**5.3 Identifying attributes of the alternatives**

The main hypothesis is that a course can be represented by a set of attributes that would define its attractiveness to a particular trainee (see Table 1). Courses are considered to be heterogeneous alternatives where decision makers may have different choice sets, evaluate different attributes, and assign diverse values for the same attribute of the same alternative.



#### **5.4 Choice model**

44 Security Enhanced Applications for Information Systems

In order to develop the *Training Advisor*, we used theories of individual choice behaviour analysis (Ben-Akiva & Lerman, 1992; ChoiceStream, 2004; Chaptini, 2005). Discrete choice analysis is the modelling of individuals' choices from a set of mutually exclusive and collectively exhaustive alternatives. A decision maker is modelled as selecting the alternative with the highest utility among those available at the time the choice is made. An operational model consists of parameterised utility functions in terms of observable independent variables and unknown parameters the values of which are estimated from a sample of observed choices made by decision makers when confronted with a choice

The framework for choice theories can be viewed as an outcome of a sequential decisionmaking process that includes the following steps: 1) Definition of the choice problem; 2) Generation of alternatives; 3) Evaluation of attributes of the alternatives; and 4) Choice

In our case, the decision makers are seafarers and employees of the shipping industry. These

The recommender system acts as an automated training advisor. It facilitates a bundle of courses choice selection task by recommending courses that would satisfy employees'

The underlying hypothesis is that trainees perceive courses as a bundle of attributes. The utility of a course to a particular individual is a function of its attributes. Once those attributes are defined, discrete choice models can be used to calculate the utility of a set of

All the courses offered at various training institutions define the universal choice set of alternatives. This includes courses that are feasible during the decision process. The feasibility of an alternative is defined by a variety of constraints such as course offering and

The additional complexity of the problem stems from the fact that the training advisor needs

The main hypothesis is that a course can be represented by a set of attributes that would define its attractiveness to a particular trainee (see Table 1). Courses are considered to be heterogeneous alternatives where decision makers may have different choice sets, evaluate different attributes, and assign diverse values for the same attribute of the same

to recommend a combination of courses that may be offered by different institutions.

decision makers face different choice situations and have widely different tastes.

courses, and the bundle of courses with the highest utilities would be selected.

personal preferences and suit their abilities and interests.

**5. The model** 

situation.

Model.

**5.1 The choice problem** 

**5.2 Generation of alternatives** 

alternative.

scheduling requirements and prerequisites.

**5.3 Identifying attributes of the alternatives** 

The choice model involves the estimation of a preference function, based on the attributes presented above. The estimation is based on stated preferences data, which are expressed responses to hypothetical scenarios presented to the employees.

With regards to the development of the training advisor it is important to estimate models in a relatively short time frame in order to deliver online recommendations.

Therefore, two models are developed:

*Training Needs Module* - An offline model system: that relies on advanced choice models to estimate the base parameters for different trainee's profiles; and

Development of an e-Learning Recommender System

Using Discrete Choice Models and Bayesian Theory: A Pilot Case in the Shipping Industry 47

Training methodology Classic lecture 552

Training material Printed 811

Duration of Training Academic semester 432

**Attributes Levels Distribution of** 

On-the-Job 430

 Digital 548 Audiovisual 305

Flexible 795

Institute Governmental 1355 Cost No fee 705

Location Ashore 739 On-board 476

Certification Certificate 1123 Assessment procedure Exams 783

**Rating Level Frequency Percent**  1 Most Unlikely 159 9,6 2 More Unlikely 112 6,7 3 Unlikely 137 8,2 4 In The Middle 413 24,8 5 Likely 298 17,9 6 More Likely 269 16,2 7 Most Likely 276 16,6

Regression models were run with as dependent variable the choice of course and as

..... 0 11 22 *<sup>y</sup> XX Xk k*

500 &

 

0 1 3 4 5

*y SelfLearning OnTheJob VR Govermental Flexible NoFees GT Euros Certificate Exams Ashore onBoard*

 

6 7 8 9 10 2

 

Table 2. Number of Observations Administered at Each Level

**Preference** 

Table 3. Choice of Course

or

**5.7 Model estimation results** 

 

independent variables the attributes of the course.

