taupesalmonInternet και Εφαρμογές Web

21 Οκτ 2013 (πριν από 3 χρόνια και 5 μήνες)

74 εμφανίσεις


Levent V. Orman

Cornell University
, Ithaca NY 14853


Commercial transactions have

rich semantics that is largely ignored by
the commercial data base

ing the

semantics in a machine
searchable form would create an infrastructure

that can
support a variety of novel
applications, ranging from recommendation systems and consumer
communities, to experimental and virtual organizations. Recommendation systems can be reduced to
mere queries against rich semantic transaction databases. Co
nsumer tasks and consumer goals can be
expressed unambiguously as collections of

and semantic descriptions of these large
regates allow

consumers to subscribe to the aggregates

instead of executing individual transactions,
significantly r
educing transaction costs.

Incentives can be designed to encourage consumers, businesses,
and third party experts to contribute to the creation of such databases, by creating an environment that
generates economic opportunities from small contributions. Fi
nally, organizations can be described
unambiguously as collections of transactions they executed in the past, and as collections of transactions
they plan and anticipate for the future. Such descriptions can be used by the various stakeholders to exert
ict controls over the organization at various levels of detail.


commercial transactions, transaction semantics, semantic data model, semantic queries,
recommendations, virtual o



A commercial transaction is an
exchange of resources. In modern economies, transactions typically
take place in physical or electronic marketplaces, and they involve

exchange of cash or credit for
various good
s and services. C

is often

provided by third parties such as banks and credit card
companies through
payment systems. More complex transactions involving multiple parties, and
delayed or repeated exchanges of various resources

are also common. A commercial transaction may
nvolve individuals, firms,
nonprofit or government organizations, or even take place entirely within an
A transaction can be a spot transaction with the exchange taking place completely at a
certain point in time, as in cash purchases of cons
umer items
. I
t c
an be a delayed transaction

where the
two sides of the exchange are not synchronized, as in credit sales or deliveries where the payment or the
delivery of goods comes at a later date than the sale. It can
be a protracted transaction w
ith multiple
exchanges taking place over a period of time,
as in employment transactions, where the value is generated
over time, and the payments are made periodically


Commercial transactions are recorded carefully in the databases of the participati
ng firms for
reporting and auditing purposes. Even individuals tend to track their transactions, and sometimes keep
records, for the purposes of tax computation and reporting, settling disputes, or budgeting and planning.
But the information captured about

transactions is minimal, and the rich semantics of transactions often
remains buried in the minds of the participants.
A typical commercial transaction may record what was
exchanged and who the participants were. But, the rest of its semantics is not like
ly to be recorded
anywhere, such as the purpose of the exchange (i.e.
did you buy

that book as a gift

or for yourself), the
context of the exchange (i.e. was it a birthday gift for your ten
year old nephew), the expectations from
the exchange and the exte
nt to which those expectations were satisfied (i.e.
did you expect it to be a
children’s book, and was it), the relationship to other transactions, past, present, and future (i.e. did
have the book shipped, did you have the cover engraved, did you also

he second volume later when
your nephew liked the first), and other transactions that were considered in lieu of this one, but rejected,
and the reasons for it (did you consider buying a CD instead, but rejected it, and why). Such semantic
content ca
n be very useful, if captured and recorded in a searchable format, in a variety of applications

ranging from recommendations, advice, and planning, to reducing transaction costs, automating some
predictable transactions, and building flexible and experimen
tal organizations.
That is the objective in this
article through semantic mode
g of commercial transactions [23



Semantics has received a great deal of attention in the context of the Semantic Web [
of web documents has been
critical to the development of WWW, and more effective description

semantic models is an ongoing concern to facilitate

more effective search and utilization of the web
resources [
11, 12

Extended Manipulation L
anguage XML was a cr
tical developmen

It created a
tagging system for document components, attaching keywords to document components, which in turn
allowed querying for document compon
ents on the basis of those tags


The tags can be viewed as
attribute names attached to values where th
e values are document components, leading to a collection of
value pairs. For example the text “Ford Focus” can be tagged with the keyword “car model”,
leading to an attribute
value pair of “car model
Ford Focus”. Such attribute
value pairs have
been the
basis of many semantic models ranging from semantic networks to
data models

22, 23, 24

value pairs of XML can be arranged hierarchically to represent the hierarchical arrangement
of document components

. For example, a car document and its Ford subcomponent are in a
hierarchical relationship. These document components are stored as XML databases and queried using
XML query languages such as XQuery and XPath that retrieve document components

when related
components in an

XML hierarchy satisfy some given attribute
value conditions.
For example, the fuel


of Ford Focus hatchback can be retrieved from an XML document about cars with the tag
sequence of

car. ford. focus. hatchback.mpg=?


a stylized language similar to XPath

where every node on the path is a node on a type hierarchy
or an attribute name
. A number of languages such as XPath are ava
ilable to query XML databases, and
imilar attribute
value based query languages have been ver
y successful with relational and object

databases also [

XML databases are supported by Data Type Descriptions (DTD’s) which
act as schemata to ensure consistent use of tags. DTD’s can also be extended to universal ontologies that
define t
erms, place them in type hierarchies, and attach constraints to limit the acceptable values [

