In this chapter the authors describe novel methods now available to mobile phone operators to optimize targeting and improve profitability from VAS offers... In such delivery systems per
Trang 1Chapter 8 Mobile Marketing:
The Challenges of the New Direct
Marketing Channel and the Need for Automatic Targeting and Optimization Tools
The mobile phone market is becoming
increas-ingly saturated and competitive (Leppaniemi &
Karjaluoto, 2007) In several European countries
mobile phone penetration is now over 100% and first-time customers (new users that enter the market and expand the business) are practically inexistent (The Netsize Guide, 2009) In the US, similar competitive intensity has also become the norm after the introduction of wireless number
ABSTRACT
In most developed countries competition among mobile phone operators is now focused on switching customers away from competitors with extremely discounted telephony rates This fierce competitive environment is the result of a saturated market with small or inexistent growth and has caused operators
to rely increasingly on Value-Added Services (VAS) for revenue growth Though mobile phone tors have thousands of different services available to offer to their customers, the contact opportunities
opera-to offer these services are limited In this context, statistical methods and data mining opera-tools can play
an important role to optimize content delivery In this chapter the authors describe novel methods now available to mobile phone operators to optimize targeting and improve profitability from VAS offers.
DOI: 10.4018/978-1-60960-067-9.ch002
Trang 2portability by the Federal Communications
Com-mission in November 2003 Facing saturated and
stagnant markets, mobile service operators are
now focused on attracting competitors’
custom-ers Because one of the main factors influencing
customers’ operator choice is the availability of
a more convenient telephony rate plan, (Eshghi,
2007), mobile operators are relying increasingly on
price competition for customer acquisition while
revenue expansion comes mostly from
Value-Added Services (VAS) Examples of these services
include the provision of sports information, news,
and weather forecasts, download of ring-tones,
games, music, short movies, and even TV shows,
all for a fee Occasionally some of these services
are offered for free In such cases the objective
of the service is not generating revenue directly
but doing so indirectly For example, revenues
can be generated indirectly through the charges
related with the data transmission services or the
browsing of additional web pages over the phone
In the case of free viral videos aimed at building
brand awareness and word-of-mouth, firms usually
wish to build or sustain future revenue streams and
long-term goals which are even more difficult to
assess (future revenues could be associated with
product sales both via the mobile phone or offline,
depending on the firm that launches the videos)
In addition, services may be offered for free in
order to improve users’ experience, satisfaction,
and loyalty These products or services are
pro-duced by the mobile service provider itself or by
external content providers, in which case revenue
sharing contracts are established: mobile
opera-tors and content producers each take a percentage
of the revenue generated, with the share of each
depending on the type of content and on the power
split between organizations
Push Versus Pull Delivery Systems
In a Pull delivery system (one of the types of
VAS delivery system), mobile phone users
initi-ate on their own a search for a product or service
they might be willing to buy (e.g., browse sites through the mobile phone to download videos, games, or a new ring-tone) Currently one of the most popular and successful Pull delivery system is the App Store, developed by Apple
in conjunction with the iPhone launch Anyone can now produce applications for the iPhone to
be sold worldwide through the App Store once Apple approves the application The App Store is
a “moderated” type of services, that is, Apple has
to make sure all material sold through its store is legal, does not violate operator restrictions (these differ from country to country), does not include offensive material, and so on Apple is ultimately responsible for the applications sold at the store These applications are also value-added services and the revenues obtained from their sale are split between Apple and the developer who designed and produced the application
Notice that Apple does not send messages to iPhone users selling (“pushing”) these applica-tions, instead mobile users go to the App Store and search for the applications of their interest These systems can be very successful and generate significant revenue As a matter of fact, recently Apple announced (Kerris & Bowcock, 2009) that
a total amount of more than 1.5 billion tions have been downloaded since its inception and more than 65,000 different applications are today available on its App Store
applica-Alternatively, in a Push delivery system (the other type of VAS delivery system), the mobile phone operator is the initiator of the communica-tion with the user (i.e., actually it sends an offer
