The functionality of recommendation systems is implemented in many popular services and marketing platforms, as they provide an opportunity to analyze and filter information in choosing the most suitable products, content, or services for the user according to his preferences and capabilities. However, in education, such functionality is also in great demand, especially when choosing educational content according to preferences and learning styles to create an educational trajectory based on previous experience and academic performance. Lack of understanding of their capabilities and preferences or inability to build a structured learning vector is inherent for applicants and first-year students. That is why this article analyzes the main problems in the implementation of a recommendation system for building an individual study plan for a higher education student, identifies the main ways to implement such a system, and builds a basic algorithm for our recommendation system that can provide top disciplines for a student to study based on his or her previous learning outcomes, the learning outcomes of students of the same speciality for a certain period, and the factor of similarity of subjects and the choice of his or her specialization.
According to draft law No. 10177, “On the Development of Individual Educational Trajectories and
Improvement of the Educational Process in Higher Education” [
However, students face many problems, including a lack of understanding of how to properly build their path to professionalism. As a result, the applicant may choose subjects that do not best suit his or her preferences and capabilities or make a wrong choice of subject based on a small sample of data provided by friends or family.
Higher education institutions are also interested in the correct construction of the curriculum for their students because many important factors may directly depend on this, such as academic performance, satisfaction with the university, the qualification level of the specialist, and even the level of happiness of the person.
Based on the above, we are faced with finding mechanisms and means to build an individual study plan for a higher education student, which must be resolved to meet the educational institution’s and its students’ needs.
Let us identify the main problems that appear on the way to solving this problem:
• Scalability – each student needs to build a personalized learning schedule. For example, about
26 thousand students study at the National University of Life and Environmental Sciences of
Ukraine and its separate structural subdivisions [
Many researchers described how to implement recommendation systems. For example, in Li and Ye [
Mondal et al. [
Weber et al. [
Also, there is a diferent way to build your recommendation system. Cem Dilmegani, the principal
analyst at AIMultiple, says in [
To predict students’ future grades based on past grades, authors in the article [8] tried three diferent algorithms: a Linear Regression model based on least squares, a Random Forest Regressor with 100 decision trees, and an Artificial Neural Network with four dense layers and ReLU activation functions. All of them were implemented in their web application and could be used by students.
However, after analyzing many articles[
Considering the abovementioned problems, we can conclude that implementing an automated system is the most crucial way to provide personalized student recommendations in building their curriculum. Therefore, the goal we set for ourselves is to identify technologies and develop algorithms that can be used to implement the basic logic of a recommendation system that will be able to solve the problem of forming individual educational plans based on processing large amounts of data.
There are two options for implementing a system for building an individual study plan: 1. Writing our algorithm for reading and processing data from scratch could satisfy our problems.
In this case, the advantage is that we can fully customize the work of such a model to our needs. However, the disadvantage is the need to solve problems and tasks that have long been solved using the second method of implementing an automated system. 2. Using a recommendation system based on machine learning, which is a subtype of artificial intelligence that is widely used to predict the best choice of a particular product or service for a user, based on numerous factors, such as personal preferences of the user, preferences of his or her potential associates, or with the calculation of specific filters and restrictions.
The second method already contains a set of ready-made solutions that will help filter and process data and provide the most accurate forecast for the user. A recommendation system can be implemented in many frameworks and programming languages, including Python, Java, .NET, etc. That is why, when choosing a way to implement an automated system for building a curriculum for a higher education student, you should pay attention to this technology, its capabilities and its family of algorithms in general.
In article [11], a recommendation system is defined as an artificial intelligence algorithm, usually associated with machine learning, that uses a large amount of data to make a recommendation or guess for a particular user. Recommendation systems began their development as much as 40 years later [12] and gained such a powerful development to a greater extent to solve the problem of users in the corporate segment. Companies quickly realized that having such a system could be a significant advantage over competitors, so they began to invest heavily in this area. For example, Netflix and YouTube, one of the most popular content viewing platforms as of 2024 [13], actively invested resources in developing recommendation systems for their platforms in the first decade of the 21st century [ 14].
