Predicting the Impact of High-Speed Rail on Population Change in Local Cities by using a Naive Bayesian Classification-based Artificial Intelligence Model

Ⅰ. Introduction

The population decrease is the primary concern in most developed countries. Korea, especially, is experiencing a low birth rate, with a total fertility rate of 0.840 in 2020 [1]. Population decline is expected to have a huge ripple effect on all societies and economies. In particular, the decrease in population tends to be concentrated in local small cities rather than in the metropolitan area, which is expected to cause a crisis of urban shrinkage. Local small cities of Korea, in particular, are suffering from the dual effects of natural population decline and population outflow. Although the Seoul Capital Area (Seoul, Gyeonggi, and Incheon) occupies only 11% of the total land area in Korea, the population of Seoul Capital Area accounted for more than 50% of the total population of Korea [2]. In this situation, the Korean government has been suggesting numerous policies that can prevent the shrinkage of local cities. For example, the government has relocated administrative capital from Seoul to Sejong City, and Innovation Cities has been made. As one of these various policies, the Korean government tried to strengthen access to local and metropolitan areas by constructing several high-speed railways. As a result, in addition to KTX (Korea Train eXpress), SRT (Super Rapid Train) has launched in December 2016.

Transportation infrastructure plays an important role as an intervention event in this population movement [3]. High-speed rail, in particular, is known to play an important role in logistics and population movement by influencing socioeconomic conditions in the region and fundamentally altering the concept of accessibility [4][5]. There are two perspectives on the role of high-speed rail in population movement. One is the view that high-speed railroads generate so-called straw effects, accelerating the outflow of local populations to large cities [6][7]. The other is that high-speed rail performs a function of alleviating population concentration in the metropolitan area [8].

Deng et al. (2018) examine population movement by high-speed rail by dividing it into three causes. These include the agglomeration effect, diffusion effect, and siphoning effect. The agglomeration effect refers to the concentrated effect of the population and economy according to the benefits of accumulation due to improved accessibility by high-speed rail. High-speed rail is known to cause considerable agglomeration as an important SOC (Social Overhead Capital) [9]. The diffusion effect refers to the effect of spreading the population and economy along the high-speed railway line. Large-scale transportation infrastructure contributes to regional development linked to routes, as in the case of Gyeongbu Expressway, which has also been proven in the case of KTX [10]. Siphoning effect refers to a phenomenon in which the population and economy move from a relatively depressing place to a thriving place as if absorbed. Most previous works of literature that showed high-speed rail accelerates the outflow of local population point to siphoning effect [4][6][9][11][12].

In other words, the agglomeration effect and the diffusion effect cause regional imbalances between regions where high-speed rail passes and regions where high-speed rail does not pass, and the siphoning effect acts as a driving force for regional imbalances between large cities and small cities. In sum, there have been many previous works of literature on the high-speed railways and the population change; however, the result showed different results depending on the effect. Most of the previous studies in Korea focus on the effect of KTX. Considering the government’s regional rebalancing and a new launch of SRT railway, it is needed to empirically examine the impact of population changes in local cities by high-speed railway. In this situation, a question arises whether this high-speed railway increase or decrease the population in local cities; therefore, in this study, we suggest an analytical model for predicting population growth of local cities using a Naive Bayesian Classification-based Artificial Intelligence model.

Ⅱ. Data and Methodology

2.1 Study Area

The nationwide spatial range is targeted to examine the population movement across the country considering the wide-area transportation network. The spatial range of the research data is 227 districts in South Korea, except for Jeju, which is an island region. Considering the SRT launch year (December 2016), the temporal range of this study was constructed from 2014 to 2019 to examine three years before and after the SRT launch. Also, the relocation of administrative capital to Sejong city was full-fledged in 2014, and the Innovation city project also covers this temporal range. Fig. 1 shows the study area of this study.

Fig. 1.

Study area

2.2 Methodology: Naive Bayesian Classification

In order to predict the outputs, machine learning techniques are trained by data. The types of machine learning are divided into three: supervised learning, unsupervised learning, and reinforcement learning [13][14].

First, the input data of the supervised learning algorithm is provided as a labeled dataset, which shows the correct and incorrect solutions. Second, the unsupervised learning algorithm is not complete and not a clean labeled dataset. The unsupervised learning algorithm aims to explore the patterns and predict the output. Finally, the reinforcement learning algorithm is neither based on supervised nor unsupervised. It learns through the reward or feedback of the environment [15].

