> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1
Digital Libraries have become one of the most important web
services for information seeking. One of their main drawbacks is
their global approach: in general, there is just one interface for all
users. One of the key elements in improving user satisfaction in
digital libraries is personalization. When considering
personalizing factors, cognitive styles have been proved to be one
of the relevant parameters that affect the way in which a user
interacts with an interface. This justifies the introduction of
cognitive style as one of the parameters of a web personalized
service. Nevertheless, this approach has one major drawback:
each user has to run a time-consuming test that determines
his/her cognitive style. In this paper we present a study of how
different classification systems can be used to automatically
identify the cognitive style of a user using the set of interactions
with a digital library. These classification systems can be used to
automatically personalize, from a cognitive-style point of view,
the interaction of the digital library and each one of its users.
Introduction
Digital Libraries (DL) are collections of information that
have associated services delivered to user communities using a
variety of technologies (Callan et al., 2003). The collections of
information can be scientific, business or personal data and
can be represented as a digital text, image, audio, video or
other media. Due to the amount and great variety of
information stored, DL has become, with search engines in
general, one of the major web services (Liaw & Huang, 2003).
Typically, DL have a global approach in which all users are
presented with the same interface regardless the diversity of
users in terms of preferences or skills. Nevertheless, different
studies in information seeking have shown that matching the
interface with users’ preferences can help them to achieve their
tasks in a satisfactory way (Marchionini, Plaisant & Komlodi,
1998; Bladfor, Stelmaszewska & Bryan-Kinns, 2001). From
this perspective personalization is a key tool to increase user
satisfaction.
Personalization is defined as the ways in which information
and services can be tailored to match the unique and specific
needs of an individual or a community (Callan et al., 2003).
The key element of a personalized environment is the user
model. A user model is a data structure that represents user
behavior and captures human factors. The more information a
user model has, the better the content and presentation will be
tailored for each individual user. A user model is created
through a user modeling process in which unobservable
information about a user is inferred from observable
information from that user; for example, using the interactions
with the system (Zukerman, Albrecht, & Nicholson, 1999).
User models can be created using a user-guided approach, in
which the models are directly created using the information
provided by each user, or an automatic approach, in which the
process of creating a user model is hidden from the user. The
personalization elements constructed using a user-guided
approach are usually called adaptable (Fink, Kobsa, & Nill,
1997), while the ones produced using an automatic approach
are usually called adaptive (Fink, Kobsa, & Nill, 1997;
Brusilovsky & Schwarz, 1997).
When considering human factors for personalization,
cognitive style, which influences an individual’s preferences
for organizing and processing information, has been largely
ignored. Nevertheless, in recent years, different studies (Graff,
2003; Chen & Macredie, 2002) have found that users’
cognitive styles significantly influence their reaction to the
application interface in terms of user control, multiple tools
and nonlinear interaction. Some of these studies have focused
on information searching interfaces (Moos & Hale, 1999).
This relevance implies that cognitive style can be a very
relevant factor to include in a user model for personalization
purposes, especially for information searching purposes (Moos
& Hale, 1999). The main inconvenience of including cognitive
style in a user model is that, in order to identify the cognitive
style of a user, each user will have to run a questionnaire. This
process is a time-consuming activity that not all users of a
personalized system would be willing to undertake.
While there are different applications, mainly learning
environments, that personalize presentation and content using
a cognitive approach (Triantafillou et al., 2002, Papanikolau et
al., 2003; Bajraktarevic, Hall & Fullick, 2003), all of them
assume that the cognitive style of each user of the system is
known in advance, which implies that all users of that system
have run a cognitive style test. While this assumption can be
valid for testing environments (like the applications
commented) or systems with a reduced number of users, the
approach is not valid for large-scale personalized systems. In
this context, the idea of automatically create user models that
identify users’ cognitive styles is essential, because it saves the
user time consuming activities and directly allows the
Enrique Frias-Martinez, Sherry Y. Chen and Xiaohui Liu
Department of Information Systems & Computing, Brunel University, Uxbridge, UB8 3PH, United Kingdom. E-mail:
{enrique.frias-martinez, sherry.chen, xiaohui.liu}@brunel.ac.uk.
Automatic Cognitive Style Identification of
Digital Library Users for Personalization
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 2
personalized system to present a cognitive personalized
interface.
This paper presents, within the context of an information
seeking environment known as the Brunel Library Catalogue,
an automatic approach to identify users’ cognitive styles to
create adaptive user models. The user models created, which
will contain the cognitive styles, can be used to personalize the
digital library interface according to each user’s cognitive
style.
The rest of the paper is organized as follows: first we
introduce the concept and architecture of personalized DL.
Second we present the relationship between cognitive styles
and user interface design. The third section of the paper
presents the design of the experiments run for Brunel Library
catalogue. The interactions captured in these experiments will
be used to characterize and identify each user’s cognitive style.
The fourth section presents the design of different cognitive
classification systems that automatically identify each user’s
cognitive style. We conclude the paper in section 5.
Personalized Digital Libraries
In general, DLs are made up of four components (Theng,
Duncker, & Mohd-Nasir, 1999): (1) information; (2) structure,
describing the syntactic and semantic characteristics of the
information; (3) interaction elements, referring to the searching
interface, screen design, etc.; and (4) properties, referring to
security, copyright issues, etc., of the information available in
the DL. The services provided by DL through their interaction
elements (interface) can be classified into three groups:
• Mechanisms for content selection. These mechanisms make
it possible for each user to create a personal DL that contains
only the information that is interesting and relevant to that user
• Mechanisms to help in the process of navigation. These
services present each user with an environment that better suit
the way in which that user interacts with the DL.
• Information filtering (IF) and information retrieval (IR)
mechanisms. These services provide ways to find and filter the
vast amount of information that a user accesses and receives.
Although these three basic types of services provide the
basic functionality needed by a DL user, they can be improved
by the introduction of personalization. Personalization will
create more tailored services that help and simplify the process
of finding relevant information by using the content of each
user model. Formally, a user model is as a set of information
structures designed to represent one or more of the following
elements (Kobsa, 2001): (1) representation of assumptions
about the knowledge, goals, plans preferences, tasks and/or
abilities about one or more types of users; (2) representation of
relevant common characteristics of users pertaining to specific
user subgroups (stereotypes); (3) the classification of a user in
one or more of these subgroups; (4) the recording of user
behavior; (5) the formation of assumptions about the user
based on the interaction history and/or (6) the generalization of
the interaction histories of many users into stereotypes. In the
context of user modeling a stereotype is defined as a cluster of
users that share a common behavior.
Typically, personalization in DL has been user-driven. In
this approach the user specifies his/her preferences directly to
the DL, from the color background of the page, to the layout of
the components or to the content of the information presented.
