Social studies on birth rates

Regarding the economic and social effects of low fertility rates, we found that the sustained decline in birth rates has brought adverse effects to economic and social development, such as population crisis, labor shortage, declining productivity, and increased social burden of care. Firstly, low fertility rates have self-reinforcing mechanisms in demography, sociology, and economics that may lead to a “low fertility trap” with population crises [2]. Secondly, the decrease in fertility rates implies a reduction in the labor force [18], leading to the Lewis turning point and the disappearance of demographic dividends [3]. Additionally, the aging of population, especially the aging of the workforce further reduces labor productivity, hindering industrial structural upgrading and technological progress. [6] Maesta et al.’s study [5] also shows that two-thirds of the total impact of aging on per capita GDP growth comes from the slowdown in labor productivity growth and one-third from the slowdown in labor force growth. Aging is accompanied by innovation effects and labor force effects. As the degree of aging deepens, the innovation effect decreases while the labor force effect strengthens, which is unfavorable for economic development [19]. Lastly, the decrease in fertility rates means an increase in the elderly dependency ratio, further exacerbating the social financial burden [4]. Zheng and Rui’s study [20] using theoretical models and provincial panel data in China also found that an increase in the elderly dependency ratio is unfavorable for economic growth.

Regarding factors affecting birth rates, we focus on the impact of economic development on fertility, combining education, housing prices, and female labor participation factors. With the impact of economic development on fertility intentions, scholars generally believe that there is a negative correlation between fertility rates and economic development levels [21, 22], as the utility of having children decreases as family income decreases, while the cost of raising children continues to rise. Therefore, families in highly developed economies have fewer children and invest more in each child’s education [23]. Murat, F et al. [24] linked fertility, economic development and human capital and found that economic development increases the stock of human capital and raises the opportunity cost of childbearing, thus inducing individuals to delay childbearing.

Education mainly affects fertility rates in three ways: first, an increase in educational attainment means longer years of study, which delays women’s childbearing age and reduces fertility rates [25, 26]. Second, an increase in average years of education in society raises children’s education costs and lowers fertility rates [27]. Finally, an increase in female education raises the opportunity cost of having children, delaying childbirth and reducing fertility rates [28].

Housing prices also have a significant impact on fertility intentions. On the one hand, housing costs compete with childcare costs and squeeze out fertility [29]. Couples who practice family planning may delay childbirth due to a lack of suitable housing [30]. On the other hand, rising housing prices bring wealth effects that increase fertility rates. Dettling et al.’s study [31] on the income and crowding-out effects of rising housing prices found different impacts on homeowners and non-homeowners.

Female labor participation affects fertility rates from two aspects: on the one hand, an increase in female labor participation and wages raises household income and increases fertility rates; on the other hand, an increase in female labor participation means that the “shadow price” of having children rises, lowering fertility rates. Most studies show that as women’s income increases, childbirth occurs later and fertility levels decrease [32]. Some studies show that with improved childbirth support measures and women’s status, an increase in female wages can bring “income effects” and raise fertility rates [33].

In addition to the aforementioned reasons, population mobility has a negative impact on fertility desire by increasing the cost of living and influencing marriage and childbearing habits, particularly in the process of rural population migration to urban areas [34]. Furthermore, the social support system, mainly comprised of medical and social security levels, also affects the birth rate. Wang et al. [35] used the CHNS database to study the influence of new rural cooperative medical insurance on fertility desire. The study showed that social security has a crowding-out effect and an income effect on fertility, and empirical evidence indicates that the crowding-out effect is dominant, which decreases residents’ fertility desire. Holmqvist et al. [36] also found a negative impact of pension coverage on the fertility rate using partial national data from South Africa.

In summary, this paper used comprehensive data covering the aforementioned indicators, encompassing 12 dimensions of urban development in prefecture-level cities, as input variables for BRP-Net. However, in addition to the above indicators, policies are also an important factor influencing population birth rates in China. Therefore, we have proposed a policy-induced birth impact indicator, PBQ, and provided calculation methods for different regions. Subsequent experimental results have also demonstrated the effectiveness of this indicator.

