUSTERING – COMBINING FIELDS INTO DATA COMPONENTS

The number of original segmentation fields exceeds 50. Using all these fields as direct input in a clustering algorithm will produce a complicated solution. Therefore the approach followed was to incorporate a data reduction technique to reveal the underlying main data dimensions, prior to clustering.

This approach was adopted to eliminate the risk of deriving a biased solution due to correlated original inputs. Moreover, it also ensures a balanced solution, to which all data dimensions contribute equally, and it simplifies the tedious procedure of understanding the clusters.

Specifically, a PCA model with Varimax rotation was applied to the original segmentation fields. The solution finally selected after many trials included 12 extracted components and was based on the “Eigenvalues over 1” criterion. The main reason for retaining this solution was the fact that it produced a relatively low number of meaningful components without sacrificing much of the information in the initial fields.

Before using the derived components and substituting more than 50 fields for just a dozen ones, the data miners of the organization wanted to be sure that the PCA solution carried over most of the original information. Therefore, they started to examine the model results by looking at the table of “variance explained,” Table 7.6.

Table 7.8 Interpretation of the extracted principal components.

	Derived Components
Component	Label and Description
1	Incoming voice calls The high loadings in the first column of the rotated component matrix denote a strong positive correlation between component 1 and the original fields which measure incoming voice traffic such as the number and minutes of incoming voice calls (VOICE IN CALLS, VOICE IN MINS) and the size of the incoming voice community (IN COMMUNITY VOICE). Thus, component 1 seems to be associated with incoming voice usage Because generally voice calls constitute the majority of calls for most users and tend to dominate the total incoming usage, a set of fields associated with total incoming usage (total number of incoming calls, incoming calls by day/hour) are also loaded high on this component
2	Outgoing voice calls Likewise, component 2 measures the outgoing voice traffic. Therefore, users with heavy outgoing voice usage and high values in the corresponding original inputs (VOICE OUT CALLS, OUT COMMUNITY VOICE, VOICE OUT MINS, etc.) are expected to also show proportionately high values in component 2 The fact that some voice outgoing fields are also moderately loaded on component 1 implies an intercorrelation between outgoing and incoming voice calls
3	SMS calls Component 3 measures SMS usage, incoming as well as outgoing. The rather interesting negative correlation between component 3 and the percentage of voice out calls (PRC VOICE OUT CALLS) denotes a contrast between voice and SMS usage. Thus, users with high positive values of this component are expected to have increased SMS usage and increased SMS to voice calls ratio. This does not necessarily mean low voice traffic, but it certainly implies a relatively lower percentage of voice calls and an increased percentage of SMS calls.
4	Day/hour of peak incoming usage Component 4 indicates the day/hour of peak incoming usage. Users with high positive values in component 4 receive the majority of their calls in non-work days/hours. Perhaps this signifies a type of “social” use.
5	Day/hour of peak outgoing usage Likewise, component 5 represents the day/hour of peak outgoing usage. Users with high positive values in component 5 make most of their calls in work days/hours. Perhaps this signifies a type of “professional” use.
6	Roaming usage Component 6 measures roaming usage (making calls when abroad
7	International calls Component 7 is associated with calls to fixed international networks
8	Events and GPRS Component 8 measures event calls and traffic. GPRS traffic is also moderately loaded high on this component
9	Internet usage Component 9 is associated with Internet usage and GPRS traffic
10	MMS usage This usage is measured by component 10
11	ACD (Average Call Duration) Fields denoting ACD for incoming and outgoing voice calls are related and combined to form component 11
12	Frequency of usage The frequency of usage, that is, the days in a month with any type of outgoing or incoming usage, is represented by component 12

The resulting 12 components retained more than 80% of the variance of the original fields. This percentage was considered more than adequate and thus the only task left before accepting the components was their interpretation. In practice only a solution consisting of meaningful components should be retained.

So what do these new composite fields represent? What business meaning do they convey? As these new fields are constructed to substitute for the original fields in the next stages of the segmentation procedure, it is necessary that they be thoroughly decoded, before being used in upcoming models.

The component interpretation phase included an examination of the rotated component matrix, a table that summarizes the correlations between the components and the original fields, that is, Table 7.7.