The main inconveniences that this approach has are: (1) the
concept of personalization cannot be necessarily understood
by all the users of a DL, (2) users are not usually willing to
give feedback to the system, even if it is for receiving a better
service, and (3) users do not necessarily know what their
interests are and how they change over time, and cannot
provide information to the system. In constructing a user
model that contains the cognitive style, there is an added
inconvenience: users need to run a time-consuming test to
identify the cognitive styles.
We think that a better approach to provide personalized DL
services will be based on using user models that are
automatically constructed using machine learning techniques,
i.e. adaptive user models, because the application of these
techniques will remove the limitations that a user-driven
approach entails. Although machine learning techniques have
been extensively used in e-commerce sites (mainly for
recommendation purposes), their implementation in DL has
been very limited up to now.
Figure 1 presents the architecture of an adaptive DL. As
can be seen, the interaction between the User and the DL is
handled by a Decision Making & Personalization Engine,
which takes into account the user models and the interaction,
personalizes the interface and the results to each user. The
adaptive characteristic is given by the “User Modeling
Generation” module, which has as input a database containing
the interactions between the set of users and the library, and
automatically produces the set of user models. This automatic
approach allows to observe users in an unobtrusively way and
solves the problem that the user-guided approach has.
In this paper we are going to focus on the two modules
responsible for automatically creating user models: The “User
Modeling Generation” module and the “Database of
Interaction” module. First we will design a set of experiments
aimed at capturing user interactions with Brunel Library
Catalogue, the output of which will be stored in the “Database
Interaction Module”. Then, using this information we are
going to design a “User Modeling Generation module” that
User
Decision Making & Personalization Engine
User
Models
Hypermedia
Database
Interaction Elements
Content
Personalization Navigation IF/IR
Information
Structure &
Semantics
Properties
User
Modelling
Generation
Database of
Interactions
Query Output
FIG. 1. Generic Architecture of an Adaptive DL.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3
automatically constructs a set of user models that describe the
cognitive style of each user. This set of user models makes it
possible to implement a personalized DL from a cognitive
style point of view while at the same time avoids the userdriven
approach which will imply that each user of the library
would need to run a test to identify his/her cognitive style. The
following section highlights why cognitive style is a very
relevant parameter for personalization.
Relevance of Cognitive Styles for Personalization
Although there are a lot of different definitions of cognitive
style in the literature, it can be defined as an individual’s
preferred and habitual approach to organizing and representing
information (Riding and Rainer, 1998). Cognitive style is a
personality dimension, which influences the way individuals
collect, analyze, evaluate, and interpret information (Harrison
and Rainer, 1992). There are a variety of dimensions of
cognitive styles, but among these dimensions, Field
Dependence versus Field Independence has significant impacts
on users’ information processing, because it reflects how well
an individual is able to restructure information based on the
use of salient cues and field arrangement (Weller et al., 1994).
Their different characteristics are:
Field Dependence(FD): Field Dependence describes the
degree to which a user’s perception or comprehension of
information is affected by the surrounding perceptual or
contextual field, that is, “the extent to which the organization
of the prevailing field dominates perception of any of its parts”
(Witkin, et al., 1971). Field Dependent individuals typically
see the global picture, ignore the details, and approach a task
more holistically. Field Dependent individuals are considered
to have a more social orientation than Field Independent
persons since they are more likely to make use of externally
developed social frameworks. They tend to seek out external
referents for processing and structuring their information. They
are more readily influenced by the opinions of others, and are
affected by the approval or disapproval of authority figures
(Witkin et al., 1977).
Field Independence(FI): Field Independent individuals
tend to discern figures as being discrete from their
background, to focus on details, and to be more serialistic in
their approach to learning. These individuals tend to exhibit
more individualistic behaviors since they are not in need of
external referents to aide in the processing of information.
They are better at processing impersonal abstract material, are
not easily influenced by others, and are not overly affected by
the approval or disapproval of superiors (Witkin et al., 1977).
Recent studies have found that users’ cognitive styles
significantly influence their reaction to the user interface in
terms of user control, multiple tools, and non-linear
interaction. With respect to user control, several studies have
suggested (Chuang, 1999, Chanlin, 1998) that FI individuals
could particularly get benefit from the control of media choice.
Other studies (Marrison and Frick, 1994) have suggested that
FD users prefer to have auditory cues in the systems.
Regarding multiple tools, Ford and Chen (2000) showed that
FD individuals tend to build a global picture with the
hierarchical map when interacting with web services, while
Palmquist and Kim (2000) found that FD novices tend to
follow links prescribed by a web page. Regarding non-linear
interaction, Dufresne and Turcotte (1997) investigated the
effect of cognitive style within a searching information
environment. They found that FD students who used the
system with non-linear structure spent more time completing
the test than those who used the system with linear structure.
FI individuals consulted the user guide for a longer period than
FD individuals in the linear version, while FD individuals
consulted it for longer in the non-linear version.
Results from these studies suggest that different cognitive
style groups prefer different interface functionalities and
structures provided by web-based applications. These results
indicate the relevance of cognitive styles for personalization.
In this paper we present a system that automatically identifies
the cognitive style of a library catalogue user for
personalization purposes. This approach is even more relevant
because different studies have shown that FD/FI is consistent
across domains and stable over time (Goodenough, 1976;
Messick, 1976; Witkin, et al, 1977; Witkin and Goodenough,
1981). This basically implies, that once a user has been
assigned a cognitive style type, that type is constant in that
environment, Brunel Library Catalogue in our case, and not
only that, but that it is constant in other domains, i.e. if we
identify the cognitive style of a user for Brunel Library
catalogue, the identification of his/her cognitive style may be
able to be used for other applications.
Experiment Design
This section describes the different characteristics of the
experiments that were designed to capture interaction data to
automatically create cognitive user models. The following
subsections present the characteristics of the participants, the
research instruments used, including the DL in which this
study focuses, and the tasks designed and data collection
techniques used.
Cognitive Style
Field DependetIntermediateField Independent
Count
30
20
10
0
Gender
female
male
FIG. 2. Number of participants in each cognitive style.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4
Participants
The study was conducted at Brunel University’s
Department of Information Systems and Computing. A total of
54 students participated in this study. All participants had the
basic computing and Internet skills necessary to use Brunel
Library Catalogue. Figure 2 illustrates the number of
participants in each cognitive style.
Research Instruments
The research instruments used include: (1) Cognitive Style
Analysis (CSA) to measure participants’ cognitive styles, (2) a
digital library catalogue, Brunel Library catalogue, which is
the focus of the study, and (3) a tool for capturing user
interaction and storing a user questionnaire, WebQuilt.