There are two types of problem-solving methods of supervised learning: classification and regression [16]. Regarding the classification methods, it helps to predict a discrete value. Naive Bayesian Classification, Support Vector Machines, and Logistic Regression can be examples [17][18].

The Naive Bayesian Classification is based on Bayes Theorem, which states the equation (1). This equation can be regarded as the probability that A will occur when B occurs. In this formula, A and B can be regarded as the output and input variables, respectively [17].

$\[ P\left(A|B\right)=\frac{P\left(B|A\right)\times P\left(A\right)}{P\left(A\right)} \]$

(1)

If more than one condition corresponds to B, it can be represented as the following equation (2).

$\[ \begin{array}{c}P\left({AvertB}_1\cap B_2\cap \cdots B_n\right)=\hfill \\ \frac{P\left(B_1|A\right)\times P\left(B_1|A\right)\times \cdots \times P\left(B_n|A\right)\times P\left(A\right)}{P\left(B_1\right)\times P\left(B_2\right)\times \cdots \times P\left(B_n\right)}\hfill \end{array} \]$

(2)

Regarding equation (2), we set up and propose the research model with the variables of Table 1. PG is assigned as the output variable, and all the other 16 variables are assigned as the input variables for the proposed model.

Table 1.

Variables and their sources

2.5 Research Flow

The research flow of this study is shown in Fig. 2. First, we set up the artificial intelligence model based on Naive Bayesian Classification to predict the population change. Second, according to the proposed research model, we collected the related data. Third, the descriptive statistics was performed with the research data. Forth, we performed Naive Bayesian Classification with the research data. Finally, the proposed research model was validated.

Fig. 2.

Research flow

Ⅲ. Result

3.1 Descriptive Statistic

The descriptive statistic results are shown in Table 2. The total number of records is 1,362, which is 227 multiplied by 6 because the number of districts is 227 and the temporal range is 6 years. Looking into the variables, the average PG is 0.3950, which indicates that 39.5% of the districts had population growth. Other dummy variables such as KTX, SRT, KS, MC, IC, and HI can also be considered in the same way.

Table 2.

Descriptive statistics(n=1,362)

Besides the dummy variables, there are categorized groups found: groups A and B. The variable NOP is included in group A, whose data are widely scattered around the mean because the standard deviation is bigger than the average. The other variables are included in group B.

3.2 Results of Naive Bayesian Classification

According to the variables of Table 1 and equation (2), we performed the Naive Bayesian Classification based on Python and got the results as shown in Table 3. One of the factors affecting the performance of machine learning models is the training data size. If the size of training data is too big, the machine learning model is usually overtrained, which is called overfitting. The overfitting phenomenon comes from the high error rates on test data. On the other hand, if the model cannot know the relationship between the input and output data because of a lack of training data, it could generate high error rates on both the training data and test data. This phenomenon is called underfitting. It is not easy to avoid underfitting and overfitting phenomena when performing the machine learning models by training them and predicting the results. Another factor affecting the performance of a machine learning model depends on the test data for validation.

Table 3.

Validation results of Naive Bayesian classification

In this study, therefore, the machine learning was trained with the default size of data, which is 0.25, and predicted the results with randomly collected test data. The randomly collected test data consists of 400 records, 200 population growth-related data, and 200 population decline or population- no-change-related data. Therefore, we performed and conducted this research model 200 times repeatedly and got the results as shown in Table 3.

First, the accuracy average is 0.6369, which indicates that the proposed research model classified correctly by 63.69%. As mentioned earlier, the results depend on the training and test data. Also, this is an inevitable phenomenon when dealing with artificial intelligence models. Second, the average precision is 0.6645, which indicates that 66.45% of correctly predicted positive observations among the total predicted positive observations by the proposed research model. Third, the average recall is 0.5692. This value is the lowest among the average values of those four validation results: accuracy, precision, recall, and F₁-score. This value depends on the value of FN, as shown in equation (5). Therefore, precision and recall are challenging to complement each other. Finally, the average F1-score is 0.5973.

