Introduction

Information communication technologies (ICTs), especially the internet and the web, have changed every aspect of human life. These aspects range from individual social life to health outcomes and from the modernization of industry to the economic growth of nations [1, 2]. Owing to ICTs, many people have become capable of efficiently communicating with others, even in a noncontact environment, and easily acquiring various types of information. Despite the prospect of ICTs for enhancing the everyday lives of people, the inaccessibility of ICTs has resulted in a significant gap between those who can access, use, and benefit from these interventions and those who cannot [3–5]. This gap is now emerging as a new type of inequality in the era of ICTs [6–8]. In particular, as elderly individuals fail to keep pace with the development of ICTs and become alienated in various daily living activities, including economic, social and cultural activities, a digital divide between the older and younger generations of society emerges.

The specific discussion about the digital divide is known to have begun in 1995 when the National Telecommunications and Information Administration (NTIA) first mentioned the ‘digital divide’ in its report. Since then, this term has been defined in various ways. [9] defined the digital divide as the inequality that enhances the economic and social gaps between those who have access to new ITs and those who do not have such access. [10] defined the digital divide as the gap between individuals, households, businesses and geographic areas with regard to both those who can access or who use ICTs and those who cannot access or use ICTs. Similar to these examples, the early discussion on the digital divide was focused on physical access to ICTs [8]. Since then, physical access has increased due to the development of internet speed and the spread of ICTs, but the digital divide among individuals still remains. Therefore, arguments have been made that improving physical access to ICTs may not resolve the digital divide [6, 11], and the causes of the digital divide have been extended from physical access to multidimensional causes [12].

Most previous studies on the digital divide have focused on identifying its causes by using regression analysis or structural equation modeling. [13] argued that the factors of the digital divide are technological access, autonomy, social support, skill and types of uses. [14] reported that the accumulation of capital depends on the purpose and method of using digital information, which are considered factors in the digital divide. [15] showed that access to the internet and interest in the usage of information are higher in elderly individuals who have a higher education level and a higher income level, so a digital divide can occur even among older people. [16] conducted a study with older people and reported that the major factors of the digital divide included age, educational level, and the recognition of the need to be closer to family members. [17] argued that the major factors in the digital divide are demographic variables, including socioeconomic status, education and age. [18] showed that despite the availability of abundant information through the internet, a digital divide can occur among individuals depending on their purpose and method of using the internet and their smart devices. [19] conducted a study with older people and identified their education, income, interests in technology, computer usage before their retirement and social support as the major factors in the digital divide. [20] reported that age, education, income and digital device experience may be the major factors in the digital divide.

As described above, previous studies on the digital divide suggested that demographic and socioeconomic characteristics are the major factors [21]. One notable point is that older people are frequently mentioned as one of the social classes that experience a digital divide. According to the ‘2020 Report on Digital Information Divide Survey’ by the National Information Society Agency, the digitalization level (access, capability and usage) of older people is 68.6% and that of the age group of 70 years or higher is 38.8%, which is the lowest among the information vulnerable classes (disabled people, lower-income people, farmers, fishermen, women, and marriage immigrants). As Korea is expected to be a ‘superaged society’ in 2026 [22], the digital divide of older people is becoming more serious. In particular, as ICTs have penetrated many parts of society since the outbreak of the COVID-19 pandemic, being incapable of using ICTs not only causes inequality but also threatens one’s survival. Previous studies may have considered older people as composing a social group that undergoes a digital divide due to their limitations in vision, decreased cognitive ability and decreased social relationships [23]. This viewpoint is supported by the fact that studies have been conducted to investigate the status and seriousness of the digital divide of older people [24] to identify the relationship between the digital divide and life satisfaction [25, 26] and to explore the factors of the digital divide [19, 27]. In addition, a digital divide can occur even among older people [28, 29] because the digitalization level of elderly individuals can differ depending on their educational level, access to the internet and smart devices and interest in ICTs [15].