The “interpretation” results are summarized in Table 7.8.

The explained and labeled components and their respective scores were subsequently used as inputs to the clustering model. This brings us to the next phase of the application: the identification of useful groupings through clustering.

IDENTIFYING, INTERPRETING, AND USING SEGMENTS

The generated components represented all the usage data dimensions of interest, in a concise and comprehensible way, leaving no room for misunderstandings about their business meaning. The next step included the use of the derived component scores as inputs in a clustering model.

The clustering process involved the evaluation of different solutions obtained by trying different clustering algorithms and different model settings. All these trials resulted in overall similar, yet not identical, solutions, so it was up to the data miners and the marketers involved to select the optimal solution for deployment. The solution finally adopted was based on a TwoStep clustering model and was chosen because it seemed to best address the marketing needs of the organization.

The model automatically detected five clusters which, after extensive profiling, were assessed as meaningful and potentially useful for building the differentiated marketing strategies.

The modeling options used for the development of the adopted segmentation solution are summarized in Table 7.9.

The distribution of the derived clusters is shown in Table 7.10. These clusters were not known in advance, nor imposed by users, but were uncovered after analyzing the actual behavioral patterns recorded in the usage data.

Each revealed cluster corresponds to a distinct behavioral typology. This typology had to be understood, named, and communicated to all the people in the organization in a simple and concise way, before being used for tailored interactions and targeted marketing activities. Therefore the next phase of the project included the application of simple reporting techniques to profile each cluster and identify what it stands for.

Table 7.9 Modeling options used to produce the segmentation solution.

Data reduction
Modeling option	Setting
Model	Principal components (PCA)
Rotation	Varimax
Criteria for the number of factors to extract	Eigenvalues over 1. Resulting 12 components
Clustering
Modeling option	Setting
Model	Two step
Input clustering fields	12 component scores derived from PCA
Number of clusters	Automatically calculated
Exclude outliers option	On
Standardize numeric fields option	On

Table 7.10 Distribution of the derived clusters.

		Percentage of MSISDNs
Clusters	Cluster 1	16.6
	Cluster 2	31.0
	Cluster 3	19.3
	Cluster 4	6.6
	Cluster 5	26.5
Total		100.0

This “recognize and label” procedure started with an examination of the clusters with respect to the component scores, providing a valuable first insight into their structure, before moving on to their profiling in terms of the original usage fields. The latter results are given in Table 7.11. This table summarizes the means of the majority of the original clustering fields across the derived clusters. Large deviations from the marginal mean characterize the respective cluster and denote a behavior that differentiates the cluster from the typical behavior.

Inferential statistics have also been applied to flag statistically significant differences from the overall population mean. These results (based on one-sample t-tests with a significance level of 0.01 and Bonferroni adjustment for comparisons across clusters) are illustrated in the heat map of Table 7.12. Cluster values larger than the average of the total population are represented in dark gray, and values lower than the total average in light gray.

Based on the information presented so far, the project team involved started to outline a first rough profile of each cluster:

• Cluster 1 includes users with basic usage. They are characterized by increased ACD and very low incoming and outgoing communities.
• Cluster 2 contains average voice users. They have medium SMS usage and a relatively low percentage of traffic in work days/hours.
• Cluster 3 mainly consists of SMS users who seem to have an additional inclination toward technical services like MMS, Internet, and event calls. They also present very low voice usage and ACD.
• Cluster 4 includes roamers and users with increased communication to international destinations. They are accustomed to Internet services and make most of their calls during work days/hours. They also have the largest outgoing communities.
• Cluster 5 contains heavy voice users with traffic peaks in work days/hours. Furthermore, they have the largest voice incoming communities. Perhaps this segment includes residential customers who use their phones for business purposes (self-employed professionals).

This first rough profiling of the clusters was supplemented with complementary information concerning demographic characteristics and the relationship of the derived clusters with MARPU and the value segments, already defined by the relevant segmentation.

As indicated by Table 7.13, cluster 5 includes most of the high-value users with an average MARPU of almost 48. This cluster contains about 26% of the users who account for about 34% of the total MARPU. On the other hand, cluster 1 mainly consists of mass customers who contribute a poor 5% of the total MARPU.

