Learning Path Recommendation using Hybrid Particle Swarm Optimization

A R T I C L E I N F O A B S T R A C T Article history: Received: 31 October, 2020 Accepted: 13 January, 2021 Online: 28 January, 2021 Revised Bloom's Taxonomy (RBT) is proposed in general to look more forward in responding to the demands of the developing educational community, including how students develop and learn and how teachers prepare Learning Objects (LO). The variety of characteristics of students' abilities in a class has always been a problem that is often faced by a teacher. Unfortunately, cognitive classifications to develop students' knowledge to a high level have not been used to plan a learning path that is appropriate for their cognitive level. The purpose of this study is to recommend a learning path that matches the cognitive abilities of students from a learning object ontology. The method used in this research is Hybrid Particle Swarm Optimization (HPSO) which integrates Binary Particle Swarm Optimization (BPSO) and Discrete Particle Swarm Optimization (DPSO). The Connection Weight (CW) function is used to test the quality of the connection between the learning objects of an ontology subject controlled by the cognitive class. Based on experimental studies, the HPSO method can recommend a suitable learning path for cognitive classes, namely Low Cognitive (CL), Medium Cognitive (CM), and High Cognitive (CH). The similarity of the sequence of learning paths based on population in CL-class is 87.5%, CM class 75%, and CH class 87.5%.


Introduction
In general, learning groups are designed to be heterogeneous. That is, a class is inhabited by students with various types of learning and characters. There are students with a fast learning type. While others are slow and even very slow. Therefore, it is difficult for teachers to follow an approach that is suitable for every student. In an effort to reach all students, teachers often design class lectures and activities. To remedy this situation requires personal support and guidance so that individual needs and difficulties can be addressed. However, given the number of students a class has, it is not easy for a teacher to deal with the diversity of levels and needs of students at all times [1].
Curriculum Sequencing (CS) is a technique to give students flexibility in planning the most appropriate sequence of individual learning tasks [2]. Produce an individual learning method in which each student dynamically chooses the most optimal teaching operation through presentations, examples, questions, or problems. Meanwhile [3] found the optimal learning sequence through the pretest to identify the weaknesses of students. Thus, CS aims to replace the rigid, generic, modeled learning structure set by the instructor or pedagogical team, with a more flexible and personalized learning path.
CS development according to [4] must consider students' learning abilities, background, and motivation. Even for certain students, the requirements can change according to their increased knowledge due to learning. According to [5] states that CS not only helps students find the most efficient and appropriate learning path but also helps instructors to organize program structures, content, or learning objects and make improvements. Meanwhile [6] stated that individualization of teaching materials was a challenge in choosing the right LO and making the LO sequence easy to learn. Thus, optimal results will be obtained because the learning path of learning is following students' abilities. Learning paths that have been developed have not used cognitive classification as a controller of the learning object ontology of a subject. The Evolutionary Computing (EC) algorithm approach has been widely used in solving Curriculum Sequencing problems.
According to [7] the EC method that is most widely used to solve CS problems is grouped into two main approaches, namely the social sequencing approach and the individual sequencing ASTESJ ISSN: 2415-6698 approach. Genetic Algorithm (GA) is a metaheuristic method which is famous because it has good performance for various types of optimization problems. There have been many GA studies in the CS domain including [8]- [11].
In [12] present a Personalized e-Course Composition approach based on Particle Swarm Optimization (PC2PSO). A binary particle encoding is used, in which each dimension determines whether the object of learning is included in sequence (1) or (0). The initial particles are generated randomly. The fitness function is used to determine Gbest and Pbest based on the level of compatibility between the pedagogical targets of students and the learning concept: learning material, level of compatibility, difficulty level, the time limit for learning, and student checks on the material that has been studied. Comparative evaluation in terms of average fitness and stability functions shows that the PSO method is better than the GA method. While the newer optimization method is swarm intelligence.
Swarm intelligence is an optimization method based on a distributed agent system, where each agent has simple abilities that can solve complex problems together. The approach to swarm intelligence is very different from EA because swarm intelligence emphasizes cooperation rather than competition. To support the concept of cooperation, each agent has a simple ability to learn from experience and also communicate with fellow agents. Metaheuristic methods based on the concept of swarm intelligence are Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO).
The PSO algorithm was originally proposed in 1995 [13]. In the PSO, each particle represents the solution to the problem. Particle learning consists of two factors, namely particle experience called cognitive learning and the combination of learning from the whole swarm is called social learning. Cognitive learning as Pbest is the best position that a particle has ever reached, while social learning as Gbest is the best position of all particles in the swarm. One of the uses of PSO in the e-education domain includes research conducted by [14] using the PSO method to offer formative value in e-assessment and turn it into a learning tool.
This research proposes a learning path model based on the cognitive classification of Revised Bloom's Taxonomy and an ontology of learning objects. Cognitive classification is based on RBT into three cognitive classes (Cognitive Low (CL), Cognitive Medium (CM), and Cognitive High (CH)). Hybrid Particle Swarm Optimization (HPSO) algorithm is used to solve combinatorial problems, namely learning object ontology with Discrete PSO which is controlled by cognitive class using Binary PSO. So that a learning path model will be produced that is following the cognitive abilities of students through optimization of the assessment of the LO relationship between RBT and the ontology of a subject.

