DYNAMIC INTEREST PROFILES
Tracking User Interests Using Personal Information
Justin Lessler, Stefan Edlund, Joann Ruvolo, Vikas Krishna
IBM Almaden Research Center, 650 Harry Road, San Jose California, USA
Keywords: Personal Information Management, E-Mail
Abstract: When building applications it is usually the case that developers are forced to focus on “one size fits all”
solutions. Customization is often burdensome for the user, or so complex that it would be unrealistic to ask
an end user to undertake this task. In the areas of personal information management and collaboration there
is no reason to accept this limitation, as there is a body of information about the user that reflects their
interests: namely their personal documents. The Dynamic Interest Profile (DIP) is a system intended to track
user interests, allowing for the creation of more intelligent applications. In this paper we discuss our
approach to implementing the DIP, challenges that this implementation presents, as well as the security and
privacy concerns that the existence of such an application raises.
1 INTRODUCTION
One of the biggest challenges in developing software
is that all users are different. For example when
retrieving email some users make meticulous use of
folder structures to file and find email, while others
keep all email in one location and use search or
sorting to find needed documents (Ducheneaut,
2001). Because of these differences, attempts to add
intelligence to applications often fail. For instance
due to its tendency to give advice to users who
neither want nor need it, the Microsoft Office
Assistant (a.k.a. Clippy the paperclip) was hated by
many.
Personal information management (PIM)
applications are not free from the problem of dealing
with user difference. Lacking any better course of
action these applications tend to cater to the lowest
common denominator (e.g. address completion
applications suggest names in alphabetical order and
search returns results based on lexical similarity).
But in the case of personal information management
there is a ready store of information that can be used
to tailor applications to individual users, namely
their personal documents. The Dynamic Interest
Profile (DIP) is our implementation of a system that
tracks user interests and provides an interface for
applications to easily access this information.
The DIP recognizes four different types of
entities: people, documents, terms, and collections.
For each of these the DIP assigns both an all time
importance and detects entities of emerging
importance. The DIP is able to determine these
values based on a user’s personal documents.
Tracking user information for the purpose of
making these determinations carries with it a set of
challenges. The algorithms for determining
importance involve a classic computation vs. storage
trade-off. Depending on the PIM system,
maintaining the information necessary to keep the
DIP up to date may require additional storage. This
is an especially important challenge as the
infrastructure requirements for enterprise email
systems are usually quite large already.
In addition to the practical implementation
challenges, privacy and security concerns are raised
by the very existence of a process that tracks user
interests. Many users may balk at the idea of their
interests being tracked, and mistrust assurances that
the information is not being shared. Even if users
trust that their information will not be misused, it is
necessary to provide the same level of security for
this summary information to that which is provided
for the documents characterized.
It is our belief that despite these challenges the
functionality the Dynamic Interest Profile provides
is well worth the effort. Applications such as search
and address completion, when effectively tailored to
a user’s interests, can greatly ease their lives.
Collaborative applications can be automated to
281
Lessler J., Edlund S., Ruvolo J. and Krishna V. (2004).
DYNAMIC INTEREST PROFILES - Tracking User Interests Using Personal Information.
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 281-288
DOI: 10.5220/0002650702810288
Copyright
c
 SciTePress

indicate when a user’s interests provide an
opportunity to leverage the skills of others. An inbox
can be sorted not just by date or sender, but by
importance as well. In an era when many of us spend
much of our day dealing with email and other
documents, seemingly small improvements such as
these can significantly improve both productivity
and job enjoyment.
2 ALGORITHMS
User profiling has been an area of interest for many
years. A variety of techniques have been developed,
based on different data sources (shopping habits,
browsing habits, user feedback (Widyantoro et. al.,
1999), etc.) and utilizing different algorithmic
techniques (machine learning, keyword analysis,
social filtering, etc. (Soltysiak and Crabtree, 1998)).
In our approach we use user data and actions to
determine the overall and emerging importance of
four different entity types: people, terms, documents,
and collections. For each instance of one of these
types (e.g. a person, a term) we use a variety of
statistical techniques specific to the type in order to
determine the instance’s importance to the user.
2.1 People Importance
The importance of a person to a user is determined
by analyzing the relationship between the user and
the person as indicated by the user’s email. Those
people who are most important are the ones with
whom a user has the strongest relationships.