Table 7.14 presents the age distribution of each cluster. Cluster 1 contains the most aged part of the customer base. On the contrary, cluster 3 mainly consists of younger users. What is interesting is the relatively increased percentage of persons between 45 and 54 years old in this cluster. Perhaps this can be explained by the fact that the demographic information refers to contract holders and not users. Maybe the age data recorded upon contract registration refers to parents who bought the contracts for their children. Remember that the scope was to build segments based on usage and not on demographic information. Cluster assignment is based on usage patterns. Therefore, from the behavioral segmentation point of view, adults who “use” their phone in an “adolescent” way should be assigned to the relevant segment.

In order to enrich the insight into each cluster, the marketers of the organization decided to also investigate the qualitative characteristics. Therefore, they extracted a random sample from each cluster and conducted a market research survey and focus group sessions in which they examined the needs and wants and the factors that have an effect on the level of satisfaction (satisfaction drivers). The short phone interviews of the selected sample were carried out by the call center agents of the organization.

The profiling outcomes are summarized in the descriptions that follow. They present the differentiating usage characteristics, a short demographic profile, and the key market research findings on each cluster. They also present the main points of the specialized marketing approach decided for each segment. A series of summarizing charts, graphically illustrating the percentage deviations of each cluster from the overall population, are also presented as a supplement of the behavioral profile.

Segment 1: Oldies – Basic Users (Figure 7.8)

Behavioral Profile

• Lowest utilization of all services.
• Smallest outgoing and incoming communities.
• Phones used for a few days in a month.
• When phone calls are made, however, the duration is quite long, as denoted by their ACD, which surpasses the average.

Segment Size

17%

Contribution to Total MARPU

5%

Demographic Profile

Highest ratios of elderly people, most of them retired.

Market Research Findings

• Satisfaction drivers:

– Low prices.

– Simple and easy to understand rate plans.

– Network quality.

• Needs/wants:

– Basic communication (voice) mostly with a limited number of people.

– Easy communication.

– Not interested in new technologies/innovative services.

– Most calls are personal.

Marketing Approach

• Product/services offered: Offering of incentives to increase usage. Base rate plans with no free call time but with low monthly fees. FnF (Friends and Family) rate plans with reduced rates for selected “family” on-net numbers. Especially for elderly people, promotion of simple and easy-to-use handset devices
• Communication channel: Direct voice calls, direct mail.
• Brand image communicated: An established, trustworthy, and reliable operator.

Segment 2: Adults – Classic Social Users (Figure 7.9)

Behavioral Profile

• Average usage.
• Mostly voice usage.
• Medium SMS usage.
• Low usage of technical services (MMS, events, Internet).
• Average communities and days with usage.
• Relatively increased usage during no work days/hours.

Segment Size

31%

Contribution to Total MARPU

34%

Demographic Profile

Typical age profile, mainly consisting of married users.

Market Research Findings

• Satisfaction drivers:

– Competitive pricing.

– Optimum rate plan.

– Reliable services.

• Needs/wants:

– Reliable communication, mostly with friends and family.

– Seek value for money.

– Most calls are personal.

Marketing Approach

• Product/services offered: Offering incentives to promote use of non-voice services. Optimized rate plans, tailored to their usage, providing free voice minutes or reduced rates for voice calls.
• Communication channel: Direct voice calls.
• Brand image communicated: An operator that offers reliable services at competitive prices.

Segment 3: Young – SMS Users (Figure 7.10)

Behavioral Profile

• Heaviest SMS usage, almost three times above the average.
• Low voice usage.
• Mostly use handsets for sending SMS messages. By far the highest percentage of SMS calls and the lowest percentage of voice calls.
• Voice usage is predominantly incoming (perhaps as a result of receiving calls from worried parents!).
• Voice calls are brief, as denoted by their ACD, which is the lowest among all segments.
• More than comfortable with the new technologies, especially with technical services such as MMS and events.
• Internet usage is the second highest.

Segment Size

19%

Contribution to Total MARPU

18%

Demographic Profile

Young persons, most of them under 35. Predominantly students, young employees, and single people.

Market Research Findings

• Satisfaction drivers:

– Offering a wide range of innovative services.

– Competitive pricing, especially for SMS.