3.3 Discussion on the Result

Based on the validation result of Naive Bayesian Classification, we visualized the prediction value of the population in 2019 in Fig. 3. As seen in Fig. 3, the Seoul metropolitan area (including Seoul, Gyeonggi, and Incheon) and the nearby districts were predicted to increase the population from the model. Overall, districts with KTX stations showed a prediction of population decrease, while the districts with SRT stations showed a prediction of population growth. In particular, districts near Gangwon showed population decrease prediction even though having the KTX station. It can be interpreted that KTX has been built comparatively for a long time, but the newly launched SRT is more likely to increase in population in areas with SRT stations. In addition, while launching SRT a consideration of the effect of population dispersion in local cities was important, so other regional balancing projects would have been processed in consideration of SRT.

Fig. 3.

Validation result in 227 districts

The districts with both KTX and SRT stations showed higher prediction value on the population growth. However, among the districts having both KTX and SRT stations, districts in Jeolla had a lower prediction value in population growth. The result shows that the straw effect and alleviation effect varied based on the districts. For example, overall, the districts with SRT stations showed agglomeration to the district having SRT stations, while the districts with KTX stations, especially in Gangwon districts, interpreted outflow of the local population to large cities. Also, districts having both KTX and SRT stations was predicted population growth, but the effect was smaller in the districts of Jeolla.

Ⅳ. Conclusion

This study proposed a Naive Bayesian Classification-based artificial intelligence model to predict the population change. Previous studies usually performed and conducted the traditional methodologies such as regression analysis to investigate the population change. However, this study proposed and carried out the prediction of population increase and decrease.

According to the research result, the average values of all the validation indexes are between 0.5692 and 0.6645. This performance indicates that the results can predict social phenomena using artificial intelligence techniques and suggest additional research. Considering the overfitting and underfitting issues, the parameters for the artificial intelligence model can be changed to improve the validation results of the model. Also, a comparative study that can compare the performance between models such as decision tree, SVM, etc., could be delivered as a future study.

Nevertheless, the methodology used in this study is in the form of a black box, making it difficult to determine which factors influenced exactly and how. Based on city characteristics, future studies could reflect these points and expect to gain more detailed implications for population changes due to high-speed rails.

According to the results of this study, the following implications can be explained. First, this study proposed and carried out the prediction of population increase and decrease, which is one of the key topics in the field of social science, by using an artificial intelligence model that can overcome extrapolation. Second, this study conducted an empirical analysis using public open data from the web. Finally, this study investigated the relationship between opening high-speed railways and the increasing or decreasing population. This study is valuable as a preliminary investigation into population growth using artificial intelligence in transportation and urban planning.