Relationship strength is based on the relationship’s
(Whittaker et. al., 2002):
• Longevity
• Currency
• Reciprocity
• Exclusivity
• Frequency
This information is captured using a linear
combination of functions where each function
represents a particular aspect of the relationship. The
equation for calculating a person p’s importance is:
123
45
()
l
ongevity currency reciprocit
y
exclusivity frequency
score p f f f
ff
λλλ
λλ
=++
+
The component functions of this equation can be
described as follows:
121
2
( ) ( ) has older interactions
with the user than .
longevity longevity
fpfpp
p
>→
121
2
() () has more recent
interactions with the user than .
currency currency
fpfpp
p
>→
11 2 2
1
2
(,) (, ) has
more bidirecitonal interactions wit
h the user than .
reciprocity p p reciprocity p p
fSRfSRp
p
>→
1:1
()
()
exclusivity
p
p
f
p
fp
S
R
=
+
()
frequency p p
f
pS R=+
Where:
the set of communications sent to .
p
Sp
=
the set of communications received
from .
p
R
p
=
1:1
( ) the number of 1 on 1 interactions
between the user and .
fp
p
=
2.2 Term Importance
Terms are words and phrases that carry meaning in a
user’s documents. Parsing documents to find terms
falls outside the scope of this paper, but there is a
large body of literature on this topic, (Silva & Lopes,
1999), (Reyle & Saric, 2001), (Thanopolous et. al.,
2003), and (Manning and Shutze, 2000), just to
name a few. Once terms have been identified, we
define a term as important to the user when it is:
•
Associated with the user
•
Descriptive
•
Discriminative
•
Relevant
As with people importance, this information is
captured by using a linear combination of functions
where each function represents one of the above
qualities. The equation for calculating a term t’s
importance is:
12 3 4
()
u
ser subject folders tim
e
s
core t f f f f λλ λ λ =+ + +
The component functions of this equation have
the following properties:
ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
282

121
2
( ) ( ) is more associated with
the user than .
user user
ft ft t
t
>→
121
2
( ) ( ) is more associated with
document subjects than and hence co
nsidered
to be more descriptive.
subject subject
ftftt
t
>→
121
2
( ) ( ) is more useful in
determining what folder a document
containing the term resides in than
and hence considered to be more
discriminative.
folder folder
ftftt
t
>→
121
2
( ) ( ) is more associated with
the current time period than and hence
considered to be more relevant.
time time
ft ft t
t
>→
2.3 Document Importance
Documents are the basic unit in a PIM system. A
document may be a calendar entry, an email, a word
processing document, etc. Document importance is
determined by considering document content, meta-
data, and usage. Unlike the calculation of people and
term importance, document importance is dependant
upon other DIP calculations, namely people and
term importance. The exact equation varies based
on document type, in the case of an email, d, it is:
12
3
4
5
() ()
() max ()
()
usage
tS
pR
tB
score d f score a
s
core p score t
R
score t
B
λλ
λ
λ
λ
∈
∈
∈
=+ +
++
∑
∑
where a is the document author, R is the
document recipient list, S is the set of terms in
the subject and B is the set of terms in the body.
The function
f
usage
is a function with the property
that:
12
1
2
() ()
has been accessed more than .
usage usage
fdfd
d
d
>→
The astute reader may ask why we choose to
average the importance of the body terms and take
the max of subject term importance when calculating
document importance. This is a result of our
observation of the differing content of the document
subject and the document body. The document
subject or title tends to contain one or two key terms
that describe the document while the other terms
serve as window dressing. Because of this
averaging over the importance of all of the terms in
the title leads to inaccurate comparisons. For
instance, if “golf” is my most important term a
document titled “Improving your Golf” should not
be considered less important than one titled “Golf”.
Document bodies on the other hand tend to contain
similar numbers of filler words as they are generally
written in prose. Because of this, averaging over the
importance of terms in the bodies still leads to
comparable results.
2.4 Collection Importance
Collections may be of varying types, ranging from
email folders to public discussion databases. What
constitutes a collection is dependant on the specific
PIM system being used, but for all collections the
broad attributes that indicate importance are the
same: collection content and usage. For some
collections meta-information about the collection
may also be important. As with documents, the
calculations for discovering collection importance
are dependant on other DIP calculations. The
general form of the calculation of the importance of
a collection c is:
3
12
() ( )
||
c
access change
dD
c
s
core c f f score d
D
λ
λλ
∈
=+ +
∑
where D
c
is the set of documents in collection c.
