Fuzzy-based Adaptive Framework for Module Advising Expert System

Obada Alhabashneh, "Fuzzy-based Adaptive Framework for Module Advising Expert System”, Annals of Emerging Technologies in Computing (AETiC), Print ISSN: 2516-0281, Online ISSN: 2516-029X, pp. 13-27, Vol. 5, No. 1, 1st January 2021, Published by International Association of Educators and Researchers (IAER), DOI: 10.33166/AETiC.2021.01.002, Available: http://aetic.theiaer.org/archive/v5/v5n1/p2.html. Research Article


Introduction
Every academic semester, students enroll in different modules offered by their universities. Enrolment might sound like a simple process, but it involves many complications and problems for students particularly in these universities where the credit hour system is adopted. According to this system, although there are some constraints such as the academic plan, the students still have the freedom to select the sections to enroll in based on their preferences including the module, timeslot, and lecturer of the offered sections in the timetable. However, the lack of knowledge about the modules and the lecturers, coupled with the contradictory advice they receive from their colleagues and the complexity of the academic plan of their programs may affect the students' ability to select the right choice. This creates a need for professional advice as taking a wrong choice may lead to unwanted situations such as failing the module, or more seriously, academic warning, program withdrawal or even leaving the university [1][2][3].
To address this problem, universities usually have academic advisors to help the students in selecting their modules and sections for the upcoming semester based on different factors including www.aetic.theiaer.org the student's cumulative Grade Point Average (GPA), the available modules, timeslots, module's lecturer, and the academic plan of the student. However, module advising is a challenging task as it is complicated, time-consuming, and needs qualified and knowledgeable human resources. This urges the need for developing intelligent systems for module-advising to aid the students in selecting the right modules at the right time to ensure a smooth passage for students throughout their academic plans to finish their degrees with fewer problems.
Intelligent Recommender Systems (RSs) are widely used in different application areas including online shopping, movies, social networking, and others. However, they are less common in education and this area is still under development [1][2][3]. RSs can provide the students with personalized recommendations on suitable modules based on the student's academic history. Such systems can also incorporate previous student's experiences to provide an improved recommendation. Furthermore, they adapt to the changes in the data and give the recommendation based on that. The need for RSs and their requirements in the education sector has been discussed in [4][5][6]. More recently, the use of RSs in education has received more attention and their potential has been proven [7,8].
This paper discusses the development of a fuzzy-based framework to be used in recommender systems for module advising. The proposed framework builds the recommendation based on creating three main spaces which are: student-space (SS), module-space (MS), and lecturer-space (LS). These spaces are used to estimate the risk level associated with each student, module, and lecturer. The framework then associates each abnormal student case in the students' grade history with the estimated risk level in the SS, MS, and LS involved in that particular case. Fuzzy-based associationrule learning [9,10] is then used to extract and summarize a fuzzy rule base. The fuzzy rule base is used to predict the risk level associated with the combination of a specific student, module, and lecture and build the recommendation to students based on that. The proposed approach provides a novel mechanism to estimate the consequent risk associated with the student selection of a specific module that is taught by a specific lecturer. The risk estimation is based on creating and analyzing the three space elements.

Course/Module Advising in Higher Education
Academic advising is vital for a good academic experience of students as it positively impacts their success and retention [11,12]. However, it requires specific knowledge of the student's situation, history, program's academic plan, modules, and lecturers [12]. Gordon et al. [13], defined academic advising as "situations in which an institutional representative gives insight or direction to a college student about an academic, social, or personal matter". Module advising is one of the main tasks of the academic advisor, which refers to the process by which the academic advisor supports the student in selecting the right modules in which to enroll. Having the importance and the complexity of module advising, different researchers have argued that there is a need for developing intelligent systems to support this task [2,4,7,14]. Developing such systems aims to minimize the demand for human advisors and gives them more time to focus on other important advising tasks including career development and financial issues. [12][13][14][15][16].
A novel approach for module long term plaining called Interactive Decision Support for Course Planning (IDiSC+) was proposed by Mohamed [12]. The approach used optimization techniques to support both the student and the advisor in building a recommended long-term academic plan (towards graduation) to be followed by the student. Laghari [14] developed an Automated Course Advising System (ACAS) for module planning. The system distributes the modules of the academic plan on different semesters based on the history of other students.
There has been significant research effort in applying information technology for module advising. Roushan et al. [15] introduced an Internet-based approach to support the module-advising process which integrated the process of advising with the enrolment taking into account the constraints of the program plan and the university policy. However, the system did not replace the human academic-advisor but rather facilitated the advisor role by providing an automated tool for www.aetic.theiaer.org communication and information exchange. A decision-support system was proposed to support academic advisors in preparing a pre-enrolment plan for the students and assist in the offering of appropriate modules for the upcoming semester [16]. Al-Ghamdi et al. [17] developed an advisor expert system (PAS) for postgraduate students. The proposed system was designed to assist them in the selection of the most relevant modules without referring to a human module-advisor. A webbased advising system [18] was proposed which supports three types of users, including students, advisors, and heads of department to make sure that a complete picture is available for the students. Mattei et al. [19] developed a decision-theory advising tool to enhance the advisor-student relationship. The tool allowed students to browse the module offerings, possible future scenarios, and their outcomes. Shatnawi et al. [20] used the enrolment and marking history from similar cases and applied an association rule-based system to give general recommendations when selecting the modules on which to enroll.