There are two general ways to generate candidates: content-based filtering and collaborative filtering [15]. The first type can make recommendations based on the user’s preferences, and the second type can make recommendations based on the preferences of a subgroup of users who are similar to the target user. Content-based filtering considers only the student’s preferences, which is not desirable for a recommendation system for a higher education institution since, although universities can train specialists in a narrow profile, possessing basic diverse knowledge in their specialized field is mandatory to create a highly qualified specialist. The experience of previous generations is also an essential factor in choosing a future study plan, as students are likely to choose subjects that are most likely to be passed with the highest probability of success and have the best possible feedback from senior students. That is why collaborative filtering is the best fit for our requirements, as it can solve our target task of recommending subjects based on the student’s academic performance and the results of other students.
In addition, it is advisable to introduce a factor of similarity of subjects into this recommendation system. This will allow us to diversify the factors that determine the most optimal subject for a student and increase the accuracy of the forecast. Subjects are often very similar or even identical in content, but they are entered into the database as separate, unrelated entities. Suppose we introduce the similarity factor into the model for the recommendation system. In that case, we can also tell that these items are close to each other, so it is necessary to analyze them as related.
Our study used matrix factorization, a simple model for displaying a discrete data set in a vector space. It is used in joint filtering, based on which a further forecast for the end user is built. The formula for matrix decomposition can be as follows:
≈ In this formula: • (R) – rating matrix; • (U) – matrix of users; • ( ) – transposed product matrix.
Matrix factorization has both advantages, such as the ability to provide a personalized forecast based on the interaction between users and products and the ability to work with diferent types of data, but also has disadvantages, such as the mandatory need for a large amount of data to provide an accurate forecast and the complexity of computation. This set of advantages and disadvantages is satisfactory for our task since we need a personalized prediction based on the grades of the target student and other students, and a large amount of data will always be present in a system where many academic performance results are constantly recorded.
The algorithm of such a recommendation system is shown in (figure 1).
The main data sources with which our proposed algorithm works are the grades of a specific student during the period of study, the similarity of subjects, and grades from the subjects of students of the same specialty for the last 3 years. Thus, the recommendations we will receive after training the system will be relevant, as they are based on fresh data and useful in building our future vector of individual training by profession and preferences.
To evaluate the model’s accuracy, the root mean square error (RMSE) is used, which measures the average square error between the predicted actual values and the actual values. The formula for the root mean square error is as follows:
⎯ = ⎷⎸⎸ 1 ∑=︁1 ( − ˆ)2, • (n) – count of observations; • () – actual value; • (ˆ) – predicted value.
Using matrix factorization and methods for evaluating the assessment’s accuracy, the recommendation system’s basic logic was built using Microsoft and the C# programming language.ML library, which allows us to use machine learning. It includes:
1. To hook up our libraries that will provide us with data. One (e.g. Models) to make it possible to use our custom models in basic logic. Another one (e.g., Repositories) allows one to extract custom-specific data for training models. 2. To define the main models for machine data analysis (e.g. StudentCourse for training base data and CourseRecommendation for prediction result). 3. To enable client code functionality, it is essential to define a specific service (e.g. SubjectRecommender) that can be instantiated and utilized from any location. This service will operate based on the input of specific parameters. The unique student identifier will serve as the primary key in this context. Utilizing this identifier, we will be able to locate and analyze data directly associated with the user and process it to generate relevant recommendations: 4. To build a model for learning data, we need to get our custom students to score data, identify the training data, and create a pipeline based on it. For this purpose, we got custom information about the scores of students in disciplines of the same specialization as our target student. Then, we convert specific custom information to ML training data. After that, we split the dataset into the train set and test set according to the given fraction. Moreover, finally, we create a matrix factorization trainer with advanced options, which predicts element values in a matrix using matrix factorization: var studentDisciplinesScoresOfTheSameSpecialization = _studentDisciplineRepository