XML provides an infrastructure on top of which more elaborate semantic structures can be built.
XML’s simple attribute
value pairs proved inadequate for
niversal identification and description of web
resources, and reasoning and inference with such widely distributed objects. Resource Description
Framework (RDF) was built on top of XML, and extended
with Universal Resource I
(URI) for unique
identification of all resources, and with entity
value triplets where entities are
uniquely identified by URI’s and the values can be other entities [
]. Such a recursive definition
increases the semantic power of the model, and the complexity o
f the resources it can describe.
the resource descriptions have to be placed in central repositories to facilitate search, selection and
retrieval. This capability requires
earchable resource description databases, and many of these were
d such as the Universal Description Discovery and Integration (UDDI) framework

, 25
. They
all rely on XML for their database structure, and XML languages for search, selection, and retrieval.

This article follows in the tradition of semantic resource
description, but focuses on a specific
resource: commercial transactions. It sketches out a
semantic model to describe c
ommercial transactions
using an

RDF type model; it
develops the arguments for developing such a model, including its
advantages both for

business and consumer applications; and finally it lists technical and organizational
difficulties in developing such an extensive semantic system, and proposes solutions. As a first attempt in
designing such a large scale semantic system, it focuses on m
otivation, advantages, and general design
principles, rather than specific implementation
, scale, and efficiency issues.


Transaction Semantics

Consider the purchase of an automobile.
The transaction is recorded by the seller
, and the record
includes a desc
ription of the automobile, the sale price, the payment method, and the buyer’s identity. The
buyer may also record the transaction in his personal records for tax and record keeping purposes
including the price, the payment method, and the seller’s identit
y. The third parties such as insurance
companies and government agencies may also record the same data for insurance and registration
purposes. Such transaction databases of large businesses and government agencies grow quickly to
terabyte sizes, and becom
e difficult to use. But, despite their formidable size, they fail to capture some
important components of the transaction semantics, which contributes to their usability problems. For
example, they fail to capture and record why the automobile was purchase
d, how it was intended to be
used, and for what purpose

i.e. its goal semantics.
They fail to capture and record how it was used after
the purchase, and which objectives were satisfied and to what extent, i.e. its outcome semantics. They fail
to record an
d capture secondary transactions such as insurance, repairs, maintenance, accessories, and the
eventual replacement, and the relationships between the secondary transactions and the original
transaction, i.e. its bundle semantics. They fail to capture and
record what other


but rejected by this consumer

in lieu of this transaction, and why they were rejected, i.e. its
alternatives semantics. Finally, they fail to capture and record the conditions and the events that triggered
is automobile purchase, i.e. its time semantics.

Capturing and recording such information
in a searchable format
can have significant benefits.

semantic information can be queried by consumers in their search for detailed product descriptions and
commendations. A typical query would involve finding the most appropriate automobile for a given set
of requirements and consumer characteristics, by utilizing the goal and outcome semantics.

The semantic
information can also be aggregated to predict futur
e needs of consumers and the future demand for the
products and services of businesses under various economic conditions and triggers. Typical aggregate
information could determine the demand for a particular model of cars, in a particular community, at
mes of increasing oil prices, or increased sensitivity to environmental degradation, by utilizing the time
d alternatives semantics. The s
emantic information can be organized to identify and bundle groups of
related transactions, for specific consumers,
to reduce transaction costs.
A typical bundle would cluster an
automobile purchase with insurance and maintenance contracts, or even the eventual replacement of the
automobile, into a single transaction by u
tilizing the bundle semantics. The s
emantic infor
mation can be
analyzed to aid in the design of new products,
services, and new organizational forms. Some design
opportunities for an automobile dealer could be new leasing arrangements, new pricing schemes, or new
car sharing or customized transportation
systems, by utilizing the goal and alternatives semantics. The
semantic information can be
used to identify select consumers and third party experts who provide useful
relevant information to others, and reward them with leadership opportunities. Typical l
opportunities could be in the form of role models, consumer community leaders, and expert advice
generators, by utilizing the outcome and alternatives semantics.

A typical commercial transaction

records what resources are exchanged, and who the participants are.
These can be called “what” and “who” semantics. A complete semantic description would have to
elaborate on what and who semantics by describing the resources and the participants in detai
l, and would
have to add many more components

such as the goal of the transaction, called the “goal” or “why”


the use of the resources involved and the resulting satisfaction level, called the “outcome” or


the location of the t
ransaction, called the “location” or “where” semantics; the payment
mechanism, called the “payment” semantics; secondary transaction

supporting the original transaction,
called the “bundle” or “what else” semantics; and the alternative transactions that h
ave been considered
and rejected, called the “alternatives” semantics. All of these components can be implemented as the
attributes of an RDF resource, where the resource is a commercial transaction. The values attached to
these attributes can be quite com
plex, and require the full recursive power of RDF, where the attribute
values can themselves be entities with their own attributes, and with complex constraints and
relationships imposed on them. Such complexity is rare in
corporate data stores,



and it motivates the complex model proposed in this article.

