The use of artificial intelligence systems for data processing in the educational process

Kopysheva Tatyana Nikolaevna ORCID: 0000-0003-3392-1431 PhD in Physics and Mathematics Associate Professor; Department of Mathematical and Hardware Support of Information Systems; I.N. Ulyanov Chuvash State University 428015, Russia, Republic of Chuvashia, Cheboksary, Moskovsky ave., 15, office B-305 tn_pavlova@mail.ru Mitrofanova Tatiana Valerievna ORCID: 0000-0002-5750-7991 PhD in Physics and Mathematics Associate Professor; Department of Mathematical and Hardware Support of Information Systems; I.N. Ulyanov Chuvash State University 428015, Russia, Republic of Chuvashia, Cheboksary, Moskovsky str., 15, office B-304 mitrofanova_tv@mail.ru Smirnova Tatiana Nikolaevna ORCID: 0000-0001-6687-9415 PhD in Physics and Mathematics Associate Professor; Department of Mathematical and Hardware Support of Information Systems; I.N. Ulyanov Chuvash State University 428015, Russia, Republic of Chuvashia, Cheboksary, Moskovsky ave., 15, office B-304 smirnova-tanechka@yandex.ru Other publications by this author

Khristoforova Anastasiia Vladimirovna ORCID: 0000-0003-3534-8747 PhD in Physics and Mathematics Associate Professor; Department of Mathematical and Hardware Support of Information Systems; I.N. Ulyanov Chuvash State University 428015, Russia, Republic of Chuvashia, Cheboksary, Moskovsky str., 15, office B-304 dlya.nastenki@mail.ru

Introduction

In the Russian Federation, much attention is paid to the development of end-to-end digital technologies, including artificial intelligence technologies. Digital transformation is one of the five national goals of the Russian Federation for the period up to 2030, defined by Decree of the President of the Russian Federation dated July 21, 2020 No. 474 "On the National Development Goals of the Russian Federation for the period up to 2030". For this reason, Decree of the President of the Russian Federation No. 490 dated October 10, 2019 "On the development of Artificial Intelligence in the Russian Federation" approved the National Strategy for the Development of Artificial Intelligence for the period up to 2030. As part of the implementation of the National Strategy in 2019, the federal Artificial Intelligence project was developed and approved, which aims to create conditions for enterprises and citizens to use products and services created using mainly domestic artificial intelligence technologies that provide a qualitatively new level of efficiency.

Currently, solving various problems based on artificial intelligence systems is a common topic and is used in various fields – in information technology ^[1], finance ^[2], sports ^[3].

Many interesting works are related to agriculture. The article ^[4] describes the developed models for estimating the yield of winter wheat using multiple linear regression and a neural network. The issues of tracking the behavior of farm animals using neural networks are reflected in ^[5], the classification of porcini mushrooms in order to automate the sorting process based on a neural network model is ^[6].

Many works are devoted to helping and supporting people suffering from depression. For example, the article ^[7] describes the process of creating an effective audio-based automatic depression detection system using a convolutional neural network. The use of deep learning to determine stress levels based on continuous audio signals in the database is reflected in ^[8].

The issue of improving the generation of knowledge-based dialogues through a two-stage selection of knowledge and a network of pointers focused on the choice of knowledge was addressed by the authors of ^[9]. The article ^[10] is devoted to the development and application of an interactive information system in college personnel management.

The authors of this article focused on subparagraph (c) of paragraph 51.5 of the National Strategy, which states that the main directions for improving the level of competence in the field of artificial intelligence and the level of awareness of citizens about artificial intelligence technologies are "the development of skills in using artificial intelligence technologies among graduates of educational institutions of higher education through the inclusion of artificial intelligence modules in each educational program." the program (taking into account the specifics related to industry affiliation and training areas)".

The problem that this study is aimed at solving is the lack of adapted techniques that allow students who do not have prior training in this field to teach the basics of AI application. Standard approaches to teaching artificial intelligence and machine learning assume that students have in-depth knowledge in mathematics, programming, and modeling. However, students in non-core fields need a methodology that allows them to master AI tools in stages and in a practice-oriented format that is understandable and accessible.

This article is devoted to the methodology of conducting laboratory and practical classes, as well as organizing independent work of students of 1-2 bachelor's and specialist's courses in areas of training not directly related to AI ^[11].

The scientific novelty lies in the development of a methodology for using the available Neural Network Wizard and Loginom Community platforms to solve approximation and classification problems. The peculiarity of the methodology is the adaptation of the approach for teaching students whose fields of study are not related to artificial intelligence. This contributes to the development of skills in working with artificial intelligence, which is important in the context of digital transformation and increased attention to end-to-end digital technologies. The proposed assignments allow students studying fundamental and applied sciences to master the use of artificial intelligence methods for data analysis. Use cases include tasks such as forecasting, data analysis in business and economics, performance evaluation, and research related to processing large amounts of data. The proposed approach allows AI training in the format of practical laboratory work and independent development of the principles of building and applying models for classification and approximation, which contributes to the expansion of professional skills and meets the challenges of the digital economy. Within the framework of the research, the following were developed: learning goals at each stage, principles of interaction with platforms and methodological recommendations, which together provide a logical, step-by-step system for mastering AI tools.