Content-based
Different researchers have used a content-based academic recommendation system for module selection. For example, Lin et al. [21] utilized a multi-agent approach and ontology to provide a dynamic and personalized recommended module list. The multi-agent approach, which included various agents, used a preference-driven planning algorithm supported by the ontology to build the recommendations. Darimola et al. [4] integrated Case-Based Reasoning (CBR) with Rule-Based Reasoning (RBR) techniques to provide an intelligent approach for module advising. The approach also used historical data to build a list of recommended modules for the following semester.

Collaborative filtering
Collaborative filtering was also used for academic advising. Huang et al. [22] proposed a scorebased prediction approach for course recommendation. The approach used a user domain collaborative filtering to create the recommended course list. Courses were clustered based on the student feedback in [23] and the resulted clusters were combined with fuzzy based rule association technique to create the course recommendation. Nafea et al. [24] developed a learning-style-based collaborative-filtering approach for module recommendations which utilized different metrics to identify similar profiles including k-means, cosine, and person correlation. Chang et al. [25] proposed a user-based collaborative-filtering approach to predict student grades. Mortenson et al. [26] introduced a collaborative filtering approach for module selection which utilizes an artificial immune mechanism for the prediction. Bydžovská [27] investigated the effect of student and module features on the enrolment patterns and designed a collaborative-filtering-based system to predict the module grade. Yao et al. [28] attempted to increase the fairness of the module recommendation by addressing the biased-recommendation problem against minority groups of students. They developed four different fairness metrics that can be optimized using the learning objectives

Knowledge-based
Different knowledge-based recommendation approaches have been proposed for module selection. Xu et al. [29] developed a knowledge-based approach to offer a personalized module sequence to the new students. This approach utilized a dynamic learning algorithm that learns from the performance of other students on a specific module. Koutrika [30] argued that recommendation methods should not be 'hard-wired', but it should be flexible. In that sense, a new paradigm for the recommendation was introduced in which a recommendation approach can be defined declaratively as a high-level parameterized workflow comprising traditional relational operators and new operators that generate or combine recommendations. Keston et al. [31] utilized semantic web expert system technologies to build a knowledge base that is used by an intelligent web-based application to provide the required recommendations. Engin et al. [32] developed an expert rule-based system for module recommendation. The rules were captured from the real advisors and then injected into a rule-base using Oracle Policy Automation (OPA). www.aetic.theiaer.org Hashemi and Blondin [33] included several factors to be taken into consideration when recommending modules for students such as the frequency of the module offering, balancing the module load, and shortening the graduation path. All these factors and others were included in a rule base which was used for recommending modules. Ayman [34] proposed an expert system for module selection which included both prescriptive and developmental advising models and utilized objectoriented databases for data and rule representation. Abdullah Al-Ghamdi et al. [17] proposed a rulebased advising system. The system was designed for postgraduate students and built recommendations to support the students in the selection of modules related to the topic of their research thesis. Nambia and Dutta [35] introduced a dynamic and flexible rule-based advising system in which the rules were separated from the execution to enable the student to try different scenarios by updating the XML file where the rules are stored. Nguyen et al. [36] proposed a knowledge-based framework that utilized a learning data-warehouse for discovering patterns in the student behavior including module selection and achievements. These patterns were then used to make the recommendations.

Hybrid
Other researchers proposed Hybrid approaches in which more than different perspectives were integrated. Daramola. et al. [4] designed an intelligent expert system for module advising which integrated rule-based reasoning (RBR) with Case-Based Reasoning (CBR) using the academic history of the students. Sobecki [37] applied Ant Colony Optimization (ACO) to provide an efficient module advisor system. The system predicted the final grade of the students in a module, based on a domainspecific representation integrated with the ACO. Abdulwahhab [38] integrated Genetic Algorithms (GAs) and a Decision Tree for short-term module-scheduling.