The component functions of this equation have
the following properties:
121
2
( ) ( ) is accessed by the user
for reads more often than .
access access
fcfcc
c
>→
121
2
( ) ( ) is modified by the user
more often than .
change change
fcfcc
c
>→
Note that the document score score(d)
incorporates term and people importance, hence it
reflects the frequency of important people and terms
in the collection.
2.5 Emerging Items
It is not always the case that we are interested in
items that are important over all time, sometimes we
DYNAMIC INTEREST PROFILES: TRACKING USER INTERESTS USING PERSONAL INFORMATION
283

are interested in items that are of emerging
importance. In order to detect emerging items we
calculate how associated the item is with the recent
time period using point wise mutual information,
Chi-Squared, or some equivalent statistical test
(Yang and Pedersen, 1997) (Manning and Shutze,
2000) (Duda et. al., 2001). In the case of terms and
people the number of references to the person or
term is used for this calculation. In the case of
collections the frequency of access and modification
by the user is used. The general equation for
detecting the extent to which a term x is of emerging
importance is:
() A(,)
x
e
merge x x t C= where t is the current
time period, A(x,t) is the strength of association
between x and t, and C
x
is the frequency of term
x.
2.6 Efficient Calculation and Storage
In order for the DIP to be deployable in a real world
setting it is necessary for the importance and
emergence calculations to be performed efficiently.
In order to completely recalculate many of the
functions involved the entire corpus of documents
would have to be examined, a very time consuming
operation. To allow for efficient recalculation,
intermediate results are stored. These allow
recalculation based only on the delta in the
document corpus. These intermediate results take
the form of feature vectors. A feature vector is a set
of keys (name, terms, etc.) and associated counts.
These feature vectors represent those intermediate
parts of the calculation that can be incrementally
changed when documents are created, deleted, or
modified.
Once DIP scores have been recalculated the
results are stored as score vectors representing the
importance of the DIP entities to the user. Score
vectors are sets of keys associated with a floating
point score. Details of the storage framework are
discussed in section 3.
2.7 Limited DIP Implementation
In order to validate the underlying concepts of the
DIP and provide a platform for further
experimentation we have developed a limited DIP
implementation. This implementation performs the
calculations for term importance, as well as limited
versions of the document importance and people
importance calculations, based upon user emails.
When calculating term importance we treat the
subject and body as sets of terms. We do not use any
information as to the order of the words or the
frequency of their occurrence within the document.
The implementations of the sub functions making up
term importance are:
,
,
,
,
max[0,I( , )log( )]
max[0,I( , )log( )]
max [I( , )log( )]
max[0,I( , )log( )]
user t user
s
ubject t subject
folders t f
f Folders
time t year
ftuserC
ftsubjectC
ftfC
f t year C
∈
=
=
=
=
where C
t,x
is the number of occurrences of term
t in set x and I(x,y) is the point wise mutual
information between x and y. See (Yang and
Pedersen, 1997) and (Manning and Shutze,
2000) for a description of point wise mutual
information.
In order to calculate
f
user
it is necessary to have
some notion of the general frequency of use of a
term. In order to determine this we maintain an
organizational “background vector” that contains the
number of times each individual term has appeared
overall in an organization’s email. In environments
where security is a concern, it is possible to replace
this organizational background vector with one
based on the general frequency of words within the
language being used.
People importance in the draft implementation is
based upon a limited notion of frequency and
reciprocity characterized by the equation:
12
()
p
p
s
core p S R λλ =+ where S
p
is the set of
emails sent to p, R
p
is the set of emails received
from p and
1
2
.
λλ >>
Document importance in the draft
implementation is the same as the full equation
described in section 2.3 minus the usage term:
2
1
4
3
() () ()
max ( ) ( )
pR
tS
tB
s
core d score a score p
R
s
core t score
t
B
λ
λ
λ
λ
∈
∈
∈
=+ +
+
∑
∑
Both term importance and people importance are
stored in the prototype implementation, while
document importance is calculated on the fly.
ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
284

3 IMPLEMENTATION
As we set about implementing the DIP component
into a piece of enterprise collaboration software, we
initially compiled a list of requirements:
Single- and multi-user support. We had to
ensure that we were able to maintain interest profiles
for a potentially large number of users. In addition,
maintaining a DIP for a single user on a single
workstation working off-line was also required.
Abstract document model. The DIP is required
to handle information from a diverse set of domains,
including: email documents, calendar & scheduling
data, instant messaging, and personal documents. A
document model was needed for transforming data
from such domains into an abstract format, capturing
the information needed for compilation of interest
profiles.