Cognitive Style Analysis
A number of techniques have been developed to measure
Field Dependence/Field Independence, and among those we
have chosen the Cognitive Styles Analysis (CSA) by Riding
(1991). The CSA test includes two sub-tests: (1) the first
presents items containing pairs of complex geometrical figures
that the individual is required to judge as either the same or
different and (2) the second sub-test presents several items
each comprising a simple geometrical shape, such as a square
or a triangle, and a complex geometrical figure and the
individual is asked to indicate whether or not the simple shape
is contained in a complex one by pressing one of two marked
response keys (Riding and Grimley, 1999). These two subtests
have different purposes. The first sub-test is a task
requiring Field Dependent capacity, while the second sub-test
requires the disembedding capacity associated with Field
Independence. This provides a big advantage with other
methods that only measure one of the factors.
The CSA measures what the authors refer to as a
Wholist/Analytic (WA) dimension, noting that this is
equivalent to Field Dependence (Riding & Rayner, 1998). WA
dimension is a real number between the values of 0.6 and 3.0
that indicates the degree of field dependence. Riding's (1991)
recommendations are that scores below 1.03 denote Field
Dependent individuals; scores of 1.36 and above denote Field
Independent individuals; and scores between 1.03 and 1.35 are
FIG. 3. Basic Search Interface of BLC.
FIG. 4. Advanced Search Interface of BLC.
FIG. 5. Multiple Results Interface of BLC.
FIG. 6. Results Interface of BLC.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5
classified as Intermediate. In this study, categorizations were
based on these recommendations.
Brunel Library Catalogue
Brunel Library Catalogue (BLC) is a typical digital library
to access the bibliographical resources of Brunel University.
BLC has two main mechanisms that provide different
strategies for finding information: (1) Basic Search (Figure 3),
which is the one presented by default by the system, and (2)
Advanced Search (Figure 4) which is accessed through the
corresponding link presented in Figure 3. Basic Search allows
to run a quick search of the library catalogue using a set of
keywords and one of the following commands: “word or
phrase”, “author” “title” or “periodical title”. Also the user
can choose in which library he/she wants to search that
information. The help link describes briefly what each link is
supposed to do. Advanced Search, as presented in Figure 4,
presents the user with a much broader way of searching
information. The user can give value to each field (a generic
work, author, title, subject etc.), and combine these words
using and/or Boolean operators. The system also allows users
to select other information like the library, the language, the
publication year, etc.
Once a user submits a query to the system using the Basic
Search or the Advance Search, the system responds with the
items found in the database. An example of the interface
presented is given in Figure 5. The system presents a set of
buttons in the top part: “Go Back”, “Limit Search”, “New
Search”, “Backward”, “Forward”, “Prefs” and “Exit”. The
“Limit Search” option is a link to the bottom of the page where
the search mechanism used (Basic Search or Advanced
Search) is presented with the terms used and a set of options
for Search Limits (language, publication year, etc.). The limit
search is obtained adding more words to the set of terms
already introduced. The “New Search” option presents again
the interface of Figure 3. The “Backward/Forward” button
allows to move up and down the items found. Once a user
selects one item the information and interface given is
presented in Figure 6.
WebQuilt & Exit Questionnaire
WebQuilt Proxy Server (Hong et al., 2000)
(http://guir.berkeley.edu/projects/webquilt) is a proxy system
that unobtrusively gathers click stream data as users complete
specified tasks. It is designed to conduct remote usability
testing on a variety of Internet-enabled devices and provide a
way to identify potential usability problems when the tester
cannot be present to observe and record user actions.
WebQuilt utilizes Java Servlet and JSP technology to track
users' interaction and then store that data by (1) creating a log
file of each user's web use and (2) additionally caching the
pages a user accesses for later viewing. Figure 7 shows the
basic architecture of WebQuilt Proxy. Once the proxy server is
running, each user connects to any web page through the web
server. The proxy server stores any interaction between the
user and any web pages and a snapshot of each page visited by
each user. These snapshots are given a number that is the same
one used to describe the sequence of pages visited by the user.
The Web proxy server has the possibility of adding a task box
that can be used to indicate when a task has been finished.
Once a user finishes each task and uses the task box links to
finish it, Web Quilt allows to present to each user a set of
questions regarding the task.
All the information captured is stored in the proxy server
creating in each case a file using and id for each user. This
allows to centralize all the information in the same place and at
the same time being able to access the information of each user
independently. The use of a proxy server architecture allows
us to easily capture all the interaction between users and BLC,
which otherwise would be far more difficult due to the changes
needed to be implemented in BLC.
Task Design
The purpose of this experiment is to collect enough data to
create an automatic classification system capable of identifying
each user’s cognitive style. The main behaviors that a user that
access a web library catalogue has are two: browsing and
searching (Bryan-Kinns, 2000). In this context browsing is
defined as the search of ill-defined information while
searching is defined as the localization of specific well-defined
Proxy Server Brunel Library
User
User
Fig. 3. Generic Architecture of an Adaptive DL.
TABLE 1. Set of tasks designed for the experiment and their type.
Task Type
1 Find the Call Number of the book “The Man in the High Castle” by Philip Kendred Dick. Search
2 Find the title of any book related with applications of fuzzy logic. Browse
3 Find the number of books written by Aldous Huxley that are part of TWICKENHAM Library Search-Browse
4 Find a book about how to implement data mining with Java. Browse
5 Find a Java book written by Hugh Vincent. Search
6 Find a book about 20th
century American Drama in TWICKENHAM campus. Browse
7 Please find an IEEE journal on consumer electronics. Search
Proxy Server Brunel Library
User
User
FIG. 7.. Generic Architecture of an Adaptive DL.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6
information. Theng et al. (1999) presents an example of this
approach, with two experiments designed to capture user
interaction with a digital library in order to evaluate the degree
of satisfaction of the users with the library interface.
In order to identify users’ real perceptions, participants were
asked to perform a set of seven practical tasks (Table 1). The
design of the task was interface dependent: the set of tasks was
designed to involve all the functionalities that BLC provides to
each user and the different behaviors (searching and browsing)
that a user can show.
The first question captures a searching behavior, as it has a
clear well-defined answer contained in the library catalogue. It
is also designed to capture if the user uses the “Search
Everything”, “Author” or “Title” options (which are different
ways of approaching the problem) or if a Power Search is
used. In case the power search is used it will be interesting to
capture which elements are used (title, author, year), and if any
search limit is introduced. The second task is a browsing
question designed to test if the user uses the “Subject” option
or prefers an approach using “Title” or “Search Everything”.
The third question is a combination of searching-browsing task
designed especially to see if users use the option of selecting a
specific campus or do it manually. It is also interesting to see
how the user approaches the problem (searching by author or
by general search). The rest of the tasks are designed to
replicate some of the functionalities and/or behaviors, in order
to have more data to effectively construct a cognitive
classification system.