These result in regression-type models of the following form:

 

Self learning 682

Short seminars 437

100-500 Euros 528 >=500 Euros 431

Both ashore and on-board 449

**Total** 1664 100,0

 

  **Observations** 

*Training Advisor Module* - A real time model system: that is based on the above models and in real time, estimates and customizes the parameters to each individual using Bayesian Techniques, and give fast recommendations. Bayes's theorem calculates the probability of a new event on the basis of earlier probability estimates which have been derived from empirical data. A key feature of Bayesian methods is the notion of using an empirically derived probability distribution for a population parameter. The Bayesian approach permits the use of objective data or subjective opinion in specifying a prior distribution. With the Bayesian approach, different individuals might specify different prior distributions. Bayesian methods have been used extensively in statistical decision theory. In this context, Bayes's theorem provides a mechanism for combining a prior probability distribution for the states of nature with new sample information, the combined data giving a revised probability distribution about the states of nature, which can then be used as a prior probability with a future new sample, and so on. The intent is that the earlier probabilities are then used to make ever better decisions. Thus, this is an iterative or learning process, and is a common basis for establishing computer algorithms that learn from experience (Greene, 1993).

#### **5.5 Training needs module**

This section presents the development of the training needs module and estimation results of the training advisor offline models that are used in our approach.

#### **5.6 Data collection**

In order to identify the seafarers' needs and develop the SLIM-VRT training advisor a field study was conducted. The target sample included members of seafarer's unions, shipping office's personnel and people working at shore based activities related to shipping sector. A total of 5000 questionnaires were sent to crew and shore based personnel, as well as to students of marine academies. 1195 completed questionnaires were received, corresponding to a 24% response rate. From these completed "employee's questionnaires", 59% (710 seafarers and employees) were Greek, 10% (115 seafarers and employees) were from the U.K, 7% (85 seafarers and employees) were from Spain and the rest 24% (285 seafarers and employees) were from other countries (Norway, Ukraine, Egypt, the Philippines, India, etc.). The multinational and multicultural character of the sample represents the decision making behaviour of major nationality groups in the shipping industry. The questionnaire included one to two Stated Preferences Experiments for Self-Learning for work. In each scenario individuals were presented with a course, described by several attributes and were asked to state their preference for following such a course.

A total of 1664 observations were used for estimating the preference models. Table 2 presents the distribution of the observations of the course attributes included in model estimations as independent variables.

We can see a very good distribution of the observations among the levels of attributes. This suggests a successful distribution of the stated preferences experiments between subjects.

Table 3 presents the distribution of observations of the dependent variable, or preference rating, taking the value of 1 if the individual is most unlikely to take the course and 7 if the individual is most likely to take the course.


Table 2. Number of Observations Administered at Each Level


Table 3. Choice of Course

#### **5.7 Model estimation results**

Regression models were run with as dependent variable the choice of course and as independent variables the attributes of the course.

These result in regression-type models of the following form:

$$y = \beta\_0 + \beta\_1 X\_1 + \beta\_2 X\_2 + \dots + \beta\_k X\_k$$

or

46 Security Enhanced Applications for Information Systems

*Training Advisor Module* - A real time model system: that is based on the above models and in real time, estimates and customizes the parameters to each individual using Bayesian Techniques, and give fast recommendations. Bayes's theorem calculates the probability of a new event on the basis of earlier probability estimates which have been derived from empirical data. A key feature of Bayesian methods is the notion of using an empirically derived probability distribution for a population parameter. The Bayesian approach permits the use of objective data or subjective opinion in specifying a prior distribution. With the Bayesian approach, different individuals might specify different prior distributions. Bayesian methods have been used extensively in statistical decision theory. In this context, Bayes's theorem provides a mechanism for combining a prior probability distribution for the states of nature with new sample information, the combined data giving a revised probability distribution about the states of nature, which can then be used as a prior probability with a future new sample, and so on. The intent is that the earlier probabilities are then used to make ever better decisions. Thus, this is an iterative or learning process, and is a common basis for establishing

This section presents the development of the training needs module and estimation results