The studies on birth rate prediction

In the past few decades, scholars have proposed various methods for forecasting birth rates, including traditional time series analysis, machine learning, and deep learning methods. Traditional time series analysis methods include ARIMA and VAR models, which are capable of handling non-stationary time series data and making relatively accurate predictions of future trends over a certain period. For instance, H et al. [7] accurately predicted the birth rates in the United States from 2002 to 2008 based on data from 1975 to 2001, using an ARIMA-based method. Guo et al. [8] used the ARIMA method to forecast the total population of Yunnan Province from 1973 to 2018. Hasan et al. [9] constructed a multivariate stepwise linear regression model to forecast the birth rates in Bangladesh. Wang et al. [37] used a vector autoregressive model to forecast birth rates in China and discussed its feasibility in the Chinese healthcare sector. Collantes-duarte et al. [38] compared three time series forecasting techniques, including traditional ARIMA statistical forecasting methods, artificial neural networks, and emerging intelligent methods such as Neo Fuzzy Neurons, to forecast birth rates in Spain. However, birth rates are influenced by various factors such as social policies, economic development, and cultural changes, which may extend beyond the scope captured by ARIMA models, highlighting a limitation of traditional methods.

Compared to traditional methods, machine learning methods are more suitable for handling non-linear and non-stationary time series data. These methods typically involve feature engineering, selecting and constructing features related to birth rates such as age structure, education level, and economic development level. Then, regression analysis, decision trees, or support vector machines are used for modeling and prediction. These methods are advantageous in terms of being easy to interpret and implement, suitable for relatively simple data patterns. For example, Otoom et al. [11] compared 17 machine learning techniques in predicting population growth rates in the United Nations dataset, where random forest outperformed all other techniques in predictive performance and robustness to missing features. Huang et al. [39] used the LassoCV method and cluster analysis machine learning model to predict the birth rates of 31 provinces and cities in China based on birth rates and regional economic development levels. Zhang et al. [40] used various algorithms such as random forest, association rules, local variable importance, perceptual maps, and analogy analysis to explore the intrinsic relationships influencing the fertility levels of women of childbearing age and predict birth rates. However, traditional machine learning methods cannot accurately extract coupled features among complex multivariate variables, and due to the large number of parameters, their performance remains suboptimal.

With the rapid development of neural networks and computing resources, deep learning methods have been widely applied in recent years. These deep learning methods are capable of capturing complex nonlinear relationships within data, and they have shown good performance in tasks such as time series prediction of birth rates. Compared to traditional methods and machine learning, deep learning methods are better at handling long-term dependencies and nonlinear patterns among multiple variables or deep features, thereby improving prediction accuracy and robustness. Jia et al. [12] and Wu et al. [13] used BP neural networks to predict birth rates. Kumar et al. [41] proposed a deep learning model based on Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) for predicting birth rates in India, achieving excellent predictive results. Johann et al. [42] proposed a random CNN model for predicting birth rates and labor supply in Germany. However, traditional CNN-based deep learning methods have not effectively utilized temporal information among multiple variables and have not fully exploited the correlation within time series data, thus exhibiting certain limitations for tasks such as birth rate prediction.

The studies on LSTM and attention mechanism

LSTM, a special network within the Recurrent Neural Network (RNN), is known for its unique gating structure and mechanism that enables it to capture long-term dependencies and handle non-linear patterns in time series data. This makes it suitable for complex multivariate time series forecasting tasks such as predicting birth rates. LSTM has been widely applied in various time series prediction tasks. Lu et al. [43] proposed a stock price prediction method based on CNN-LSTM, selecting eight input features including opening price, highest price, lowest price, closing price, trading volume, turnover, price fluctuation, and change. Kim et al. [44] presented a CNN-LSTM neural network that effectively extracts complex features of energy consumption for accurate residential energy consumption prediction. Guo et al. [45] utilized the LSTM method to model and detect the health status of bridges, enabling accurate prediction of potential anomalies in the future. Shahid et al. [46] introduced a bidirectional long short-term memory network (Bi-LSTM) for forecasting the COVID-19 pandemic. While these LSTM-based methods have shown promising results in their respective prediction tasks, traditional LSTM methods still face challenges in accurately extracting deep-seated clues and feature coupling relationships among complex multivariate time series data with specific objects, and may not better focus on variables with higher contributions.