Experimental Procedures
The experiment was conducted using Brunel Library
catalogue. The experiment comprised three different steps:
(1) Participants were given a task sheet, which described the
task activities that they needed to complete with BLC. One
participant carried out the experiment at a time.
(2) The CSA was used to classify participants’ cognitive styles
into Field Independent, Intermediate, or Field Dependent.
(3) Participants were observed while they were carrying out
the tasks, and clarifications were given when requested.
(4) At the end of each task the user answer the following
questions: (1) was the user able to solve the task?, (2) was the
task difficult?, and (3) what is the answer found?.
Data Collection & Summarization
The data collected from each user was centrally stored in
the proxy server. For each user, the interaction data captured
for solving the seven given tasks was summarized into six
dependant variables, variables 1 to 6 in Table 2, that formed
one vector containing the elements describe in Table 2. Each
variable was then normalized “to one task” by dividing each
value by seven. After normalization, each vector captured the
way in which each user interacts with BLC to solve one
generic task. Each vector had also two independent variables:
users’ cognitive styles and WA ratio (variables 7 and 8 in
Table 2). The final database contains one vector for each user
describing his/her interactions with BLC and his/her cognitive
style and WA ratio. This provides an environment in which
machine learning theories can be easily applied.
Cognitive Automatic Classification System
In this section we detail the construction of an efficient
cognitive identification system. We start with a traditional
machine learning approach using classification with neural
networks and decision trees. After studying why the results
are not satisfactory, we propose the construction of
classification systems based on regression.
Decision Trees and Neural Networks
Two traditional approaches for constructing classification
systems are decision trees and neural networks. Both theses
approaches can also be used for regression (function
approximation).
Decision tree learning (Mitchell, 1997; Winston, 1992) is a
method for approximating discrete-valued functions with
disjunctive expressions. There is a great variety of different
decision tree algorithms in the literature: Classification &
TABLE 2. Table of variables that compose a user behaviors’ vector.
Variable
Name
Information
1 BS Number of times that the user used the Basic Search functionality to solve the seven tasks
2 AS Number of times that the user used the Advance Search functionality to solve the seven tasks.
3 SE Number of times that the user used the “Word or Phrase” option from the Basic Search Interface to solve the
seven tasks.
4 ATS Number of times that the user used the “author”, “title” and “periodical” options from the Basic Search interface
to solve the seven tasks.
5 NS Number of times that the user pressed the New Search functionality to solve the seven tasks.
6 GB Number of times that the user pressed the Go Back button to solve the seven tasks.
7 CS User cognitive style obtained using CSA test (Field Dependent, Intermediate of Field Independent)
8 WA WA ratio of the user provided by the CSA test.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7
Regression Trees (C&RT) (Breiman et al, 1984), Chi-squared
Automatic Interaction Detection (CHAID) (Kass, 1980), C4.5
(Quinlan, 1993), or ID3 (Quinlan, 1986). In the context of user
modeling, decision trees can be used to classify users,
according to the level of expertise, or the interests, or the
cognitive style, etc., in order to use this information for
personalization purposes (Tsukada et. al, 2002; Beck et al.,
2003; Paliouras et al, 1999; Zhu et al., 2003; and Webb et al.,
1997). Classification rules are an alternative representation of
the knowledge obtained from classification trees. Algorithms
such as CART and C4.5 include methods to generalize rules
associated with a tree.
An Artificial Neural Network (ANN) is an information
processing paradigm that is inspired by the way biological
nervous systems process information (Fausett, 1994; Haykin,
1999). NNs are able to derive meaning from complicated
and/or imprecise data. Also, NN does not require the definition
of any metric which makes them completely application
independent. No initial knowledge about the problem that is
going to be solved is needed. These characteristics make NNs
a powerful method to model human behaviour and an ideal
technique to create user models for personalization (Bidel et
al., 2003; Sas et al., 2003; Beck et al., 2003; and Sheperd et
al., 2002). Depending on the architecture of the network,
different NN can be defined. One of the most typical
architectures is Multi-Layer Perceptrons (MLP), which are
fully connected feed-forward nets with one or more layers of
nodes between the input and the output nodes (typically there
are three layers, called the input, hidden and output layer) and
where each layer is composed of one or more artificial neurons
in parallel that use the backpropagation training algorithm.
Result Analysis using Classification: C4.5 and MLP
Two cognitive classification systems were constructed
using: (1) C4.5 as an example of classification trees and (2)
MLP as an example of neural networks. Both systems were
constructed using Weka Data Mining Software (Witten &
Frank, 1999), http://www.cs.waikato.ac.nz/~ml/weka/. MLP
was designed with a three-layer network with a sigmoidal
transfer functions and was run for 500 epochs with a
backpropagation algorithm. The training vectors consisted of
six dependent variables (variables 1-6 in Table 2) and one
independent variable, the cognitive style (variable 7 in Table
2). The output in each case was a classification system that
identifies the cognitive style of a BLC user (FD, Intermediate
or FI) by considering the set of interactions of that user. In
order to test the classification system, two testing techniques
were applied: (1) splitting and (2) cross-validation. Splitting
divided the file into 66% for training and 33% for testing and
3 cross-validation, was applied. Table 3 present the cognitive
classification results.
Classification results are not satisfactory: basically only
one out of two users is assigned their correct cognitive style. In
our opinions, the possible reasons are: (1) although it may
seem that the problem we are dealing is a traditional
classification problem in which all the instances (users) have
assigned a class (cognitive style), this is not entirely true,
because each user, originally, is assigned a number (WA ratio)
which is then translated into a class, (2) the definition of
cognitive style is something fuzzy which is not completely
understood how it translates into the context of a library
catalogue, (3) the behaviour of users within a class (cognitive
style) is not necessarily constant: two users with the same
cognitive style can have very different behaviours if for
example one of them has a WA ratio near a border with
another cognitive style, with which will share some behaviour
patterns, and the other has a WA value not near any border,
showing a pure behaviour of that cognitive style. This
characteristic makes it harder to construct an efficient
cognitive classification system.
Taking into account the previous conclusions we think that
a regression approach (function approximation), in which we
construct a system that predicts the WA ratio of a user and
from that his/her cognitive style, instead of obtaining his/her
cognitive style directly, would produce better results because
the system would avoid the ill-definition of the concept that is
trying to classify.
Result Analysis using Regression: C&RT and MLP
Two regression systems were constructed using: (1) C&RT
as an example of classification trees used for regression and
(2) MLP as an example of neural networks architecture using
for regression. MLP was designed with a two-layer network
with backpropagation learning, with a tan-sigmoid transfer
function in the hidden layer and a linear transfer function in
the output layer, which is a typical structure for regression
problems (Demuth & Beale, 1998), and was trained until the
root mean square (RMS) error was smaller than 0.01. Both
systems were designed with MATLAB
(www.mathworks.com/products/matlab), using the NN
Toolbox (www.mathworks.com/products/neuralnet) for MLP
and the Computational Statistics Toolbox (Martinez &
Martinez, 2001) for C&RT. The training vectors consisted of
six dependent variables (variables 1-6 in Table 2) and one
independent variable, the WA ratio (variable 8 in Table 2).
