The year 2020 was a critical moment for the global society, as the COVID-19 pandemic exposed the vulnerabilities and opportunities of various sectors and domains [1,2,3,4,5]. The education sector was one of the most affected by the pandemic, as it had to undergo a rapid digital transformation, a shift to online learning, and a suspension of educational activities [6,7,8,9,10]. The field of economics also faced significant changes, such as the digitalization of processes, the adoption of remote work, and the alteration of service and communication with customers [11,12]. The fast-paced world has become more digital than ever, and the demand for data literacy, data-driven decision making, and data science skills has increased accordingly.
Data science is an interdisciplinary field that uses tools, algorithms, and knowledge of mathematics and statistics to extract insights from data. Data science has a wide range of applications in various domains, such as business, marketing, banking, insurance, medicine, tourism, etc. However, the potential of data science in education has been relatively underexplored, and many opportunities for advancing the field have not been fully exploited.
Data science can be used in education to address scientific problems, such as in the study of behavior in economics, in macro-and microeconomics, marketing, finance, agriculture, environmental and ecological economics, and so on. Data science can also be used to enhance the teaching and learning of economics by providing students with relevant skills and competencies for the modern and technologically advanced society.
Data science has a big list of tools: linear regression, logistic regression, density estimation, confidence interval, test of hypotheses, pattern recognition, clustering, supervised learning, time series, decision trees, Monte-Carlo simulation, naive Bayes, principal component analysis, neural networks, k-means, recommendation engine, collaborative filtering, association rules, scoring engine, segmentation, predictive modeling, graphs, deep learning, game theory, arbitrage, cross-validation, model fitting, etc. Some of these tools were used in the next researches.
Teaching data science, for example, were introduced in [13], Big data and data science methods presented in [14,15,16,17,18,19,20], machine learning used in [21,22,23,24,25,26,27,28,29,30,31,32,33,34,35], Monte Carlo method presented in [36], Artificial Intelligence presented in [37,38,39,40]. Data science is fast developing. A large volume of information that grows with each passing year makes it possible to build high-precision models that simplify and partially automate the decision-making process. Models are being developed that implement the key data science algorithms for different areas of economics: financial data science [41,42,43,44,45,46,47,48,49,50,51,52], for institutional economics - [53,54,55,56,57,58], for agriculture - [59,60,61], for taxation - [62], and labor market - [63].
Data science developing for education discussed in [64,65,66,67,68].
Data science in education is a multidisciplinary approach to technologies, processes, and systems for extract knowledge, understanding of data, and supports decision-making under uncertainty. Data science deals with mathematics, statistics, statistical modeling, signal processing, computer science & programming, database technologies, data modeling, machine learning, natural language processing, predictive analytics, visualization, etc. Data science in education has two aspects of the application: (i) the management and processing of data and (ii) analytical methods for analysis and modeling, and includes nine main steps (figure 1). The first aspect includes data systems and their preparation, including databases facilities, data cleansing, engineering, visualization, monitoring, and reporting. The second aspect includes data analytics data mining, machine learning, text analytics, probability theory, optimization, and visualization. The basis of the learning process is the availability of relevant data that is of sufficient quality, appropriately organized for the task. Primary data often requires pre-processing.
First of all, it is necessary to investigate the availability of the necessary data and how they can be obtained. The data search ends with the creation of a data set in which data coexistence is to be provided. Data science has a wide range of tools for data evaluation and preparation, in particular for data mining, data manipulation (value conversion, data aggregation and reordering, table aggregation, breakdown or merge of values, etc.) and validation of data (checking format, ranges of test values and search in legal values tables). The problem of missing values is solved by using different analytical methods: simulation, inserting default values, statistical simulation. Data science provides broad opportunities for text analytics. In addition, the use of data science tools facilitates work with big data. The main approaches in data science are supervised learning models and unsupervised learning models.
Supervised learning is one of the methods of machine learning, in which the model learns on the basis of labeled data. Using Supervised learning is possible to decide on two types of tasks: regression and classification. The main difference between them is the type of variance that is predicted by the corresponding algorithm. In regression training, it is a continuous variable, in the classification, it is a categorical variable. To solve these problems, many algorithms have been developed. One of the most common is a linear and logistic regression, a decision tree. Linear regression. Regression analysis can be considered as the basis of statistical research. This approach involves a wide range of algorithms for forecasting a dependent variable using one or more factors (independent variables). The advantage of applying such an approach to modeling is the simplicity and clarity of the results, the speed of learning, and the release of the forecast. The disadvantage is not always sufficiently high precision (since in economics and finances, the linear relationship between changes is rare).