– “Cool” brand image.

• Needs/wants:

– Interested in new technologies/innovative services.

– Use of phone not only for communication but also for fun.

– Identify with the phone, so it must be “cool.”

– Seek and compare competitors’ offers.

– Most calls are with friends and family.

Marketing Approach

• Product/services offered: Offering of “Bring a friend” incentives. Promotion of rate plans and bundles with SMS and MMS discounts. Promotion of cell phones with advanced technical capabilities. Promotion of operator’s web site and e-billing services. This is the ideal target group for promoting all high-tech services: event downloads (ringtones, games, MP3s), Internet, mobile TV, video calls, and so on.
• Communication channel: SMS, MMS, e-mails.
• Advertising channels: Internet banners, TV channels, radio stations, and magazines influential to young people.
• Brand image communicated: Cool and stylish brand providing fun/entertainment and innovative services.

Segment 4: International Users (Figure 7.11)

Behavioral Profile

• Citizens of the world! Seem to travel a lot so show the highest roaming usage.
• Highest number of roaming calls and roaming traffic minutes.
• Frequent calling of international numbers. Have the highest number of international calls and minutes.
• Most calls made during work days/hours.
• Accustomed to the new types of communication, with high Internet, events, and MMS usage.
• Top users, particularly in terms of the Internet. Probably use it a lot when traveling abroad.
• Largest outgoing community and with the highest ACD.
• Voice usage is also high.

Segment Size

7%

Contribution to Total MARPU

9%

Demographic Profile

Young and middle-aged persons, most of them under 45. This cluster mainly consists of male employees and self-employed people with a college degree.

Market Research Findings

• Satisfaction drivers:

– Network coverage.

– Efficiency of customer care.

– Roaming quality.

• Needs/wants:

– The phone is a business tool that should provide support around the world.

– Interested in new technologies/innovative services.

– “Always connected.”

Marketing Approach

• Product/services offered: Offering of rate plans and bundles with roaming and international call benefits. Teaming up with airlines and travel agencies, in order to offer “air miles and more” reward schemes with point offerings according to phone usage. Promotion of company’s web site and Internet usage.
• Communication channel: Direct voice calls, e-mails, MMS.
• Advertising channels: Internet banners, news channels and stations, newspapers, travel magazines, airport advertisements.
• Brand image communicated: A modern operator, with a presence and reliability around the world, which offers a wide range of innovative services.

Segment 5: Business Users (Figure 7.12)

Behavioral Profile

• Heavy voice usage, incoming as well as outgoing.
• Majority of calls made during work days and peak hours.
• By far the largest incoming community.
• Highest number of voice calls and minutes. This segment includes many residential customers using their phone as a business tool as well.
• Although the phone is used on a daily basis, voice calls are quite short, as denoted by the relatively low ACD.
• Low SMS usage.

Segment Size

26%

Contribution to Total MARPU

34%

Demographic Profile

Middle-aged persons. Mainly includes self-employed professionals and employees in high-level positions.

Market Research Findings

• Satisfaction drivers:

– Network coverage.

– Reliable services.

– Efficiency of customer care.

– Competitive pricing.

– Optimum rate plan.

• Needs/wants:

– Phone used as a business tool.

– High-quality services and guaranteed availability at competitive prices.

– “Always connected.”

Marketing Approach

• Product/services offered: Offering of business rate plans with free minutes and discounts on voice usage. Offering of incentives for handset replacement.
• Communication channel: Direct voice calls, direct mail.
• Advertising channels: News channels and stations, newspapers, business magazines.
• Brand image communicated: Trustworthy, high-status company, the choice of successful people.

SEGMENTATION DEPLOYMENT

The final stage of the segmentation process involved the integration of the derived scheme into the company’s daily business procedures. From that moment all users were characterized by the segment to which they were assigned. The deployment procedure involved a regular (on a monthly basis) segment update for the needs of which a specialized classification model was developed. More specifically, a decision tree model (C5.0) was trained, using all the initial clustering fields as inputs and the segment membership field as the target. The generated model was used not only to gain insight into the segments, but also as a scoring engine for allocating new records to the established clusters. Although the generated cluster model has been saved and could also be used for scoring, the decision tree solution was preferred due to its transparency. Decision trees translate the distinct profiles of the segments into a set of intuitive rules, similar to common business rules. Therefore they facilitate a clear understanding of the defining characteristics of each cluster and enable simpler communication within the organization. Moreover, they are easier to handle, allowing possible interventions in and modifications to the assignment procedure, if considered necessary.