to the user) to stimulate the purchase of a specific product/service, or to have the user respond to
an offer In such delivery systems periodically mobile phone operators send text (SMS) and/or multimedia (MMS) messages to mobile phone us-ers that contain typically one or more commercial offers These offers invite users to subscribe or acquire services and/or to download digital prod-ucts (e.g., ring-tones, TV shows, video clips) that can be purchased directly from the mobile phone
Trang 3in a few clicks Messages sent to mobile phones
might also direct users to browse additional web
pages or download data over the phone, which can
also produce additional revenues depending on
the type of service contract Hence, in such Push
systems mobile phone users are not the initiator of
the communication and do not search for specific
applications or products they might need or desire
Mobile phone operators are actively engaged
in targeting users with specific offers (Wray &
Richard, 2009), and users only need to respond
to such offers Figure 1 presents an example of an
MMS commercial message sent to mobile phone
users that offers a wallpaper image for download
Mobile users can simply click on the message to
download the image and set it as wallpaper on
their mobile phone The cost of the service will
be added to their monthly bill or deducted from
their pre-paid account
Push Delivery System
Push delivery system is focused mainly by the
authors in this chapter Their objective is to
review and discuss how mobile operators can
actively optimize the delivery and targeting of
offers to their customer base The goal of
opera-tors is to maximize revenues by delivering the
offers with the highest profit potential From the
mobile operators point of view, it is noted that
the Push delivery system is in general very cost
effective Whereas lower telephony rates that attract new telephony customers place a direct negative pressure on company revenues, and may even produce a (tolerated) loss This type
of Value-Added Services represent an additional revenue source and tend to be associated to sig-nificant profits when properly managed The cost
of operations is often dominated by the one-time investment on the message-delivery infrastructure and, subsequently, each message can be sent at zero (or close to zero) marginal cost As a result, operators can easily reach millions of potential buyers at little cost making the profit potential
of these advertising-related services very high.Despite the great benefits mobile phone opera-tors can extract from these Push value-added ser-vices, their effective management poses significant challenges: operators need to target users with a selection of messages from a massive catalogue
of offers while facing limited testing capacity and heterogeneity in the content production process Recently, and in response to these challenges, researchers have developed new tools and meth-ods specific to this direct marketing channel that allow a more profitable use of value-added offers These tools and methods take advantage of the detailed logs of customer interaction with the offered services kept by current infrastructures These logs track all the messages and offers sent
to a customer and the corresponding feedback (e.g., whether the customer opened a message, viewed a page, bought a video, or clicked on a link) The information contained in these logs can then be used by an automated targeting system
to aid message selection and customer targeting.The chapter reviews and analyzes the chal-lenges faced by mobile operators in managing their VAS systems and discusses some of the methods available to improve profitability for the direct targeting activities of mobile phone operators engaged in the delivery of value-added services Based on the vast experience in implementing optimization systems in this area, the authors describe many of the experiments they carried out
Figure 1 An example of a Multi Media Message
(MMS) offer as shown on the mobile phone screen
Trang 4Also the findings, which the authors believe, can
aid mobile phone operators in the management
and design of their offers is also explained The
remaining of this chapter is organized as follows
Next the challenges faced by this new direct
marketing channel is described Then the findings
from previous research and from the authors own
experiments regarding the management of these
services is presented The chapter concludes with
discussing future areas of research in the mobile
marketing domain
CHALLENGES IN THE
MANAGEMENT OF MOBILE
VAS SYSTEMS
The management of mobile phone value-added
services presents several significant challenges,
which will be discussed in this section In the
following sections alternative methods that can
be employed to deal with such challenges will
be described
Massive Number of Value-Added
Service Offers and the Need
for Fast-Learning Methods
Because VAS are now a significant revenue
source, and central to profitability, mobile phone
operators and external production companies have
become increasingly creative and extremely fast
in generating new services and offers Virtually
anyone with computer skills can create digital
content to be offered to mobile users As a result,
production businesses have proliferated in the