Learning Object
Primary reference sources in the form of journals and dissertations, and secondary sources in the form of books, many have provided explanations about learning objects. However, in general, there is a common view of the concept of learning objects. The definition of a learning object is broadly described as all entities that are related or used as learning and training materials.
From this understanding, all general and conventional learning materials such as worksheets, modules, handouts, etc. are also included in the learning object category. Researcher [15] as well as a learning practitioner defines learning objects as each learning resource in the form of a digital entity that can be used repeatedly as a learning resource. While other researchers [16] presented various definitions of LO, the simplest description refers to data objects, content, or information.
The Institute of electrical and electronics engineer (IEEE) defines learning objects, namely as any digital or non-digital entity that can be used, then reused and can become a reference for technology-assisted learning. Thus LO is any entity (digital or nondigital) that can be reused or referenced including multimedia content, instructional content, software, people, organizations, or events for technology-supported learning. This definition leaves room for the entire curriculum to be seen as an LO.   Figure 1 shows that in one topic or one subject, the description of the content of the course is followed by the related learning material. In other words, subject content is usually arranged based on ontology, where the yellow circle represents the learning content while the blue rectangle represents LO [17].

Particle Swarm Optimization
The PSO algorithm works based on particles in the population that work together to solve existing problems disregarding the physical position. The PSO algorithm combines local and global search methods that balance exploration ability to conduct investigations in different areas of the search area to get the best optimal value and exploitation ability to concentrate around the search area for fix solution. The similarity of PSO and GA is that the system starts with a population formed from random solutions, then the system seeks optimization through random generation changes [18]. Each particle holds traces of position in the search space as the interpretation of the best solution (fitness) that had been achieved.
The basis of the PSO algorithm consists of three steps, namely determining the particle position and velocity, updating the velocity, and updating the position [19]. Each particle tries to update its position using information such as the current position, the current velocity, the distance between the current position to Pbest, and the current position to Gbest which mathematically updates the particle velocity calculated (1).
The achievement of the results is obtained from a new velocity calculation for each parallel based on the distance from the Pbest it has and the distance from the Gbest position updating the particle position , +1 is shown in (2).
Personal best (Pbest) according to [20] is the best value that a particle has, the value in the next time step, t + 1, where t ∈ [0,…, N], update Pbest is shown in (3), While the Global best (Gbest) is the best value that takes into account all particles in the population calculated using (4) and > 1 (4) Each iteration, each particle is given information about the latest Gbest value so that a one-way information sharing mechanism occurs to carry out the process of finding the best solution with a fast convergence shift.