The complexity is exacerbated by
the difficulty of collecting, structuring, searching, and utilizing such complex and detailed information
about transactions.

nsaction 655876



when where
payment what else alternatives








action object action object why not


Cinemapolis watch movie


ntic Cinemapolis cash


popcorn watch football





Bills Cowboys


too expensive too violent

Figure 3.1:
The model of a simple transaction of

John Smith going to the movies where the labeled edges
are attributes and the unlabelled edges are subtypes.

The basic model
of Figure 3.1
is a simple set of entity
value triplets such as transaction
entertainment that can be implemented as a relational database with commercial software that stores
and maintains the data and facilitates retrieval with its query lan
guage [
]. But there are a number of
complications. The first complication is the need to disambiguate the terms used

for universal
consistency in a large scale distributed system. The solution for en
tities is to use Universal Resource
dentifiers (URI’
s) as suggested

by Resource D
escription Framework (RDF)

. Figure 3.1 contains
URI’s at all unnamed nodes and also at the transaction node. The solution for attribute
value pairs is to
use ontologies, such as Web Ontology Language (OWL)

attached to RDF

. Ontologies place each
value into a type hierarchy, and attach to it a dictionary definition and constraints


Casablanca f
or example points to a specific movie in the movie ontology. The movie is identified
uniquely by a URI; it

is characterized by the type hierarchy of the ontology as to its genre and type; and it
is described by a set of attributes of the movie entity to identify its actors, directors, producers, awards it
won, and the reviews it received.

These features can be

implemented with a relational database for the
value pairs, OWL ontologies for defining attribute values unambiguously

in dictionaries and
type hierarchies
, and special pointers inserted into the database linking database values to the ontolog
entries, as shown in Figure







love story

adventure story




Humphrey Bogart

Ingrid Bergman



Figure 3.2:
Extending the Relational database with pointers to ontologies to disambiguate terms

to describe them

The second complication is
the existence of ad
hoc and multi
valued attributes. For example, some
movies may have awards and reviews as attributes; others may have musical score and animation
technology as attributes; some do not have any, and some have them as multi
valued. Some at
may be very sparse, or even unique to individual movies such as the cost of the airport scene in
These ad
hoc attributes can be implemented as RDF entity
value triplets, as
relational databases would have great difficulty acc
ommodating them. An additional difficulty here is the
need to link the relational and the RDF components, and to search them simultaneously to provide
complete responses to questions. This task can be accomplished by creating a high
level RDF model that
compasses both the relational and the RDF components, and the queries are posed to this RDF model
first, and then translated into queries against the components for more efficient processing. This is a
classic distributed and heterogeneous data store examp
le which characterizes most web services, and the
research in this area is very active [



award award airpoort scene


best film

best music



Cannes Oscar $6M

Figure 3.3:
The model of a movie containing ad
hoc, multi
valued, sparse, and unique attributes.

Finally, the third complication is the attribute values being entities or even transactions
themselves, leading to a recursive model.
Both bundle semantics and alternatives semantics are
implemented as attributes of transactions which are transactions them
selves, and they point back to the
transaction database to identify related transactions. Alternatives semantics consists of virtual
transactions, since alternative transactions are transactions that have never taken place, but merely
considered. The
se alt
ernative transactions

also have “why not” semantics that explains the reasons for the
rejection of those transactions.
These recursive relationships with ad
hoc and sparse attributes can be
implemented as RDF triplets, and stored in RDF data stores


transaction 655876

transaction 876722


action object why not why not



too expensive

too violent

Figure 3.4:
The model of alternative transactions considered by a consumer in lieu of watching a

The complete semantic model then is an RDF schema

whose implementation contains relational and
RDF components and ontologies with linkages among all three.

In effect,
the implementation involves a
relational database extended with an RDF data store to accommodate ad
hoc, sparse, and recursive
. It also involves pointers embedded into the relational database and the RDF data store, pointing
to OWL ontologies or
recursively to
other relational
RDF components.

The pointers to OWL
ontologies are implemented as XML path expressions that point t
o specific locations in the XML type
hierarchies of the ontologies, which identify attribute values unambiguously, and attach constraints and
dictionary definitions to them. The pointers to other relational records or to RDF triplets contain merely
the URI
’s of those entities since all entities, whether they are relational or RDF, are identified uniquely by
their URI.

Efficient implementation of such an extensive semantic database is a challenge, especially because its
large scale exacerbates its structura
l complexity.
It could be implemented wholly as an RDF data store of
value triplets and associated OWL ontologies, but RDF data stores do not scale up
efficiently. Implementation as a combination of relational databases and RDF data provid
es a more
efficient solution by separating the highly structured transaction data with pre
specified attributes from
hoc data with unique attributes and sparse content.
Highly structured standard transactions can rely on
relational databases and their e
fficient indexing structures. The only source of inefficiency
here is the
data values that are XML path expressions pointing to OWL ontologies.

