Adaptive neuro-fuzzy pedagogical recommender

Expert Systems with Applications 39 (2012) 9797–9806

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

Adaptive neuro-fuzzy pedagogical recommender Zoran Sevarac ⇑, Vladan Devedzic, Jelena Jovanovic Department of Software Engineering, Faculty of Organizational Sciences, University of Belgrade, Jove Ilic´a 154, 11000 Belgrade, Serbia

a r t i c l e

i n f o

Keywords: E-learning Recommender systems Neural networks Fuzzy logic

a b s t r a c t Neuro-Fuzzy Pedagogical Recommender (NFPR) is adaptive recommender based on neuro-fuzzy inference, that can be used to create pedagogical rules in Technology Enhanced Learning (TEL) systems. NFPR is domain independent, provides easy to use API for integration with other systems, and comes with specialized tool (wizard) for creation of NFPR software components. The most important feature of NFPR is its flexibility, that allows teachers to create their own set of pedagogical rules. The proposed model has been implemented and tested with simulated data. Our effort in bringing neuro-fuzzy recommender to the field of TEL (Technology Enhanced Learning) seems to be the first attempt of its kind. Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction In general, recommender systems are used for recommending some items that might be of interest to the users. Recommendations are typically given based on information such as user profiles, item properties (content based recommenders), and users preferences (collaborative filtering) expressed explicitly (e.g., by user ratings and ‘likes’) and/or implicitly (e.g., by the frequency of visits/ downloads) (Jameson, Konstan, & Riedl, 2002). By combining this information with a set of recommendation rules, a recommender system tries to predict which items will be of interest to the user, so that he/she can achieve some predetermined goals. Important questions to be addressed when designing recommender systems include (but are not limited to): 1. What are the most effective techniques for recommendation in specific domain? 2. What information about the users is needed; how to collect and represent it? 3. What information about the items is needed; how to collect and represent it? 4. How to evolve and adapt recommendations in order to make them continuously effective (i.e., to sustain their effectiveness despite the changes e.g., in users preferences or any other requirements)?

In the context of Technology Enhanced Learning (TEL), recommender systems are used for suggesting learning activities, materials and/or topics to students in order to assist them in achieving their desired learning goals – in general, to increase their level of knowledge on some subject (Tang & McCalla, 2003). In this case, the recommendation problem can be defined as the student’s request to the system: ‘‘Given a representation of my current knowledge and preferences, recommend me the next topic/content/activity in order to help me learn the given subject’’ (Basu, Hirsh, Cohen, & Nevill-Manning, 2001). To address this request, the system generates recommendations based on the student model (i.e., its internal representation of the students’ knowledge and preferences), and the teaching model (i.e., the chosen pedagogical strategy usually defined by the teacher). Student model typically contains information about the student’s knowledge, preferences, learning style and the accomplished learning activities. This information is often extracted from history of interaction between the student and the learning environment. Teaching model defines a pedagogical strategy typically as a set of rules that determines the optimal way for learning some topic for certain type of students. The challenging issues in educational recommender systems are the same as those recognized in other recommender systems:

In addition to these challenges, information gathered about the users is often incomplete or unreliable, and that makes the generation of useful recommendations even more difficult.

1. How to collect and represent relevant information about students, and how to structure the student’s model? 2. How to use the student-related data (stored in the student model) to generate useful recommendations, i.e., how to define and evolve pedagogical recommendation rules? 3. From the practical point of view, how to implement the recommendation rules in the most efficient way?

⇑ Corresponding author. Tel.: +381 63 8427 533, +381 11 3950 853; fax: +381 11 2461 221. E-mail address: [email protected] (Z. Sevarac).

The first two questions can be addressed by leveraging the research work and the results achieved by other researchers in the

0957-4174/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2012.02.174

9798

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

field (see Section 2). The last question is especially challenging considering that at the moment there is a lack of open implementations of general pedagogical recommenders that could be reused in different domains and TEL systems. Aiming to address this technical challenge, we have developed an open and adaptive software component, named Neuro Fuzzy Pedagogical Recommender (NFPR), for creating pedagogical recommenders in learning environments. NFPR, which is the central topic of this paper, provides a wizard-style user interface and an easy-to-use API, which makes it suitable for easy integration with various learning environments. The paper is organized as follows: related work which includes recommender and neuro-fuzzy systems for TEL is given in Section 2; Section 3 outlines current challenges in the field of pedagogical recommender systems and clearly states the problem this work aims to address; the overall architecture of the proposed software system and the algorithms it is based upon are given in Section 4; these are followed by a basic working example (Section 5) and a sample application (Section 6); usability and pedagogical evaluation is given in Section 7, whereas Section 8 outlines conclusions of this research.