Binary Particle Swarm Optimization (BPSO)
In [21] developed PSO to operate in binary search space because real value domains can be converted into binary valued domains. The proposed algorithm is called the PSO binary algorithm (BPSO) in which particles represent positions in binary space and the particle position vectors can take a binary value 0 or 1 where ∈ {0,1}. In this case, it maps from the binary space of the dimension (an eg string of bits of length n) to the real number = → (where f is the fitness function and is the set of real numbers). That means that the position of the particle must belong to to be calculated by the fitness function [22]. In BPSO, the particle velocity is connected to the probability that the particle position is 0 or 1. Equation (1) for updating the particle velocity is still used in BPSO. Then the sigmoid function shown in (5) is used to update the particle velocity.
The position of the particle is influenced by (6), where is a random number selected from the distribution (0, 1) and of the sigmoid function.

Discrete Particle Swarm Optimization (DPSO)
In 2000, Clerc modified the DPSO algorithm formulated by Kennedy and Eberhart. Clerc modifies the representation of the position of the particle, the shape of the velocity produced by the particle, and the effect of velocity on the position of the particle. This modification hopes that it can be applied to problems with discrete models, especially combinatorial types [23]. The velocity update formula that has been modified by Clerc is defined in (7), while for fixed position updates using (1).

Methodology
The Hybrid Particle Swarm Optimization algorithm is applied to overcome combinatorial problems in a more practical and orderly manner in determining learning pathways. Determination of the order of learning objects through LO ontology based on cognitive classes from the learning vector quantization method and using RBT to assess connection quality. The expected result is that each student gets a recommendation for a learning path that is appropriate to their cognitive level.
The model used in this study consists of three components, cognitive classification with LVQ, learning object ontology based on RBT, and hybrid discrete particle swarm optimization. The general architecture of the proposed model can be seen in Fig. 1  Meanwhile, the value of particles with a value of 1 is determined as LO which must be followed and the calculation is carried out which is not affected by the particle velocity. The updated particle velocity and position of the CL cognitive class are presented in Figure 3.

Particle Representation DPSO
LO particles consist of eight particles arranged based on objectoriented programming lessons that have been determined based on the depth and breadth analysis of RBT, namely Table I. The representation of LO-LO sequence particles is shown in Figure 4 randomly from three groups of particles.

BPSO -DPSO Integration
HPSO is the part that performs the process of determining the learning path, consisting of BPSO and DPSO. Cognitive classes (CL, CM, and CH) are represented as particles in the BPSO method and ontology learning objects are represented as particles in the DPSO method. The integration of BPSO and DPSO through the fitness function will produce a learning path model that is suitable for students' cognitive abilities. Figure 5 shows the process of updating the speed and position of particles for DPSO and BPSO in each iteration so that it will ensure changes in LO and cognitive class (CL, CM, CH). Update the velocity and position of DPSO particles on the learning object update through (7) to update the particle velocity, and use (2) to update the position. Meanwhile, the update of the speed and position of BPSO particles in the cognitive class is through (1) to update the speed, (5) to calculate the sigmoid value, and (6) to update its position.

The Proposed HPSO Algorithm
The fitness function is used as a measuring tool to select the best object from a set of existing objects. The fitness function is identical to the objective function of the optimized problem, which is to make the best individual learning route according to their cognitive abilities. The objective function is determined based on the highest fitness value. The higher the fitness value, the results are closer to the objective function.
The objective function defined in this research is to create individual learning paths or routes based on student cognitive data, namely CL, CM, and CH with the learning object ontology of a subject. Thus, the fitness function is to find the best Connection Weight (CW) value between LO in ontology-based on students' cognitive classes, namely CL, CM, and CH as illustrated in Figure  6.
Connection Weight ( ) was used to assess the relationship of LO in RBT ontology as cognitive level evaluators (8).