market and provide new offers to mobile phone
operators on a daily basis In addition, traditional
media companies (music labels and TV networks)
quickly transform their existing products into
content to be delivered via mobile phones
As a consequence of these market features,
the number of alternatives that mobile operators
have available to send to mobile phone users is
now extremely large and growing quickly It is not unusual in this context to have tens of thousands
of possible products or services to advertise at any moment and, in most cases, the content catalogue grows by dozens of new items a day, a growth rate that is not likely to be reduced This massive num-ber of offers to be tested and studied poses some difficulties in terms of knowledge discovery For example, previously the direct marketing industry had used human-intensive methods to classify, optimize, and test different offers and then target these to specific individuals In the case of the thousands of multimedia messages available in current catalogues to be advertised to mobile users,
it is simply too costly, thus prohibitive to rely on human experts for their content classification and testing Instead, automatic systems that require minimum human intervention become essential.Finally, because of the sheer size and growth rate of content catalogues and because of the limited life of many of the offers (e.g., many of the offers expire in a matter of few days; some expire on the same day of their release or even
in a matter of few hours, as in the case of news videos), mobile operators face significant difficul-
ties in the implementation of standard pre-testing
methods Traditionally, companies have relied
on pre-testing to determine the best offers to be sent to specific target groups whenever facing a low cost of contact and a large target population (e.g., email marketing) (Nash, 2000) In such contexts, pre-testing is a simple and economical procedure that, in a nutshell, works as follows: alternative executions of a specific persuasive message are sent to different sub-samples from the target population; after a certain period of time, the responses from each execution are compared among themselves and the best ones are chosen for use with the rest of the population Because
of the massive number of offers that needs to
be tested quickly (before they expire), this task becomes either not feasible or ineffective in the context of mobile marketing
Trang 5Limited Contact Opportunities per
Customer and the Need for Targeting
Even though most mobile operators can contact
millions of customers, the number of
opportuni-ties to contact each customer is quite small In
order to send commercial offers to mobile phones,
many countries require the advertiser, content
producers, or the telephony provider to obtain
the receivers’ permission in advance (though the
requirements for opt-in or opt-out systems vary
from country to country) (Barwise & Strong, 2002;
Salo & Tahtinen, 2005) This factor significantly
reduces the total available customer base for
targeted offers
In addition, mobile devices are highly personal
instruments that users take with them almost
everywhere at all times Mobile operators have
recognized that if messages are not accepted in
advance, are not relevant to the receiver, arrive
at an inconvenient time, or too frequent, the
re-ceiver can easily regard mobile offers as illegal,
intrusive, and irritating (Wehmeyer, 2007; Ngai
& Gunasekaran, 2007; Barnes & Scornavacca,
2008; Barwise & Strong, 2002) As a result,
operators have now understood that offers sent
to mobile phones should not be based on a mass
communication paradigm Instead, in order to
avoid service cancellation or an operator switch,
only a limited number of messages should be sent
to individuals and these should be targeted and
personalized to the receiver’s needs Confirming
this belief, previous research has demonstrated
that few well-targeted messages are more
effec-tive than many generic ones (Bauer, Neumann,
& Reichardt, 2005)
As a result, today operators follow very strict
business rules that limit the number of messages
sent periodically to users In many typical real-life
applications operators have restricted to one per
day the number of messages that could be sent to
each user, though each company sets its own limits
and often adjusts these to the country in which
it is operating Some operators are
experiment-ing new business models in which the telephony service is provided free of charge in exchange for advertising exposure (i.e., mobile users can make calls and send text messages if they are willing
to be exposed to a certain number of daily ads) However, at the time this chapter is being writ-ten, reports from companies like Blyk in the UK and Mosh Mobile in the US that have adopted this business model are not extremely positive Recently, Blyk has been acquired by Orange who reportedly plans to offer students a range
of promotions, such as tickets and possibly free calls and texts, in return for receiving advertising
on their mobile phones (Wray, 2009) Even when
a message can contain more than one offer, the total number of offers per message varies typically from one to four due to the limited screen size of users’ handsets Hence, each person can only be exposed to no more than a very small fraction of all possible offers