2. Related work Related work for this research includes recommender systems in general, recommender systems in TEL and neuro-fuzzy systems in TEL. Accordingly, this section gives a brief overview of these three research areas. Currently, a widely used state of the art approach in recommender systems is Matrix Factorization (Bell, Koren, & Volinsky, 2009) which belongs to collaborative filtering family of recommender systems (Bobadilla, Hernando, Ortega, & Bernal, 2011). The most general data representation technique applied in this type of recommender systems is a matrix of n users and m items, where each matrix cell corresponds to the rating given to item i by the user u (Melville & Sindhwani, 2010, chap. 00338). The Matrix Factorization algorithm is used to predict which item will have the highest rating for some user, based on the ratings of other items by that user and ratings of other users. The main issue with this approach is the so called ‘cold start’ problem, which means that in the beginning there is not enough data (ratings) to make good recommendations, and it is not possible to give recommendations for new users (before they provide some ratings) and new items (before they get some ratings). In practice these issues are resolved using simple average ratings, by creating hybrid recommenders in combination with content filtering techniques (Hummel et al., 2007), or using some more sophisticated methods (Gantner, Drumond, Freudenthaler, Rendle, & Schmidt-Thieme, 2010; Preisach, Marinho, & Schmidt-Thieme, 2010). A wide range of other techniques, including statistics and machine learning based techniques, are also used in order to analyze data and give recommendations (Melville & Sindhwani, 2010, chap. 00338). In the area of recommender systems for Technology Enhanced Learning (Manouselis, Drachsler, Vuorikari, Hummel, & Koper, 2010), research is focused on the construction of recommender systems for recommendation of learning resources (materials or peers to provide help) or learning activities to the learners (Ghauth & Abdullah, 2010; Manouselis et al., 2010). Recommender systems for educational purposes are challenging research direction (Drachsler, Hummel, & Koper, 2009) since preferred learning activities of students might pedagogically not be the most adequate (Tang & McCalla, 2004) and recommendations in e-learning should be guided by educational objectives, and not only by the user’s preferences (Santos & Boticario, 2010). Also, there are a number of specific features that have to be taken into account, such as (Drachsler et al., 2009):

– the importance of context (which is not taken into account in common recommender systems); – the inherent novelty of most learning activities; – the need for a learning strategy; – the need to take changes and learning processes into account. There are many different approaches for recommenders in TEL, from collaborative and content filtering to hybrid approaches and each of them has some advantages and disadvantages depending on the context in which they have been used and how they are evaluated (Manouselis et al., 2010). For example, the above mentioned Matrix Factorization technique which has already proved to be very successful in e-commerce and movie recommendation domains (Melville & Sindhwani, 2010, chap. 00338), is promising for educational domain, as well (Thai-Nghe et al., 2011), but it lacks one important feature – the ability to adapt to teacher’s pedagogical strategy. Neuro-fuzzy techniques have been used in TEL so far, mainly for student modeling. For instance, Kinshuk, Nikov, and Patel (2001) have used fuzzy back-propagation to develop a neuro-fuzzy system aimed at maximizing adaptability in business education tutoring. The system learned from the data that came from the interaction with students. Al-Hammadi and Milne (2004) designed a student classification method using neuro-fuzzy technique called NEFCLASS, in order to obtain learners’ performance prediction. Stathakopoulou, Magoulas, Grigoriadou, and Samarakou (2005) attempted to create and update a student model, using a neural network-based fuzzy model, where the fuzzy component was aimed at simulating human decision-making process and the neural networks were incorporated to provide the system with learning and generalization abilities. Mir Sadique and Ghatol (2004) used the architecture of the adaptive neuro-fuzzy inference system (ANFIS) in the field of intelligent tutoring systems. Learners were tested in concept understanding, memorizing skills and possible misconceptions, and results of these tests are given as input to the inference system. Finally, learners’ performance is categorized as poor, fair, good or excellent. Neuro-fuzzy techniques have also being used in recommender systems as means for discovering and learning meaningful recommendation rules. For example, Castellano, Fanelli, and Torsello (2008) proposed a neuro-fuzzy approach for the extraction of a recommendation model from the usage data encoding user navigational behavior. The model is expressed as a set of fuzzy rules which may be exploited to provide personalized link suggestions to the users visiting a web site. In particular, a neuro-fuzzy network is trained using information about user categories to discover a set of fuzzy rules capturing the associations between user behavior models and relevance degrees of pages to be recommended. To the best of our knowledge, at the moment there are no implementations of neuro-fuzzy recommenders in TEL, and this research tries to fill that gap. Although various data mining and statistical based techniques used by the aforementioned systems do provide efficient ways for extracting recommendations from data, they do not provide a way for teachers to implement their own pedagogical strategies. Furthermore, recommendation given by these systems often cannot be explained. By utilizing neuro-fuzzy approach, this research aims at creating a more user-centered adaptive recommender system for TEL. 3. Problem statement The work presented in this paper addresses three challenges related to pedagogical recommender systems:  A pedagogical recommender should support any number of criteria for recommendation.

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

 A pedagogical recommender system should be adaptable, in order to support different pedagogical approaches and different domains. This may be the key for a wider adoption of such recommenders. However, to the best of our knowledge, most (if not all) implementations of pedagogical recommenders at the moment are very domain- and problem-specific, and they cannot be reused in different environments. This also means that the development of every pedagogical recommender starts almost from scratch.  Last but not the least, a pedagogical recommender system should be intuitive and easy-to-use for end users (pedagogical experts). Once a pedagogical strategy is implemented with the recommender, it should be easy for an end user to extend, modify it to suit the changes in his/her teaching practice. To this end, the paper proposes Neuro Fuzzy Pedagogical Recommender (NFPR), a general pedagogical recommender system based on neuro-fuzzy techniques. It can be used to create pedagogical recommendation rules for different domains. In addition, NFPR allows for adapting the recommendation rules to include specific learning and teaching preferences, such as prerequisite knowledge, learning style, and other elements that can be extracted from the student and domain model. Accordingly, for students, NFPR can provide personalized learning experience by providing learning content recommendations based on their knowledge, learning style and other properties from student model. For teachers, it aims to provide great flexibility in defining pedagogical models that support their own teaching style and experience.

