On Modelling Cognitive Styles of Users in Adaptive Interactive

Systems using Artificial Neural Networks

Efi Papatheocharous, Marios Belk, Panagiotis Germanakos and George Samaras

Department of Computer Science, University of Cyprus, Nicosia, Cyprus

Keywords: User Modelling, Cognitive Styles, CAPTCHA, Artificial Neural Networks.

Abstract: User modelling in interactive Web systems is an essential quality to optimally filter, personalise and adapt

their content and functionality to serve the intrinsic needs of individual users. The mechanism for obtaining

the user model needs to be intelligent, adaptive and transparent to the user, in the sense that user experience

should not be disrupted or compromised. Human factors are extensively employed lately for enriching user

models by capturing more intrinsic perceptual characteristics of the users. Accordingly, this paper proposes

the use of Artificial Neural Networks (ANNs) for attaining cognitive styles of users in adaptive interactive

systems. One of the main benefits is the automatic prediction of cognitive typologies of users by avoiding

psychometric tests, which are among the typical ways of constructing user profiles and are particularly time-

consuming. Furthermore, ANNs can efficiently model the relationship between cognitive styles and user

interaction. The experimental setup and the results obtained show that ANNs are suitable for predicting the

cognitive styles ratio of users in respect to their actual cognitive style ratio value.

1 INTRODUCTION

Adaptive interactive systems (Brusilovsky et al.,

2007) have become progressively popular since the

late 2000s due to the exponential increase of users

and availability of digital information, mainly with

regards to interactive systems on the World Wide

Web. Based on various definitions given to date

(Brusilovsky et al., 2007); (Perkowitz and Etzioni,

2000); (Frias-Martinez et al., 2005) an adaptive

interactive system is capable to automatically or

semi-automatically adapt its information architecture

and functionality to the needs and preferences of its

users with the aim to provide a personalised and

positive user experience.

Adaptation of the functionality and content in

interactive systems heavily depends on successful

user modelling. The user model is the representation

of static and dynamic information about an

individual, and it represents an essential entity for an

adaptive interactive system. User models aim to

provide or guide adaptation effects in adaptive

interactive systems (i.e., the same system can look

different and provide diverse functionalities to users

with dissimilar user models).

Popular user characteristics considered in user

modelling of adaptive interactive systems are the

user’s knowledge, interests, goals, background, and

cognitive styles (Brusilovsky et al., 2007). This

work focuses on modelling cognitive style of users,

which represents an individually preferred and

habitual approach to organising and representing

information (Riding, 2001).

Psychometric tools have been primarily used for

classifying users to particular cognitive typologies

(Brusilovsky et al., 2007), which could be further

used to adapt the content and functionality of

interactive systems. However, systems that offer

personalised content based on cognitive styles

heavily depend on the users’ willingness to dedicate

a considerable amount of time for participating in

the user modelling process. Therefore, an imperative

need has been identified for intelligent user

modelling in adaptive interactive systems to offer

automatically personalised content but without

requiring any effort on behalf of the user.

In this context, nature-inspired intelligent

methodologies such as Artificial Neural Networks

(ANNs) could be used as a powerful technique to

dynamically and transparently model human

behaviour in Web-based applications. ANNs

comprise of emerging effective modelling

techniques especially in nonlinear conditions and

where the development of conventional relations in a

563

Papatheocharous E., Belk M., Germanakos P. and Samaras G..

On Modelling Cognitive Styles of Users in Adaptive Interactive Systems using Artiﬁcial Neural Networks.

DOI: 10.5220/0004158905630569

In Proceedings of the 4th International Joint Conference on Computational Intelligence (NCTA-2012), pages 563-569

ISBN: 978-989-8565-33-4

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

particular context becomes impractical and

cumbersome. ANNs have been extensively used for

user modelling, mainly for classification and

recommendation in order to group users with the

same characteristics (Frias-Martinez et al., 2005).

Accordingly, ANNs could be used to predict the

users’ cognitive characteristics based on their

interactions with the system, something which has

not been investigated thoroughly yet, to the best of

the authors’ knowledge.