According to [30], people over the age of 70 in Korea have difficulty accessing digital devices, such as personal computers and smartphones. According to the authors’ survey, the smartphone possession rate of elderly individuals aged 70 years or older in Korea is 44.9%, which is much lower than that of the whole Korean population (92.3%). The low access rate of elderly people to digital devices represents the severe alienation of the elderly group over the age of 70 years. In addition, the digitalization level of those over the age of 70 years is 38.8%, which is considerably lower than that of 60–69-year-olds (78.8%), who are also considered elderly [30]. Individuals in their 60s now have consumed various media as adults since 2010, when the ICT era was advocated and Web 2.0 emerged. Therefore, this generation is often considered ‘young seniors’ rather than ‘elderly people’ because many in this generation are still active members of the community. The smartphone possession rate of those in their 60s is as high as 89.7%, which is close to that of the whole Korean population (92.3%) [30]. The percentage of 60–69-year-olds who enjoy online shopping is also very high. According to [31], the amount of online shopping by those in their 60 s increased by 171% from 2014 to 2019. Fifty to fifty-nine-year-old baby boomers, who will soon be referred to as the ‘silver generation,’ have a smartphone possession rate of 98.8% [30], which is the highest among older people. Individuals in their 50s are ranked third among all the generations who use YouTube, and the number of silver creators is continuously increasing [31], thus indicating that this generation has a high digitalization level. Therefore, the generations comprising those over the age of 50 include both the generation that has difficult access to digital devices, such as computers and smartphones, and the generation that can use these devices on command.

Considering the recent noncontact environment, elderly individuals who are undergoing a digital divide are more likely to experience difficulties in daily living. Therefore, it is necessary to screen and assist these individuals. However, most of the previous studies on the digital divide of older people have focused on the factors of the digital divide, and few studies have been conducted on how well the identified factors predict the actual occurrence of the digital divide of older people and how to assist these individuals. Studies on how to address the digital divide of elderly individuals are necessary, considering that this issue has become more serious since COVID-19 and that it is more difficult for older people to benefit from online services due to their low digitalization level [32]. Although the digital divide can occur even among older people, previous studies have focused only on the digital divide between the general population and elderly individuals and on identifying the factors related to the digital divide. However, since the noncontact environment is applied to daily living worldwide, more studies need to be conducted on how to rapidly screen those who are experiencing the digital divide and how to predict those who may undergo the digital divide.

This study was conducted to provide a method for predicting the digital divide of older people by establishing digital divide models based on public data and machine learning and performing relevant analyses. To improve prediction accuracy while following the traditional research methodologies of previous studies, this study was conducted by using logistic regression analysis as the classification method for the prediction model, XGBoost, which is a boosting method evolved from the decision tree, and CNN, which is an artificial neural network. This study presents the variables significant in predicting the people who will experience the digital divide and the considerations needed to address these variables. The highlights of this paper are as follows.

• This study confirms the existence of a digital divide, even among elderly individuals, and proposes a method for making predictions through machine learning techniques.
• Important variables in predicting the digital divide within the elderly population are the demographic perspective, internet usage perspective, self-efficacy perspective, and social connection perspective.
• CNNs, which are widely used in fields such as images and videos, have been confirmed to be effective in predicting tabular data, such as that used in this study.

Theoretical framework

Use of machine learning

Contrary to multiple linear regression, machine learning methods in artificial intelligence (AI) are increasingly used for prediction-related research [38, 44–50]. Machine learning is one approach to implement artificial intelligence, where computers discover new rules and patterns or make predictions about new data through data learning using algorithms [51]. Machine learning approaches can be mainly divided into supervised learning and unsupervised learning. Supervised learning is a method of labeling data and predicting future results. Supervised learning approaches can be divided into regression and classification tasks according to the characteristics of the results. In unsupervised learning, hidden patterns or structures in data are found by providing data without labels. Unsupervised learning approaches include clustering and principal components analysis. In this study, classification techniques were used in supervised learning to classify given data based on discrete expectations for the digital divide. Using the classification technique, a model to distinguish data was built using different data dimensions according to specific criteria to predict discrete results for new data [52].

The advantage of machine learning is the ability to use both categorical and numerical predictors to generate models by assessing linear and nonlinear relationships between variables and the importance of each predictor. In regression analysis, which is a traditional statistical method, when many variables are used simultaneously, the basic assumptions about independent variables, such as exogeneity and homoscedasticity, are difficult to maintain, and the high correlation between variables can cause multicollinearity [53]. In contrast, for the analysis of the accuracy of the machine learning-based prediction model, it is assumed that dependent and independent variables are associated with each other. In addition, the roles that dependent variables play in predicting independent variables are analyzed, so the predictive power is unaffected even when multicollinearity occurs [54]. Therefore, an analysis can be performed even in the presence of many variables. These machine learning classification algorithms include logistic regression (LR), support vector machine (SVM), K-nearest neighbor (KNN), decision tree (DT) ensembles, and eXtreme gradient boosting (XGBoost), convolutional neural network (CNN) [55].