The output in each case was a regression system that obtained
and approximated value for the WA ratio of each user
considering the set of interactions of that user. Again, in order
to test the classification system two testing techniques were
applied: (1) splitting and (2) 3-cross validation. Table 3
TABLE 3. Classification and Regression results.
3- Cross Validation 66% Split
Classification
C4.5 45.8 % 52.9 %
MLP 60.4 % 41.1 %
Regression
C&RT 70.2% 71.4%
MLP
68.2% (5 Neurons)
56.5 % (10 Neurons)
66% (5 Neurons)
58% (10 Neurons)
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 8
presents the cognitive classification results for C&RT and
MLP, considering for MLP two different architectures with 5
and 10 neurons in the hidden layer. In order to obtain the
correct classification rate when using the regression approach,
first, the WA ratio predicted by the system was transformed
into the cognitive style, and after that, the comparison with the
correct cognitive style was made.
In general, we can see an increment in the correct
classification rate, for both MLP and decision trees. This
demonstrates that the regression approach outperforms the
classification approach, which may be because of the ability of
regression to manage the ill-definition of the concept being
classified. This result is in accordance with what Peterson,
Deary & Austin (2003) stated, that the use of category
information on the CSA is considerably less reliable than the
use of the WA ratios. When using MLP it seems that an
increment in the number of hidden neurons reduces the correct
classification rate. The best results are obtained using the
decision tree approach with C&RT, which achieves a 70%
correct classification rate, 5% higher than using MLP. We
consider that one of the limitations of using a decision tree
approach is that we lose the ability to capture the inherent
uncertainty that modelling human behaviour has. In this
context, a combination of decision trees with a soft computing
technique (like NN for example) could increase the correct
classification rate. Considering that decision trees can also be
expressed in the form of classification rules, there are in the
literature a variety of algorithms that combine classification
rules with neural networks, generally called neuro-fuzzy
systems. Neuro-fuzzy systems provide an excellent framework
to automatically create rules that learn from examples using
neural networks and that are able to handle the uncertainty and
fuzzyness of the concepts and data being used.
Neuro-Fuzzy System
Neuro-Fuzzy systems combine a knowledge representation
framework, fuzzy logic, with the learning capabilities of neural
networks (Jang & Sun, 1995; Jang and Sun, 1999).
Fuzzy Logic defines a framework in which the inherent
uncertainty of real information can be captured, modeled and
used to reason with uncertainty (Klir, 1995; Yang et al. 1994).
Fuzzy Inference Systems (FIS) are constructed using a set of
membership functions for each input (also called linguistic
labels) and Fuzzy Inference rules. Fuzzy Inference Rules take
the form ``IF x is a, THEN y is b'', where x and y are inputs of
the system and a and b are membership functions defined in x
and y respectively. Under classical logic, the THEN
implication is true if the antecedent is evaluated to be true. For
fuzzy rules, the implication is set to be true to the same degree
as the antecedent. A traditional FIS is divided into three steps:
(1) fuzzification; (2) fuzzy inference; and (3) defuzzification.
The basic idea of combining fuzzy systems and neural
networks is to design an architecture that uses a FIS to
represent knowledge and the learning ability of a neural
network to optimize its parameters. Neuro-Fuzzy systems
(NFS) use NNs to learn and fine tune rules and/or membership
functions from input-output data. NFS automate the process of
transferring expert or domain knowledge into fuzzy rules. One
of the most important NFS algorithms is ANFIS, AdaptiveNetwork-based
Fuzzy Inference Systems, (Jang, 1993), which
has been used in a wide range of applications (Bonisone,
Badami & Chiang, 1995). The combination of NN and fuzzy
sets offers a powerful method to model human behavior which
allows NFS to be used for a variety of user modeling tasks
(Lee, 2002; Stathacopoulou Grigoriadou & Magoulas, 2003;
Drigas et al. 2004; Magoulas, Papanikolau & Grigoradou,
2001).
One of the main limitations of NFS is the training time
needed, which is exponential with the dimension of the input
space. This dimensionality problem also appears with the
rules, the number and size of rules exponentially increments
with the dimensionality of the input space (with a factor given
by the number of membership functions). This complexity in
rules will also affect the execution time of NFS, i.e. the time
needed to identify the cognitive style of a user. In our case this
is of critical importance because we want to implement a
system that is able to identify the cognitive style of a user in
real-time in order to present a cognitive personalized interface.
These problems imply that, in order to efficiently implement a
NFS approach, we first need to reduce the input space of the
system.
Feature Selection by Information Gaining
The objective of this section is to identify the subset from
the six original dependent variables that better characterize
each cognitive style within the context of a neuro-fuzzy
classifier in order to avoid the dimensionality problem. To
identify which variables are more relevant we have selected
all subsets of one, two, three and four variables from the
original six, and for each combination a neuro-fuzzy system
has been trained using data splitting, with 66% for training
and 33% for testing (NFS of higher dimensionality were not
able to be trained due to the dimensionality problem). All
neuro-fuzzy systems were implemented using MATLAB’s
Fuzzy Logic Toolbox
(www.mathworks.com/products/fuzzylogic). The training
process was run for one epoch, the original fuzzy logic
knowledge base was automatically generated using grid
partition and each input was assigned two labels. For each
system we collected its Root Mean Square (RMS) training
error (the error when the testing file is the same 66% used for
training) and its RMS testing error (the error when the testing
items used were the 33% of the file not used for training).
Figure 8 presents the training and testing error for each
subset of one variable (6 subsets) ordered using the training
error. Figure 9 to Figure 11 present the same results for each
subset of two variables (15 subsets), three variables (20
subsets) and four variables (showing the 15 subsets with
smaller training error). From Figure 8, we can obtain that IN2
(AS in Table 2) provides the better training error, 0.4448, and
IN1 (BS in Table 2) the best testing error, 0.3499. Figure 9
corroborates the importance of IN2 for training error, because
the smallest training errors are always obtained using IN2 as
one of its inputs. Also, again, the smallest testing error is
obtained in combination with IN1, by the pair of inputs IN1IN6
(GB in table 2). Figure 10 and Figure 11 show that,
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9
although the incorporation of a third and fourth variables can
produce subsets of inputs with similar training errors, testing
errors are much higher. Typically the set of variables would
be chosen according to the smallest testing error, which
produces IN1-IN6 as input variables (testing error of 0.3442).