Therefore, the industry has begun to attempt to combine attention mechanisms and LSTM methods to overcome the above-mentioned issues, and has further started to apply them in more complex multivariate time series data processing and prediction tasks. Lin et al. [47] proposed a double-stage encoder-decoder attention-based long short-term memory (LSTM) network for short-term regional load forecasting. Jiang et al. [48] combined LSTM with attention mechanisms and used TPE Bayesian optimization for hyperparameter optimization, achieving accurate indoor temperature prediction. Teng et al. [49] introduced a multi-scale local clue and hierarchical attention-based LSTM model (MLCA-LSTM) to capture potential stock price trend patterns, resulting in broader prediction coverage and more accurate results compared to the study [43]. Shi et al. [50] employ complementary attention to accurately predict and understand user intentions in social networks within cross-media multimodal data. Therefore, inspired by the above research, we first attempt to apply the advantages of combining attention mechanisms and LSTM to the prediction task of the birth rate in prefecture-level cities, proposing a BRP-Net dual-branch attention network model to extract deep temporal coupling features of specific prefecture-level city multivariate time series data, aiming to achieve accurate and robust birth rate prediction in prefecture-level cities.

Overview

Recurrent neural network (RNN) is a machine learning algorithm used for processing time series data. Hochreiter et al. [51] proposed the long short-term memory (LSTM) model, which improved the generic RNN. The basic unit of the LSTM model is the memory cell, which includes four neural network layers: the cell state, forget gate, input gate, and output gate, instead of the simple tanh layer repeated in generic RNN. The LSTM model solves the problem of long-term dependencies that exist in the generic RNN. It can handle and predict important events with long time intervals and event sequence delays, and apply remote contextual information to the current moment. Therefore, using LSTM as the core model for predicting prefecture-level city birth rates can fully connect information data of the same region in different years, establish cross-scale feature connections, and achieve better prediction results. However, traditional LSTM models can only handle univariate time series data and often fail to effectively extract deep coupling clues and features among variables when handling multivariate time series data.

Therefore, we design a dual-branch attention structure based on LSTM. The advantage of this attention mechanism is that it overcomes the long-term dependency and information loss of RNN, better characterizes input data by assigning different importance to each element of the input network and focusing on more relevant inputs, and effectively learns the long-term temporal dependency and spatial correlation of complex multivariate time series. In practical terms, although economic development, education, healthcare, and population structure all have significant impacts on the birth rate, their respective influence weights are certainly different. By adaptively allocating different weights to each factor through the attention mechanism, it better aligns with practical significance and enhances the interpretability of the proposed network.

The proposed network

The BRP-Net model we designed is a hybrid neural network consisting of three parts: the feature attention module, the temporal attention module, and the LSTM recurrent prediction module. Its complete structure diagram is shown in Fig 1. We selected 11 factors with the highest correlation with the birth rate in prefecture-level cities, such as house prices, the proportion of women of childbearing age, and the GDP level of prefecture-level cities, as the multidimensional input variables of the network model. The feature attention module learns the feature correlations in the multivariate input time series data. The temporal attention module and the feature attention module are in parallel structure, modeling the temporal relationships of multivariate input data based on transposed input data. The outputs of the two attention modules are generated by element-wise multiplication (denoted by *) of attention weights with input data, and the resulting output is combined and used as the input to the LSTM recurrent prediction module, ultimately outputting the predicted result value.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. The structure of the BRP-Net proposed in this article.