4. Proposed solution An important feature of NFPR is its flexibility: it can be used on custom input and output data sets (which correspond to, e.g., the student model and the recommended learning content, respectively), and allows for the creation of personalized recommendation rules. Fuzzy set theory is used to transform high-level pedagogical rules into a computational model, whereas a neural network is used to provide adaptivity to the teacher’s preferences. Thanks to the wizard-style user interface, using the system does not require in-depth knowledge of fuzzy sets and neural networks. NFPR is available as an open source implementation1 that can be easily integrated with almost any TEL system. Fig. 1 illustrates how NFPR can be used with the student model (comprising the student’s knowledge and learning style) as the input, and the recommended learning content as the output. In what follows we present the main building blocks of NFPR and their role in this recommender system.

4.1. Domain model NFPR’s domain model contains structured knowledge of the domain in the form of a topic map (Amruth, 2006). Topics are related to one another with the prerequisite relation which means that one topic is a prerequisite for learning another topic. This type of domain model is chosen since it is intuitive to the end users (i.e., teachers). Other, more complex techniques for domain modeling (such as ontologies), do offer advantages, but have one significant disadvantage: they are difficult to accept by the end users and thus pose problems to wider adoption in educational practices (Hatala, Gasevic, Siadaty, Jovanovic, & Torniai, 2009). This model is the base for creating pedagogical recommendation rules. 1

NFPR is available for download at http://neuroph.sourceforge.net/NFPR.zip.

9799

Fig. 1. The conceptual model of neuro-fuzzy pedagogical recommender.

4.2. Student model Student model stores information about student’s current state of knowledge and personal characteristics (Stathacopoulou, Magoulas, & Grigoriadou, 1999). The student model used in NFPR is the overlay student model which represents student’s knowledge as a subset of expert/system’s knowledge of domain (Kass, 1989). The most important information it contains is a list of topics to learn and corresponding test results for those topics which represent the current state of student’s knowledge. 4.3. Inputs The inputs for NFPR are extracted from the student model. They can be, for example, the student’s knowledge of some topics and the preferred learning style. The student’s knowledge can be evaluated with tests, while the learning style can be elicited through an appropriate questionnaire (Kinshuk et al., 2001). 4.4. Outputs The output of NFPR is the recommended learning content that corresponds to some domain concept (i.e., a concept from the domain model). Possible outputs are identified by relating the available learning content to the appropriate concepts from the domain model. 4.5. Recommendation rules Recommendation rules define the mapping of the inputs to the outputs of a recommender system and are based on the following set of high-level pedagogical assumptions: IF (Student has good knowledge of Topic1) THEN Student should learn Topic2 Topic1 and Topic2 are topics (concepts) defined in the Domain model and they are related through the prerequisite relationship. Student’s knowledge of these topics is stored in the Student model. If Topic2 is related to some learning content, e.g., LearningContent2, then the above rule gets the following form: IF (Student has good knowledge of Topic1) THEN Student should study LearningContent2 The conditional part of the rule may be more complex and include several conditions, like: IF (Student has good knowledge of Topic1) AND (Student has excellent knowledge of Topic2) THEN Student should learn Topic3

9800

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

Or, if Topic3 is related to some learning content (e.g., LearningContent3): IF (Student has good knowledge of Topic1) AND (Student has excellent knowledge of Topic2) THEN Student should study LearningContent3 If the student’s learning style is also considered, then the rule takes the following form: Fig. 2. Fuzzy membership functions for student’s knowledge.

IF (Student has good knowledge of Topic1) AND (Student has excellent knowledge of Topic2) AND (Student learning style is SomeLearningStyle1) THEN Student should learn LearningContent4 These are high-level rules, typically used and understood by teachers. In NFPR such rules are converted to a computational model using the fuzzy set theory. Students’ knowledge and learning style are considered to be linguistic variables, which can take values of the corresponding fuzzy sets. It is also assumed that each learning topic has the corresponding learning content. In the current implementation, three fuzzy sets are used to express the student’s knowledge of some domain topic: POOR – insufficient knowledge of the topic GOOD – basic understanding of the topic EXCELLENT – advanced understanding of the topic The implementation allows for an easy extension to a more fine-grained representation (e.g., five or more fuzzy sets). In the following example two fuzzy sets are used to express the learning style: EXTENSIVE_READER – student who needs to read all available material in detail and understand prerequisites before proceeding to the next topic; PRACTITIONER – student who learns by doing practical exercises. These learning styles are intuitively defined based on our own teaching experience. However, a teacher can define students learning styles based on his/her own experience or can opt for an existing, well-founded learning style theory. For example, in our previous work (Jovanovic´, Gaševic´, & Devedzˇic´, 2009), we have successfully applied Felder–Silverman model of learning styles (Felder & Silverman, 1988) which categorizes a student’s learning style on four dimensions: sensing-intuitive, visual-verbal, active-reflective and sequential-global. Thanks to the great flexibility of the proposed solution, it can support both formal and intuitively defined learning styles, which is very important for wide adoption of the proposed solution in different domains. Considering all these fuzzy sets, the above recommendation rules can be expressed in the following way: IF (Knowledge of Topic1 is GOOD) AND (Knowledge of Topic2 is EXCELLENT) THEN Topic3 IF (Knowledge of Topic1 is GOOD) AND (Knowledge of Topic2 is EXCELLENT) THEN LearningContent3 IF (Knowledge of Topic1 is GOOD) AND (Knowledge of Topic2 is EXCELLENT) AND (LearningStyle is EXTENSIVE_READER) THEN LearningContent5