Logistic regression is used when it is necessary to predict the release of a binary variable using a dataset of continuous or categorical variables. Situations, where the parent variable has more than 2 possible values, can be simulated by a one-vs-all approach when constructing a logistic classifier for a possible output, or one-vs-one when constructing logistic classifiers for each possible combination of categories of the original variable. The dependence between the independent and the logarithmic variable in logistic regression is linear, the only difference with linear regression is sigmoidal functions, which converts a linear result in the probability of belonging to a class within [0; 1]. The advantages and disadvantages of logistic regression are due to the advantages and disadvantages of linear regression. This is the speed of the algorithm and the possible interpretation of the results, on the one hand, and a little accuracy -on the other. Logistic regression is often used to construct vote-counting models. An important factor in this is the interpretation of its results. The influence of each factor is clearly expressed by the magnitude of the coefficient 𝑏, which allows it to be clearly defined which of them positively and to what extent influence the decision.
A decision tree is an approach to both regression and classification. It is widely used in intelligent data analysis. The decision tree consists of "nodes" and "branches". The tree nodes have attributes that are used to make decisions. In order to make a decision, it is needed to go down to the bottom of the decision tree. The sequence of attributes in a tree, as well as the values that divide the leaves into branches, depends on such parameters as the amount of information or entropy that the attribute adds to the prediction variable. The advantages of decision trees are the simplicity of interpretation, greater accuracy in decision-making simulation compared with regression models, the simplicity of visualization, natural modeling of categorical variables (in regression models it is needed to be coded by artificial variables). However, the decision trees have one significant drawback -low predictive accuracy [69].
Unsupervised learning describes a more complex situation in which, for each observation 𝑖 = 1, ..., 𝑛, observation of the measurement vector 𝑥 𝑖 , but without any variables in the output 𝑦 𝑖 . In such data, the construction of linear or logistic regression models is impossible, since there are no predictive variables. In such a situation, a so-called "blind" analysis is conducted. Such a task belongs to the class of tasks of unsupervised learning, due to the absence of an output variable that guided the analysis. Unsupervised learning algorithms can be divided into algorithms for space reduction and clustering algorithms. The main task of clustering is to find patterns in the data that allow you to divide the data into groups and then in a certain way analyze them and give them an interpretation.
K-means is one of the most popular clustering algorithms, whose main task is to divide 𝑛 observations into 𝑘 clusters. The minimum sum of squares is the distance of each observation to the center of the corresponding cluster. This algorithm is iterative, at each step the cluster centers are re-indexed and redistributed observation between them until a stable result is achieved. The benefits of such an algorithm of clustering are the simplicity, speed, and the ability to process large amounts of data. But the user must specify the number of clusters he wants to use for clustering before computing; the instability of the result (it depends on the initial separation of points between the clusters).
Hierarchical clustering is an alternative approach to clustering, which does not require a preliminary determination of the number of clusters. Moreover, the hierarchical clustering ensures the stability of the result and gives the output an attractive visualization based on the tree-like structure of observations/clusters -dendrogram. This clustering algorithm uses different distance metrics and cluster agglomeration cluster criteria, which makes it very flexible to the data on which clustering is performed. However, the disadvantage of hierarchical clustering is the need to calculate the matrices of the distance between observations before agglomeration, which complicates the application of this algorithm for large data and data with many dimensions.
A time series is built by observations that have been collected with a fixed interval. It could be daily demand, or monthly profit growth rates, number of flights, etc. The time series analysis takes an important part in the analysis of data that covers the region, from the analysis of exchange rates to sales forecasting [70,71]. One of the tasks of time series analysis is the allocation of trend and seasonal components and the construction of the forecast. There are many algorithms that have been developed, and we consider models such as ARIMA and Prophet.
The ARIMA algorithm is one of the most common algorithms for forecasting time series. The basic idea is to use the previous time series values to predict the future. This can use any number of lags, which makes such an approach difficult in setting because it is necessary to select the parameter so as to minimize the error and not override the model. ARIMA is often used for short-term forecasting. A disadvantage is the complexity of learning a model in many seasonal conditions.