The LR is an analysis technique to prove the causal relationship between the independent variable and the dependent variable. Here, the form of the dependent variable is categorical data; when there are two categories, the variable is classified as a dichotomous variable, and when there are more than three categories, the variable is classified as a polynomial variable. Categorical variables are used to solve various classification problems. The formula is as follows.

(1)

The SVM is a method of separating data by finding optimal boundaries in three-dimensional space [56]. These boundaries are referred to as hyperplane boundaries, and given a set of data belonging to either category, the SVM is used to determine the category of the new data. The SVM is used to solve classification problems, such as pattern recognition. The formula is as follows.

(2)

The KNN is an algorithm that looks at the closest k data around given data and determines the group to which the data belong. The distance is measured in terms of the Euclidean distance. Compared to other classifiers, the KNN is simple and easy to implement, and the training speed is fast. The formula is as follows.

(3)

The DT is a commonly used data mining method for establishing classification systems based on multiple covariates or for developing prediction algorithms for a dependent variable. The inference rule is similar to a tree shape, so the decision-making process can be visually and clearly determined, and the DT is used to solve various classification problems.

XGBoost, a model developed by improving the boosting method of a decision tree, has an internal function to regularize overfitting, and internal cross-validation is performed for each trial [57]. Due to its excellent classification performance, XGBoost is often used in competitions such as Kaggle. Above all, the greatest advantage of XGBoost is its high practical usefulness. XGBoost allows for the derivation of important indices, which indicate relatively more important variables among various independent variables, so that the relative predictive power of various independent variables can be reviewed. Therefore, XGBoost was used in this study. The formula is as follows.

(4)

CNNs have emerged as a solution to the problems (learning time, network size, and number of variables) of conventional multilayered neural networks (MNNs). Despite the simple structure of MNNs, MNN-based machine learning requires many data and an excessively long learning time; these problems can be addressed by using CNNs. Concerning previous studies [58], conducted a study on sales forecasting by using tabular data, such as e-commerce transaction history, and the highest prediction accuracy among all classifiers used in the study (ARIMA, FE+GBRT, DNN, and CNN) was achieved using the CNN. [59] researched stock price prediction by using tabular data, such as stock trading volume, closing price, and market price, and the highest prediction accuracy among all classifiers used in the study (MLP, RNN, LSTM, and CNN) was achieved using the CNN. [60] conducted a study on predicting treatment behavior in patients by using tabular data, such as patient information, and the highest prediction accuracy among all classifiers used in the study (ANN, LR, SVM, DT, RFT, CNN) was achieved using the CNN.

These machine learning techniques have been used in studies for credit card fraud detection, student satisfaction prediction, cyberbullying detection model construction, and youth suicide risk prediction. [61] developed a predictive model for credit card fraud detection using public data on credit card transaction records and machine learning algorithms (LR, naïve Bayes, KNN). The analysis results proved that the highest prediction accuracy was achieved using the KNN algorithm. [62] developed a youth suicide risk prediction model using public data from the Korean adolescent risk behavior survey and machine learning algorithms (LR, RF, SVM, ANN, XGBoost). The highest prediction accuracy was achieved using XGBoost. In the study of [63], a cyberbullying detection model was developed using Twitter data and machine learning algorithms (Naïve Bayes, KNN, DT, RF, SVM), and the highest prediction accuracy was achieved using the SVM. [64] developed a student satisfaction prediction model using traditional regression analysis and machine learning algorithms (KNN, SVM, Light GBM, RF, ENet), and the highest prediction accuracy was achieved using ENet. Recently [65], highlighted the frequent use of machine learning techniques for data mining, including LR, SVM, DT, and ANN. Therefore, using machine learning algorithms to solve the digital divide can be a future exploratory direction.

Machine learning techniques have gradually been used in digital divide research. Among relevant studies [66], compared the results of logistic regression and classification tree techniques to analyze the individual level of ICT adoption (computers, internet, and mobile phones), determining that ICT adoption was mainly influenced by income, computer and internet skills, and age. The authors preferred the classification trees technique because it outperformed logistic regression with a more parsimonious model. Specifically, approximately half as many variables were included in the classification tree model as in the logistic regression model. Therefore, the classification tree model is recommended for classifying an individual as an ICT adopter or nonadopter. Similarly [67], used a technique derived from the C4.5 algorithm to describe similarities and differences among a series of municipality classes that present different percentages of internet presence in households. For the internet presence analysis, some classification rules describing cities’ digital divide profiles were generated. Additionally [68], used public data from the Digital Skill Survey of the Spanish National Institute of Statistics and a decision tree to establish a model for predicting the digital divide and performed relevant analyses. The authors identified educational level, age, occupation and household income as the major factors in the digital divide and reported that retirees and homemakers are affected by the digital divide.