To this end, the work presented introduces an

intelligent user modelling approach for eliciting the

cognitive styles of users based on ΑΝΝs, combined

with a psychometric measurement that highlights

differences in cognitive styles of users as well as

interaction data of users within a Web-based

environment. In particular, an ANN has been trained

based on the cognitive style of users and interaction

data with the aim to predict the cognitive style of

newly entered users’ based on their interaction data.

As interaction data we have considered specific

metrics of CAPTCHA (Completely Automated

Public Turing test to tell Computers and Humans

Apart) (Von Ahn et al., 2004) since it is a common

mechanism used online daily by millions of users

(e.g., reCAPTCHA (http://google.com/recaptcha)

estimates that over 200 million reCAPTCHAs are

completed daily). The main objective of the paper is

to investigate whether specific metrics of

CAPTCHA mechanisms could be used by an ANN

to predict the users’ cognitive characteristics. The

identification of users having specific cognitive and

interaction style/pattern will ultimately help in

defining various adaptation mechanisms required to

assemble and target a different user interface

experience in Web-based environments for various

cognitive typologies of users.

The paper is organised as follows: In Section 2,

we provide related work on adaptive interactive

systems that make use of ANNs. In the same section

we present the related theoretical background. The

experimental setup and data metrics are presented in

Section 3. In Section 4, we analyse the experimental

results and consequently, we conclude the paper and

describe our directions of future work in Section 5.

2 RELATED BACKGROUND

Numerous researchers have attempted to use ANNs

in the context of adaptive interactive systems,

primarily for classification of users with the same

characteristics and creation of user models with the

aim to recommend and adapt Web content. For

example, Kim et al. (2004) have proposed an ANN-

based collaborative filtering method that investigates

the possibility of identifying and predicting the

correlation between users or items in a Web

environment using a Multi-Layer Perceptron (MLP).

Chou et al. (2010) aim to identify the users’ prior

knowledge for specific products in e-commerce

applications by analysing their navigation patterns

through Web mining and constructing a Back-

Propagation Network (BPN) (Wu et al., 2006) that

uses a supervised learning method and a feed-

forward architecture, in order to predict the users’

potential future needs. Magoulas et al. (2001) use

ANNs to learn and fine tune rules and/or

membership functions from input-output data to be

used in a Fuzzy Inference System (FIS). In

particular, they have proposed a

classification/recommendation system with the aim

to plan the learning content of a course according to

the student’s level of knowledge.

Taking into consideration the abovementioned

works, this paper examines the potential of ANNs

for predicting users’ cognitive characteristics based

on their preference and ability in solving CAPTCHA

challenges in order to offer an automatic way for

capturing their typology and adapting accordingly

the content and functionality of interactive systems.

The overall benefit of modelling users’ cognitive

style through an automatic mechanism would be to

minimise the effort of users performing specially

designed psychometric tests (Brusilovsky et al.,

2007) and instead model the users’ cognitive styles

with a dynamic, and not visible to the users, user

modelling mechanism. The proposed method elicits

similar groups of users based on their interaction

with CAPTCHA mechanisms and investigates how

these groups may be related to cognitive styles. To

the best of the author’s knowledge, this is among the

first attempts to study the relation between the

cognitive style of users and their interaction data in

Web-based systems, apart from sporadic attempts,

the first of which utilised a number of clustering

techniques to understand human behaviour and

perception in relation with cognitive style, expertise

and gender differences of digital library users (Frias-

Martinez et al., 2007), and a second more recent

research attempt which studied the connection

between the way people navigating in a museum and

electronic encyclopaedia system and the way they

preferred to process information cognitively

(Antoniou and Lepouras, 2010); (Belk et al., 2012).

The rest of this section presents the theoretical

background on cognitive styles and ANNs.

IJCCI2012-InternationalJointConferenceonComputationalIntelligence

564

2.1 Cognitive Styles

Among the numerous proposed theories of

individual styles (e.g., (Riding, 2001; (Felder and

Silverman, 1988); (Witkin et al., 1977), the proposed

work utilises Riding’s Cognitive Style Analysis

(CSA) that classifies users to the cognitive

typologies of Verbal-Intermediate-Imager (Riding,

2001). The so called Verbal/Imager dimension refers

to how individuals process information. Users that

belong to the Verbal class can proportionally process

textual and/or auditory content more efficiently than

images, whereas users that belong to the Imager

class the opposite. Users that belong in between the

two end points (i.e., Intermediate) do not differ

significantly with regards to information processing.