Because of these limitations and constraints,
message targeting, which was once heralded as an
advantage of mobile marketing, has now become
a requirement in any VAS Push Management
System together with systems that allow for the optimization of message design However, with the reduced number of contact opportunities, these tasks (message targeting and design optimization) are also more challenging
Structural Limitations and the Need to Cluster Users
A third challenge associated with the targeting and knowledge discovery in the context of mobile value-added services relates to structure limita-tions Though each infrastructure might have different constraints, from the experience of the authors, current systems are typically restricted
to sending no more than a few hundreds of
dif-ferent messages a day Because each message
can be sent to thousands of different individuals, message delivery systems can reach millions of customers a day as long as individuals are grouped
Trang 6in a meaningful way (e.g., in clusters based on
previous response to offers) and all individuals in
a cluster receives a common message
These constraints might ease over time
How-ever, full customization and personalization (one
customized message sent to each individual) is not
yet feasible in existing infrastructures and it is far
from becoming feasible As a result, methods to
adequately cluster individuals and decide which
message to send to each cluster are central for
revenue optimization
Content Categorization
and the Need for Automatic
Categorization Systems
A final challenge that mobile operators face in
managing VAS relates to the different
catego-rization of offers used by each content provider
with whom the company contracts Because each
producer provides his own content, created
inde-pendently, each producer has also developed their
unique categorization schema and is not always
willing to change it For instance, a java game
from producer A might be classified in a category
called “Entertainment.” A similar java game from
producer B could instead be classified by that
producer as “Online Games.” Hence, the offers
coming from multiple producers can be assigned
to categories with very different names and with a
very different breadth (e.g., “Entertainment” as a
category will include many other types of offers,
not only online games)
The differences in name and scope of
vendor-specific categories pose another optimization
challenge Content categories could be powerful
predictors of purchase for specific groups or
in-dividuals given their previous purchase history
(similarly to applying collaborative filtering to
categories and users) Despite this potential, given
the way the category information is currently
collected by mobile phone companies, this
vari-able introduces mostly noise into the analysis It
is then necessary to develop approaches that can
overcome this problem to better learn message performance and decide on targeting and message optimization
In sum, the challenges that any Push VAS optimization and management system needs to overcome are significant However, the authors experience reveals that it is possible to design and implement systems that can deal with such challenges by relying on recent statistical and data-mining (Close, Pedrycz, Swiniarski, & Kurgan, 2007) techniques The authors have also conducted several experiments whose results can help mobile operators in the development of such systems and the design of their offers In the next section previous research in this area and the methods proposed to overcome the challenges discussed above is reviewed, and the results of some of the experiments is described
CUSTOMER CLUSTERING
One of the challenges in managing Push VAS services is that current systems cannot send a cus-tomized offer to each mobile phone user Instead,
in order to reach millions of customers, current systems need to deliver a common message to groups of users Clustering customers in a mean-ingful way is then essential to the management
of such Push systems The objective would be to group together customers with similar interests and then proceed to knowledge discovery, testing, and message targeting by taking into account and relying on these user clusters (Giuffrida, Sismeiro,
& Tribulato, 2008)
Behavioral Clusters
User clustering can be achieved using efficient clustering algorithms that rely on non-supervised classifiers and on customer-centric data, which might include demographic information and the previous response to commercial offers (i.e., previous behavior) As noted, however that in
Trang 7many real-life mobile applications demographic
information is often too noisy and sparse and, as
in the case of mobile phone pre-paid accounts,
might not be available all together
Hence, from the experience, clustering and
optimization systems that rely on demographic
information are often unreliable, especially when
compared with systems that rely on previous
re-sponse and behavior This result follows closely
what researchers in marketing have found both
in the online and bricks-and-mortar
environ-ments Indeed, previous research has concluded