Most previous studies employed only traditional logistic regression analysis or decision trees. However, as various classification algorithms have emerged through the development of artificial intelligence (AI), other classification methods can be added to traditional methods to increase the accuracy of prediction models. Therefore, in this study, we attempted to achieve higher prediction accuracy by using SVM, KNN, XGBoost, and CNN, as well as LR and DT, which were used in previous studies.

Additionally, given the different structures and natures of datasets, including the number of variables, dimensionality, and cardinality of predictors, that can substantially influence the accuracy of each algorithm, there is no best machine learning or statistical method for prediction accuracy [38, 44, 47–50, 69]. Although previous studies have shown machine learning algorithms to outperform multiple linear regressions, especially in handling complicated models or datasets with high complexity, most machine learning methods are considered black boxes and are uninterpretable [70–72]. Consequently, there is a controversial tradeoff between prediction accuracy and a model’s interpretability for making decisions using simple and transparent models such as multiple linear regression or potentially more accurate but complicated black box machine learning models. In social sciences, prediction is important, but interpreting the results is also critical. Therefore, to increase the convenience of model analysis while using many independent variables, this study was performed by extracting the variables that are highly important in predicting the ‘digitalization level’. For variable extraction, XGBoost’s variable importance calculation algorithm was used (more details are introduced in the methodology below). The results of this study can be used as basic data for policy establishment to solve the digital divide as a social problem.

Materials and methods

This study aims to predict the digital divide of the older population. Therefore, this study was conducted in three stages: data preprocessing, data analysis, and interpretation of the results, as shown in Fig 1 below.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Research process.

https://doi.org/10.1371/journal.pone.0281291.g001

Data

The purpose of this study was to predict the digital divide of elderly individuals. The data from the ‘2020 Report on Digital Information Divide Survey’ by the National Information Society Agency were used in this study. The survey was conducted as a face-to-face survey with 15,000 subjects in Korea for 3 months from September to December 2022. The questionnaire, shown in Table 1, included questions about the level of access to digital information, the level of digitalization capability, the level of digital information utilization, internet usage, the attitude toward using digital devices, and the change in internet usage due to COVID-19. The data included 2,300 samples, and the demographic characteristics are shown in Table 2. We also collected 2019 data for comparison between post coronavirus and precoronavirus. The 2019 data included a total of 15,000 samples and 2,300 elderly people.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Measurement items.

https://doi.org/10.1371/journal.pone.0281291.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Demographic characteristics.

https://doi.org/10.1371/journal.pone.0281291.t002

Dependent variable

In this study, a binary dependent variable, whose values are 0 and 1, was prepared to predict the digital divide within the elderly population. Eighteen questions were used to measure the level of access to digital information, the level of digitalization capability and the level of digital information utilization, as shown in Table 1, to prepare the dependent variable. All the questions had a 4-point scale (strongly disagree; disagree; agree; strongly agree). The average score of the 18 questions was calculated (average: 2.0000749), and those respondents whose score was lower than the average were classified as a group with a low digitalization level and assigned a value of zero (0) (1,193, 51.87%). Those respondents whose score was higher than the average were classified as a group with a high digitalization level and assigned a value of one (1) (1,107, 48.13%). Those respondents who were assigned a value of zero are affected by the digital divide. The dependent variable was named the ‘digitalization level,’ as used by the National Information Society Agency, to conceptualize the 18 questions. Fig 2 shows the ratio of the dependent variable.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Digitalization level.

https://doi.org/10.1371/journal.pone.0281291.g002

Prediction model analysis

In this study, LR, SVM, KNN, DT, XGBoost, and CNN were used as classification algorithms. Since this study attempted to compare the accuracy of classifiers, the parameter values of each classifier were set and analyzed as default values for a fair comparison. In the CNN, the Conv1D layer of the Keras open source library was used. A total of three layers were formed; the filter had structures of 10, 20, and 10; and the kernel size was 1. The rectified linear unit (ReLU) was used as the active function of each layer, and the sigmoid function was used as the active function of the last output dense layer. All other parameters used default values. In this study, 70% of the raw data were used as training data, and 30% were used as the test data. After training the prediction model with the training dataset, the accuracy of the prediction model was analyzed using the test dataset. The prediction model accuracy indices used in this study were accuracy, precision, recall and the F1-score. The definitions and formulas of the individual indices are shown below.