In this regard, we consider that Riding’s CSA

implications can be mapped on Web environments,

since they consist of distinct scales that respond

directly to different aspects of the Web space. The

CSA implications can provide guidelines in the

context of Web design (i.e., selecting how to present

visual or verbal content) and is probably one of the

most inclusive theories, since it is derived from the

common axis of a number of previous theories

(Riding and Cheema, 1991).

2.2 Artificial Neural Network Models

The use of Artificial Neural Network (ANN)

(Haykin, 1999) models in this work was based on

the advantage they have in offering automatic

computations that may dynamically provide

guidelines in adapting the content of interactive

systems in a way that is not disruptive to the user.

The ANN model has been successfully used across

many interdisciplinary areas and extensively in cases

of automatically extracting patterns, decisions, or

transformations based on human behaviour due to

the ability they have to approximate reasoning

processes in a model-free manner, i.e., without

requiring knowledge of a function or relation a priori

(Haykin, 1999).

This work utilises a Multi-layer Perceptron

(MLP) model which is a popular and flexible

mechanism for the representation of common

characteristics of users and for obtaining predictions

based on these characteristics. The resulting

prediction may ultimately be used for providing

personalised content and functionality.

3 EXPERIMENTAL SETUP

3.1 Sampling and Procedure

A total of 93 individuals participated voluntarily in a

study carried out within February 2012. All

participants were undergraduate students and their

age varied from 17 to 20. A Web-based

psychometric test, exploiting Riding’s CSA (Riding,

2001), was developed that measures the response

time of specific statements (i.e., identify whether a

statement is true or false) and computes the ratio

between the response times for each statement type

in order to highlight differences in cognitive style.

Furthermore, one text- and one picture-based

CAPTCHA mechanism were developed using

available open-source software (Elson et al., 2007);

(Golle, 2008); (Secureimage, 2012). In Figure 1 and

Figure 2 we illustrate an example of the text- and

picture-based CAPTCHA mechanisms used during

the study, respectively. The text-based mechanism

produced distorted images of random characters

whilst the picture-based mechanism produced

pictures. During the experiment participants were

asked to reproduce the distorted random characters

and to select the appropriate pictures belonging to a

specific group (e.g., select pictures that illustrate

cats) in order to solve the CAPTCHA challenge.

Both CAPTCHAs contained a refresh button that

initialised the CAPTCHA with a new sequence of

characters or pictures.

Figure 1: Text-based CAPTCHA used in the study.

Figure 2: Picture-based CAPTCHA used in the study.

The participants visited a Web-page but before

entering the Web-page they were asked to solve a

number of CAPTCHA challenges. In addition, the

users were first required to choose between the two

variations of CAPTCHAs (i.e., text- vs. picture-

based) and then solve the preferred CAPTCHA

challenge. The same task of choosing between the

two CAPTCHA types was repeated 10 times in

order to offer the chance to the users to try out the

OnModellingCognitiveStylesofUsersinAdaptiveInteractiveSystemsusingArtificialNeuralNetworks

565

variations of CAPTCHAs offered and also increase

the chance of optimum measurements of

performance and ability of the users by taking

average values of their overall effort in solving the

CAPTCHA challenges, rather than taking just one-

off measures. After solving the CAPTCHA

challenge, the users were redirected to the

psychometric test aiming to identify the users’

cognitive styles which will be used by the ANN

model for verification purposes (i.e., whether the

prediction obtained from the ANN model regarding

cognitive style ratio is correct).

3.2 Data Metrics’ Definition

The following metrics were utilised for monitoring

the usage of CAPTCHA in the study.

 Preference; how many text- and picture-based

CAPTCHAs were solved by each user,

 Processing Time; average time (in seconds)

required per user to solve a text- and picture-based

CAPTCHA,

 Ability; how many attempts and how many

attempts on average were needed by each user to

successfully solve a text- and picture-based

CAPTCHA.