A part of the resulting decision tree is shown in Figure 7.13. Although only the top two levels of the tree are presented here, the beginning of segment separation is evident. Additionally, a subset of the tree’s rules is shown in Figure 7.14.

Classification rules were identified for all segments. All records passing through the generated model are allocated by these rules to the segment that best matches their behavior. For example, we notice that users with a relatively high percentage of SMS usage (above 10%) and an increased frequency of inbound usage are assigned to segment 3 as “SMSers.” Similarly, records with a high ratio of outgoing voice calls (above 90%), high frequency of both incoming and outgoing usage (higher than 22 and 29 days respectively), and high number of outgoing calls in peak hours (59 and above) land in segment 5 as “Business users.”

Subsequent steps in the deployment included the loading of the derived segment information into the operational and front-line systems. Segmentation information and guidelines for the specialized handling of each customer according to the profile of its segment were made available to all points of service.

THE FIXED TELEPHONY CASE

Nowadays, fixed phone operators are offering much more than plain voice services. They provide a full range of telecommunication products, from Internet access and broadband services to Internet Protocol TV (IPTV) services. The diverse wants and needs of the customers are also reflected in different usage behaviors. Fortunately these differentiations are recorded in traffic data which can be mined to reveal the distinct customer typologies. In times when the market changes dramatically, offering new challenges, new possibilities, and new fields for development, a thorough understanding of an evolving customer base is necessary and the role of segmentation is vital.

The behavioral segmentation procedure in fixed telephony is in many ways similar to the mobile telephony case which was described in detail in the previous sections of this chapter. It aims at identifying natural groupings of customers according to their usage patterns and utilization of the services offered. The main data dimensions which should be covered by such an analysis are relevant to the ones presented for mobile telephony and include:

• Product/service utilization and usage traffic:

– Voice usage

– Internet usage

– Broadband usage

– Digital TV usage

– Voice mail usage, and so on.

• Community size:

– Outgoing community

– Incoming community.

• Usage by origination/destination network:

– Long-distance calls

– International calls

– On-net calls

– Off-net calls, and so on.

• Usage by peak/off-peak days and hours:

– Peak/off-peak hours calls

– Work/non-work days calls.

• Tenure.

Customer typologies expected to be found include:

• Young families:

– New connections

– Ages between 27 and 35

– Relatively high broadband penetration.

• Basic users:

– Low usage

– Price sensitive, low ACD

– Ages between 35 and 55

– Classic voice usage only.

• Classic users:

– Normal/average usage

– Ages between 35 and 55.

• Professionals:

– High usage at peak hours

– Large incoming community

– Use of voice mail and broadband

– International calls.

• Superstars:

– Very high levels of usage at off-peak hours

– Highest broadband penetration

– High ACD

– International calls

– Digital TV and video-on-demand packs.

• Seniors:

– Basic voice usage

– Low incoming community

– Ages over 60.

SUMMARY

In this chapter we have followed the efforts of a mobile telephony network operator to segment its customers according to their usage patterns and their value. The business objective was to group customers in terms of their profit and their behavioral characteristics and to use this insight to deliver personalized and value-driven customer handling. The organization initially focused on residential postpaid customers, who were further segregated into five behavioral and five value-based segments, as summarized in Table 7.15.

Residential customers	Postpaid – contractual
	Value-basecl segments	Behavioral segments
	Platinum	Oldies – basic users
	Gold	Adults – classic social users
	Silver	Young – SMS users
	Bronze	International users
	Mass	Business users
	Prepaid
Business customers	Large business – corporate
	Small and medium enterprise (SME)
	Small office, home office (SOHO)

The procedure followed for the behavioral segmentation included the application of a PCA model for data reduction and a cluster model for revealing distinct user groups. Furthermore, a decision tree model was also applied to profile the segments and compile a set of understandable scoring rules for updating the segments. The whole process is depicted in Figure 7.15.