In order to identify the seafarers' needs and develop the SLIM-VRT training advisor a field study was conducted. The target sample included members of seafarer's unions, shipping office's personnel and people working at shore based activities related to shipping sector. A total of 5000 questionnaires were sent to crew and shore based personnel, as well as to students of marine academies. 1195 completed questionnaires were received, corresponding to a 24% response rate. From these completed "employee's questionnaires", 59% (710 seafarers and employees) were Greek, 10% (115 seafarers and employees) were from the U.K, 7% (85 seafarers and employees) were from Spain and the rest 24% (285 seafarers and employees) were from other countries (Norway, Ukraine, Egypt, the Philippines, India, etc.). The multinational and multicultural character of the sample represents the decision making behaviour of major nationality groups in the shipping industry. The questionnaire included one to two Stated Preferences Experiments for Self-Learning for work. In each scenario individuals were presented with a course, described by several attributes and were asked to

A total of 1664 observations were used for estimating the preference models. Table 2 presents the distribution of the observations of the course attributes included in model

We can see a very good distribution of the observations among the levels of attributes. This suggests a successful distribution of the stated preferences experiments between subjects.

Table 3 presents the distribution of observations of the dependent variable, or preference rating, taking the value of 1 if the individual is most unlikely to take the course and 7 if the

computer algorithms that learn from experience (Greene, 1993).

state their preference for following such a course.

estimations as independent variables.

individual is most likely to take the course.

of the training advisor offline models that are used in our approach.

**5.5 Training needs module** 

**5.6 Data collection** 

$$\begin{aligned} y &= \beta\_0 + \beta\_1 \text{Self/Learning} + \beta\_2 \text{OnThe} \text{Ide} \text{lob} + \beta\_3 \text{VR} + \beta\_4 \text{Government} \text{al} + \beta\_5 \text{Flexible} \\ &+ \beta\_6 \text{NoFees} + \beta\_7 \text{GT500} \text{Eurros} + \beta\_8 \text{Certificate} + \beta\_9 \text{Exans} + \beta\_{10} \text{Ashore} \text{ \& on Board} \end{aligned}$$

Development of an e-Learning Recommender System

saved to be used in the Training Advisor:

*s*

variance-covariance matrix of the prior

*g g*

*N* number of experiments presented to the individual, and

*s* 

**5.8 Training advisor module** 

The steps followed are: **Step 1.** Subject profile

**Step 2.** Elicitation

Assume:

is asked of his preferences.

in which each individual belongs.

Assume *g* 1,...,*G* number of groups.

*<sup>n</sup> y* (Nx1 vector) = the ratings of course n. **Step 3.** Creation of Individualized Data

*<sup>n</sup> y* (Nx1 vector) = the ratings of the courses. *Xn* (1xK matrix) = attributes of course n

The new data is ( *<sup>n</sup> y* , *Xn* ) where:

K= number of attributes

are appended.

equation as follows:

Using Discrete Choice Models and Bayesian Theory: A Pilot Case in the Shipping Industry 49

The basic models developed for each group of trainee, as described in the previous section, need to be customized for each respondent. Bayesian theory is used to provide the

For each Group the following information, the following outputs of survey regressions are

*g*

A number of questions are asked at the beginning of the session regarding the characteristics of the trainees. A number of these characteristics (the *X* 's) define the Group

The individual is presented with a sample of courses (using always the same attributes) and

In the new table the respondent has *N* ratings. To each rating the K attributes of the course

Bayesian updating is used to calculate the personalized coefficients of the preference

1 2 1 2 1 *n g <sup>g</sup> g g s XX s Xy*

   

**Step 4.** Develop the personalized preference equation for each individual

( )

estimated coefficients for each group g

suggestions of the courses that match the preferences of the individuals.

standard deviation


Table 4 presents a generic model estimated with all the available observations.

Table 4. Model Estimation Results

The estimated results demonstrated the following:


The above-mentioned results have been tested by a panel of experts and were found consistent with the current situation and emergent trends in the maritime education and employment environment, and our a priori hypothesis regarding the behaviour of seafarers.

The equation implemented with regards to the preference rating of each course (y) is therefore the following:

$$\begin{aligned} \text{by = 3.54 + 0.36} & \text{SelfLearning} + 0.42 \text{OnThe} \newline 0 + 0.25 \text{VR} + 0.31 \text{Government} + 0.23 \text{Flexible} \\ + 0.10 \text{NoFees} - 0.84 \text{GCFS} 00 \newline \text{Euroos} + 0.78 \text{Credit} \newline \text{date} - 0.26 \text{Examus} + 0.35 \text{Ahorze} \newline \text{8 on Board} \end{aligned}$$

Similar equations are estimated for different user groups. These groups were defined based on the opinion of the experts' panel used for this purpose. The categorization is based on the following characteristics: (1) Age; (2) Education; (3) Years of working experience; (4) Current Job (Engine, Deck, other); (5) Learning styles; and (6) Soft skills.