https://doi.org/10.1371/journal.pone.0307721.g001

Inspired by the self-attention mechanism, we constructed the feature attention module and the temporal attention module using only input values and simplified single-layer perceptrons. Attention weights are applied to each input variable of each layer in the feature attention module and to each time step in the time attention module. The outputs X_f of the feature attention module and X_t of the temporal attention module are concatenated and used as the input to the LSTM model. The LSTM recurrent prediction module consists of a single-layer stateful LSTM and two fully connected layers. A stateful LSTM model means that the hidden state h_t learned at the current time step is transferred to the initial state in the next learning period. To prevent overfitting and ensure the robustness of the prediction results, we also applied a dropout layer after the output of the LSTM. After flattening, the output of the dropout layer is input to a fully connected layer for the final prediction output. Table 1 shows the hyperparameter settings used in each layer. To predict the final single true value, the number of neurons in the last fully connected layer is set to 1. The principles and details of the feature attention module, the temporal attention module, and the LSTM recurrent prediction module will be elaborated in detail in the following content.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Model configurations.

https://doi.org/10.1371/journal.pone.0307721.t001

The LSTM recurrent prediction module.

The LSTM architecture used in this article is shown in Fig 2. An LSTM unit consists of three gate units. Long short-term memory (LSTM) is a well-known deep learning architecture used for processing time series data. It incorporates gate mechanisms and internal states to enable selective accumulation of historical information and retention of new information [52]. Fig 2 illustrates the structure of an LSTM unit. The LSTM unit comprises three gates, namely the forget gate f_t, input gate i_t, and output gate o_t. The forget gate f_t decides how much historical information to retain in the internal state when new information is introduced. The input gate i_t determines which new information should be updated to the internal state, while the output gate o_t selects which updated information can be passed as input to the network for the next time step [7]. When dealing with time series prediction problems, the LSTM network has faster convergence speed and higher accuracy compared to other neural network models. The following equation represents the internal operation of the LSTM unit: (11) here, h is the hidden state and x is the input variable. represents the candidate memory information. W denotes the weight of the neural network and b denotes the bias of the neural network.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. The structure of the LSTM.

https://doi.org/10.1371/journal.pone.0307721.g002

Loss function and evaluation index

To further optimize the model parameters, it is necessary to minimize the reconstruction error. We have used mean squared error as the loss function, which is expressed as follows: (12) where represents the predicted value of the prefecture-level city’s birth rate, and x_i represents the true value.

In this article, we have evaluated the predictive performance of the model using three evaluation metrics: root mean squared error (RMSE), mean relative error (MRE) and coefficient of determination (R-squared score). (13) (14) (15) where represents the predicted value, y_i represents the real value, and n represents the number of data points. A smaller RMSE value indicates a smaller prediction error of the prefecture-level city’s birth rate prediction model. The closer the R-squared value is to 1, the better the fit between the predicted prefecture-level city’s birth rate and the actual value.

Design of indicators for policy-based birth quota impact

As the strict family planning policy has encountered resistance in some regions, the government allows each province to formulate different birth policies based on their population status to ensure policy implementation. To control the normative impact of different birth policies in different regions, this article proposes a method for calculating the Policy-Based Birth Quota (PBQ) inspired by Guo (2003). We calculate the PBQ allowed for each family under the “one-child policy” based on different family types and ethnic structures, with specific variables shown in Table 2. We classify China’s prefecture-level cities into three categories by province: the first category includes Beijing, Tianjin, Jiangsu, Chongqing, and Sichuan, where the “one-child policy” is widely implemented. The second category includes Jilin, Liaoning, Hebei, Shanxi, Shaanxi, Gansu, Shandong, Zhejiang, Henan, Anhui, Hubei, Hunan, Fujian, and Guangdong, where the proportion of ethnic minorities is relatively small, and the PBQ only considers regional differences. The calculation formula is as follows: (16)