DIP object model The object model captures
summary interest profiles for entities. As discussed
in Section 2, summary interest profiles contain
importance scores associated with terms, people, and
collections.
Incremental updates. The DIP is compiled
from a large body of documents generated over
potentially long periods of time. The DIP
infrastructure had to be able to incrementally update
the object model to gradually refine results over
time.
DIP persistence model. Incremental updates
require a persistence model to store interest profiles
constructed over long periods of time. Also, since
maintaining the DIP for multiple users consumes
space, persistence allows us to bring a smaller
“active” working set of the interest profiles into
server memory on demand.
Privacy and Security. As mentioned earlier,
documents can contain sensitive information that
people aren’t always willing to disclose.
Mechanisms for preventing access to sensitive
personal information were needed.
We now proceed to discuss the details of the
implementation.
3.1 DIP Document Model
The document model adopted in the implementation
summarizes the original document with regard to
people and terms. As discussed in Section 2,
information on people is extracted from fields within
the original document, preserving the relationships
that exist between documents and people. This
allows us to give a higher importance to (for
instance) a person on the “to:” list of an email versus
a person on the “cc:” list.
For email documents, people are represented by
their email addresses. However, since multiple
addresses can map to a single email account,
ambiguity must be resolved. When translating
between original documents and abstract DIP
documents, we attempt to normalize email addresses
to their unique id maintained by a profile
management system. In our target enterprise
collaboration software, WebSphere Member
Manager™ (WMM) is the global solution for
managing member profiles.
Terms encountered in header fields and the body
of documents are compressed into a set of unique
terms as discussed in Section 2. Bags of unique
terms are maintained for each field deemed relevant
for compilation of interest profiles. Also, any source
collection information included in the original
document is kept.
3.2 DIP Object Model and
Persistence
We adopted the vector model discussed in Section 2
for the DIP object model, where a collection of
feature and score vectors comprise the interest
profile for a user. This allows us to extend the model
by adding additional features as needed. The two
main features captured are term and people
occurrence, but the model also contains 24 smaller
vectors representing terms and people for the last 12
months. Counts computed from documents
processed during a more limited timeframe can be
more informative than the global vectors since they
incorporate a recency factor into the calculations,
and they allow applications to track how interests
change over time.
Feature vectors are either personal, such as the
vector describing people importance, or shared. The
vectors contain a user identifier (or an identifier for
the global user), a vector id, and a set of features,
where each feature has a feature name – for instance
a term – and a count/score. The feature vectors are
persisted in a relational database (DB2 UDB™) in a
table having these four columns. To optimize for
quick retrieval of vector features, a B-tree index on
user id and vector id is maintained. The table is a
read-only table most of the time, minimizing lock
contention and maximizing concurrency. While the
DIP is being updated, applications are forced to wait
until the update is complete.
DYNAMIC INTEREST PROFILES: TRACKING USER INTERESTS USING PERSONAL INFORMATION
285

3.3 Offline Operation
We have also implemented an offline version of the
DIP, running on a workstation for a single user.
When in offline mode, a client runs the DIP
algorithms locally on a smaller set of personal
documents.
As mentioned in Section 2, shared information is
used, though minimally, to compute the important
terms for a user. Shared information in the form of a
background vector is stored on a common server,
and it is replicated to the client whenever the client
goes online. The replica of the background vector on
the client improves the accuracy of the term
importance calculation. The client keeps track of
local changes to the background vector, and it
updates the shared representation when going back
online.
3.4 Security and Privacy
A DIP server falls under a category of systems
described as “user modeling servers” (Fink and
Kobsa, 2000), that is servers that support
applications requiring user models for
personalization features. As such, it needs
mechanisms to protect personal information from
unauthorized access and potential eavesdropping.
Many factors are taken into consideration when
determining the level of security and privacy to
provide, for instance the privacy preferences of the
user, self-regulatory privacy principles within an
organization, or the privacy laws of the country or
state.
A combination of anonymization and encryption
techniques can be deployed to ensure the privacy of
the individual. In the DIP system, personal
information is encrypted at the database level to
ensure privacy. This protects from both
eavesdropping and from individuals, such as
database administrators, having high levels of
authorization. Methods for anonymization (i.e.
concealment of the relationship between a particular
user and data about him) are currently being
investigated for future implementation.