The main purpose of the methodology is to teach students the basics of using neural networks and machine learning algorithms using examples of specific tasks: approximation (on the Neural Network Wizard platform) and classification (on the Loginom Community). Tasks include:

• mastering the skills of setting up and training neural networks;

• consolidation of practical knowledge based on the performance of typical tasks;

•Development of data analysis skills and estimation of model accuracy.

Expected learning outcomes include:

1. Mastering the skills of building neural networks. Students will be able to create and configure simple neural networks for approximation and classification tasks using available platforms.

2. Gain experience working with AI tools. Students will be able to prepare data for analysis, adjust model parameters, and interpret the results.

3. The ability to analyze and adjust models. Based on error analysis, students will learn how to adjust model parameters, add training data, and improve forecast accuracy.

4. The application of knowledge to solve applied problems. Students will be able to apply the basic principles of machine learning to solve specific applied problems in their future professional activities, such as forecasting indicators or analyzing economic data.

Description of the technique of using Neural Network Wizard for function approximation

Function approximation is a basic example of using AI in various tasks. The methodology assumes that students create a simple neural network architecture aimed at approximating functions with a pre-prepared training sample. In laboratory conditions, a training sample is formed by students for an example of calculating the product of two multipliers.

The learning process is represented by the following stages:

1. Data preparation. Creating a text file with triples of numbers (two multipliers and their product).

2. Configuring the neural network. On the Neural Network Wizard platform, students build a network architecture consisting of one or more layers.

3. Learning process and interpretation of results. The network is trained using the example of a product of numbers, the accuracy of calculations and ways to improve it are analyzed (additional training, increasing the sample size, changing network parameters).

Function approximation using neural networks

Function approximation is a common application of artificial neural networks. In ^[12], a theorem was proved according to which, using linear operations and cascade coupling, it is possible to obtain from an arbitrary nonlinear element a device that calculates any continuous function with some predetermined accuracy.

It is proposed to start studying the material in Neural Network Wizard, an environment for visual modeling of the architecture of a convolutional neural network and its training.

The use of Neural Network Wizard is reflected, for example, in works for short-term weather forecasting ^[13], modeling prices for passenger cars ^[14], building models of the smartphone market ^[15], forecasting some currency quotes on the Forex market ^[16].

First, we consider an example of an approximation of a function designed to calculate the product of two factors. A txt file is uploaded to the Neural Network Wizard environment with a training sample containing triples of numbers – two multipliers and a product (Fig. 1). We will train the neural network only on this sample. When substituting pairs of numbers that are not included in this set, the error can be significant.

Fig. 1. The file with the training sample

Fig. 2. The result of multiplication by a trained network

As can be seen from Fig. 2, the accuracy of calculations is quite high – the deviation from the correct answer is less than one percent.

Ways to increase the accuracy of calculations:

1) go back one step and train the network additionally;

2) change the training parameters of the neural network used;

3) increase the training sample.

The tasks require building networks with one, two, or three neurons on a single inner layer and training them and evaluating the accuracy of calculations.

Description of the methodology for using Loginom Community for classification tasks

Loginom Community is used to solve classification problems based on data analysis. Students are offered the task of determining the profitability of enterprises based on historical data. The initial data includes the values of two indicators that divide enterprises into profitable and unprofitable groups.

Task completion stages:

1. Create a project. Students upload the initial data and configure the Neural Network (Regression) component.

2. Network training. The settings specify the number of restarts (up to 1000), and the selection parameters are configured, which allows you to automate the learning process.

3. Visualization and analysis of the results. Students use visualizers to evaluate the quality of neural network prediction, analyze errors, and form conclusions about classification accuracy.

Application of the Loginom Community platform for function approximation

The Loginom platform (developed by Loginom Company) is included in the register of Russian programs and is designed for advanced analytics by business users. For educational purposes, we have reviewed the desktop edition for non-commercial use of Loginom Community, which can be downloaded for free. Visual builder allows you to customize the entire analysis process.: integration, data preparation, modeling, and visualization. As a result, the time between hypothesis testing and creating a working business process will be reduced.

The Loginom Community platform is used to solve a variety of tasks. For example, processing large amounts of data when solving transport problems ^[17], analyzing the product range ^[18], developing a social rating model ^[19], and analyzing the store's customer base ^[20].

To train the neural network, we will use the initial training sample (Fig. 1). Let's add the Neural Network (regression) component to the project (Fig. 3). Specify the number of neural network training restarts equal to 1000. Let's pay attention to the settings of the "Setting up automatic selection of Neural network parameters" window. As the complexity of the dataset increases, you will have to use them to find the optimal neural network and its settings. Next, you need to train the neural network using the context menu item "Retrain node". The learning process can be lengthy on computers with low performance.

Fig. 3. Adding the "Neural Network (regression)" component

Fig. 4. Analysis of the results

Next, it is necessary to analyze the effectiveness of the Neural Network (Regression) component training using a visualizer. It is possible to add up to five different neural network output visualizers ("Diagram", "Data Quality", "Cube", "Statistics", "Table"), based on which you can analyze the quality of neural network training.