Entering such complex values into a relational database, and indexing and querying them efficiently
are signif
icant challenges.
Such an expression can be entered directly into a database,
or it can

be entered
as a pointer to
the right

in the ontology
A value in the transaction database then can be an XML
path such as outcome.feeling.romantic

pointing to
a location i
n an OWL data type
hierarchy about
feelings and their causes.
Entering a pointer would have to be supported by a user interface that allows
navigating type hierarchies of ontologies. Entering path expressions would require resolving ambiguities

when only a partial path expression is entered, or superfluous nodes are specified. For example the
complete path above may be transaction.outcome.feeling.romanatic, or feeling.romantic depending on
ow the ontologies are organized, and the correct expres
sion has to be derived automatically from the
partial expression entered by the user.
Alternatively, the user may have to be guided in entering such path
expressions, whether they contain new values, or point to existing ontologies. Path expressions may ha
o be entered one node at

and at each step, the users are presented with a set of possibilities from
the existing ontologies, and also given an opportunity to define new values that do not violate the type
restrictions. For example, the outcome
attribute of a movie watching transaction has a number of possible
values, one of which is “feeling”. The users need to be presented with a good collection of possible values
when they arrive at the outcome node. For a different transaction involving groce
ry shopping,
feeling may
not be an existing relevant attribute. The outcome for that type of transaction may involve dinner recipes,
and they in turn may involve a feeling attribute. The options presented to the user, and the sequence of
those options, are

specific to each transaction type, and the choices at each step limit the choices available
in subsequent steps. Such navigation of ontologies is critical to populating the database with correct path
expressions and pointers.


Queries and Recommendations

transaction databases is always a challenge because of their size, and the ex
aggregation needed

to derive meaningful
Increased semantics proposed in Section 3 not only
adds to their already significant size, but also dramatica
lly increases their complexity, especially by
combining relational and RDF databases and ontologies in one system. The user interface does not have
to reflect that complexity, but query processing poses many challenges. For example, a simple query to
all transactions of John Smit
h can be expressed as

a simple path expression like

transaction. (who=”John

Smith”, what=?)

but requires significant processing against a complex structure. The first task of a query processor is
generate complete and unam
biguous path queries from the incomplete user expressions, by utilizing the
ontologies. This could be done either by searching for user path segments in the existing ontologies, or by
interacting with the user to navigate through the relevant portions
of a
n ontology, or both. Such a process
may convert the simple expression above to a complete path such as

transaction. (who = employee. engineer. electrical

. employee216. (name=John Smith”), what=

The second task of query processing is translating such path expressions
that are
based on a
global schema, to queries against individual RDF of Relational databases, and combining the results from
individual databases to respond to the global query. This
task is the standard distributed and
heterogeneous database processing problem [

The third important task of the query processor is
the need for large scale aggregation.
arehousing technology has been successful in transforming transaction databa
ses i
nto smaller aggregate
databases, called warehouses,
for efficient querying [

Such pre
processing is effective, but requires a
deep knowledge of aggregation requirements, and a separate design for the aggregate database
ntologies can be used to c
apture the aggregation requirements, alleviating the need for a separate
warehouse design. They can also be used to guide the users in querying the aggregates, since ontologies
capture the semantics of useful aggregates in their type hierarchies.

For example, all transactions
executed by engineers can
very simply located by the expression


by first finding all engineers from an employee ontology, and then finding the transactions executed by
those employees. Similarly,

he total amount of transactions by engineers can be expressed by

+ transaction. (payment.amount=?, who=engineer)

The three tasks of the query processor
can be observed in a simple query to find all transactions
executed by engineers whose spouses have
health problems:

transaction. (what=?, who=engineer.



which would first locate all types of engineers from the ontology of engineers;
then it would extract all
transactions for those engineers from the relational database of transaction
s; and finally, it

would locate
the ones with unhealthy spouses from the RDF data store, since the spouse’s health is an ad
hoc attribute
not included in the relational database.

Such extensive semantics can be queried
to get answers to complex questions
about the
consumers’ expectations and preferences with various products and services, and as such it can be used as
an elaborate recommendation system

, 16
. This is a fundamentally new approach to building a
recommendation system, since the standard rec
ommendation systems start with transactions with minimal
semantics, and recommend new transactions. The
re is an

intermediate step of computing additional
semantics by aggregating rat
ings and rankings of consumers which is invisible to the consumers. As suc
standard recommendation systems often appear mysterious, and have difficulty explaining their
recommendations, except in statistical terms.
The approach in this article starts with extensive semantics
and recommends new transactions directly from that
semantics. As such, there are no invisible
intermediate steps, and the recommendations are easily explained in terms of the transaction semantics.

The ability to query semantic structures leads to novel and sophisticated recommendation
systems. Existing r
ecommendation systems aggregate rankings to generate recommendations.
Aggregation can be over all transactions involving the item, or it can be over a subset of transactions
relevant to a particular consumer, such as the transactions by similar consumers.

Similar consumers can
be defined by matching attribute values, or as those who have executed similar transactions, or ranked the
same items similarly. All common recommendation systems rely on rankings, but rankings have serious
problems. Their semantics
is ambiguous, as it is not at all clear what a rating of 5 means as opposed to a
rating of 4. Moreover, they are often attached to entities rather than specific attributes, increasing the
ambiguity about their meaning

, 18
A rich semantic structure moves the evaluation to the attribute
level, and provides a detailed description of the attribute values to replace numeric rankings. In this
the consumers can pick the relevant attributes, and specify the desirable
values for those
attributes, rather than mere numeric rankings of the objects. For example, a movie can be evaluated
specifically on the basis of its story line, and the desired value attached to that attribute can be as specific
as a deeply romantic story

line, or as a slapstick funny story line, as opposed to a mere numeric value.