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Explanations of factors represented in the calculation process of PBQ.

https://doi.org/10.1371/journal.pone.0307721.t002

The third category includes Inner Mongolia, Tibet, Xinjiang, Qinghai, Ningxia, Guizhou, Yunnan, Guangxi, and Hainan, where the population has different structures. The calculation formulas are as follows: (17) (18) (19) (20) (21) (22) (23) (24)

Eq 17 represents Inner Mongolia and Guangxi, while Eqs 18–24 represent Guizhou, Yunnan, Ningxia, Xinjiang, Hainan, Qinghai, and Tibet, respectively.

It is worth noting that after the second-child policy was implemented in 2016, the PBQ value for all prefecture-level cities was 2. The introduction of the PBQ indicator can more effectively measure the impact of local government fertility policies on the birth rate, making the overall prediction results more accurate and interpretable. The results of the ablation experiments in the next section also confirm the effectiveness of the proposed PBQ indicator.

Dataset

To validate the effectiveness of the BRP-Net method in predicting the birth rate of prefecture-level cities, we collected comprehensive data from 361 prefecture-level cities in China from 2000 to 2020, forming a widely covered and regionally diverse dataset of birth rates. The birth rate data for 2010 and 2020 are derived from nationwide population census data, and the data for 2005 and 2015 are from 1% population sample surveys. All the data were collected and compiled from the statistical bulletins of the National Bureau of Statistics and various prefecture-level cities. Data for other years are sourced from the provincial statistical bureaus to which the prefecture-level cities belong.

Table 3 presents the birth rate situation from 2000 to 2020. From the content of the above figures and tables, it can be seen that there is significant heterogeneity in the fluctuation of birth rates among different prefecture-level cities in China. For example, in 2000, the lowest birth rate in Haikou City was only 3.77‰, while the highest in Wenzhou City reached 27.25‰, resulting in a difference of 23.48‰. By 2019, the birth rate in Baicheng City, Jilin Province, was only 3.4‰, while in Shenzhen City, it reached 21.68‰, resulting in a range of 24.34‰. These significant differences are caused by various reasons such as local economic development and educational levels.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. The birth rate situation of prefecture-level cities published from 2000 to 2020.

https://doi.org/10.1371/journal.pone.0307721.t003

Hence, it can be seen that there are numerous and extensive factors that affect the birth rate of prefecture-level cities, and they are interrelated to a certain extent. Therefore, in selecting the input variables for the network, we have chosen twelve factors based on their significant impact on the birth rate of prefecture-level cities. These factors include GDP, PBQ, registered population relative to de jure population, primary school teachers per 10,000 people, number of practitioners per 10,000 people, number of university students per 10,000 people, old-age dependency ratio, per capita social security and employment expenditure, housing price relative to GDP, female education level, and percentage of employed women. These factors essentially cover various aspects of a prefecture-level city, such as economic development, educational level, and age-appropriate population structure. The detailed descriptive statistical data for these factors are shown in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Descriptive statistics of main variables.

https://doi.org/10.1371/journal.pone.0307721.t004

Implement details

We partitioned the data into training, validation, and test sets in an 8:1:1 ratio. Adam [8] was utilized as the optimization algorithm, and the network was implemented in the Pytorch [9] framework. For testing, a system with one Core i9–12900KF CPU, one NVIDIA RTX3090ti GPU, and 64GB of memory was used. We used a batch size of 32 and a learning rate of 0.001. Our model was trained for a total of 1000 epochs, taking a total of 156 minutes.

Experimental results and comparative analysis of birth rate prediction in prefecture-level cities

To validate the predictive accuracy of the proposed BRP-Net model, we established several benchmark models and compared our model’s performance with other similar models. These benchmark models include classic traditional methods such as ARIMA, VAR, as well as learning-based methods like CAE, LSTM, GRU, CNN-LSTM:

ARIMA: Based on the autocorrelation and moving average properties of time series, it models data through differencing, autoregression, and moving averages. The ARIMA model can identify trends and seasonality in the data for prediction.