Choosing the set of variables with smallest testing error does
not imply that these variables will produce the best solution,
because each neuro-fuzzy system was run for just one epoch,
and there is no indication about how both testing and training
error will evolve. Also, considering that in this case IN1-IN6
has the highest training error for the set of two variables, we
also decided to choose a set of variables that, while having a
testing error similar to IN1-IN6, also had one of the smallest
training errors, in other words, the combination of variables
with the smallest testing error among the ones with the
smallest training error. Considering this approach we selected
the pair IN1-IN2, which has a testing error of 0.3597 but a
smaller training error than IN1-IN6. Also, this combination of
variables is very promising because it combines the variable
that minimizes the training error (IN2), with the variable that
minimizes testing error (IN1).
We reach the conclusion that an optimum system in the
sense of: (1) size of the fuzzy knowledge base, (2) training
and testing time and (3) efficiency of the classification system,
would be achieved by a two dimensional system with a
possible combination of BS and AS or BS and GP as inputs.
The following sections checks which one of these
combinations produces better results.
Result Analysis: Neuro-Fuzzy Cognitive Classification
Two neuro-fuzzy systems were constructed using the Fuzzy
Logic Toolbox implementation of ANFIS as learning
algorithm. The training vectors were composed of two
variables: BS and AS (variables 1 and 2 of Table 2) in the first
case, and BS and GP (variables 1 and 8 of Table 2) for the
second case. The independent variable was in both cases the
WA ratio (variable 8 of Table 2). Both neuro-fuzzy systems
were designed not to directly give the cognitive style, but to
obtain the WA ratio. In order to test each neuro-fuzzy system
two testing techniques were applied: (1) 66%-33% splitting
and (2) 3-cross validation. In both cases, the original fuzzy
system was automatically generated using grid partition, and
learning run for 50 epochs. The training algorithm selected the
fuzzy inference system that minimized the testing error within
FIG. 8. RMS Error for one-dimensional systems.
FIG. 9. RMS error for two-dimensional systems.
FIG. 10. RMS error for three-dimensional systems.
FIG. 11. RMS error for four-dimensional systems.
FIG. 12. Training error of the neuro-fuzzy system.
FIG. 13. Comparison between the testing WA ratios (+ signs)
and the predicted WA ratios (* signs).
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 10
the first 50 epochs. Figure 12, where dots represent testing
error and asterisks the training error, shows the learning
process and how the minimum testing error is obtained in
epoch 21 and has a value of 0.38 when using BS-AS as inputs.
Both fuzzy system are quite simple with two inputs (BS-AS or
BS-GP), one output (WA value), three membership functions
per input, three output singletons and only three rules.
Figure 13 shows the output of the testing file when using
66%-33% splitting for the system constructed with BS-AS. As
can be seen, although the exact WA ratio is not predicted, the
system is able to give very approximate values that actually
classify the user, in general, in the correct cognitive style.
Nevertheless, users that have a WA ratio near cognitive style
borders have a higher probability of being incorrectly
classified. Using splitting the correct classification rate
obtained is 82% with BS-AS and 75% with BS-GP while 3cross
validation provides a 76% correct classification rate with
BS-AS and 73% with BS-GP. The results show, that for our
problem, BS-AS captures better that BS-GP the characteristics
of each cognitive style. These results are better than the ones
provided using just a rule-based approach, as done by
C&ART, which implies that the soft computing approach
captures, to some extent, the uncertainty of modelling
cognitive styles. Also the reduction of the input space allows
to implement an ANFIS system with a real-time response,
which is specially critical for personalization of applications.
Cognitive Classification from a Personalized Application
Perspective
As we said previously, the concept of cognitive style (Field
Dependent / Intermediate / Field Independent) is actually
constructed using the concept of WA ratio (a real number in
the range of 0.6-3.0), in which WA scores below 1.03 denote
Field Dependent individuals; scores of 1.36 and above denote
Field Independent individuals; and scores between 1.03 and
1.35 are classified as Intermediate. Such classification is given
by Riding (1991), but other values for classification of
cognitive styles are also possible, i.e. the borders between
cognitive styles are fuzzy. Taking that idea into account and
also considering that our cognitive identification system will
be part of a personalized environment for BLC, it can be
possible that some of the users that have been assigned to an
incorrect cognitive style can find the interface assigned to
them by personalization useful. Users with WA ratio near
cognitive borders have a higher probability of being
incorrectly classified; nevertheless it also implies that these
users, to some extent, share the behaviours of its neighbours,
so they can also find useful their personalized interface.
Taking into account the above observation, we can define
the concept of “being a user near a cognitive border” as a user
whose WA ratio is within (-0.1, +0.1) of the border of a
cognitive style. Considering the previous assumption, those
users would find useful the two possible personalized
interfaces associated with each one of their two valid cognitive
styles. For example, a user with a WA ratio of 1.01 could be
classified as Field Dependent, or considering that 1.01 is
included in [1.03±0.1], with 1.03 the border value between FD
and Intermediate, it can also be classified as Intermediate. It
would be interesting to find out the correct classification rate
of the proposed regression systems taking into account this
definition. Table 4 presents the correct classification rates
using MLP, C&RT, ANFIS with BS-AS (which provided
better results) and the new concept of classification rate. Note
that when using ANFIS we achieve, in the worst case, a 91%
correct classification rate. These results, which arouse from an
application perspective, show that the automatic identification
of a user cognitive style within BLC is feasible, and opens the
door to automatically personalize BLC to each one of its users
from a cognitive perspective without the need to run a
cognitive style test.
Conclusions
The cognitive style of a user is a very relevant factor to
determine the way in which a user interacts with a web-based
service. This importance implies that cognitive personalized
services can be very useful to tackle the different problems that
users have when interacting with Internet, especially in
environments as relevant as digital libraries. The main
drawback of considering a cognitive personalized interface is
that in order to assign a cognitive style to a user, each user
needs to take a cognitive style test. This process is time
consuming and some users would not be willing to take it.
In this paper we have proposed an approach to overcome
the inconvenience of implementing a cognitive interface by
automatically identifying each user’s cognitive style. We have
reached two main conclusions: (1) In order to better identify
the CS of a user, due to the fuzziness of the definition of
cognitive style, a regression approach, in which we obtain the
WA ratio, outperforms a classification approach, in which the
cognitive style is directly identified, and (2) in general, and
considering that we are modeling human behavior, the use of a
soft computing approach improves the classification rate. Also
we have proposed a correct classification rate definition from
an application perspective, in which users near cognitive
borders can have two correct cognitive styles. We have
focused our study in what we consider very relevant Internet
tools such as digital libraries, using Brunel Library Catalogue
as a testing environment. The results obtained by using a
neuro-fuzzy approach in this context have shown that the
system can be applied to automatically generate user models
TABLE 4. Classification and Regression results from an application
perspective.