Figure 6: Illustration of Determining the Fitness Function
The objective function value is calculated for all the particles generated earlier. The PBest position is determined from the objective function value obtained. The initial velocity of the particles is the absolute difference between the initial position and the personal best position of the particles. The particle with the maximum objective function value is the GBest particle and the corresponding position is the global best position [24]. The calculation of the fitness function stops until the iteration value limit has been determined.

Results
Testing the HPSO algorithm to recommend a learning path model that suits the CL, CM, and CH cognitive classes through steps, i.e: a. Testing the next computational program, test the consistency of the PBest fitness function for each population to become GBest through five particle populations, ten particle populations (ten LO groups), and fifteen populations (fifteen LO groups). b. Groups of students in cognitive classes CL, CM, and CH will get the best learning path recommendations based on their respective GBest.

Connection Weight for Fitness Fuction
A CW calculation process occurs when one LO is paired together with CL of value 1. The calculation starts with finding the DBO, DBB, and CW values of each LO. In the state of one LO paired together with CL is 0 then the LO is not used in the connection weight calculation process. The mechanism for testing connection weight according to in accordance with the procedure shown in Figure 7.   The fitness function in the CL,CM,CH-class is calculated using (10), (11), (12) ,the fitness value data with 15 groups of particles is shown in Figure 8. Pbest tests through the fitness function in the CL, CM, and CH classes show that the higher the iteration used, the more optimal the solution will be from the system, marked with the increasing fitness results. This is due to the increasing number of iterations used to make the particles move to find the optimal solution so that the particles can find the optimal solution.

Learning Path Recommendation
Gbest is the best value that takes into account all particles in Gbest the population in each CL, CM, and CH class. Each iteration, each participant is given information about the latest Gbest value so that there is a mechanism for sharing one-way information to carry out the process of finding the best solution with a fast convergence movement.
Learning path recommendations based on the Gbest values shown in Table 2 indicate that an increase in the number of particles affects the value of the resulting learning path sequence, but is still within the schema of each cognitive class.  The HPSO algorithm can create learning pathways by CL, CM, and CH cognitive classes. The cognitive class CL tendency of its learning path (1-3-2-5-4-6-7-8) shows the order of the basic LO to a higher LO. Whereas the cognitive class CH of the learning path tendency (8-5-7-6-4-3-1-2) shows a sequence of high LO to the basic LO. In CM cognitive class the tendency of learning path is (3-5-7-6-4-2-8-1). The similarity of the learning path sequence based on the number of CL cognitive particles is 87.5%, CM is 75%, and CH is 87.5%, so the average similarity of the learning path sequence is 83.3%.
The results of the learning path model research using HPSO have been presented and explained to three education experts. The recommendation of three education experts stated that cognitive classification with LVQ and the logical flow learning path model was acceptable. Researchers can prove the use of RBT through the results of learning paths based on cognitive low (CL), cognitive medium (CM), and cognitive high (CH). The HPSO learning pathfinder is a very important part of combining the output of the LVQ cognitive classification with the ontology learning object.
The advice given from education experts is as follows, the dynamics of the interrelation between the depth of the material and the breadth of the material can be determined by the teacher or teacher based on competency targets to produce a learning path in accordance with the characteristics of students. This research can be developed in the realm of attitudes and skills in accordance with learning evaluation competencies, namely, attitudes, knowledge, and skills.

Conclusion
The Hybrid Particle Swarm Optimization method, which consists of DPSO for learning object ontology and BPSO in cognitive classes, can be applied more practically and regularly to produce a learning path recommendation. Based on experimental studies and verification of education experts, the HPSO algorithm in this research can recommend a suitable learning path for students classified in the CL, CM, and CH cognitive classes.
Further research is the integration of the Learning Vector Quantization method with Hybrid Particle Swarm Optimization in a platform that is supported by the provision of standard learning object ontologies that can be the right and efficient solution in the development of smart e-learning mode during the Covid-19 pandemic.