that standard demographics information is rarely
predictive of consumer decision making Instead,
past purchase and consumption behavior provides
far better predictions of future purchases and
con-sumption (Eshghi et al., 2007; Montgomery, 1999)
In the previous applications, the authors have
relied successfully on user behavior, in the form of
purchase histories, to cluster successfully mobile
phone users Purchase histories can be represented
as a vector of dummy variables that specifies if
an item has been bought, or not, by the user in the
past; previous behavior can also be represented
as a vector of integers reflecting how many times
the user has bought from a specific offer category
Hence, it can be assumed that two customers
are similar (and should be placed together in a
cluster) if they buy similar content over time or,
more precisely, if they shop in similar categories
in a similar proportion Different strategies
ex-ist to discover customer behavior patterns from
such type of data (Sarwar, Karypis, Konstan, &
Reidl, 2001) but any fast and efficient clustering
algorithm with good scalability like the spherical
k-means algorithm (Dhillon & Modha, 2001a;
Dhillon, Fan, & Guan, 2001b; Zhong, 2005) can
be used (this is a particular version of the
histori-cal k-means (Mac Queen, 1967) and is based on
dot-product metrics that nicely fit with the mobile
marketing domain as discussed in Giuffrida et
al (2008))
Delta Clustering
The set of mobile phone customers that needs to
be clustered is not static or stable: new customers join the service, others discontinue the service, and still others make purchases; all on a daily basis Naturally that this will require that any system based on customer clustering takes into account these dynamics In the limit, customers might need
to be re-clustered on a daily basis, which might
be a costly operation depending on the algorithm used, the number of customers, and the number
of categories or items in the purchase history Based on the authors experience, changes in the customer based are very low probability events Because customer histories and customer status change very slowly, it is possible to overlook the evolution in the customer base over short periods and perform delta clustering without any signifi-cant loss in precision (Giuffrida et al., 2008) It can be re-assigned, each day if necessary, those users with new purchasing activity in the previous day; it can be started from the status of the latest cluster execution and use the centroids found in the latest run as a starting point (after the new purchase data is collected)
Cluster centroids, and a truly full clustering run, are conducted only over larger periods of time (e.g., every two weeks) This allows the considerable reduction of the execution time needed to analyze the data The new clustering schema will include the recent users’ activities, and depending on the purchasing of a specific content, a user might switch to a different cluster that in this new run shows a greater affinity with her new purchase history Keeping clusters stable (or almost) for longer periods of time also provides additional benefits: not only does it reduce computation time,
it also reduces the likelihood of sending multiple exposures of the same message to a significant number of users Indeed, when customers with different past viewing histories are re-grouped together, it becomes more difficult to satisfy the no-multiple-show condition Also, frequently
Trang 8changing customers might lead the system to
discard a good offer too frequently, just because
a significant part of the cluster has seen it before
Hence, in the applications used in this chapter,
the authors typically make a trade-off between
how often to perform a complete re-clustering
and how long to maintain the population within
each cluster (relatively) stable This is however an
empirical question that can be investigated with
some experimentation (e.g., It is able to define
an adequate frequency for re-clustering after few
trials only)
Managing Non-Clickers
One of the problems with clustering mobile phone
users based on their previous behavior is that, at
any point in time, there is always a significant
por-tion of mobile users that never buy anything, that
is, never click on the offers (called non-clickers)
For example, in one of the previous applications
only about 35% of the population had purchased
something in the past (called clickers), whereas
the remaining 65% had never purchased anything
(non-clickers) As a result, only use the activity
of a minority of the mobile users to perform the
clustering could be used For the majority of the
users (non-clickers) historical information is not
provided
To try to get usable information from
non-clickers, previous researchers have proposed
simple heuristics that have performed well in
real-life applications For example, in Giuffrida
et al (2008) the authors send good offers to
non-clickers, that is, non-clickers are targeted with
offers that tend to perform well overall, among the
entire clicker population (regardless of the
clus-tering schema) In addition, and to avoid pushing
only few offers, the authors split the non-clickers