Based on the aforementioned metrics, a browser-

based logging facility was implemented with JQuery

JavaScript Library (http://jquery.com) to collect the

CAPTCHA usage/interaction data. The interaction

data were used to predict the cognitive ratio of new

users (that were not used in the training phase of the

algorithms developed) utilising ANNs as explained

in the experiments section.

4 EXPERIMENTS AND RESULTS

4.1 Experimental Design

The ANN designed and utilised in this work were

implemented in Matlab R2011a (http://www.

mathworks.com/products/matlab). As already

mentioned, a typical ANN was developed for

predicting the cognitive style of users based on their

interaction with the experimental environment; a

Multi-layer Perceptron (MLP) ANN with one output

neuron for calculating the cognitive style ratio. The

nodes of the ANN were organised in layers and

forming the so-called input-hidden-output layers. At

the input layer the aforementioned metrics were

used.

Initially, the data were normalised and then they

were randomly split into three subsets, the training,

the validation and the testing subset consisting of

60%, 20% and 20% of the original data samples.

The inputs were inserted in the computational

neurons and the output (cognitive typology) was

compared with the desired output (actual cognitive

typology measured using the psychometric test). The

predictive power of the ANN was measured on the

testing subset, the validation subset was used as a

pseudo-test set in order to evaluate the quality of the

network during training and the training subset was

used for obtaining the optimal network model.

The weights were adjusted using a gradient

(steepest descent) algorithm until the desired output

is achieved; this process is called the training of the

network. Various architectures were evaluated in

order to find the optimal one, which comprise of

single hidden layer ANN with varying number of

hidden nodes (i.e., nodes in the internal layer or

Number of Hidden Nodes (NHN)). Specifically, the

varying number of internal neurons examined varied

from 6 to 16, with step size equal to 1, in order to

identify the optimum performing networks. An early

stopping of the training process was also performed

to stop training when the validation error was

increased.

In addition, the following parameters were

selected as they did not demonstrate any instability

in the training process of the networks: The

algorithm used for training was the Levenberg-

Marquardt (trainlm) backpropagation algorithm

(Rumelhart et al., 1986) which is usually very

efficient, but it requires a lot of memory to run. The

maximum number of epochs was set to 100 and the

performance was evaluated using the MSE with reg

performance function, which takes the weight sum

of two factors the mean squared error and the mean

squared weight and bias values.

Finally, a cross validation process was followed

(namely holdout cross validation (Weiss and

Kulikowski, 1991)) to ensure the generalisation of

the model on larger datasets, which is necessary to

be performed especially in cases of small datasets.

4.2 Results and Discussion

Initially, a statistical analysis of the metrics showed

that most users irrespective of their cognitive

typology preferred to solve the picture-based

CAPTCHAs. This preference may be interpreted by

taking into consideration that the majority of Web

application providers utilise text-based CAPTCHA,

and thus, users wanted to try out the different

CAPTCHA type (i.e., the picture-based CAPTCHA)

IJCCI2012-InternationalJointConferenceonComputationalIntelligence

566

out of curiosity and due to the attractiveness images

have over text. In addition, this preference seemed to

have affected the observations obtained based on the

processing time and ability metrics of the users,

since an indirect relation between the CAPTCHA

metrics was observed.

Therefore, the model-free ANN method was

utilised next to investigate the existence of patterns

among users’ cognitive typologies preference,

performance and ability in solving CAPTCHA

challenges. The ANN offered an automatic way for

predicting the typology of users and adapting

accordingly the content and functionality of

interactive systems.

The results of the ANNs employed are explained

in this section. The best performance of the ANN

(with MSE=0.01) was obtained at the 8th

experimental repetition and at the 9th epoch. The

results for the training, validation and testing sets of

this network are presented in Table 1 for the

performance metrics of Mean Magnitude of Relative

Error (MMRE), Mean Absolute Error (MAE), Mean

Z Error Ratio (MZ), Median Magnitude of Relative

Error (MdMRE), Mean Absolute Error (MdAE) and

Mean Z Error Ratio (MdZ). The performance

metrics are calculated based on the local measures

(described in equations (1)-(3) in Table 2) by using

their means and medians respectively.

Figure 3 shows a comparison of the actual

cognitive style ratio of the samples used during the

testing phase compared to the predicted ratio

obtained from the ANN model with architecture 8-

13-1.