3- Cross
Validation
66% Split
Regression
C&RT 77.4% 79.2%
MLP 73.5% (5 Neurons) 71.4% (5 Neurons)
NFS
ANFIS – (BS,AS) 91.5% 100%
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 11
for cognitive personalization, thus avoiding the main
inconvenience of constructing cognitive personalized
interfaces. We consider that these improvements open the
door to actually implement high-scale personalized DL based
on cognitive styles.
We plan to study how this methodology applies to, first,
other digital libraries/search interfaces and, second, to other
applications/interfaces. Initial studies have shown that FD/FI is
consistent across domains and stable over time (Witkin, et al,
1977; Witkin and Goodenough, 1981). This implies that once
we have constructed a cognitive user model, the same model
can be applied to any application and it will not change over
time. We plan to check these theoretical results from an
application perspective to see if our approach can actually be
applied to a variety of applications.
Acknowledgments
The work presented in this paper is funded by the UK Arts
and Humanities Research Board (AHRB grant reference:
MRG/AN9183/APN16300).
References
Bajraktarevic, N., Hall, W., & Fullick, P. (2003). Incorporating
Learning Styles in Hypermedia Environment: Empirical
Evaluation. In The Fourth Conference on Hypertext and
Hypermedia, Workshop on Adaptive Hypermedia and Adaptive
Web-based systems AH2003.
Beck, J., Jia, P., Sison, J., & Mostow, J. (2003). Predicting Student
Help-Request Behavior in an Intelligent Tutor for Reading. In
Proc. of the 9th Int. Conf. on User Modeling, LNAI 2702, 303-
312
Bidel, S, Lemoine, L., & Piat, F. (2003). Statistical machine learning
for tracking hypermedia user behavior. In 2nd Workshop on
Machine Learning, Information Retrieval and User Modeling, 9th
Int. Conf. in UM.
Blandford, A., Stelmaszewska, H., & Bryan-Kinns, N (2001). Use of
multiple digital libraries: a case study. In Proceedings of the
JCDL’01, ACM Press.
Bonissone,P., Badami, C., & Chiang, X. (1995). Industrial
Applications of Fuzzy Logic at General Electric. In Proceedings
of the IEEE, 83(3), 450-465.
Breiman, L., Friedman, J.H., Olshen, R.A., & Stone, C.J. (1984).
Classification and Regression Trees. Wadsworth Int.
Group/Probability Series, Belmont, California, USA.
Brusilovsky, P., & Schwarz, E. (1997). User as Student: Towards an
Adaptive Interface for Advanced Web-Based Applications. In A.
Jamesson, C. Paris and C. Tasso (Eds.), User Modeling:
Proceedings of the Sixth International Conference, UM97, 177-
188.
Bryan-Kinns, N., Blandford, A., & Thimbleby, H. (2000). Interaction
Modelling for Digital Libraries. In Workshop on Evaluation of
Information Management Systems.
Callan, J., Smeaton, A., Beaulieu, M., Borlund, P., Brusilovsky, P.,
Chalmers, M., Lynch, C., Riedl, J., Smyth, B., Straccia, U., &
Toms E. (2003). Personalization and Recommender Systems in
Digital Libraries, Joint NSF-EU DELOS Working Group Report ,
URL http://www.ercim.org/publication/ws-proceedings/Delos-
NSF/Personalisation.pdf.
ChanLin, L. (1998). Students' cognitive styles and the need for visual
control in animation. Journal of Educational Computing Research,
19(4), 351-363.
Chen, S. Y., & Ford, N.J. (1997). Towards adaptive information
systems: individual differences and hypermedia. Information
Research, Vol.3 (2).
Chen, S.Y., & Macredie, R. (2002). Cognitive Styles and
Hypermedia Navigation: Development of a Learning Model.
Journal of the American Society for Information Science and
Technology, 53(1), 3-15.
Chuang, Y-R. (1999). Teaching in a Multimedia Computer
Environment: A study of effects of learning style, gender, and
math achievement. Interactive Multimedia Electronic Journal of
Computer-Enhanced Learning, 1(1) 1999. Available online:
http://imej.wfu.edu/articles/1999/1/10/.
Demuth, H., & Beale, M. (1998). Neural Network Toolbox User’s
Guide. Ver. 3.0, The Mathworks, Inc.
Drigas, A., Kouremenos, S., Vrettos, S., Vrettaros & J., Kouremenos,
D. (2004). An expert system for job matching of the unemployed.
Expert Systems with Applications 26, 217-224.
Dufresne, A. & Turcotte, S. (1997). Cognitive style and its
implications for navigation strategies. In B. Boulay and R.
Mizoguchi (eds.) Artificial Intelligence in Education Knowledge
and Media Learning System. IOS Press, Amsterdam, 287-293.
Fausett L., Fundamentals of Neural Networks, Prentice-Hall, 1994
Fink, J., Kobsa, A., & Nill, A. (1997). Adaptable and Adaptive
Information Access for All Users, Including the Disabled and the
Elderly. A. Jamesson, C. Paris and C. Tasso (Eds.), User
Modeling: Proceedings of the Sixth International Conference,
UM97, 171-173.
Ford, N. & Chen, S. Y. (2000). Individual differences, hypermedia
navigation and learning: an empirical study. Journal of
Educational Multimedia and Hypermedia, 9(4), 281-312.
Goodenough, D. (1976). The role of individual differences in field
dependence as a factor in learning and memory. Psychological
Bulletin, 83, 675-694.
Graff, M (2003). Cognitive Style and Attitudes Towards Using
Online Learning and Assesment Methods. Electronic Journal of elearning,
Vol. 1, Issue 1, pp. 21-28
Harrison, A. W. & Rainer, R. K. (1992). The influence of individual
differences on skill in end-user computing. Journal of
Management Information Systems, 9(2), 93-111.
Haykin S., Neural Networks, 2nd Edition, Prentice Hall, 1999
Hong, J., Heer, J., Waterson, S., & Landay, J.A. (2001). WebQuilt: A
Proxy-based Approach to Remote Web Usability Testing. ACM
Transactions on Information Systems, Vol. 19 (3), 263-385.
Jang, J.S. (1993). ANFIS: Adaptive-Network-Based Fuzzy Inference
Systems. IEEE Transactions on Systems, Man, and Cybernetics,
Vol. 23(3), 665-685.
Jang, J.S.R. & Sun, C.T. (1995): Neurofuzzy modelling and control.
Proceedings IEEE (3,) Vol. 83, 378-406.
Jang, J.S.R., Sun, C.T. & Mizutani E. (1997). Neurofuzzy and Soft
Computing: a Computational Approach to Learning and Machine
Intelligence. Prentice Hal.
Kass, G.V. (1980). An Exploratory Technique for Investigating
Large Quantities of Categorical Data. Applied Statistics 29: 119-
127.