3.5 Implementation Overview
Figure 1 shows an overview of the DIP implemented
as a processing agent in a pipeline of agents
operating on a document stream. Processing agents
are designed to carry out specific tasks, for instance
indexing documents for search, flagging documents
that contain viruses or – in the case of the DIP –
updating interest profiles. Some agents modify
documents as they travel down the stream, for
instance an agent expanding group names into a list
of members. While the DIP agent does not modify
documents, it is easy to imagine a system where the
agent tags documents with information such as
sender importance, text importance, or an
aggregated overall importance score. Observe that
such tagging provides a snapshot of the importance
calculated when the document was first received,
and that the importance will change when the
underlying interest profile is updated. Sometimes it
is useful to recall the importance of a particular
document at the time it was received, but as a
person’s interest profile changes, it is often desirable
to bring old documents to the forefront. For this
reason, we do not put scores into the documents.
Documents on the stream originate from a
variety of applications, including email, calendaring
& scheduling, and instant messaging. A flag may be
associated with documents indicating whether they
have been updated or deleted from their
corresponding data repositories. This allows us to
update the DIP object model correctly. For instance,
spam ignored by spam filters is typically deleted
immediately from an inbox. Often we can detect that
an email has been deleted early enough to remove it
from the queue before it has been incorporated into
an interest profile. Efficient spam filters also help
keep DIP profiles more accurate.
In real-world mail systems, message transfer
agents (MTAs) ensure that incoming messages are
delivered locally into the mail store, and outgoing
messages are sent using SMTP to their destinations.
Mail transfer agents are operating on queues of
incoming/outgoing email, and the DIP processor is
simply another agent for such queues. Similar
mechanisms exist for other messaging systems (e.g.
instant messaging).
Figure 1: Agents operating on a stream of
documents
Indexing Agent
DIP Agent
ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
286

Figure 2 shows the agent passing documents to
the DIP translator, which handles the conversion
from the stream document format to the DIP
document format. A translated document is pushed
onto internal queues, where they are stored until a
task is triggered to update the interest profiles.
Translation, name normalization, and queue
operations are handled quickly to prevent the DIP
agent from becoming a bottleneck in the system.
DIP documents on the queues are periodically
popped and fed to the processor, which
incrementally updates the interest profiles. There is a
many-to-one mapping between documents and
profiles, enabling the processor to update one profile
at a time by batching documents together. Taking
email as an example, the batch contains all emails
sent by or received by an author during the
timeframe since the last DIP update was carried out.
As a consequence, documents are tagged to indicate
which person (or group) they are associated with.
For instance, an email sent by person A to person B
with a carbon copy to person C generates three new
documents (but not necessarily on the same mail
system). To assist in the task of batching documents,
we use multiple queues internally, one for each
unique person.
A batch of documents is processed in memory
building a corresponding set of delta vectors used to
incrementally update the underlying interest profile.
In our current implementation, a persisted interest
profile is stored in a relational database, and the
implementation of the delta vectors allow us to track
the SQL inserts, deletes, and updates needed to
incorporate the new set of documents. To ensure
database consistency, all updates to an interest
profile are treated as a single transaction.
Applications access interest profile information
through a public application programming interface
(API) in a read-only mode. The API has methods for
retrieving people and terms ranked by importance,
and also methods for computing aggregated
importance scores for documents. Search
applications can use this functionality to refine
search results according to personal interest profiles.
The DIP is implemented as an enterprise
business component based on the J2EE
recommended multi-tiered architecture. The four
tiers in our implementation are, bottom up, the
Resource tier housing the data store and data
accessor classes, the Service tier containing
Enterprise Java Beans, the Workspace tier
containing the client side representation of the server
classes, and the User tier in which DIP clients
operate.
4 DISCUSSION
Our early experiences with testing the DIP have
been quite encouraging. Even the relatively simple
versions of our algorithms in the described
implementation yield quite good results, and
developing useful applications based on the DIP is
quite easy. Early tests of DIP technology have led to
some interesting observations on the use of DIP
technology.
It is often the simplest DIP based applications
that prove the most compelling. For instance: inbox
sorting and address completion are always a hit at
demos, while the more complex displays of the
people associated with important terms receives a
more muted response.
Displaying raw DIP results, such as a user’s n
most important people, is not recommended. Terms
and people that may be computationally important in
determining importance may seem “wrong” when
viewed by the user. For instance, a rare abbreviation
commonly used by my manager may be valuable in
discovering document importance, but I would not
perceive it as a term that is important. We have
begun research on filtering based on controlled
vocabularies and ontological information to help in
the production of DIP results that are more human
friendly.