Figure 3: Actual vs. Predicted cognitive style ratio of

ANN 8-13-1 during testing phase.

From the results obtained we observe that the

proposed ANN model is suitable for recognising the

cognitive type of users that did not participate in the

training of the model (i.e., this can be observed in

the prediction evaluation results of the Test set in

Table 1). Observing the rest ANN models

constructed we conclude that they present a robust

method to distinguish Verbal/Imagers and

Intermediates users. This observation is obvious

from the results of Table 3 which show on average

the performances of various ANN architectures

(varying the Number of Hidden Neurons (NHN))

during the testing phase.

The main result obtained from this work is that

ANN models can be effectively trained, even with

using just a small sample of users, and may reach to

very accurate predictions of the cognitive ratio of

users that have not participated in the training and

construction process of the models. This proposes

that utilising only CAPTCHA-related metrics the

probability of reaching to accurate approximations

of the cognitive styles of users is quite high.

Table 1: MLP ANN 8 -13-1 performance results.

Set MMRE MdMRE MAE MdAE MZ MdZ

Train 0.087 0.056 0.084 0.059 1.024 1.023

Val 0.135 0.095 0.127 0.091 1.053 1.045

Test 0.144 0.102 0.135 0.130 1.032 1.002

Table 2: Local performance metrics.

MRE



(1)

EiAii

YYAE 

(2)

Z 

(3)

Table 3: Cross validation testing performance results of

MLP ANN.

NHN

MRE

MAE

MZ MdZ

8 0.180 0.153 0.173 0.147 1.071 1.046

9 0.160 0.128 0.152 0.130 1.061 1.027

10 0.166 0.137 0.155 0.141 1.096 1.073

11 0.141 0.102 0.135 0.102 1.032 1.002

12 0.176 0.136 0.164 0.140 1.087 1.046

13 0.143 0.113 0.135 0.110 1.049 1.024

14 0.141 0.111 0.133 0.112 1.053 1.025

15 0.135 0.108 0.130 0.109 1.023 1.002

16 0.190 0.172 0.183 0.166 1.074 1.036

5 CONCLUSIONS

The purpose of this paper was to present results of

OnModellingCognitiveStylesofUsersinAdaptiveInteractiveSystemsusingArtificialNeuralNetworks

567

an experimental setup, in order to increase our

understanding on automatically attaining cognitive

styles based on specific Web interaction data of

users. Specific data metrics of CAPTCHA

mechanisms have been proposed and utilised by an

Artificial Neural Network (ANN), with the aim to

predict the users’ cognitive style.

The experimental process of the ANN yielded

very promising results for the sample examined. In

particular, the results obtained with ANNs for

predicting the cognitive styles ratio of individuals

were particularly successful in respect to their real

cognitive style ratio value. This indicates that

techniques such as ANNs are suitable for predicting

users’ cognitive typology using their interaction data

with CAPTCHA-related challenges.

The practicality and significance of this work is

that the suggested ways of capturing intrinsic

characteristics of users, like cognitive styles, and

their analysis through intelligent techniques may be

more effective and less time and effort consuming

than traditional instruments since they might

optimise the user modelling process. However, in

order to build a cohesive user model psychometric

tests are not yet to be replaced since this study is still

in its very early stages.

The meaning of the relation between cognitive

styles and interaction data needs to be further

examined to reach to a more cohesive user model

and effectively guide adaptation in interactive

systems. Future work includes further

experimentation for investigating these relations and

further employing fuzzy algorithms or other

Artificial Intelligence (AI) methods to determine the

degree of adaptation based on user profiles and the

correlations obtained by ANN models such as the

ones used in this work.

ACKNOWLEDGEMENTS

The work is co-funded by the EU project Co-

LIVING (60-61700-98-009), smarTag (University of

Cyprus) and PersonaWeb (Cyprus Research

Promotion Foundation).

REFERENCES

Antoniou, A., Lepouras, G., 2010. Modelling Visitors'

Profiles: A Study to Investigate Adaptation Aspects

for Museum Learning Technologies. J. Computing

Cultural Heritage. 3(2), 1-19.