There are three methods of building
a recommenda
tion system on top of transactio
semantics. The easiest is to directly query the individual experiences of consumers with
arious products
and services. A more complex approach is to
compute and
query aggregates that reflect the general
sentiment of the consumers about various products and services.
The third and the most sophisticated
approach is to construct communities of c
onsumers, and to aggregate experiences within the relevant
community for each consumer query.

The first approach would involve simple queries about individual experiences, such as the
movies that JohnSmith thought had a romantic story:

.movie=?, who=John Smith,





Aggregates can also be similarly expressed by defining aggregation operators such as ALL, SOME,

MOST, and FEW. MOST(x,y,z) is defined as returning x, for which most y satisfy z.

To find the movies
most people thought had a romantic story, one would use the MOST operator as follows:

x, y=transaction.


movie=x, z=







Partial aggregation over communities of consumers is more difficult to compute and express in general,
but ontologies provide natural aggregation at multiple levels, and query processors can take advantage of
that structure. For example, the movies that mo
st engineers thought had a romantic story can be expressed
similar to the previous query:

MOST(x, y=transaction.


movie=x, who=engineer), z=







nd the query processor resolves the value “engineer” by consult
ing the ontology and identifying all
instances of engineers.

The major disadvantages of such a semantic recommendation system are the difficulty of
aggregating non
numeric values, and the difficulty of querying such complex semantics.
nonnumeric values can be accomplished by utilizing aggregation functions such as MOST, SOME, ALL,
NONE, ANY, FEW, and MANY as introduced above. They allow calculating the general sentiment about
an attribute

aggregated along various dimensions and a differ
ent levels
, allowing for example
to find

movies that most people found to be
romantic, or to find

movies most college students found to be

or even
the movies
most Cornell sophomores on a date found to be romantic. The querying of
aggregations can

be greatly simplified
by pre
computing the most common aggregate attributes

attaching them to the
objects in

For example, movies in
a movie

ontology can
have an attribute called “generated
feeling” that can be computed by aggregation from the
outcome.feeling attribute of all the transactions involving those movies. Such aggregate attributes can be
queried quite simply as if they were inherent

attributes of the objects involved. For example romantic
movies can be located simply by the expression

movies. generated

Even multi
attribute aggregations can be partially pre
computed along each dimension, and then
combined on
the fly to respond to compound queries. For example the movies that are found to be
romantic by college students can be expressed as:

movies. feeling
generated. (what=romantic, who=student. college)

as long as what and who attributes are pre
ed for movie feelings as a two
dimensional data
aggregation. Such attributes can capture general sentiment about goods and services in great detail and in
various levels of aggregation. Efficient implementation of such partial aggregates
has been a topic o
study in Data Warehousing. The literature on Data Warehousing is extensive on efficient methods for
instantiating and pre
computing such aggregate data, and that literature provides many effective solutions
and commercial products to resolve this problem



Virtual Transactions

A virtual transaction is a fictitious
transaction, relevant to specific consumers under specific
conditions. Virtual transactions can be computed from past transactions as recommendations, or they can
be created directly by vend
ors or task experts as advice or advertisement.
The identity of these creators of
virtual transactions is captured by an “author” or “by who” semantics, which is implemented as an
additional attribute of virtual transactions. “
y who” attribute is differen
t from the “who” attribute, since
the person creating the virtual transaction,

i.e. making the recommendation, is distinct from the person
expected to execute the transaction, i.e. the receiver of the recommendation.

Virtual transactions

can be aggregated

into various collections called “stereotypes” or “role models”.

A stereotype is a collection of virtual transactions that are
typically needed

to execute a given task.
collection can be easily c
aptured by utilizing the why,

what else
, and the “by who”

semantics of
, where the “why” semantics captures the task; and the “what else” semantics captures all the
other transactions needed for that task. The creators of stereotypes are typically experts on the task, and
their identity is captured b
y the “

or “by who”

semantics of
transactions, if they choose
to reveal it.

Stereotypes not only capture all
the transactions that are necessary to execute a task, but also the
alternative transactions, and the conditions under which
those alternatives

would be more appropriate,
using the alternatives semantics. Consider for example, t
e task of attending a professional conference.
requires many transactions ranging from registration, flight, and taxi arrangements, to hotel and r
reservations. All could be captured in a single stereotypical aggregate transaction, which allows the
conference attendees to execute the whole aggregate as a single transaction, and as a result,
it can reduce

the transaction costs considerably.
Variations on the stereotype can be captured by the alternatives
semantics, to accommodate the individualized needs of the attendees, such as the departure city of the
individual attendee, a different class or type of hotel for the individual lifestyle, or

a different cuisine of
food for the individual taste. Such individual variation

can be accommodated by starting with the
stereotypical solution, and incrementally modifying
to suit the indi
vidual needs, by using the
alternatives semantics. For example
, the stereotypical solution for the conference attendee might involve a
flight from New York, yet those who are arriving from Chicago can modify the stereotype minimally to
change the departure city to Chicago, and find another aggregate that is increment
ally different yet
internally consistent. Alternative