Even though the first two rules look almost the same, they have different purpose. The first rule recommends a topic to study, while the second rule recommends learning content related to the recommended topic. The expressions in the IF part of the given recommendation rules (like Knowledge of Topic1 is GOOD) are evaluated to the value of the fuzzy membership function, which is then used for fuzzy reasoning and rule calculation. Specifically, the level of the student’s knowledge of a certain topic (originating, e.g., from test results) is turned into the corresponding fuzzy value using the fuzzy membership functions shown in Fig. 2. Fuzzy reasoning used for calculating the value of the expressions in the IF part of the rule is a simple MIN function, which is commonly used as the fuzzy logical AND operator (Boukezzoula, Galichet, & Foulloy, 2007). 4.6. Neural network A neural network is used to customize the mappings between the premise (IF part) of the rule and the corresponding recommendation. Each input (neuron in the input layer) corresponds to some domain topic, and will be fed with a test result for that topic. Each output (neuron in the output layer) also corresponds to some domain topic, and after the network input is set and the output is calculated, the recommended neuron/topic in the output layer will be activated according to the recommendation rule. Fig. 3 shows the architecture of the neural network. It has four layers of neurons, and each layer plays a specific role. The first two layers represent the fuzzy interface, which transforms real inputs into the corresponding fuzzy values, while the third and the fourth layers form a perceptron network (Rosenblatt, 1962) capable of learning a set of IF–THEN rules (Sevarac, 2006). More precisely, the roles of the network layers are as follows: Input layer – just distributes inputs to the next, fuzzyfication layer. Each input neuron is connected only to the corresponding neurons in the fuzzyfication layer. Fuzzyfication layer – each neuron in this layer corresponds to one fuzzy set, and calculates the value of fuzzy membership function for the given input. Premise layer – each neuron in the premise layer corresponds to one combination of fuzzy sets from the fuzzyfication layer. It represents the Cartesian product (all possible combinations) of fuzzy sets that belong to different inputs. Output layer – each neuron in the output layer corresponds to one recommendation, either a topic or a learning content. The neural network has two modes of operation: adaptation mode and recommendation mode. In the adaptation mode, the network creates relations between the premise and the output layers based on the data presented to the network (the so-called training set). These data take the form of input–output pairs that

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

9801

process inputs and desired outputs are presented to the network, and adaptation algorithm automatically adjusts the weights so the output responses to the inputs will be as close as possible to their desired responses. After adaptation, when in recommendation mode, network will provide desired input–output mapping as defined with training set. In the recommendation mode, some input is presented to the network, and after the calculation is done, the active neuron in the output layer is considered to be the recommendation. Continuing the previous example, if the network input is [85, 57, 0, 0] the trained network will output [0, 0, 1, 0]. The third neuron is in activated state (outputs 1) so the recommended topic is Arrays. 4.7. Advantages of the proposed solution

Fig. 3. The architecture of the NFPR neural network.

consist of the input and desired network output for that input. For example, if there are four learning topics (Variables_and_Data_ Types, For_Loops, Arrays, Classes_and_Objects), there will be four inputs and four outputs where each one of them corresponds to one learning topic. These are considered to be the input and desired output vectors. Each element of the input vector takes as its value the test result for the corresponding learning topic, while the elements of the output vector have the value 1 if the corresponding topic is recommended according to some rule, or 0 otherwise. Each rule is presented with one input/output pair of vectors. For example, consider the following simple pedagogical rule for a basic programming domain: IF (Knowledge of Variables_and_Data_Types is EXCELLENT) AND (Knowledge of For_Loops is GOOD) THEN Arrays This rule says that the learning topic Arrays will be recommended to the student if he/she has a high level of understanding of the topic Variables and Data Types and a basic understanding of the topic For Loops. For this simple rule the excerpt from the corresponding training set is: Input: [80, 50, 0, 0] Output [0, 0, 1, 0] where the elements and their positions in the vectors correspond to the following learning topics, respectively: [Variables_and_Data_Types, For_Loops, Arrays, Classes_and_Objects]. Input values come from characteristic points of fuzzy membership functions. Adaptation is accomplished by using simple single-layer supervised learning rule, known as Delta Rule (Widrow & Lehr, 1995). The Delta Rule, also known as Widrow–Hoff Rule or Least Mean Squares, is an iterative learning rule for updating the weights of artificial neurons in a single-layer perceptron. During the learning

The proposed model of pedagogical recommender system is very flexible as it allows for various customizations in order to support individual pedagogical strategy. It supports intuitive, non-formal pedagogical models that can be created by teachers based on their teaching experience, and can also be adapted to the teacher’s preferences. Verbal pedagogical model can be easily translated to the corresponding fuzzy model by using fuzzy sets and rules. Furthermore, a neural network can be automatically generated and trained thanks to its straightforward architecture and learning rule. This means that different pedagogical recommenders can be automatically generated without a need to change low level implementation details. Accordingly, it is possible to create very sophisticated tools with intuitive and easy-to-use user interface, that can produce ready to use neuro-fuzzy pedagogical recommenders. Possible customizations include: 1. any number of inputs (theoretically), which means that it can support any number of criteria for recommendation (e.g., customized number of learning topics, learning styles, and even other criteria constituting learning context); 2. each recommendation criterion can have a customized set of corresponding fuzzy sets, so the translation from the verbal pedagogical model to fuzzy computational model can also be customized (for example, instead of POOR, GOOD and EXCELLENT, some may want to have five levels of grading with different naming); 3. any number of outputs, which means customized number of learning topics for recommendation; 4. adaptation of recommendation rules to some specific preferences; this is possible due to the fact that high level pedagogical rules transformed to fuzzy domain are automatically learned by the neural network. An additional benefit lies in the fact that regardless of all the above mentioned customizations, the internal operations of the proposed pedagogical recommender remain the same. This further means that the same implementation can be applied to a wide range of learning domains. Possible constraints could be faced when working with large number of inputs and fuzzy sets, which can cause rule layer to grow fast, so it may require more memory than usual (than standard configurations provide). However, having in mind the amounts of memory that modern systems provide, this can be easily resolved through appropriate system configuration (e. g., by assigning more heap memory to Java Virtual Machine). 5. Basic working example Basic principles of operation of NFPR are described on a simple but realistic example of a pedagogical recommender for an intro-