Belk, M., Papatheocharous, E., Germanakos, P., Samaras,

G., 2012. Investigating User Models through

Cognitive Styles and Navigation Behavior in Web 2.0

Environments using Clustering Techniques. J. Web 2.0

engineering: new practices and emerging challenges,

Elsevier (submitted).

Brusilovsky, P., Kobsa, A., Nejdl, W., 2007. The Adaptive

Web: Methods and Strategies of Web Personalization.

Springer, Heidelberg, Berlin.

Chou, P., Li, P., Chen, K., Wu, M., 2010. Integrating web

mining and neural network for personalized e-

commerce automatic service. Expert Systems with

Applications, Elsevier. 37(4), 2898-2910.

Elson, J., Douceur, J., Howell, J., Saul, J., 2007. Asirra: A

CAPTCHA that Exploits Interest-Aligned Manual

Image Categorization. In CCS’07, International

Conference on Computer and Communications

Security, ACM Press. 366-374.

Felder, R., Silverman, L., 1988. Learning and Teaching

Styles in Engineering Education. Engineering

Education 78(7), 674-681.

Frias-Martinez, E., Chen, S. Y., Macredie R. D., Liu, X.,

2007. The Role of Human Factors in Stereotyping

Behavior and Perception of Digital Library Users: A

Robust Clustering Approach. J. User Modelling and

User-Adapted Interaction, Springer, 17(3), 305-337.

Frias-Martinez, E., Magoulas, G., Chen, S., Macredie, R.,

2005. Modelling Human Behavior in User-Adaptive

Systems: Recent Advances Using Soft Computing

Technique. Expert Systems with Applications, 29(2),

320-329.

Golle, P., 2008. Machine Learning Attacks Against the

Asirra CAPTCHA. In CCS’08, International

Conference on Computer and Communications

Security, ACM Press. 535-542.

Haykin, S., 1999. Neural Networks: A Comprehensive

Foundation, Prentice Hall. Upper Saddle River, NJ.

edition.

Kim, M., Kim, E., Ryu J., 2004. A Collaborative

Recommendation Based on Neural Networks. In Lee,

Y., Li, J., Whang, K., Lee, D. (Eds.), Database

Systems for Advanced Applications. LNCS, Springer,

2973, 425-430.

Magoulas, G. D., Papanikolaou, K. A., Grigoriadou, M.,

2001. Neuro-fuzzy synergism for planning the content

in a web-based course. Informatica, 25(1), 39-48.

Perkowitz, M., Etzioni, O., 2000. Adaptive Web Sites.

Communications of the ACM, ACM Press. 43(8), 152-

158.

Riding, R., 2001. Cognitive Style Analysis - Research

Administration. Learning and Training Technology.

Riding, R., Cheema, I., 1991. Cognitive styles - An

Overview and Integration. Educational Psychology.

11, 3/4, pp. 193-215.

Rumelhart, D. E., Hinton, G. E., Williams, R. J., 1986.

Learning Internal Representations by Error

Propagation. In Parallel Distributed Processing:

Explorations in the Microstructure of Cognition, MIT

Press. 1, 318-362.

Securimage v.3.0., 2012. http://www.phpcaptcha.org.

Von Ahn, L., Blum, M., Langford, J., 2004. Telling

IJCCI2012-InternationalJointConferenceonComputationalIntelligence

568

Humans and Computers Apart Automatically.

Communications ACM, ACM Press. 47(2), 56-60.

Weiss, S., Kulikowski, C. A., 1991. Computer Systems

that Learn: Classification and Prediction Methods

from Statistics, Neural Nets, Machine Learning, and

Expert Systems, Machine Learning Series, Morgan

Kaufmann Publishers, Inc., San Mateo, CA.

Witkin, H. A., Moore, C. A., Goodenough, D. R., Cox, P.

W., 1977. Field-dependent and Field-independent

Cognitive Styles and their Educational Implications.

Review of Educational Research. 47(1), 1-64.

Wu, D., Yang, Z., Liang, L., 2006. Using DEA-neural

network approach to evaluate branch efficiency of a

large Canadian bank. Expert Systems with

Applications, Elsevier. 31, 108-115.

OnModellingCognitiveStylesofUsersinAdaptiveInteractiveSystemsusingArtificialNeuralNetworks

569