## **5.8 Training advisor module**

48 Security Enhanced Applications for Information Systems

**Number Variable Estimated Coefficient t-stat**  0 Constant 3.54 10.9 1 Self-Learning 0.36 1.8 2 On-the-job Training 0.42 2.0 3 Digital 0.25 1.5 4 Governmental 0.31 1.6 5 Flexible 0.23 1.7 6 Without fees 0.10 1.0 7 Greater than 500 Euros -0.84 -2.8 8 Certificate 0.78 5.3 9 Exams -0.26 -1.8 10 On Board and Ashore 0.35 1.6

Table 4 presents a generic model estimated with all the available observations.

Individuals prefer self-learning and on-the-job training over classical lectures

There is a preference over flexible courses adjusted to the user needs rather than courses

Individuals prefer not to pay for receiving the courses and are especially negative

There is negative attitude towards courses that have exams as the assessment procedure

The above-mentioned results have been tested by a panel of experts and were found consistent with the current situation and emergent trends in the maritime education and employment environment, and our a priori hypothesis regarding the behaviour of seafarers. The equation implemented with regards to the preference rating of each course (y) is

*NoFees GT Euros Certificate Exams Ashore onBoard*

*y SelfLearning OnTheJob VR Govermental Flexible*

Similar equations are estimated for different user groups. These groups were defined based on the opinion of the experts' panel used for this purpose. The categorization is based on the following characteristics: (1) Age; (2) Education; (3) Years of working experience; (4) Current

0.10 0.84 500 0.78 0.26 0.35 & 3.54 0.36 0.42 0.25 0.31 0.23

Job (Engine, Deck, other); (5) Learning styles; and (6) Soft skills.

Individuals would prefer courses that are offered both on-shore and off-shore.

Trainees prefer studying at governmental institutions, such as universities

towards following a course with cost more than 500 Euros

Getting a certificate is very important for the trainees

**Coefficient** 

**Summary Statistics** 

Rho-bar squared = 0.2

therefore the following:

Number of observations: 1664

Table 4. Model Estimation Results

The estimated results demonstrated the following:

Digital material are favoured over printed ones

offered for a full academic semester

The basic models developed for each group of trainee, as described in the previous section, need to be customized for each respondent. Bayesian theory is used to provide the suggestions of the courses that match the preferences of the individuals.

For each Group the following information, the following outputs of survey regressions are saved to be used in the Training Advisor:

$$s = \begin{cases} \overline{\beta}\_{\mathcal{g}} = \text{estimated coefficients for each group } \mathbf{g} \\ s\_{\mathcal{g}} = \text{standard deviation} \end{cases}$$

 variance-covariance matrix of the prior *g*

The steps followed are:

**Step 1.** Subject profile

A number of questions are asked at the beginning of the session regarding the characteristics of the trainees. A number of these characteristics (the *X* 's) define the Group in which each individual belongs.

Assume *g* 1,...,*G* number of groups.

**Step 2.** Elicitation

The individual is presented with a sample of courses (using always the same attributes) and is asked of his preferences.

Assume:

*N* number of experiments presented to the individual, and *<sup>n</sup> y* (Nx1 vector) = the ratings of course n. **Step 3.** Creation of Individualized Data

The new data is ( *<sup>n</sup> y* , *Xn* ) where:

*<sup>n</sup> y* (Nx1 vector) = the ratings of the courses.

*Xn* (1xK matrix) = attributes of course n

K= number of attributes

In the new table the respondent has *N* ratings. To each rating the K attributes of the course are appended.