semantics would capture the alternative citie
s of departure such as
Chicago. What else semantics

automatically change

all the related transactions
in the bundle
such as the
ground transportation to th
e airport
. It
ins the consistency of the aggregate by utilizing the


attributes that link together the consistent bundles of transactions

A role model

is a collection of
transactions that
are associated with a specific type of
sumer. Role models are useful in quickly identifying all transactions that are relevant to
a lifestyle,
demographic, or profession
. Consumers can subscribe to role models and easily locate all transactions that
are relevant to that type of consumer, or
even automatically execute them. The role models are captured
by the “
” semantics of virtual transactions. For example all transactions that are relevant to
Computer Science sophomores are


identified by
the ‘

semantics that


o the “
computer science sophomores” entry in

the student ontology.
computer science sophomores can follow the links from the ontology to the transaction database, to find
all relevant transactions.
Not all trans
actions associated with a particular role model have to be executed
at the same time, nor are they all relevant to all in the group, and under the same conditions. Where,
when, who, and alternatives semantics can capture a variety of transactions under the

same role model,
but appropriate at different times, under different conditions, and to different individuals within the group.
“When” and “where” semantics captures variation with respect to location and time; who semantics
captures the source of the inf
ormation defining this role model. Further v
ariations on the role model can
be captured by the alternatives semantics, to accommodate the individualized needs of the sophomores,
such as the specific disabilities or talents they may have, and specific advan
ce courses they may have
taken as freshmen.
The search starts with a collection of standard transactions
provided by

the role model,

to customize,

incremental modification

are made through when, where, who, and alternatives
semantics. At each step “wh
at else” semantics returns a consistent bundle, but slightly modified from the
original bundle.

For example, a Cornell Computer Science sophomore may start with the standard set of
transactions defined by the role model, and then modify the set slightly by

selecting the concentration as
, or

the emphasis as Java programming from the

limit the location to

ampus by using where semantics, and picking the transactions entered only by graduate students
by using the “
” semantics.

The transactions of the role models can be viewed as recommendations or subscriptions. They can be
presented as recommendations for individual consumers to review and select; or they can be
automatically executed for the subscribers. For exam
ple, university professors need to attend conferences,
buy books, and subscribe to scholarly journals. All those transactions can be aggregated in the persona of
a role model in each field, and the specific professors can subscribe to the transactions of t
hose role
models, either to receive as recommendations or as automatic executions.

An existing

example of a role
model is a mutual fund manager. Individual investors subscribe to mutual funds on the basis of their
consumer characteristics and investment go
als, and as such they authorize the fund manager to execute
transactions on their behalf. In this example, all transactions of the role model, i.e. the fund manager, are
automatically executed for each subscriber.

Transactions have complex and detailed s
emantics, and defining and entering them into automated
systems is not a trivial task. The consumers need to have strong incentives to define and enter such
complex semantics into automated systems; and they need intelligent and helpful interfaces designed

ease the task.
Stereotypes and role models can be used to aid in this process, by designing them as
economic entities, and allowing them to

from the

they define, both real and virtual.
Real transactions

are historical records; virtual


recommendations for the future. A role
model defines a consumer both in terms of the real transactions he executed in the past, but also the
virtual transactions he intends to execute in the future.

ly, stereotypes define tasks in terms of the
real transactions executed by an expert to accomplish the task in the past, and also the transactions he
recommends for this task in the future, at different times, places, and under different conditions.

In all

these cases, it is possible to provide incentives for the consumers and task experts to enter their
transactions into automated systems, by creating opportunities for them to convert themselves into
economic entities
, and financially benefit from their

. These opportunities are analogous to political
activism and

that provide
incentives to participate in the political process as foot
soldiers, as preparation for later leadership roles.

There are no similar opportunities for indiv
iduals in the
economic arena, short of starting a new business venture. Semantic transaction databases provide such an
opportunity for small
scale economic participation, with possibilities for expansion into new business
ventures, or

development towards l
eadership within existing businesses.

Those who provide such detailed
transaction information might do so, if there is a potential for future economic benefit, as others view
their transactions and find them useful as recommendations or advice.

The most o
bvious strategy for economic benefit
is in using the exposure to consumers for advertising
or for referrals. If a significant number of consumers find one’s transactions useful and reliable, then one
can monetize that consumer traffic either by advertising

to them and collecting fees from advertisers, or
by making referrals and taking a commission from every referral. Consider a medical doctor who
provides virtual transactions
in terms of specialists recommended for a variety of illnesses and symptoms,
real transactions describing his past referrals to specialists and their medical outcomes. Or consider a
homemaker who records her actual transactions for children’s products, and also her virtual transactions
recording her recommendations for the future.
Others may accept them as role models or stereotypes, and
follow their transaction recommendations for their specific circumstances, and the role models and
stereotypes can collect commissions from each referral.