Algorithm Prophet was developed by Facebook at the beginning of 2017 for forecasting based on time series [70]. It is based on an additive model in which nonlinear trends are of annual and weekly seasonality. This approach also allows to model holidays and weekends, thereby allowing to predict residuals in a time series. Also, the Prophet is insensitive to missed values, the bias in the trend, and significant residuals, which is an important advantage over ARIMA. Another advantage is the rather high speed of training, as well as the ability to use large-scale time series.
Under the notion of texts mining in natural language we understand the application of methods of texts computer analysis and presentation in order to achieve the quality, which corresponds to the "manual" processing for further usage in various tasks and applications. One of the actual tasks of automatic texts mining is topic modelling.
Topic modelling is a statistical approach to extract the hidden semantics that occurs in a collection of documents or reviews. Latent Dirichlet Allocation (LDA) model proposed by [72] is one of the most notable approach for unsupervised topic modeling, which assumes documents and the words within them are derived from a "generative probabilistic model". Within the class of unsupervised statistical topic models, themes are defined as distributions over a vocabulary of words that represent semantically interpretable "topic" [73]. 'Meaning' of those topics (usually, in the form of topic Label and topic Description) is an emergent quality of the relationship between words [74,75]. The task of topic meaning recognizing is often fraught with difficulty and requires the application of a triangular approach to its implementation, namely: (i) a literature review of existing topics found in the analyzed problem domain; (ii) independent work of experts on assigning labels to topics; (iii) conducting joint expert discussions in order to compare and revise the obtained labelling results.
As for main assumption of LDA method, there are the following [76]: (i) document is represented as a mixture of topics; (ii) each topic are present in many documents; (iii) each word within a given document belonging to exactly one topic; (iv) each document can be represented as a vector of proportions that denote what fraction of the words belong to each topic.
The basic LDA model is shown in figure 2. Figure 2 serves as a visual explanation of the model and could be described as follows: (i) we have 𝐷 documents and 𝐾 topics; (ii) each topic presented by 𝛽 𝑘 words distribution over the vocabulary within the topic 𝑘; (iii) each document is presented by 𝜃 𝑑 topic proportions within the document, where 𝜃 𝑑,𝑘 is the topic proportion for topic 𝑘 in document 𝑑. Finally, we have (iv) for each 𝑛 𝑡ℎ word in the document 𝑑 -topic assignments 𝑧 𝑑,𝑛 (depends on the per-document topic proportions 𝜃 𝑑 ) and (v) for each 𝑑 𝑡ℎ document -observed words 𝑤 𝑑,𝑛 which is an element from the fixed vocabulary (depends on the topic assignment 𝑧 𝑑,𝑛 and all of the topics 𝛽 1:𝑘 ) [77].
In is obviously that data scientist in cooperation with other science domains increasingly seek ways to apply NLP and especially LDA topic modelling techniques to extract, organize, recognize, label and classify customers opinions and experiences [78]. Next examples demonstrate the possibilities to sol the apply LDA topic modelling for solving: (i) human resources management, (ii) service quality assessment, (iii) research & development policy coordination tasks and (iv) strategic planning in universities.
Kobayashi et al. [79] used topic modelling to summarize the worker attributes and find worker attribute constructs and use these to cluster jobs. 140 main topics were identified, and such skills, as, for example, interpersonal communication (vocabulary of words: communication, written, oral, verbal, interpersonal, presentation, effective, listening); analytical and problem-solving (vocabulary of words: problem, solving, analytical, solver, troubleshooting, approach, abilities, capabilities); data analytical skills (vocabulary of words: data, analysis, quantitative, research, statistics, economics, statistical, modeling); willingness to travel and the ability to operate on a flexible work schedule (vocabulary of words: travel, willingness, willing, work, time, needed, internationally, international) and other. As authors mentioned, topic modelling showed that it is not only possible to classify job information from vacancies but that we can also derive behavioral characteristics that are valued or required by employers from potential or existing job holders. Moreover, as a further analysis of this research was planned the analysing trends of worker attributes required by organizations (i) over time, (ii) occupations, companies, and (iii) geographical regions, and also (iv) possibility to build a network of work activities to examine relationship among tasks.