.GetStudentDisciplinesScoresOfTheSameSpecialization(studentId); var trainingData = mlContext.Data.LoadFromEnumerable(
studentDisciplinesScoresOfTheSameSpecialization); var dataSplit
= mlContext.Data.TrainTestSplit(trainingData, testFraction: 0.2); var matrixFactorizationTrainerOptions = new MatrixFactorizationTrainer.Options {
MatrixColumnIndexColumnName = "StudentIdEncoded", MatrixRowIndexColumnName = "CourseIdEncoded", LabelColumnName = "Label", NumberOfIterations = 20,
ApproximationRank = 100 }; var pipeline = mlContext.Transforms.Conversion .MapValueToKey( outputColumnName: "StudentIdEncoded", inputColumnName: "StudentId") .Append(mlContext.Transforms.Conversion.MapValueToKey( outputColumnName: "CourseIdEncoded", inputColumnName: "CourseId")) .Append(mlContext.Transforms.Text.FeaturizeText(outputColumnName: "CourseTypeFeaturized", inputColumnName: "CourseType")) .Append(mlContext.Recommendation().Trainers
.MatrixFactorization(matrixFactorizationTrainerOptions)); 5. To train machine learning model using the training dataset and make some predictions using trained model on some trained data. Then we evaluate the accuracy of the calculations using Root Mean Squared Error: var model = pipeline.Fit(dataSplit.TrainSet); var predictions = model.Transform(dataSplit.TestSet); var metrics = mlContext.Regression.Evaluate(
predictions, labelColumnName: "Label", scoreColumnName: "Score"); Console.WriteLine(
$"Root Mean Squared Error: {metrics.RootMeanSquaredError}"); var predictionEngine = mlContext.Model.CreatePredictionEngine
<StudentDiscipline, CourseRecommendation>(model); 6. Moreover, by developing a mechanism for providing recommendations, we can provide an inevitable result of the algorithm’s work for the student based on a specific sample of user data. For example, we can get for our calculation already passed disciplines by the target student, disciplines that the target student passed and all available not passed disciplines. Based on this data, we can provide some predictions of the top 20 courses that were target student did not learn: var studentPassedDisciplines =
_studentDisciplineRepository.GetPassedDisciplines(studentId); var studentSelectedDisciplines =
_studentDisciplineRepository.GetSelectedDiscipline(studentId); var notSelectedDisciplines = studentDisciplinesScoresOfTheSameSpecialization .Where(x => studentPassedDisciplines.All(y => y != x.CourseId) && studentSelectedDisciplines.All(y => y != x.CourseId)) .GroupBy(x => new { x.CourseId, x.CourseType }) .Select(x => x.First()); var courseInfoForPredicition = notSelectedDisciplines .Select(notSelectedDiscipline => (notSelectedDiscipline.CourseId, notSelectedDiscipline.CourseType)).ToList(); var predictionResults = new Dictionary<int, float>(); foreach (var (courseId, courseType) in courseInfoForPredicition) { var prediction = predictionEngine.Predict(new StudentDiscipline { StudentId = studentId, CourseId = courseId, CourseType = courseType }); Console.WriteLine($@"Predicted preference for student {studentId} for course {courseId} ({courseType}): {prediction.Score}"); predictionResults[courseId] = prediction.Score; } var topPredictedCourses = predictionResults
.OrderByDescending(x => x.Value).Take(20).ToList(); foreach (var course in topPredictedCourses)
Console.WriteLine($@"The course that suits you best and
that we recommend you study: {course.Key}:{course.Value}"); }
The proposed basic logic of the recommendation system obtains custom-specific information about the scores of students in disciplines with the same specialization as our target student. Next, the algorithm contains the main logic of the recommendation system, creates a matrix factorization trainer with advanced options that predicts element values in a matrix using matrix factorization, evaluates scored regression data, and uses the generated prediction engine to make several predictions about the top 20 courses that the target student have not passed yet. This described logic contains custom classes and repositories for presentation only and can be replaced with your own.
The article analyzes the task of building a curriculum that best suits the student’s preferences and capabilities, identifies the main problems that arise when solving such a task, and considers ways to overcome such dificulties. A basic algorithm has been built that will give a start to the work of a recommendation system, the purpose of which is to provide a student with recommendations of subjects to study based on his or her previous learning outcomes, the learning outcomes of students of the same speciality for a certain period, and also on the factor of similarity of subjects. Based on this data and the proposed basic algorithm, the system can provide the top subjects for a student to study at a basic level. These subjects will be specialized, will be responsible for the next level of qualification of a young specialist, and will contain the highest probability of their successful completion. To implement such a system, it is proposed to use the ML.NET machine learning library for the C# programming language.