9802

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

ductory programming course. Fig. 4 shows a tiny portion of the domain model which consists of domain concepts/topics, as well as the relations between them. Relations describe the minimum level of knowledge of some topic that is required to learn some other topic. These relations are defined by the teacher, based on his/her own teaching experience. This example considers domain topics only; for simplicity reasons learning styles are not considered, but could be easily added (as explained in the previous section). Note that different learning styles may require different levels of knowledge of some topics. A set of fuzzy pedagogical rules extracted from this model is:

ommended to learn the topic Classes and Objects, the student must have good knowledge of Variables and Data Types. This kind of tabular representation is suitable for visual overview of all possible combinations of domain topics. Irrelevant combinations are marked with x, while the empty fields are optional and are left open, so they could be customized later by the user. If the number of relevant concepts in the domain model grows beyond a certain limit, the matrix should be divided to display related concepts in several groups. The NFPR’s neural network for the set of recommendation rules defined in Table 1 has:

IF (Variables_and_Data_Types is GOOD) THEN For_Loops IF (Variables_and_Data_Types is GOOD) THEN Classes_and_ Objects IF (Variables_and_Data_Types is EXCELLENT) AND (For_Loops is GOOD) THEN Arrays IF (Arrays is GOOD) AND (Classes_and_Objects is EXCELLENT)AND (Inheritance is GOOD) THEN Collections

– 7 input neurons which correspond to seven learning topics (given in Table 1). – 21 neurons in the fuzzification layer, where each neuron corresponds to a knowledge-level fuzzy set (see Fig. 2) related to each learning topic (since there is three fuzzy sets, 7  3 = 21). – 441 neurons in the premise layer (7  3  7  3 = 441) – for seven topics and three fuzzy sets for each topic. – 7 output neurons, where each neuron corresponds to a learning topic that could be recommended to the student.

This set of rules can be represented with the recommendation matrix as shown in Table 1. This matrix provides a tabular representation of the prerequisite relationship among topics in the domain model and how it determines the recommended topics. It relies on the fact that a pedagogical strategy can be modeled as a relation on the Cartesian product of topic sets (Xu, Wang, & Su, 2002). Each column in the matrix is related to one possible recommendation (i.e., one recommended topic) and contains a fuzzy representation of the knowledge level for all the prerequisite topics. For example, the second column indicates that in order to be rec-

Training data is automatically generated from the recommendation rule matrix (Table 1) and corresponding fuzzy sets (Fig. 2). It is assumed that the level of knowledge lower than the one given in the matrix does not fulfill the prerequisites, while a higher level satisfies the prerequisites. For example, to learn For Loops, a student must have at least good knowledge of Variables and Data Types, so if he/she has excellent level of knowledge that will also satisfy the prerequisite.

Fig. 4. A part of the domain model for an introductory programming course.

Table 1 Recommendation rule matrix. Recommendation

Prerequisite knowledge Variables and data types Classes and objects For loops Arrays Inheritance Interfaces Collections

Variables and data types

Classes and objects

For loops

Arrays

x x x x x x x

GOOD x x x x x x

GOOD

EXCELLENT

x x x x x

GOOD x x x x

Inheritance

Interfaces

EXCELLENT

x x x

Collections

EXCELLENT

GOOD x x

GOOD GOOD x

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

9803

Fig. 5. NFPR wizard Step 1: Creation of recommendation rules.

The adaptation of the rules is accomplished by training the neural network with the generated training data set. Training adjusts the connections between the premise and output/recommendation layers, so the network gets the desired behavior. Since there is only one layer to adapt (the output layer), the simple DeltaRule algorithm is used. If the learning style is also taken into account in a pedagogical strategy, then one recommendation matrix is required for each learning style. From the perspective of the used neural network, it will be just another input for the network, and will need corresponding neurons in the fuzzification and premise layers.

6. Application The figures bellow show NFPR’s wizard-style user interface for creating recommendation rules (Fig. 5) and neural network that implements those rules (Fig. 6). Neural network is created with Neuroph, an open source Java framework for neural network development.2 Neuroph provides simple Java API for using neural networks within Java applications, and a tool, called easyNeurons, which offers rich and intuitive GUI (Graphical User Interface) for creating and training neural networks. NFPR is created as an application sample within easyNeurons tool. Figures bellow show how to create and test NFPR using two step wizard. Step 1. Define recommendation rules. In this step user (teacher) loads all domain topics, prerequisites and possible recommendations, from QTI files, and system generates recommendation matrix (as the one given in Table 1). QTI (Question and Test Interoperability specification) defines a standard format for the representation of assessment content and results.3 Generated recommendation matrix contains all possible combinations of prerequisite relationships between domain topics, and the teacher selects recommendation for each combination (Fig. 5), thus creating a recommendation rule. Each row in the recommendation matrix represents one recommendation rule. Each field contains the name of a domain topic appearing in the corresponding rule, whereas its color indicates the knowledge level of the topic (expressed as a fuzzy set): green 2 Neuroph is available for download at http://neuroph.sourceforge.net/ download.html. 3 QTI specification is available online at http://www.imsglobal.org/question/.

for EXCELLENT, yellow for GOOD and red for POOR. Once rules are defined, user clicks the Next button, and neural network and training set are automatically created. Step 2. Train and test NFPR (Fig. 6). In this step, the user (teacher) just has to click the Train button to train the neural network with the training set (created at the end of the previous step), and that is how the neural network learns the rules. When network is trained, the user can load some student’s test results from a QTI file with test results and see the recommendations. The trained neural network can be serialized as a Java object and used as a Java component in any TEL application. It provides a simple API with only two methods for setting the input and getting the output (i.e., recommendation) from the network. The following sample code (Listing 1) illustrates how easy it is to use the created neural network with an external, e.g., TEL application:

Listing 1. The use of NFPR’s API for calculating recommendations based on the given set of test results.