group into smaller sets (in their case each subset
had about fifty thousand users) Then, the authors
target each set of non-clickers following the
empirical purchasing likelihood computed from
the clicker population By doing this the authors
also reduce the risk of picking one bad offer and sending it to a large number of customers Note also that each new customer, upon arrival, needs
to be first inserted into a non-clicker set The customer will then be assigned to clicker groups (through full clustering or delta clustering) as soon
as he/she makes a purchase The results reported
in Giuffrida et al (2008) show this method works extremely well
Number of Clusters
The task of choosing the right number of clusters
k is always a challenging one (Sugar & James,
2003) This depends on many factors such as customer base size and number of categories In general, a large number of clusters produces a more precise targeting However, a large number
of clusters requires a longer clustering execution time and data preparation time, larger storage space, and a longer message delivery process Notice that sending messages to many clusters
is time consuming, as the delivery engine has to pause for few seconds (or even minutes) between two consecutive deliveries (for technical reasons)
In addition, for marketing reasons, most mobile operators require that all customers receive mes-sages within a well-defined time frame Hence, any optimization and targeting system needs to make sure that the number of cluster is small enough not
to extend for too long the delivery phase.The final choice on the number of clusters depends upon the available storage, computation power, and the gains that adding further clusters might provide in terms of predictive accuracy In the previous applications authors have weighed all these factors and monitored the clustering per-formance as a function of the number of clusters
to make a decision of how many clusters to use For example, the spherical k-means clustering algorithm has an objective function one wants
to minimize The authors graph the value of this function for different numbers of clusters and then decide on how many clusters to use Fig-
Trang 9ure 2 shows the value of this k-means objective
function for the clustering of a real database of
mobile-phone users periodically targeted with
commercial messages The commercial messages
could be classified in one of 12 mutually exclusive
categories (these categories were obtained using
a text-mining method similar to the one which is
described in the subsequent section) The
catego-ries considered are: ring-tones, vocal ring-tones,
wallpapers, videos, songs, news, games, calendars,
services, promotions, sports, and multimedia
(User-specific 12-dimentional vector of purchase
frequencies is used to cluster individuals.)
As it can be seen from Figure 2, using about
20 to 30 clusters provides very good results:
performance improvements beyond the 11-cluster
solution are minimal, and improvements beyond
a 20 cluster solution are practically inexistent In
an application like this, unless there were
techni-cal problems of relevance (e.g., storage and
de-livery time) one would select about 20 clusters to
be used in a real system
Visualizing and Interpreting Clusters
To get a better understanding of the clusters tained, it is possible to use several visualization tools Figure 3 provides an example of a graphical representation of the outcome of the user cluster-ing with 20 clusters
ob-In Figure 3, the first line represents the clusters Each column, coded with two shades of green for easy differentiation, represents a cluster and the width of the column represents its size There are
20 columns, one for each cluster, and clusters are listed from the smallest to the largest The remain-ing lines represent the product categories and in the intersection of a cluster and a category the authors have coded the affinity between the two
Hence, given a row r and a column c, the element [r,c] represents the affinity of cluster c to catego-
ry r, and the darker the stronger this affinity
(af-finity is coded in different shades of grey, from almost white to almost black) For example, the darker elements of the matrix indicate a very strong affinity, meaning that all the users of that cluster have bought from the corresponding cat-egory Very light grey indicates a weak affinity—
Figure 2 Clustering quality as a function of the number of clusters
Trang 10customers of that cluster were not interested in
that category
Wider clusters are strongly associated to one
category and, as a result, are well defined in terms
of possible targeting strategies Clusters depicted
in columns 15, 17 and 20 are good examples:
us-ers in these clustus-ers bought products in only one
category Smaller clusters present strong affinity
with at least one category, though often with more
than one Only the clusters depicted in second
and sixth columns have less defined targeting
strategies: their users bought products in almost
every category
To get a better understanding of how
custom-ers cluster together as a result of their purchasing
history, user clusters have been depicted using
Self-Organizing Maps (SOM), (De Hoon, 2002)
which depict the customers’ vectors from an