Klir, J. & Yuan,B. (1995). Fuzzy Sets and Fuzzy Logic. Theory and
Applications. Prentice Hall.
Kobsa, A. (2001). Generic User Modeling Systems. User Modeling
and User-Adapted Interaction 11, 49-63.
Lee, R.S.T. (2002). iJADE IWShopper: A New Age of Intelligent
Web Shopping System Based on Fuzzy-Neuro Agent
Technology. Web Intelligence: Research and Development,
Zhong, N., Yao, Y., Liu, J., Ohsuga, S. Editors, Lecture Notes in
Artificial Intelligence, vol 2198, 403-412.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 12
Liaw, S. & Huang, H. (2003). An Investigation of User Attitude
toward search engines as an information retrieval tool. Computers
in Human Behavior, Volume 19, Issue 6, November 2003, 751-
765
Magoulas, G.D., Papanikolau, K.A. & Grigoriadou, M. (2001).
Neuro-fuzzy synergism for planning the content in a web-based
course. Informatica Vol. 25(1), 39-48.
Marchionini, G., Plaisant, C. & Komlodi, A. (1998) "Interfaces and
tools for the Library of Congress National Digital Library
Program". Information Processing & Management, 34(5), 535-
555
Marrison, D. L. & Frick, M J. (1994). The Effect of Agricultural
Students' Learning Styles on Academic Achievement and their
Perceptions of Two Methods of Instruction. Journal of
Agricultural Education, 35(1), 26-30.
Martinez, L. W., & Martinez, A.R. (2001). Computational Statistics
Handbook with MATLAB. Chapman & Hall / CRC.
Messick S. (1976). Individuality in learning. San Francisco: Jossey-
Bass.
Mitchell, T. (1997). Decision Tree Learning. In Machine Learning,
McGraw-Hill, Inc., 52-78.
Moss, N, & Hale, G. (1999). Cognitive Style and its Effect on
Internet Searching: A Quantitative Ingestigation. In European
Conference on Educational Research, Lahti, Findland 22-25.
Paliouras, G., Karkaletsis, V., Papathedorou, C., & Spyropoulos, C.
(1999). Exploiting Learning Techniques for the Acquisition of
User Stereotypes and Communities. In Proceedings of the
International Conference on User Modelling (UM '99).
Palmquist, R. A. & Kim, K.-S. (2000). Cognitive style and on-line
database search experience as predictors of Web search
performance. Journal of the American Society for Information
Science, 51(6), 558-66.
Papanikolau, K.A., Grigoradiou, M., Kornilakis, H., & Magoulas,
G.D. (2003). Personalizing the Interaction in a wed-based
Educational Hypermedia System: the case of INSPIRE. User
Modelling and User-Adapted Interaction 13: 213-267.
Peterson, E. R., Deary, I. J., and Austin, E. J. (2003) The reliability
of Riding’s Cognitive Style Analysis test. Personality and
Individual Differences, 34(5), 881-891.
Quinlan, J.R. (1986). Induction of Decision Trees. Machine Learning
(1), 81-106.
Quinlan, J.R. (1993). C4.5: Programs for Machine Learning. Morgan
Kaufmann.
Riding R. J. & Grimley, M. (1999). Cognitive style, gender, and
learning from multimedia materials in 11 year-old children.
British Journal of Education Technology, 30(1), 43-56.
Riding, R. & Rayner, S. G. (1998). Cognitive Styles and Learning
Strategies. David Fulton Publisher, London.
Riding, R. J. (1991). Cognitive Styles Analysis. Learning and
Training Technology, Birmingham.
Sas, C., Reilly, R., & O’Hare, G. (2003). A Connectionist Model of
Spatial Knowledge Acquisition in a Virtual Environment. In 2nd
Workshop on Machine Learning, Information Retrieval and User
Modeling.
Sheperd, A., Watters, C., Marath, A.T. (2002). Adaptive User
Modeling for Filtering Electronic News. In Proc. of the 35th
Annual Hawaii Intl. Conf. on System Sciences (HICSS-02), Vol.
4.
Stathacopoulou, R., Grigoriadou, M., & Magoulas, G.D. (2003). A
Neuro-fuzzy Approach in Student Modeling. Proceeding of the
9th Int. Conf. on User Modeling, UM2003, Lecture Notes in
Artificial Intelligence, vol. 2702, 337-342.
Theng, Y.L., Duncker, E., & Mohd-Nasir, N. (1999). Design
Guidelines and User-Centred Digital Libraries. Proceedings of the
Third European Conference on Research and Advanced
Technology for Digital Libraries, LNAI 1696, 167-183.
Triantafillou, E., Pomportsis, A., & Georgiadou, E. (2003). AEC-CS:
Adaptive Educational System based on Cognitive Styles. In
WASWBE’02, Workshop on Adaptive Systems for Web-based
Education, 2nd Intl. Conf. on Adaptive Hypermedia and Adaptive
Web Based Systems.
Tsukada, M., Washio, T., & Motoda, H. (2001). Automatic WebPage
Classification by Using Machine Learning Methods. In Web
Intelligence: Research and Development, LNAI 2198, 303-313.
Webb, G. I., B. C. Chiu, & M. Kuzmycz (1997). Comparative
Evaluation of Alternative Induction Engines for Feature Based
Modelling. International Journal of Artificial Intelligence in
Education, 8:97-115.
Weller, H. G., Repman, J., & Rooze, G. E. (1994). The relationship
of learning, behavior, and cognitive styles in hypermedia-based
instruction: Implications for design of HBI. Computers in the
Schools, 10(3/4), 401-420.
Winston, P. (1992). Learning by Building Identification Trees.
Artificial Intelligence. Addison-Wesley Publishing Company,
423-442
Witkin, H. A. & Goodenough, D. R. (1981). Cognitive styles:
Essence and origins: Field dependence and field independence.
New York, NY: International Universities Press
Witkin, H. A., Moore, C. A., Goodenough, D. R., & Cox, P. (1977).
Field-dependent and field independent cognitive styles and their
educational implications. Review of Educational Research, 47(1),
1-64.
Witten, I.H., & Frank E. (1999). Data Mining. Practical Machine
Learning Tools and Techniques with JAVA Implementations.
Morgan Kaufman Publishers.
Yan, J., Ryan, M. & Power, J.: (1994).Using Fuzzy Logic. Prentice
Hall.
Zhu, T., Greiner, R., & Haubl, G. (2003). Learning a Model of a
Web User’s Interests. In Proceedings of the 9th International
Conference on User Modeling, LNAI 2702, 11-21.
Zukerman, I., Albrecht, D.W., & Nicholson, A.E. (1999). Predicting
Users Request on the WWW. In Proceedings of the 7th
International Conference on User Modeling, UM99, 275-284.