The "Table" visualizer was selected, which displays the original training sample (columns x, y, res), as well as the forecast of the neural network (column rez|Forecast). The analysis showed that more than 90% of the predictions are correct, but there are also erroneous ones (Fig. 4). This result indicates either errors in the training sample or insufficient training of the neural network. You can improve the quality of the neural network by adding more elements to the training sample, as well as deleting rows containing incorrect values of the training sample.

Approximation of two functions using a single neural network

Let's consider a training example for performing two operations – addition and multiplication of numbers x and y on the interval ^{[0; 10]}. Since two different formulas are planned to be used in the training sample, unlike the previous task, it is necessary to consider not two, but three input variables (x, y, operation sign). Using the programming tools (Fig. 5), fill in the Excel spreadsheet containing the training sample (Fig. 6).

Fig. 5. A fragment of the program code

Fig. 6. Fragment of the training sample

Let's create a new project in the Loginom Community. Let's connect the output of the "Excel file" component with the "Neural Network (regression)" component. Let's configure the neural network according to Fig. 7. Set partitioning parameters: training – 80%, test – 20%. The number of restarts is 5,000. Next, we will check the operation of the neural network on an arbitrary sample (Fig. 8).

Fig. 7. Fragment of the settings window

Fig. 8. Fragment of the visualizer

The same task needs to be solved using Neural Network Wizard.

As an independent work, it is proposed to solve the following tasks using the Loginom Community and Neural Network Wizard software products:

1. Finalize the project for adding and multiplying two real numbers on the segment ^{[0; 10]} (or on an arbitrary segment).

2. Create and train a neural network to perform calculations using four formulas (Fig. 9). To create a training sample, use the interval of values x, y, z on the segment ^{[0; 10]}. The calculation error should be no more than 1-2%.

Fig. 9. Fragment of the task options table

Classification analysis using neural networks

The study of this issue is devoted to independent work, as a result of which students must learn how to use neural networks to solve classification problems.

Here is one of the options for the formulation of the assignment. Based on the results of the previous year, the values of two indicators are known and two groups of enterprises have been identified: group A – profitable, group B – unprofitable. Several enterprises have fulfilled their development forecast for the current year and presented it in a similar form (two indicators, Fig. 10). Using the values of the previous year as a training sample, classify enterprises based on profitability using the Loginom Community and Neural Network Wizard software products.

Fig. 10. Fragment of the table with the initial data

Conclusion

The study was conducted with students of 1-2 courses of bachelor's degree and specialty of UGSN 09.00.00, 10.00.00, 15.00.00. During the implementation of the study, the principles of consistent presentation of the material and "from simple to complex" were observed. The results of the control measures showed that the skills of using artificial intelligence technologies were formed at a sufficient level among students of training areas not related to the field of artificial intelligence.

The author's methodology provides students with the opportunity to learn the basic principles of working with neural networks and AI tools adapted to educational needs. In contrast to existing approaches, the proposed methodology:

1. It is focused on non-core specialties and is designed for students who do not have deep knowledge in mathematics and programming.

2.It includes step-by-step training with a transition from simple to more complex tasks, which reduces the entry threshold and allows you to develop data skills at a convenient pace.

3. Offers practice-oriented assignments with an emphasis on specific approximation and classification tasks, which enables students to master the applicability of AI methods using real-world examples.

Thus, the author's methodology closes a gap in educational practice by providing teachers and students with a simple and effective tool for initial development of data analysis and AI methods, which is of practical importance for students of a wide range of training areas.

The results obtained can be used to build practice-oriented training courses and tasks in various fields not directly related to artificial intelligence, which require knowledge of basic data analysis methods and artificial intelligence algorithms.

The proposed approach allows for AI training in the format of practical laboratory work and independent development of the principles of building and applying models for classification and approximation, which contributes to the expansion of professional skills and meets the challenges of the digital economy. Within the framework of the research, the following were developed: learning goals at each stage, principles of interaction with platforms and methodological recommendations, which together provide a logical, step-by-step system for mastering AI tools.

The main purpose of the methodology is to teach students the basics of using neural networks and machine learning algorithms using examples of specific tasks: approximation (on the Neural Network Wizard platform) and classification (on the Loginom Community). Tasks include:

• mastering the skills of setting up and training neural networks;
• consolidation of practical knowledge based on the performance of typical tasks;
• develop data analysis skills and model accuracy assessment.

Expected learning outcomes include:

1. Mastering the skills of building neural networks. Students will be able to create and configure simple neural networks for approximation and classification tasks using available platforms.
2. Gain experience working with AI tools. Students will be able to prepare data for analysis, adjust model parameters, and interpret the results.
3. The ability to analyze and adjust models. Based on error analysis, students will learn how to adjust model parameters, add training data, and improve forecast accuracy.
4. The application of knowledge to solve applied problems. Students will be able to apply the basic principles of machine learning to solve specific applied problems in their future professional activities, such as forecasting indicators or analyzing economic data.