N-dimensional space into two dimensions (the
representation is such that if two items are close
to each other in the N-dimensional space they
will be close also in the two-dimensional space)
Figure 4 represents the user density in a 2D space
with respect to all the categories The color scale
shows the maximum density area in dark red and
the minimum in dark blue (each image is
normal-ized with respect to the size of the corresponding
category)
This type of graph provides further rich
infor-mation on the 20 clusters For example, the three
categories in the first column have dense areas
(dark red groups of users) that are wide and not
well defined, surrounded by low density areas
group of people (shown in cyan) All the others
categories have smaller dense areas, well defined,
and surrounded by dark blue areas Categories such as ring-tones and games have some overlaps (the big cluster of users in the Games category is
in the same area as the dense cluster of users sociated with ring-tones) meaning that a subset
as-of their customers are interested in both types as-of products In contrast, sports and news have little overlap, with few common customers This initial analysis provides the first insights into how user respond to the offers and how to possibly target them There are however other tools that can significantly help in this task
Figure 3 Affinity matrix representation
Figure 4 User concentration over categories
Trang 11LEARNING ON NEW OFFERS
Though mobile user clustering resolves some of
the challenges, it does not provide an answer to
many others One of the challenges that clustering
does not resolve relates with the need to acquire
knowledge on a very large (and growing) content
catalogue Every day dozens of new offers are
added to the catalogue of mobile operators To
optimize targeting decisions, mobile operators
need to learn how likely users are to respond to
each offer and who (or which cluster) is likely
to respond Such learning needs to be performed
while dealing with the challenges which are
described previously and using the often limited
information available to mobile operators The
mobile operator might know, for example, the
of-fer’s category, as defined by the content provider,
the content (e.g., image and text), and the price of
the product or service being featured For all new
offers mobile operators do not know how they
have performed (as they have never been tested),
though operators might know how mobile users
have purchased in the past (if exposed to an offer)
and the performance of offers previously delivered
In some cases mobile operators might know also
the demographic information of mobile users,
though such information might be too unreliable
and, as in the case of pre-paid accounts, it might
not even exist
In addition, the learning phase in these
opti-mized Push delivery systems should be as
auto-matic as possible, requiring minimal human
inter-vention and ideally, they should run unsupervised
Fortunately, recent research has proposed several
automatic methods to improve the learning on new
offers that can rely on the limited information set
available to mobile operators Next the authors
reviews some of these methods and explain how
they can be implemented in real systems
Using Heterogeneous Category Information in Performance Prediction
Category information can be highly valuable to infer the purchasing likelihood of certain groups
of mobile users in the absence of actual purchase histories specific to each new offer (mobile operators do not know how each new offer will perform before testing it or sending to the entire user population but they might know how offers
of the same “type” have performed in the past)
If mobile users have purchased in the past from specific categories (e.g., ring-tones or games), it
is likely that they will keep on buying in those categories (Fennel, Allenby, Yang, & Edwards, 2003; Montgomery, 1999) for analyses in which previous behavior is a very good predictor of future behavior) Category information also al-lows researchers to learn on “types” of offers instead of learning on specific offers by applying sophisticated statistical or data-mining models
on categories instead of individual offers Also, when learning on categories of offers (instead of specific offers) it is possible to use the acquired knowledge on other new offers of the same type and researchers can capitalize on having more information available by pooling together offers
of the same type When learning on specific fers the knowledge is lost once the offer expires.Despite the potential information contained in offer categories, there are two challenges when using these in predicting offer performance First, with the categorizations different vendors provide, mobile phone operators get diversed In addition, the library of offers is extremely large (as compared to the learning occasions) and expands
of-at a significant pace, making it difficult to use a human-based labeling to create a common labeling for all offers To solve these problems previous research has proposed the use of a common and finer categorization of all offers that is generated
by an automatic system (Giuffrida et al., 2008).To obtain this categorization the authors in Giuffrida