• Accuracy: Accuracy is the most intuitive performance measure and is simply the ratio of correctly predicted observations to the total observations.

(5)

• Precision: Precision is the ratio of correctly predicted positive observations to the total predicted positive observations.

(6)

• Recall: Recall is the ratio of correctly predicted positive observations to all observations in the actual “yes” class.

(7)

• F1-score: F1 score is the weighted average of precision and recall.

(8)

Result

Independent variables importance analysis

The importance indices (weight, cover, and gain) of the independent variables were calculated by using the ‘plot_importance’ library of XGBoost to derive the variables with high importance in predicting the digital divide of elderly individuals. The independent variables are shown in Fig 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Importance of all independent variables.

https://doi.org/10.1371/journal.pone.0281291.g003

The values of the three importance indices differ between variables. Therefore, only one of the importance indices was selected and used in many previous studies, rather than handling the 3 importance indices at the same time. To overcome this limitation, the variables with high importance in terms of the 3 indices were extracted in this study by applying the ‘rank-based average calculation.’ For example, the 3 indices of the Q7 variable are weight: 3; cover: 1; and gain: 1. Therefore, the average is (3+1+1)/3 = 1.6. A lower average value means a higher rank. In this manner, the top-10-ranked variables were derived and used as the independent variables for predicting the digitalization level. The 10 variables are shown in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Independent variables.

https://doi.org/10.1371/journal.pone.0281291.t003

Prediction model analysis results

The results of the analysis of the prediction model accuracy showed that among the 5 classifiers, the highest accuracy was achieved using the CNN-based model, followed by the XGBoost-based and LR-based models. The lowest accuracy was obtained using the DT-based model. The CNN is often used as a classifier for image type data. However, the results of this study showed that the CNN can be effectively used in classifying structured binary data, such as the data used in this study. Therefore, not only the traditional LR, SVM, KNN, DT and XGBoost methods but also the various artificial neural networks that are usually employed in images and videos can be applied to achieve interdisciplinary convergence and increase the accuracy of the machine learning-based prediction models. The analytical results are shown in Table 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Prediction model analysis results.

https://doi.org/10.1371/journal.pone.0281291.t004

Discussion

This study was conducted to predict the digital divide of older people in the current situation, where digital transformation is accelerated and noncontact environments are widespread. In particular, this study was conducted by using public data and machine learning, which have not been often used in previous studies on the digital divide, to establish digital divide prediction models and perform the relevant analyses. In this study, 10 variables with high importance in predicting the digital divide were derived, and they can be classified into (1) the demographic perspective (Variables 4 and 5), (2) the internet usage perspective (Variables 1, 2 and 7), (3) the self-efficacy perspective (Variables 6 and 8) and (4) the social connectedness perspective (Variables 3, 9 and 10). Why these 4 perspectives are important in predicting the digital divide of elderly individuals is discussed below.

First, the discussion of the digital divide basically focuses on demographic factors, such as age and education [8]. Those who are more advanced in age may have a low ICT utilization capability due to their decreased sensory functions and decreased cognitive abilities [23], and those who are more educated can easily acquire and utilize internet-based ICTs [73, 74]. Therefore, age and education have been argued to be the major factors in the digital divide, and this argument may have been reflected in the results of this study. Therefore, predicting the digital divide of older people requires careful consideration of demographic characteristics, such as age and education.

Second, the finding that a recent change in internet usage is important in predicting the digital divide of older people is related to the time when the survey that provided the data of this study was conducted. In 2020, when the survey was conducted, COVID-19 was spreading nationwide in Korea, thus inducing the application of a noncontact environment to many parts of daily living. Many activities, including business work, social activities and purchasing goods, have been carried out online; these activities can naturally increase internet usage. Those who have low internet usage despite their environment may be affected by a digital divide. Therefore, predicting the digital divide of older people requires careful consideration of the recent change in internet usage.

Third, the finding that confidence in learning and using ICTs is highly important in predicting the digital divide of elderly people suggests that confidence is the manifestation of self-efficacy. Self-efficacy is a user’s confidence that they can carry out a specific task or work by using the system [75, 76]. When people are confident that they have sufficient abilities to learn and use new skills, their resistance to acquiring these skills is decreased [77]. Therefore, self-efficacy is considered important in predicting the digital divide. In particular, according to innovation resistance theory [78], older people strongly resist acquiring and using new skills. However, high self-efficacy facilitates the adoption of new skills [79]. Therefore, predicting the digital divide of older people requires careful consideration of their attitude toward acquiring and using ICTs.