Such a model could easily create the incent
necessary for many to record and describe their transactions in great detail. Google’s advertising model
attached to its search services is similar to this model, and transaction databases may allow the individuals
with narrow expertise or specialized

knowledge to benefit from the same model


A second strategy for economic benefit is to act like a retailer, and execute bundles of transactions on
behalf of consumers who subscribe. The subscribers can delegate transaction decisions completely, or
ey can receive
them as recommendations, and

before accepting. This strategy allows every
individual to act as a retailer

for like
minded people, acquiring relevant goods and services for them,
executing their transactions, and even occasionall
y holding inventory and delivering
the goods and
services. Those who build a large subscriber base can benefit significantly, whether they make
recommendations, execute transactions on behalf of their subscribers, or deliver th
e goods and services
, because s
uch a highly distributed retail system has a number of advantages.
, it reduces
transaction costs for consumers, by providing reliable recommendations by familiar people,
and bundling
them into large collections. Second, it reduces the
transaction costs for businesses, by relying on peer
networks to advertise and inform the potential customers. Third, i
t creates a very flexible business model
with minimal start up cost, because it relies on past customers and numerous third party ex
perts to
evaluate the products and advertise them through a peer
peer network. Such a model is very responsive
to market changes and very quick and flexible to take advantage of market opportunities, because it relies
on a loosely connected network rath
er than a tightly controlled bureaucracy for marketing and
. The components can be shed very easily, and replaced by other components

even automatically with no central control from the business itself.
Fourth, the mode


a very
efficient logistics system. It allows a highly distributed and flexible logistics system, where individuals
can hold inventory, and make deliveries from their own homes. Such a flexible


with a
loosely connected network of individuals
acting as reta
ilers or intermediaries, is highly unusual.
the functions of
marketing, advertising, and even logistics
to many individuals, the system
takes full advantage of an information rich economy, and a full range of information techn

typical example would be a college student leader recommending various Spring Break vacation packages
to fellow students. It is a stereotype aggregate solution to a typical vacation need, and transaction
semantics is critical for a detailed desc
ription to capture the stereotype and its many variations that may
be appropriate for different student groups.
Yet, once the semantic infrastructure is in place, any
individual student can perform this role.
An existing example of this model is Amway’s pe
peer sales
system for
cosmetics and jewelry, where individuals act as role models and sales agents, advertise and
educate, hold inventory for demonstrations, and even make deliveries from their own inventory.
However, Amway’s model relies on face
face meetings and a very limited range of products.
Extending it to a wide variety of products and to remote web

is one possible
application of transaction semantics [
15, 19


Experimental Organizations

Organizations can be described in

terms of their past transactions (i.e. their history), and the
transactions they
anticipate (i.e. their plans). The descriptions include
both their external transactions with
other organizations and their customers, and also the internal transactions

among their sub
and employees. The descriptions may involve real or virtual transactions. Real transactions are
They are often past transactions describing the history of the organization, but they can also
be commitments to fut
ure transactions. The real transactions can be aggregated to define tasks and
organizational units in terms of the transactions they execute. Virtual transactions are normative. They are
future transactions that are planned, anticipated, and recommended. T
hey a
re us
ed by organizational
designers to describe the new and experimental organizational components.
These experimental
components may exist alongside the real components, and provide a lab to study new organizational
forms and initiatives. They can al
so be used to describe new initiatives to various organizational
constituencies to solicit feedback and support.
Various constituencies such as customers, managers,
employees, and shareholders can subscribe to the new initiatives, criticize and make sugges
tions, or even
formally commit themselves. For example, business partners can support new initiatives by subscribing to
future transactions. These subscriptions can be in the form of mere intent as in surveys, or they can be
contractual commitments or even

payments for specific future transactions

[20, 21

Virtual transactions can be used to describe new business ventures to investors, business partners and
Investors can decide whether or not to support the venture, or support only parts of i
characterized by a set of transactions. Customers may subscribe to the future products and services of a

or only some of
those products and services, or

under some conditions, or after some
modification to their design, all of which are

defined unambiguously by virtual transactions. Such
subscriptions to future products and services can be a source of funding for new businesses instead of
venture capital, if they are in the form of contractual commitments or pre
payments. Simil
consumer subscriptions can be used to gauge the potential market size for various products of the new
enterprise, in lieu of


marketing research.

Such concrete subscriptions are likely to be more
reliable than surveys and questionnaires, b
ecause they are more specific and detailed with much less
ambiguity, and may even involve contractual commitments and even pre
payments by consumers. They
can be much more refined than surveys, because the consumers can subscribe to very specific transacti
involving specific products and services, at specific time periods, and under specific
subscriptions can be used for marketing research, and specifically in the product design process, to
forecast the demand for new attributes in e
xisting products and services, or to design completely new
products and services.