Wallace et al. [80], Sharma et al. [81] captured the main positive and negative words within latent aspects (topics), which characterise interpersonal manner, technical competence, and systems issues 63-81 [82] from online physician reviews. Similar with previous work, James et al. [83] based on López et al. [82] categorization, examined unstructured textual feedback of physicians in order to determine: (i) how the extracted sentiment and topics compared to traditional identified dimensions of service quality in healthcare and (ii) what tone and topic elements were driving patients' service quality ratings. As a main finding were the following list of topics and their tone: (1) Negative system quality: Staff and Timeliness (vocabulary of words: office, staff, time, doctor, wait, appointment); (2) Positive interpersonal quality: Physician Compassion (vocabulary of words: doctor, caring, great, knowledgeable, excellent, recommend); (3) Negative system quality: Experience (vocabulary of words: told, don't, doctor, ask, bad, money, call); (4) Positive Technical quality: Family (vocabulary of words: doctor, questions, staff, practice, children, son, pregnancy); (5) Positive Technical quality: Surgery (vocabulary of words: surgery, pain, procedure, staff, hospital, knee, cancer, age); (6) Negative Technical quality: Diagnosis (vocabulary of words: years, treatment, medical, patient, conditions, test, diagnosis, time, treated). The obtained results allowed the authors to establish the dependence on the degree of influence of the identified aspects (topics) on the general perception of the physician's quality, as well as the behavioural characteristics of patients when choosing a doctor online, depending on the content of comments and overall rating.
When conducting research on the basis of textual documents or customers comments, researchers often have a more of information "about the text" than "about the content of the text". From the perspective of topic modelling as a statistical approach, the existence of such information "about the text" (metadata) allows and initiates the inclusion in the model of additional covariates that could influence the following components of the topic model: (1) Proportion of the document devoted to the topic ("prevalence of the topic"). For example, we can know that "clients who buy products online are more likely to talk about delivery problems than clients who buy offline". (2) Word rates used in the discussing of the topic ("topical content"). For example, we can clarify that "when clients talking about delivery problems, clients who buy products online are more likely discuss the problems about products returning, but patients clients who buy offline are more likely discuss staff rudeness issues" [84]. Such possibilities are proposed by Structural topic modelling (STM) as an extension of the LDA framework [74,84,85] .
Drawing analogies with LDA: (i) each document in STM arises as a mixture over 𝐾 topics; (ii) topic proportions (𝜃 𝑑 ) can be correlated (LDA limitation 1); (iii) topics prevalence 𝜃 𝑑 can be influenced by set of covariates 𝑋 through a standard regression model with covariates; (ii) for each 𝑤 𝑛 word in the document 𝑑 (iii) a topic 𝑍 𝑑,𝑛 is drawn from the document-specific distribution, and (iv) conditional on that topic, a word is chosen from a multinomial distribution over words parameterized by 𝛽 𝑑,𝑘,𝑣 , where 𝑘 = 𝑍 𝑑,𝑛 . This distribution can include a second set of covariates 𝑌 [84]. Thus, the main differences between the LSA and STM models (figure 3) are that the prevalence (content) parameters determined in the LDA by the general a priori Dirichlet parameters 𝛼(𝜂) in the STM model are replaced with prior structures specified in the form of generalized linear models parameterized by document specific covariates 𝑋(𝑌 ) [86] These covariates inform either the topic prevalence (covariates 𝑋) or the topical content (covariates 𝑌 ) latent variables with information "about the text" (metadata).
In order to study customer perception of the quality of services, assess their satisfaction with goods or services received, as well as identify factors that influence customer acceptance of new offers on the market, students were asked to use STM tools. As a data source 610 textual comments about hospitals from the site http://www.ratemyhospital.ie/ (over the past two years -2018-2019) were used. STM package allows to use all additional variables to demonstrate the power of meta-data for topic modelling. With this aim, textual comments data was extended by information about (1) hospital ownership (private, public), ( 2) sentiment (positive or negative) (table 1) [87]. After that, all steps of text pre-processing were performed.
Comments before pre-processing.
A lovely friendly patient-focussed hospital Public Positive Consultant I found seriously lacking compassion for my mother the patient. Sniggered while informing us that while my mother's condition is uncomfortable, it is not life threatening.To be frank, consultant spoke down to us.
Tullamore is a very clean hospital and looks very well. All staff I had the pleasure of meeting were lovely and very professional at all times.