References

• Statistics Korea, https://www.index.go.kr/potal/main/EachDtlPageDetail.do?idx_cd=1428, . [accessed: Nov. 30, 2021]
• 2020 Population and Housing Census, https://www.census.go.kr/eng/html/index_en.jsp, . [accessed: Nov. 29, 2021]
• J. M. Coronado, J. M. de Ureña, and J. L. Miralles, "Short-and long-term population and project implications of high-speed rail for served cities: analysis of all served Spanish cities and re-evaluation of Ciudad Real and Puertollano", Eur. Plan. Stud., Vol. 27, No. 3, pp. 434–460, 2019. [https://doi.org/10.1080/09654313.2018.1562652]
• T. Deng, D. Wang, Y. Yang, and H. Yang, "Shrinking cities in growing China: Did high speed rail further aggravate urban shrinkage?", Cities, Vol. 86, pp. 210–219, Mar. 2019. [https://doi.org/10.1016/j.cities.2018.09.017]
• B. H. SUH, G. KIM, C. H. LIM, and H.-K. HA, "Effects of High-speed Rail’s Competition Strategies on Price and Share in the Transportation Market", Journal of Korean Society of Transportation, Vol. 38, No. 2, pp. 97-111, Apr. 2020. [https://doi.org/10.7470/jkst.2020.38.2.097]
• S. W. Lee, J. K. Jeong, W. S. Zhee, and J. K. Cho, "The effects of high speed rail on population distribution", Journal of The Korean Regional Development Association, Vol. 16, No. 1, pp. 119-138, Mar. 2004. https://www.earticle.net/Article/A171974, .
• W. Li, X. Wang, and O. Hilmola, "Does High-Speed Railway Influence Convergence of Urban-Rural Income Gap in China?", Sustainability, Vol. 12, No. 10, pp. 4236, May 2020. [https://doi.org/10.3390/su12104236]
• J. W. Hur, "A critical review on the straw effects of high speed train", Journal of the Korean Urban Management Association, Vol. 23, No. 4, pp. 59–74, Dec. 2010. https://www.dbpia.co.kr/Journal/rticleDetail?nodeId=NODE01583919, .
• T. Hiramatsu, "Job and population location choices and economic scale as effects of high speed rail: Simulation analysis of Shinkansen in Kyushu, Japan", Research in Transportation Economics, Vol. 72, pp. 15–26, Dec. 2018. [https://doi.org/10.1016/j.retrec.2018.06.007]
• J. U. Jo and M. J. Woo, "The impacts of high speed rail on regional economy and balanced development: Focused on Gyeongbu and Gyeongjeon Lines of Korea Train Express (KTX)", Journal of Korea Planning Association, Vol. 49, No. 5, pp. 263–278, Jun. 2014. [https://doi.org/10.17208/jkpa.2014.08.49.5.263]
• J. LEE and Y. YOON, "A Study on the Outliers Detection in the Number of Railway Passengers for the Gyeongbu Line From Seoul to Major Cities Using a Time Series Outlier Detection Technique", Journal of Korean Society of Transportation, Vol. 35, No. 6, pp. 469–480, Dec. 2017. [https://doi.org/10.7470/jkst.2017.35.6.469]
• M. N. Zheng and J. H. Rho, "Empirical analysis for the effect of the inter-regional express rail system (KTX) on the change of the relative dependency between regions in Korea-Focused on five metropolitan cities", Journal of Korea Planning Association, Vol. 50, No. 7, pp. 141–153, Oct. 2015. [https://doi.org/10.17208/jkpa.2015.11.50.7.141]
• H. Hihn and D. A. Braun, "Specialization in hierarchical learning systems", Neural Process. Letters, Vol. 52, No. 3, pp. 2319–2352, Sep. 2020. [https://doi.org/10.1007/s11063-020-10351-3]
• T. Sasakawa, J. Hu, and K. Hirasawa, "A brainlike learning system with supervised, unsupervised, and reinforcement learning", Electrical Engineering in Japan, Vol. 162, No. 1, pp. 32–39, Sep. 2008. [https://doi.org/10.1002/eej.20600]
• N. Xu, "Understanding the reinforcement learning", Journal of Physics: Conference Series, Vol. 1207, No. 1, pp. 12014, Jan. 2019. https://iopscience.iop.org/article/10.1088/1742-6596/1207/1/012014/meta, . [https://doi.org/10.1088/1742-6596/1207/1/012014]
• B. T. Pham, I. Prakash, and D. T. Bui, "Spatial prediction of landslides using a hybrid machine learning approach based on random subspace and classification and regression trees", Geomorphology, Vol. 303, pp. 256–270, Feb. 2018. [https://doi.org/10.1016/j.geomorph.2017.12.008]
• M. Karabatak, "A new classifier for breast cancer detection based on Naïve Bayesian", Measurement, Vol. 72, pp. 32–36, Aug. 2015. [https://doi.org/10.1016/j.measurement.2015.04.028]
• Y. Ma, B. Xu, and X. Xu, "Real estate confidence index based on real estate news", Emerging Markets Finance and Trade, Vol. 54, No. 4, pp. 747–760, Oct. 2017. [https://doi.org/10.1080/1540496X.2016.1232193]
• T. Jiang et al., "Tongue image quality assessment based on a deep convolutional neural network", BMC Medical Informatics and Decision Making, Vol. 21, No. 1, pp. 1–14, May 2021. [https://doi.org/10.1186/s12911-021-01508-8]
• A. Grybauskas, V. Pilinkienė, and A. Stundžienė, "Predictive analytics using Big Data for the real estate market during the COVID-19 pandemic", Journal of Big Data, Vol. 8, No. 1, pp. 1–20, Aug. 2021. [https://doi.org/10.1186/s40537-021-00476-0]
• KOSIS (Korean Statistical Information Service), https://kosis.kr/eng/, . [accessed: Sep. 13, 2021]
• KOSTAT, http://kostat.go.kr/portal/eng/index.action, . [accessed: Sep. 10, 2021]
• Korea Train eXpress (KTX), https://www.letskorail.com, . [accessed Sep. 10, 2021]
• Super Rapid Train (SRT), https://www.srail.or.kr, . [accessed Sep. 10, 2021]
• Innovaion City, MOLIT, https://innocity.molit.go.kr, [accessed: Sep. 10, 2021]