Fuzzy Based
A few researchers have used fuzzy logic to develop module advising systems. Goodarzi and Rafe [39] developed a fuzzy-based expert system for student advising. The proposed system was a web-based module that can be integrated with the university portal. The module fuzzifies the business rules and the GPA of the students to advise them on which modules should be taken in the following term. Adak [40] used Fuzzy techniques to recommend elective modules to students. It analyzed transcripts of students to extract fuzzy rules to relate the module to the students who intend to take them. Baloul and Williams [41] used the Order of Preference by Similarity to Ideal Solution (TOPSIS) technique to develop a fuzzy model for probation students to minimize the risk of taking the wrong modules in the early stage of their study.

The proposed approach
This paper discusses the development of a fuzzy-based framework of an expert system for module advising. The proposed framework assumes that recommendations are based on three main elements which are: student-space (SS), module-space (MS), and lecturer-space (LS). SS contains the current student status which includes their accumulative average (if there is any academic warning), the closest abnormal academic status and to which extent it's close to that status. It also contains the knowledge domains of the academic program and the progress of the students in each domain. MS contains the average module mark for the last five years and the knowledge domains that the module belongs to. LS contains the average of the marks for the lecturer for all modules and the average of the lecturer's marks for each module. The three spaces and the result of the final calculation are then combined in a matrix called the case-space (CS). Fuzzy based association-rule learning is then used to extract the dominant rules to classify the consequent case for each eligible module if it was taken by the student for a specific semester. The fuzzy logic is used to handle the uncertainty involved in modeling the human decision and to provide a transparent and interpretable mechanism module taking risk estimation. www.aetic.theiaer.org The main purpose of the proposed framework is to estimate the consequent risk level (Low, Medium, and High) taking a specific module. The risk level is assigned based on a list of unwanted cases associated with the student failing or not progressing in the module. These cases include a GPA decrease, moving down from one GPA category to a lower one (degrading), receiving an academic warning, withdrawn from the academic program, withdrawn from the university, and graduation pending. In addition to the risk-level estimation, the framework provides the student with a justification (interpretation) for the estimated risk-level based on the student's situation, the targeted module, and the lecturer. The framework was developed based on real-life university data which included a historical record of enrolment associated, student's marks, cumulative GPA, module offerings, academic plans for the programs, and finally the knowledge domains for both the programs and the modules. The proposed approach is depicted in Figure 1   The proposed approach consists of six main steps:

Step-1: Creating the Three Spaces and Abnormal Case Matrix
In this step, the student-space (SS), module-space (MS), lecturer-space (LS), and abnormal-case matrix are created as follows:

Step-1.A: Creating the three Spaces
The three spaces SS, MS, and LS are created as shown in Fig. 3Error! Reference source not found.. The spaces are extracted from the university database.  Table 1.

Student Average Mark for the Program Knowledge Domains
This is a matrix that includes the knowledge domains of the student academic program and the average mark of the student for the modules which belong to each domain.

Potential Risk Situation
This refers to the approximate abnormal student situation for the current situation of the student. (i.e., the current abnormal situation of the student boundary between two GPA categories, such as "very good" and "satisfactory"). Each abnormal-situation type is given a code and a percent of 100 to indicate the level of risk which is given to the code.

Current Student Situation
This refers to the student's current abnormal situation. i.e., If the current abnormal situation of the student boundary between two GPA categories, such as "very good" and "satisfactory", or the student's GPA has dropped. Each abnormal-situation type is given a code and a percent of 100 to indicate the level of risk which is given to the code.
Module Space: this space contains information about the module as shown in Table 2.  Table 3.

1.
Lecturer Id A unique identification number of the lecturer

2.
General Average of Grades The average grade of the lecturer

Fail Rate
The percentage of students who failed the modules taught by the lecturer.

Step-1.B: Creating Abnormal Case Matrix
In this sub-step, the abnormal-case matrix is created by selecting the student id, module id, lecturer id, and the student abnormal situation for abnormal cases in the student enrolment records from the university database. The abnormal case is usually indicated by a flag or a symbol in the database as shown in

Step-2: Case Risk Analysis and Calculation
For each problematic case in the dataset, a risk weight is calculated for the three spaces: SS, MS, and LS and the associated abnormal situation. The product of this step is shown in Fig. 5. The risk was calculated for the three spaces using equations developed based on common sense and the opinion of field experts such as the academic registry members, lectures, and head of departments. The results than have been presented to the experts and evaluated by them. Where LR is the risk estimation that comes from the lecturer side out 100, LS is the lecturer-space. LAG is the general average mark of the lecturer. LID is the lecturer-identification number. LFR is the lecturer fail rate which is a percentage of the students who failed the modules taught by the lecturer.

Situation risk (SR) estimation:
The resulting situation types were provided to a set of experts (i.e., academic-registry staff, advisors, and heads of department) and they were asked to assign a specific weight which reflects the risk level. The risk level takes a value between 1 and 100. The net risk for each situation type is calculated by taking the average of the weights given by the experts.