// get test results read from a QTI file double[] someTestResults = . . . // load neural network from file NeuralNetwork nfpr = NeuralNetwork.load(‘‘MyNfpr.nnet’’); // set neural net input nfpr.setInput (someTestResults); // calculate network nfpr.calculate(); // get recommendation double[] recommendation = nfpr.getOutput(); // do something with the recommendation . . .

If end users (teachers) wish to change the pedagogical strategy, the network needs to be retrained. Existing rules can be modified, new domain topics and learning styles can be added, and even new pedagogical criteria can be introduced. To accomplish this teacher just needs to re-run the NFPR wizard and make the desired changes in Step 1 (Fig. 5).

9804

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

Fig. 6. NFPR wizard step 2: Training and testing the neural network.

7. Evaluation Despite the increasing number of systems proposed for recommending learning resources, a closer look at the current status of their development and evaluation reveals a lack of systematic evaluation studies in the context of real-life applications. As indicated in (Manouselis et al., 2010), more than a half of the analyzed systems, namely 12 out of 20 the authors considered, were still in the design or prototyping stage of development, while only 10 systems were reported as being evaluated through trials that involved human users. Another observation is that, very often, experimental investigation of the recommendation algorithms does not take place, although it is a common evaluation practice in recommender systems examined for other domains (e.g., Breese, Heckerman, & Kadie, 1998). Deshpande and Karypis (2004), Papagelis, Plexousakis, and Kutsuras (2005), and Herlocker, Konstan, Terveen, and Riedl (2004), indicate that careful testing and parameterisation has to be carried out before a recommender system is finally deployed in a real-world setting. One of the main reasons is that the performance of recommendation algorithms seems to depend on the particularities of the application context. Hence it is advised to experimentally analyze recommender system before its actual deployment. Following this advice, NFPR was evaluated at the University of Belgrade with a group of 24 teachers. The group has been introduced to the idea of pedagogical rules based on test results as well as the steps needed to create pedagogical rules and save them as expert knowledge. Then the group was introduced to the NFPR tool and its features, and asked to create a set of pedagogical rules that reflects their pedagogical strategy. Participants were then asked to assess the recommendations given by NFPR, by using the previously prepared set of test results. These test results which correspond to typical student’s knowledge levels, were created by the teachers based on their teaching experience. Having finished the given tasks, the study participants were asked to fill in a survey which included the following questions related to their interaction with NFPR:

Q1: I learned to use the software quickly Q2: I was able to use it without detailed instructions Q3: Working with this software is satisfying Q4: Sometimes I didn’t know what to do next with this software. Q5: I need to learn a lot of things before I can get going with this system Q6: Student’s knowledge estimation was satisfactory Q7: Recommendations given by the system are satisfactory Q8: I was able to create a pedagogical strategy very close to those I use in my teaching practice Q9: Students would benefit from the given recommendations Q10: I would recommend this software to my colleagues Q11: Your background Q12: Familiarity with information systems and programming Q13: Additional comments about NFPR wizard The questionnaire included three sections. The first group of questions (1–5) was related to the users’ experience with the NFPR wizard. These questions were focused on usability of the NFPR wizard, and they were based on software usability surveys given in (Arnold, Measuring usability with the USE questionnaire, & October, 2001). The second group of questions (6–10) was related to how well NFPR was estimating the students’ knowledge and the quality of its recommendations. The third group of questions were open-ended questions to gather additional info about participants and their personal experience in using NFPR. The answers to the first two groups of questions in the survey were collected using a five-point Likert scale (Strongly disagree, Disagree, Neutral, Agree, Strongly agree). To make the survey easily accessible to the study participants, we used surveymonkey.com tool to make web-based survey. Survey results are summarized in Table 2. When analyzing the collected data, we also took into account the teachers’ background and familiarity with information systems and programming, which were collected from open-ended questions.

9805

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806 Table 2 Survey results. Questions