Fourth, the finding that the online social connectedness level and satisfaction are highly important in predicting the digital divide of older people suggests the manifestation of the characteristics of social connectedness. Elderly people tend to make social relationships and interact with others in an offline environment that they are familiar with rather than an online environment that they feel is difficult to access. However, older people usually cannot make new social relationships, as their social relationship formation is diminished due to their loss of social roles and physical and psychological limitations [80, 81]. Particularly in 2020, when the survey was conducted, social distancing was intensified in Korea to prevent COVID-19, thus making it more difficult to form offline-based social relationships with families and relatives. However, according to [55], online activities can help to form social relationships. The internet enables people to communicate with their children, friends and neighbors or make new friends who have common interests through online communities beyond the physical and social limitations of middle-aged or elderly people [82]. Moreover, online social connections allow for interactions with people who have totally different backgrounds regardless of political inclination, religion, sex and age [83]. Hence, the opportunities and activities that the internet environment provides to communicate with various kinds of people can not only enhance social connectedness but also naturally increase the ICT usage level, so such opportunities and activities can be considered factors that can reduce the digital divide. Therefore, predicting the digital divide of older people requires careful the consideration of their online social connectedness level and satisfaction with the connection.

Next, data from 2020 were studied. Since 2020, with the spread of COVID-19, social distancing intensified nationwide in Korea, and the majority of people were forced to use digital devices. Therefore, it was determined to be an important time to study the digital divide. In addition, data from 2019, before the outbreak of COVID-19, were collected and analyzed to compare the differences with the data from 2020. As a result of the analysis, the differences in predictors between 2019 and 2020 were determined to be the social support and social connection perspectives. In 2019, recognition of social support and desire for social connection was important, but in 2020, social support was not revealed as an important factor, and recognition of the degree of social connection was important. These results are recognized as the result of the strengthening of social distancing due to the increase in the nonface-to-face culture in 2020, making the level of actual connection through digital devices more important for survival than individual recognition that there are people who can provide assistance. This suggests that at a time when digital transformation is accelerating in various fields, support for strengthening practical connections through digital devices of the elderly is becoming more important than ever.

Then, the highest prediction accuracy of all studied classification algorithms was achieved using the CNN (accuracy: 80.4%). The CNN basically converts three-dimensional data into one-dimensional vectors and then calculates probability values, so it is recognized as effectively predicting one-dimensional data, as it does in this study. Meanwhile, the second highest accuracy (accuracy: 79%), was achieved using XGBoost, which performs well in international machine learning competitions such as Kaggle. Logistic regression analysis, which is considered a traditional method, ranked third, with good performance (accuracy: 78.3%). Smaller accuracies were obtained using SVM and KNN compared with the top three classification algorithms (CNN, XGBoost, LR), but there was not a significant difference. In contrast, the smallest value among all classification algorithms was observed when using DT. This result is recognized as a result of the characteristic that the decision tree is overfitted to the training data and is weak in predicting new data.

Conclusion

The academic significance of this study is that the CNN, which is often employed in image and video processing, was extended and applied to a social science field using structured data to improve the accuracy of the prediction model. The Korean government announced the Digital New Deal Plan to overcome the economic crisis caused by COVID-19, and one of the major goals of the plan is to resolve the digital divide. The prediction model establishment method and the analytical method proposed in this article can be used by the relevant governmental authorities to classify elderly individuals who are affected by the digital divide. In addition, these methods can be used to predict the 40–49-year-olds who are likely to experience a digital divide as they age so that preemptive actions can be taken. The practical significance of this study is that, as a method for managing individuals who are affected by a digital divide, the self-efficacy perspective about acquiring and using ICTs and the socially connected perspective are suggested in addition to the demographic perspective and the internet usage perspective. This study is limited because it lacks comparisons with other generations because this study was conducted only with elderly individuals as the subjects and specifically focused on the digital divide of elderly individuals. Future studies need to be conducted in which older people are compared with other generations to investigate the patterns of the digital divide in different generations and extend the discussion on the digital divide.

Acknowledgments

The authors would like to thank the anonymous reviewers and editors of the journal for their helpful comments and suggestions.