Step-3: Linguistic Labelling
In this step, the final risk weights of the three spaces and the resulting situation are given linguistic labels of H-high, M-medium, and L-low. These linguistic labels are generated using Type-1 and based on predefined fuzzy sets based on the Mendel Wang method [10]. In the proposed approach, the member function of the fuzzy sets uses a triangle shape as shown in Figure 6 and is based on three parameters {a, b, c} as shown in Equation 4.

Step-4: Fuzzy Rule-Base Creation
To create the fuzzy rule base, in this step and for each abnormal case, the linguistic labels resulting from the previous step are associated with each other in a form of antecedents and consequents (If-> Then) where the linguistic labels of the three spaces are the antecedents and the linguistic label of the resulting situation is the consequent as shown in Figure 8.

Step-5: Fuzzy Rule-base Compression and Validation
The rule base extracted in the previous step could contain many rules based on the size of the dataset. It may have a large amount of repetition (i.e., Repeated rules) and contradiction. The contradiction here means rules that have the same antecedent with different consequents such as shown in the following example in Table 4: To address these issues a five-fold cross-validation [42] process was used to train the rule base to summarize it to the most dominant unique patterns. In this validation method the dataset is divided into 5 equal-size sets (D1, .., D5). For each fold the following steps are applied: 1. One of the subsets Dn is selected as a testing set Tn and the other subsets are groups in one training set D^. 2. The rule compression(summarization) technique [43] is applied to the training set to produce a summarized rule-base 3. The summarized rule-base is applied to the test set to predict the risk level. 4. The predicted risk levels are compared with the actual risk levels from the dataset, to identify the accuracy of the rule-base in predicting the risk level The compression technique: This technique uses two measures for rule quality, which are "generality" and "reliability" and are used to identify rule patterns with maximum firing strength. Generality measures the number of instances in the extracted rule base which supports each rule pattern. Reliability measures the confidence level of each rule pattern [43]. www.aetic.theiaer.org In the proposed approach, generality is calculated using scaled fuzzy support, and the reliability is calculated by multiplying the scaled fuzzy support by the firing strength of the rule pattern. The support of the rule pattern refers to the number of the rules in the rule-base that the pattern represents. The "confidence" refers to the strength of a specific rule pattern against the contradictory patterns (i.e. because other rule patterns have the same antecedent and a different consequent) [43].
Fuzzy support is used to identify the unique rule-patterns with their occurrence in the extracted rule base. The fuzzy support can be scaled for each unique rule-pattern using the total number of instances in the rule-base which have the same consequent. Equation 5 shows how the scaled fuzzy support for a unique rule pattern is calculated: The "confidence" is a measure of the uniqueness of the pattern as it indicates its strength against the contradictory patterns, which are the other patterns having the same antecedents but a different consequent. Equation 6 shows how confidence is calculated

Equation 6: Confidence
Where is the number of instances in the extracted rule base which supports the contradictory rulepatterns with . The final scaled-weight of the rule pattern is calculated as the product of the fuzzy support and the fuzzy confidence as shown in Equation 7.

= × Equation 7: Final scaled weight
Each of the generated unique rule patterns is assigned the scaled fuzzy weight measure scWi as follows:

Conclusion
This paper has presented a fuzzy-logic based approach of an expert system for module advising. The approach creates three spaces which are student-space, module-space, and lecturer space, and associate them with the abnormal situation type for each abnormal case. Each space and the abnormal situation type is then given a linguistic label using fuzzy sets. The linguistic labels are then associated with each other for each case to generate a fuzzy rule. The fuzzy rules for all cases are combined in one rule-base and then compressed to extract those rules with the highest firing strength. The fuzzy logic was used to handle the uncertainty implied in the human judgment of the student case as well as to provide a transparent and interpretable mechanism for predicting the risk level of enrolling a student on a module.
The approach was developed using real-life university data and achieved an acceptable level of accuracy of 92% which is expected to improve as more data is captured and used to train the rule base. Although the achieved accuracy might sound like it needs a bit of enhancement but having that the machine learning approach used for training the rule base the accuracy is expected to enhance as more data instances are included. Also, the accuracy level acceptance depends on the problem and context, especially with cases in which are molding human decisions or behavior. Although there are different approaches have been proposed in this area, this paper introduces a novel mechanism that creates three spaces to estimate the risk of the student situation associated with a specific module. The three spaces which are namely: Students, Lecturer, and Module provide a multi-angle view on the student case and makes the estimation more realistic. Also, applying the fuzzy logic provides a means to handle the uncertainty included in the human decision-making regarding module selection. www.aetic.theiaer.org The fuzzy rule base also provides a transparent mechanism to make the recommendation which means it doesn't provide the risk level only but also the justification of that recommendation based on the risk level of the three spaces.