1 2 3 4 5 6 7 8 9 10

Answers Strongly disagree

Disagree

Neutral

Agree

Strongly agree

Mean

Standard deviation

3 4 2 13 7 2 1 1 1 1

2 7 3 6 10 1 3 2 2 2

0 0 1 0 2 2 1 2 2 2

12 8 12 3 1 11 10 13 11 12

7 5 6 2 4 8 9 6 8 7

3.75 3.13 3.71 1.96 2.38 3.92 3.96 3.88 3.96 3.92

1.33 1.48 1.23 1.37 1.41 1.18 1.16 1.03 1.08 1.06

Some of the teachers obviously found the user interface to be very intuitive, since they were able to understand how to create pedagogical rules without detailed instructions (Q2; 3.13 out of 5, SD 1.48). Most of them find working with software satisfying (Q3; 3.71 out of 5, SD 1.23), and just few of them got confused while working with it (Q4; 1.96 out of 5, SD 1.41). Important results are that participants found students’ knowledge estimation satisfactory (Q6; 3.92 out of 5, SD 1.18) and same is with the given recommendations (Q7; 3.96 out of 5, SD 1.16; and Q9, 3.96 out of 5, SD 1.08). Also, the facts that participants where able to successfully create their own pedagogical strategy (Q8; 3.88 out of 5, SD 1.03) and that they would recommend this tool to their colleagues (Q10; 3.92 out of 5, SD 1.06) show that this kind of tool would be accepted by the intended audience (teachers). However, some of the participants pointed in their comments that they had problems with understanding colors representing fuzzy rules. Overall the possibility of creating an automatic list or a map of educational content for every student, as if he/she had a personal, ‘‘one on one’’ interaction with the teacher, was most welcomed from users of NFPR in their comments. The review of the data gathered from the survey has also showed that:  the participants with little experience with information systems and programming exhibited a better understanding of the concept of pedagogical rules after a short presentation on how to use NFPR;  the wizard used for generating pedagogical rules was easy to use for 55% (Q2) before and 80% (Q1) after help presentation given during introduction;  a general opinion about NFPR’s capability to imitate expert decisions has been shown through the survey feedback – 79% of the teachers had a positive opinion about NFPR’s recommendation of educational content based on test results (Q9); in these cases, NFPR was giving recommendations as teachers would, based on personally created pedagogical rules; the idea of saving these pedagogical rules on a computer for later use in similar or identical tests was suggested in one way or another by many participants in their comments (Q13);  overall these results show that proposed solution could successfully address adaptivity and usability issues outlined in Section 3. 8. Conclusion The Neuro Fuzzy Pedagogical Recommender (NFPR) is an open, adaptive software component for creating recommenders in learning environments. It makes use of the fuzzy set theory which allows for easy translation of high level verbal pedagogical rules into fuzzy computational model. In addition, it offers an automated procedure for the creation and configuration of neural network based recommender.

For students, NFPR can provide personalized learning experience by generating learning content recommendations based on their knowledge, learning style and other properties that might be relevant for some specific subject domain. For teachers, it provides great flexibility in defining pedagogical models that support their own teaching style and experience. Preliminary evaluation showed that the system is easy to use for teachers, and the teachers are satisfied with the recommendations it can provide. Once created, the pedagogical model can be adjusted or extended without need to change implementation details. Likewise, low level neuro fuzzy principles are domain independent, which means that same component could be used in different domains. All that is required is to run the customization wizard which will configure the neurofuzzy model according to the chosen domain and teaching preferences. In a nutshell, the main advantages of NFPR are flexibility, domain independence and easy implementation/application in different environments. Future research will be focused on development of ready to use pedagogical recommenders for different learning domains and integration in existing learning environments. References Al-Hammadi, A. S., Milne, R. H., 2004. A neuro-fuzzy classification approach to the assessment of student performance. In Proceedings of the IEEE international conference on fuzzy systems (pp. 837–841). Amruth, N. K. (2006). Using enhanced concept map for student modeling in a model-based programming tutor. In Proceedings of 19th international FLAIRS conference on artificial intelligence (FLAIRS 2006) special track on intelligent tutoring systems (pp. 527–532). Melbourne Beach, Florida, USA. Arnold, M. Lund, Measuring usability with the USE questionnaire, STC usability SIG newsletter, October 2001 issue (Vol. 8, No. 2). . Basu, C., Hirsh, H., Cohen, W., & Nevill-Manning, C. (2001). Technical paper recommendations: A study in combining multiple information sources. Journal of Artificial Intelligence Research, 1, 231–252. Bell, R., Koren, Y., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. IEEE Computer, 42(8), 30–37. Bobadilla, J., Hernando, A., Ortega, F., & Bernal, J. (2011). A framework for collaborative filtering recommender systems. Expert Systems with Applications, 38(12), 14609–14623. Boukezzoula, R., Galichet, S., & Foulloy, L. (2007). MIN and MAX operators for fuzzy intervals and their potential use in aggregation operators. IEEE Transactions on Fuzzy Systems, 15(6), 1135–1144. ISSN: 1063-6706. Breese, J., S., Heckerman, D., & Kadie, C., 1998. Empirical analysis of predictive algorithms for collaborative filtering. In: Proceedings of the 14th Conference on Uncertainty in Artificial Intelligence (Vol. 461, No. 8, pp. 43-52). San Francisco, CA. ISBN: 155860555X. Castellano, G., Fanelli, A. M., Torsello, M. A., 2008. NEWER: A neuro fuzzy web recommendation system. In Applications of first international conference on the digital information and web technologies, 2008. ICADIWT 2008 (pp. 162–167). doi:10.1109/ICADIWT.2008.4664338. ISBN 978-1-4244-2623-2. Deshpande, M., & Karypis, G. (2004). Selective Markov models for predicting Web page accesses. Transactions on Internet Technology, 4(2), 163–184 [Association for Computing Machinery]. Drachsler, H., Hummel, H. G. K., & Koper, R. (2009). Identifying the goal, user model and conditions of recommender systems for formal and informal learning. Journal of Digital Information, 10(2) [Texas Digital Library. ISSN: 1368-7506]. Felder, R., & Silverman, L. (1988). Learning and teaching styles in engineering education. Journal of Engineering Education, 78(7), 674–681. Gantner, Z., Drumond, L., Freudenthaler, C., Rendle, S., Schmidt-Thieme, L. (2010). Learning attribute-to-feature mappings for cold-start recommendations. In