The staff in all capacities do not receive enough thanks for the jobs they do
First, the STM model's setup were performed. To determine the optimal number of topics, STM models from 10 till 30 topics were built were analyzed. Semantic coherence is maximized when the most probable words in a given topic frequently co-occur together, and it is a metric that correlates well with a human judgment of topic quality. Having high semantic coherence is relatively easy, though, if we only have a few topics dominated by very common words, so we wanted to look at both semantic coherence and exclusivity of words to topics. So, the most valuable number of topics should be very coherent and also very exclusive. Looking at figure 4, we draw the conclusion that the 15 topics suit the most to these criteria. Most of the topics, in this case, are above the average of exclusivity and have high coherence, especially compared to the other number of topics which are often spread out on both axes. 15-topic STM model was selected based on subjectively optimal combination of the average semantic coherence and exclusivity outcomes.
As a result, for 15-topic model, we received the (i) topic-words distribution 𝛽; (ii) document-topic proportions 𝜃; (iii) list of Highest probability-, FREX-, Lift-and Score-keywords (Highest Prob: are the words within each topic with the highest probability; FREX : are the words that are both frequent and exclusive, identifying words that distinguish topics; Lift: give more weight to words that appear less frequently in other topics by dividing their frequency into other topics; Score: score words are weighted by dividing the log frequency of the word in the topic by the log frequency in other topics [85,88,89]); (iv) set of documents, mostly associated with this topic. The figure 5 allows us to get information on 5) two students agreed on final 15 topic labels and jointly developed the topics descriptions (short summarization of the topic content) [87]. The result of topic labelling is presented in the table 2. Third, the STM covariate analysis could be performed. In this stage, we aimed the evaluating the Sentiment effect on the formation of more positively and more negatively oriented aspects of hospitals service quality (HSQ). Thus, we use Sentiment metadata as Covariate in the STM model. Formally, we can identify an aspect as negative if, according to the results of effect estimation, the proportion of this aspect in negative comments (Sentiment = Negative) is significantly higher than in comments in positive comments (Sentiment = Positive). According to the results of our experiment, 5 topics (33.33%) are positive (right side of figure 6), and 10 topics (66.66%) are negative (left side of figure 6). The dots in the figure 6 indicated the mean values of the estimated proportion differences (power of influence, PI) with 95% confidence intervals, allows us to evaluate the relative degree of influence of sentiment on of hospitals service quality aspects. For example, the five most negative Topic of are (1) Information Exchange with Patient/Family (Topic 13) with highest power of negative influence;
(2) Communication Skills (Topic 2); (3) A&E/Admission (Topic 12), (4) Waiting Time (Topic 4) and ( 5) Patient-Focusing Service (Topic 6). In turn two most positive topics are (1) Service Rapidness (Topic 14);
(2) Personnel Reliability/Treatment (Topic 8). Knowledge about Topics with a positive and negative impact of comments Sentiment allow to indicate the strength of patient satisfaction/dissatisfaction with the hospitals service quality.
Fourth, the power of Time influence on positive and negative Topics dynamics (from 2018 to 2019) using the STM model (with Year and Sentiment as a Covariates) should be performed. In terms of the Influence of the Time Factor on the Service Quality, the following four groups of HSQ Topics can be distinguished: (1) Topics causing the growth of patient satisfaction with the Service Quality over the time: positive topics with a positive dynamic over the time; (2) Topics causing a recession in patient satisfaction with the hospitals service quality (HSQ) over the time: positive topics with a negative dynamic over the time; (3) Topics causing the growth of patient dissatisfaction with the HSQ over the time: negative topics with a positive dynamic over the time (4) Topics causing a recession in patient dissatisfaction with the HSQ over the time: negative topics with a negative dynamic over the time.
As an indicator that allows us to identify the direction and growth rate (GR) of change in the level of positive or negative comments describing the Topic, the slope of the regression (dependence between the proportion of Positive/Negative Aspects and Time) will be used. The presented four charts (figure 7 a, b, c, d As a result, student could see that the largest number of aspects (37.5%) has a negative impact on the HSQ. The highest degree of growth in patient dissatisfaction is characterized by A$E/Waiting Time topic. Moreover, this growth rate is not only the largest in the category of Negative impact, but in all analyzed topics. The most rapid (within the whole set of topics) decrease in the number of positive comments is characterized by the aspect of Maternity Unit/Care. The group of topics on which improvement in their quality is noted is 25.1%. At the same time, the Hospital Environment is characterized by the highest rate of improvement. 16.7% of topics have a positive effect on the HSQ, among which Service Rapidness and Maternity Unit/Treatment have the largest increase in the number of positive comments.