Virtual transactions can be used to describe government services, and even complete government
ies as collections of
, so that the citizens can decide wheth
er to support those agencies
and their services with their tax dollars. The citizens can subscribe to specific

and reject others,
distribute their tax payments among various services on the basis of virtual transactions.
Such a
dynamic governm
ent organization allows citizens to exert strict control over spending by dynamically
allocating their tax payments among various government services, among various agencies, and even
among multiple governments at various levels. For example, the citizens
of a municipality may direct
their tax payments to an alternative sewer system provided by a neighboring municipality, or even a
private contractor. They may direct their tax payments to a private school system, or a church
charity. They may move the
ir support from the local police to the police force of a neighboring city, town,
or the state. They may shift their support from one agency to another, in effect voting with their taxes
whether to support various foreign, domestic, and local initiatives.
This could be accomplished by
committing actual funds to specific virtual transactions, or by using virtual transactions as a simulation to
gauge support for various services and their components.
Either way, they can be effective in organizing a
responsive government, with dynamic allocation of resources, and with an empowered electorate
involved at a very detailed level with the administration of the government




Commercial transactions have

rich semantics that is often ignored

transaction databases. Capturing

semantics in a
searchable form would create an infrastructure that can support a variety of
novel applications. Recommendation systems can be reduced to queries against such a system, and the

it produces can be quite refined and customized. Aggregation of such semantic content
can reduce transaction costs by customized large
scale bundling of products and services. All transactions
necessary to execute a task or to follow a lifestyle can be cu
bundled and offered as a single
transaction. Aggregation at multiple levels can be done by taking advantage of the type hierarchies of
ontologies, and aggregate data can be queried as simply as the raw data, leading to novel structures for
Data Wareho
The consumers can be incentivized to provide such detailed semantics, and enter them
into automated systems, by creating opportunities for them to turn themselves into economic entities and
profit from even small contributions to such sem
antic transa
ction databases.
Finally, t
he semantic models
of transactions
can be used to describe complete organizational units, and government agencies in terms
of the transactions they execute. As such, shareholders

of businesses

and citizens of


ert strict control on these organizations by
supporting or rejecting specific organizational units, or even
specific sets of transactions.
A complete implementation of semantic transaction databases, and the
related issues of efficiency and scale are left
for future research. Similarly, there is much work that needs
to be done in developing effective user interfaces to such systems.



Adomavicius G.
, Sankaranarayanan R., Sen S.,

Tuzhilin A.
Incorporating Contextual Information
in Recommender
Systems Using a Multidimensional Approach,

ACM Transactions on
Information Systems
23, 1
, 103
, 2005


Lee T., Hendler J., Lassila O. The Semantic Web: A New Form of Web Content that is
Meaningful to Computers will unleash a Revolution of New Po
Scientific American

May 2001.


Collins S.R.,

Navathe S.

Mark L. XML Schema Mappings for Heterogeneous Database Access.
Information and Software Technology


4, 251
257, 2002.



J., Storey V.C., Sugumaran V. Improving Web
Query Processing Through Semantic
Knowledge. Data and Knowledge Engineering 66, 1, 18
34, 2008.


Cordella A. Transaction Costs and Information Systems: Does IT Add Up?
Journal of
Information Technology
, 21, 195

02, 2006.


Davenport T. H. and Harris J. G.
Competing on Analytics: The New Science of Winning.

Business School Press 2007.


Erl T.
Service Oriented Architecture (SOA): Concepts, Technology, and Architecture
. Prentice
Hall, 2005.


Fensel, D..
ies: Silver Bullet for Knowledge Management and Electronic Commerce
Verlag, 2001.


Camino A. et al. Electronic Institutions,






Perez A., Fernandez
Lopez M., Corcho O.
Ontological Engineering.



Hendler J., Parsi
a B., XML and the Semantic Web.

XML Journal
, Oct 2002.






The Semantic Web:
Semantics for Data and Services on the
. Springer 2008.


Khan A.
Data Warehousing: Concepts and

Iuniverse 2003.


Luger G. F.
Artificial Intelligence: Structures and Strategies for Complex Problem Solving.

Addison Wesley Pub. 2005.


McAfee A. Will Web Services Really Transform Collaboration?”

Sloan Management Review


Middleton S.E., Shadbolt N.R. Roure D.C.D. Ontological User profiling in Recommender
Systems. ACM Transactions on Information Systems 22, 54
58, 2004.


Necib C.B., Freytag J.C. Ontology based Query Processing in Database Management Systems.
Proceedings o
f ODBASE Conference on Ontologies, Databases and Applications 2003.


O’Mahoney M. et al. Collaborative Recommendation: A Robustness Analysis.

Transactions on Internet Technology

44, 11, 344
377, 2004.


Orman L.V. Design and Implementation of Virtual Org
anizations as Electronic S
ervices. Cornell
University 2008


Orman L.V. Service Semantics, Structure, and Design.
Service Journal
6, 2, 2008.


Sheth A., Verma K., Gomadam K. Services Science: Semantics to Energize the Full Services
s of ACM

49, 7, 2006


Simsion G. C., Witt G. C.
Data Modeling Essentials
. Morgan Kaufmann, 2005


Sowa J. F., Borgida A.
Principles of Semantic Networks: Explorations in the Representation of
Morgan Kaufmann



Steyvers, M.; Tenenbaum
, J.B. The Large
Scale Structure of Semantic Networks: Statistical
Analyses and a Model of Semantic Growth.
Cognitive Science

, 1, 41

78, 2005.


Umapathy K.,

S. A theoretical Investigation of the Emerging Standards for Web S
Information Systems Frontiers

, 1, 119