Two of the most compelling DIP based
applications are the dynamic address book and inbox
sorting. Our target enterprise collaboration software
uses the DIP to maintain dynamic address books, i.e.
Figure 2: High level architecture and
data flow in the DIP
Document Queue
DIP Translator
DIP Processor
DIP Agent
DIP
DB
∆
Profile Management
System (WMM)
DIP API
Appl. Server
DYNAMIC INTEREST PROFILES: TRACKING USER INTERESTS USING PERSONAL INFORMATION
287

address books that are automatically compiled from
the most important people as determined by the DIP.
In addition, the collaboration software allows users
to search documents and refine the results using
personal interest profiles. Documents deemed
important by the DIP are ranked higher, increasing
the likelihood that users find what they are looking
for.
One of the favorite demo applications we
developed was sorting the user’s inbox by
importance. The DIP proved to do a wonderful job
of giving spam and mailing list articles low
importance and important work communications
from coworkers and managers high importance. By
allowing for some adjustment of the weights
involved the sort could be further refined, sending
either documents with important topics or by
important people to the top of the list. We hope to
perform formal validation of this functionality and
make it available in products in the future.
This paper has discussed the concept of Dynamic
Interest Profiles and the integration of such profiles
into enterprise collaboration software. As important
next steps our team plans to: formalize the concepts
described, perform more validation, and see what
role standards can play in this area.
We would like to acknowledge the following
people for their contributions and support in this
work: Carl “Pooter” Kraenzel, Mike O’Brien, Kate
Glickman, Sesha Baratham, Igor Belakovskiy,
Niklas Heidloff, Gurushyam Hariharan, and Bill
Cody.
REFERENCES
Ducheneaut, N., & Bellotti, V. 2001a. Email as Habitat:
An Exploration of Embedded Personal Information
Management. Interactions, 8(5), 30-38.
Duda, R. O., Hart, P.E., Stork, D.G., 2001. Pattern
Classification. John Wiley & Sons, Inc. 2
nd
Edition.
Fink, J. and Kobsa, A., 2000. A review and analysis of
commercial user modelling servers for personalization
on the world wide web. User Mod. User-Adapted
Interact.10, 3-4, 209-249
Manning, C. D., & Schutze, H. 2000. Foundations of
Statistical Natural Language Processing, The MIT
Press. Cambridge Mass., 3
rd
printing.
Nardi, B., Whittaker, S., Isaacs, E., Creech, M., Johnson,
J., Hainsworth, J., 2002. Integrating communication
and information through ContactMap.
Communications of the Association for Computing
Machinery
Reyle, U. & Saric, J. 2001. Ontology Driven Information
Extraction. In Proceedings of the 19
th
Twente
Workshop on Language Technology, University of
Twente.
Schiaffino, S.N., Amandi, A. 2000. User Profiling with
Case Based Reasoning and Bayesian Networks.
IBERAMIA-SBIA 2000 Open Discussion Track.
Silva, J., & Lopes, G., 1999. Extracting Multiword Terms
from Document Collections. In Proceedings of the
VExTAL, Venezia per il Trattamento delle Lingu,
Universitá cá Foscari, Venezia November 22-24.
Soltysiak, S., & Crabtree, B. 1998. Knowing Me Knowing
You: Practical Issues in the Personalisation of Agent
Technology. In Proceedings of the 3
rd
International
Conference on the Practical Applications of Agents
and Multi-Agent Systems (PAAM-98).
Thanopolous, A., Fakotakis, F., Kokkinakis, G. 2003. Text
Tokenization for Knowledge-free Automatic
Extraction of Lexical Similarities. TALN 2003.
Traitement Automatique des Langues Naturelles.
Whittaker, S., Jones, Q., Terveen, L., 2002. Contact
Management: Identifying Contacts to Support Long-
Term Communication. CSCW’02
Widyantoro, D.H., Ioerger, T.R., Yen, J., 1999. An
Adaptive Algorithm for Learning Changes in User
Interests. Proceedings of the Eighth International
Conference on Information and Knowledge
Management, 405-412.
Yang, Y., Pedersen, J., 1997. A Comparative Study on
Feature Selection in Text Categorization. In
Proceedings of ICML-97, 14
th
International
Conference on Machine Learning.
ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
288