9806

Z. Sevarac et al. / Expert Systems with Applications 39 (2012) 9797–9806

Proceedings of the 10th IEEE international conference on data mining (ICDM 2010). IEEE Computer Society. Ghauth, K., & Abdullah, N. (2010). Learning materials recommendation using good learners’ ratings and content-based filtering. Educational Technology Research and Development, 1042–1629. Hatala, M., Gasevic, D., Siadaty, M., Jovanovic, J., Torniai, C. (2009). Can educators develop ontologies using ontology extraction tools: An end user study. In Proceedings of the 4th European conference on technology enhanced learning (ECTEL09), LNCS 5794 (pp. 140–153). Springer-Verlag. Herlocker, J. L., Konstan, J. A., Terveen, L. G., & Riedl, J. T. (2004). Evaluating collaborative filtering recommender systems. ACM Transactions on Information Systems, 22(1), 5–53 [Association for Computing Machinery]. Hummel, H. G. K., van Den Berg, B., Berlanga, A. J., Drachsler, H., Janssen, J., Nadolski, R., et al. (2007). Combining social-based and information-based approaches for personalized recommendations on sequencing learning activities. International Journal of Learning Technology [Interscience]. Jameson, A., Konstan, J., Riedl, J. (2002). AI techniques for personalized recommendation. Tutorial notes, AAAI-02. Edmonton, Canada. Jovanovic´, J., Gaševic´, D., & Devedzˇic´, V. (2009). TANGRAM for personalized learning using semantic web technologies. Journal of Emerging Technologies in Web Intelligence, 1(1), 6–21. ISSN: 1798-0461. Kass, R. (1989). Student modelling in intelligent tutoring systems-implications for user modelling. In A. Kosba & W. Wahlster (Eds.), User models in dialog systems (pp. 386–410). Berlin: Springer-Verlag. Kinshuk, A., Nikov, A., Patel (2001). Adaptive tutoring in business education using fuzzy backpropagation approach. In Proceedings of the ninth international conference on human–computer interaction 2001 (Vol. 1, pp. 465–468). . Manouselis, N., Drachsler, H., Vuorikari, R., Hummel, H., & Koper, R. (2010). Recommender systems in technology enhanced learning. In P. B. Kantor, F. Ricci, L. Rokach, & B. Shapira (Eds.), 1st Recommender systems handbook (pp. 1–29). Berlin: Springer. Melville, P., Sindhwani, V. (2010). Encyclopedia of machine learning (1st ed.). Springer. ISBN 978-0-387-30164-8. Mir Sadique, A., Ghatol, A. A. (2004). A neuro fuzzy inference system for student modeling in web-based intelligent tutoring systems. . Papagelis, M., Plexousakis, D., Kutsuras, T. (2005). Alleviating the sparsity problem of collaborative filtering using trust inferences. In Proceedings of the 3rd

international conference on trust management. LNCS 3477 (pp. 224–239). Berlin, Heidelberg, New York: Springer. Preisach, C., Marinho, L. B., Schmidt-Thieme, L. (2010). Semi-supervised tag recommendation-using untagged resources to mitigate cold-start problems. In 14th Pacific-Asia conference on knowledge discovery and data mining (PAKDD 2010), Lecture Notes in Computer Science (Vol. 6118, pp. 348–357). Springer. Rosenblatt, F. (1962). Principles of neurodynamics: Perceptrons and the theory of brain mechanisms. Washington DC: Spartan Books [ASIN: B0006AXUII]. Santos, O. C., & Boticario, J. G. (2010). Modeling recommendations for the educational domain. Procedia Computer Science, 1(2), 2793–2800 [Proceedings of the 1st Workshop on Recommender Systems for Technology Enhanced LearninRecSysTEL 2010]. Sevarac, Z. (2006). Neuro fuzzy reasoner for student modelling. In Proceedings of the sixth IEEE international conference on advanced learning technologies 2006 (pp. 740–744). . Stathacopoulou, R., Magoulas, G. D., Grigoriadou, M. (1999). In Neural network-based fuzzy modeling of the student in intelligent tutoring systems, international joint conference on neural networks (pp. 3517–3521). Washington DC, USA. doi:10.1109/IJCNN.1999.836233. ISBN 0-7803-5529-6. Stathakopoulou, R., Magoulas, G., Grigoriadou, M., & Samarakou, M. (2005). Neurofuzzy knowledge processing in intelligent learning environments for improved student diagnosis. Information Sciences, 170, 273–307. Tang, T. Y., Mccalla, G. (2003). In Smart recommendation for an evolving e-learning system workshop on technologies for electronic documents for supporting learning, international conference on artificial intelligence in education (AIED). Tang, T., McCalla, G. (2004). Beyond learners interest: Personalized paper recommendation based on their pedagogical features for an e-learning system. In Pacific rim international conference on artificial intelligence – PRICAI 2004 (pp. 301–310). Thai-Nghe, N., Drumond, L., Horváth, T., Krohn-Grimberghe, A., Nanopoulos, A., Schmidt-Thieme, L. (2011). In O. C. Santos, J. G. Boticario, (Eds.), Factorization techniques for predicting student performance. Educational recommender systems and technologies: Practices and challenges (ERSAT 2011). IGI Global. doi:10.4018/ 978-1-61350-489-5. ISBN13: 9781613504895. Widrow, B., & Lehr, M. A. (1995). Perceptrons, adalines, and backpropagation. In M. A. Arbib (Ed.), Handbook of brain theory and neural networks (pp. 719–724). MIT Press. ISBN 978-0-262-01148-8. Xu, D., Wang, H., Su, K. (2002). Intelligent student profiling with fuzzy models. In Proceedings of the 35th Hawaii international conference on system sciences.