Fifth, students may identify the influencing the Hospital Ownership on more positively and more negatively oriented HSQ aspects structure (using the Sentiment and Hospital Ownership factors as in the Covariates STM model). For this purpose, the following interpretation of the results could be proposed: (1) the Topics, more related to Public Hospital Ownership according to the results of effect estimation, in which the proportion of this Topics in comments about Public hospitals (Hospital Ownership = Public) is significantly higher than in comments about Private hospitals and vice versa; (2) the direction (positive or negative) of Hospital Ownership influencing on HSQ. For reaching the first purpose, the Hospital Ownership effect estimation was performed for revealing the aspects in which the proportion of the comments about Public hospitals (Hospital Ownership = Public) is significantly higher than comments about Private hospitals and vice versa.
For formalization the rules for second purpose reaching, in terms of discovering the Influence of the Hospital Ownership on the Service Quality, the following groups of aspects proposed to be distinguished: approaches which could develop students' technical and research skills in the public perception analysis. Such approaches also allow students to gain practical experience in the study and interpretation the influence of additional metadata, characterizing the comments authors, on differences in their opinions about events, companies, goods, and services.
Classical methods of statistical analysis, modeling methods, and data mining are used in economics. The analysis of data in these areas is aimed at the study of causation. In economics, current issues include policy development, determining the impact of a decision, long-term and short-term planning and forecasting, choosing the best solution from many possible, and many others. Drawing conclusions is also important in economics. In addition, the modern economy and finance are characterized using big data, so it is not always possible to use classical methods. Therefore, the methods of data science are precisely those methods that should be used in economics, which gives positive results and effect. Data science methods were first used in economic research and gradually penetrated into practice. Today, economics need specialists who have knowledge in these areas and are able to apply data science methods. In response to this market need, universities have begun to implement data science courses and programs for students of economics. The table 4 presents the courses and programs of the top 20 universities in the world.
A study programs in economic field in Ukrainian universities has shown that data science courses and programs are still being introduced in Ukraine. Currently, there are separate programs for studying Data Science, mainly for computer science. Therefore, we believe that the prospects that data science opens for modern economists necessitate the introduction of courses and programs in data science.
Data science is a rapidly growing and evolving field that has applications in various domains, such as research, society, and business. Data science requires significant investments and innovations from businesses and governments, as well as adequate education and training for students and professionals. However, as our research has shown, the integration of data science in economics education is still in its infancy. Only a few leading universities offer data science courses and programs for economics students, but this trend has not been widely adopted and needs to be further developed.
As an example of the use of data science methods in economics education, we have demonstrated the application of STM-modeling in teaching students. STM-modeling is a technique that allows analyzing textual data and identifying latent topics based on additional metadata, such as the characteristics of the text authors. STM-modeling can help students develop their technological and research skills, work with big data, and study and interpret the differences in opinions about various topics, such as events, companies, products, and services.
The STM-modeling technique is just one of the many methods and algorithms that can be used for modeling and analyzing economic processes. There are numerous examples of how data science can be applied in economics education, such as using time series analysis to predict the future value of a cryptocurrency, using regression models to determine customer loyalty or the likelihood of customer insolvency, etc. Data science offers a rich set of tools and techniques that can enhance the learning and teaching of economics.
Education should keep pace with the modern development of the digital economy, digital society,
N. P. Volkova) https://pg.edu.pl/b5968c8562_nina.rizun (N. O. Rizun); https://scholar.google.com.ua/citations?user=NkrrNKAAAAAJ (M. V. Nehrey); https://scholar.google.com.ua/citations?user=Y18aS7EAAAAJ (N. P. Volkova) 0000-0002-4343-9713 (N. O. Rizun); 0000-0001-9243-1534 (M. V. Nehrey); 0000-0003-1258-7251 (N. P. Volkova
innovation, and creative entrepreneurship. The use of data science in education should be cross-platform, that is, used not only in the study of specific subjects, but also in the teaching of all subjects, interaction of students with each other and with teachers, real experts, research, and individual learning.