**Step 4.** Develop the personalized preference equation for each individual

Bayesian updating is used to calculate the personalized coefficients of the preference equation as follows:

$$\tilde{\mathcal{B}}^{\ } = \left[ s\_{\mathcal{S}}^2 \Sigma\_{\overline{\mathcal{B}}\_{(\mathfrak{n})}}^{-1} + X^\prime X^\prime \right]^{-1} \left[ s\_{\mathcal{S}}^2 \Sigma\_{\overline{\mathcal{B}}\_{\mathfrak{s}}}^{-1} \overline{\mathcal{B}}\_{\mathcal{S}} + X^\prime y^\prime \right]$$

Development of an e-Learning Recommender System

performance of the recommender system itself .

**7. References** 

2005.

Using Discrete Choice Models and Bayesian Theory: A Pilot Case in the Shipping Industry 51

applicable career paths and respective seafarers' training needs. To formally model these requirements we developed a knowledgebase for accumulating the basic knowledge regarding the shipping work and training environment and registered the information of a 5000 users- sample population, furthermore we used statistical analysis to support the choices of each individual separately. In specific, our e-learning recommender framework is based on advanced choice models and Bayesian techniques and is considered as an intelligent system that can be tested and reused in different e-learning settings, favouring intense personalization and recommendation value-adding features. The foundational techniques used in our system offer the strong competitive advantage of a comparatively detailed, user-focused e-learning attributes modelling framework (advanced choice theory) and a competent system learning capability (Bayesian theory), that improves over time the

Adomavicius, G., Tuzhilin, A. (2005). Toward the next generation of recommender systems:

Balabanovic, M. (1998). Exploring versus Exploiting when Learning: User Models for Text

Ben-Akiva, M., S. Lerman, S. Discrete Choice Analysis: Theory and Application to Travel

Brusilovsky P., Maybury, M. T. (2002). From Adaptive Multimedia to the Adaptive Web,

Brusilovsky, P. Adaptive and Intelligent Technologies for Web-Based Education, in:

Burke, R. (2000). Knowledge-based Recommender Systems. In: A. Kent (ed.): *Encyclopedia of* 

Chaptini, B. Use of Discrete Choice Models with Recommender Systems, PhD thesis, MIT,

ChoiceStream (2004). Review of Personalization Technologies: Collaborative Filtering vs.

García-Crespo, Á., López-Cuadrado, J.L., Colomo-Palacios, R., González-Carrasco I., Ruiz-

Ho, S.Y. (2006).The Attraction of Internet Personalization to Web Users. *Electronic Markets*,

Kim, H.K., Kim, J.K., Ryu, Y.U. (2009). Personalized Recommendation over a Customer

Mezcua, B. (2011). Sem-Fit: a semantic based expert system to provide recommendations in the tourism domain. *Expert Systems with Applications*, Vol. 38,

Network for Ubiquitous Shopping. *IEEE Transactions on Services Computing,* Vol. 2,

http://www.choicestream.com/pdf/ChoiceStream\_TechBrief.pdf

*Knowledge and Data Engineering,* Vol. 17, No 6, pp. 734–749.

Demand. The MIT Press, Cambridge, MA, 1985.

Kunstliche Intelligenz, 2002, pp. 19–25.

No 10 (2011), pp. 13310–13319.

Vol.16, No 1, pp.41–50.

No. 2, pp. 140-151.

*Communications of the ACM ,* Vol. 45, No 5, pp. 31-33.

*Library and Information Systems*. Vol. 69, Supplement 32.

ChoiceStream's Attributized Bayesian Choice Modeling.

Greene, W.H. Econometric Analysis, Second Edition, Macmillan, 1993.

a survey of the state-of-the-art and possible extensions. *IEEE Transactions on* 

Representation. *User Modeling and User-Adapted Interaction*, Vol. 8, No 1, pp. 71-102.

Rollinger, C. & Peylo, C.Eds., *Special issue on intelligent systems and tele-teaching,*

where:

$$s = \overline{\beta}\_{\mathcal{S}'} s\_{\mathcal{S}'} \Sigma\_{\overline{\beta}\_{\mathcal{S}}} \text{ = outputs of the survey -- regressors of group } \mathbf{g}\_{\mathcal{V}} \text{ and }$$

( *y* , *X* ) = the new data

*y* is Nx1

*X* is NxK


A number of potential bundles of courses exist based on expert judgment. Apply ratings to each course of the bundle. Sum-up the ratings.

#### **Step 6.** Recommend

Present to the trainee the bundle of courses with the highest rating (Fig. 5).


Fig. 5. E-learning system recommendation