VAR: A multivariate time series modeling method that considers the mutual influence between variables. It involves selecting appropriate lag orders, estimating VAR model parameters, and making predictions.

CAE: Convolutional Autoencoder. It utilizes convolutional neural networks for feature extraction and compression, followed by decoder reconstruction of the input.

LSTM: A type of recurrent neural network structure capable of capturing long-term dependencies. It takes in time series input, learns long-term dependencies, and outputs predictions.

GRU: Similar to LSTM but with a simpler structure, featuring update and reset gates. It learns inherent patterns and dependencies in time series data for prediction.

CNN-LSTM: Combines the feature extraction capability of CNN with the time series modeling ability of LSTM. This model first extracts features through CNN and then inputs them to LSTM for sequence modeling and prediction.

GCN-LSTM: The GCN-LSTM model is based on the T-GCN model structure proposed by Zhao et al. [53]. T-GCN combines graph convolutional layers [54] and the GRU model, and has been applied to time series data prediction tasks such as electricity load forecasting and traffic flow prediction.

DA-RNN: The DA-RNN model proposed by Qin et al. [55] is an encoder-decoder structured model with attention mechanisms applied in both the encoder and decoder. This model has demonstrated outstanding performance in indoor temperature forecasting and exchange rate prediction, covering relatively long time spans.

In addition to traditional methods, all learning-based methods used for comparison underwent training for 1000 epochs to reduce random errors. We utilized root mean square error (RMSE), mean relative error (MRE), and coefficient of determination (R-squared score) to compare results. The results of the quantitative experiments and their visualization are shown in Table 5 and Fig 3. The experimental results indicate that overall, the prediction errors of traditional methods are higher than those of learning-based methods. Among the learning-based methods, CAE exhibits the highest average MRE and RMSE at 75.5%, possibly due to its weaker learning capability compared to other benchmark models and its inability to capture temporal relationships in multivariate input data. Subsequently, LSTM and GRU recorded the second and third highest error rates, with MRE and RMSE at 0.5986, 0.6100 and 0.8815, 0.9021, respectively, possibly due to the inability of pure LSTM and GRU to cover complex and long-span multi-input variables. Relatively, although GCN-LSTM’s performance is slightly better than LSTM and GRU, it still falls short of models like CNN-LSTM, which first extracts multivariate features before conducting time series prediction. This may be attributed to the limitations of graph convolutional layers optimized for learning spatial information in capturing strong time series features of birth rates. Encoders with input attention mechanisms and decoders with temporal attention mechanisms in DA-RNN and the proposed BRP-Net exhibit excellent performance. The reason for these results is that both models use attention mechanisms to effectively learn time series features of input data and input variable features. The BRP-Net model records the smallest error results in terms of RMSE and MRE at 0.2206 and 0.1397, respectively, with R-squared closer to 1 compared to other models, indicating robustness and higher predictive accuracy. The predicted birth rates of Chinese prefecture-level cities are also closer to the actual values with smaller errors.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Qualitative comparison of birth rate prediction results in prefecture-level cities.

Bold is best.

https://doi.org/10.1371/journal.pone.0307721.t005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Visualization of the compared methods on RMSE, MRE and R².

https://doi.org/10.1371/journal.pone.0307721.g003

Fig 4 displays the visual results of population birth rate predictions for each model at the prefecture level. From these qualitative experiments, it is evident that traditional methods exhibit significant fluctuations. In contrast, our proposed BRP-Net method achieves precise and robust prediction results.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Visualization and comparison results with traditional and learning-based methods for birth rate prediction in prefecture-level cities.

(a)ARIMA, (b)VAR, (c)CAE, (d)LSTM, (e)GRU, (f)GCN-LSTM [53], (g)CNN-LSTM, (h)DA-RNN [55], (i)BRP-Net(ours).

https://doi.org/10.1371/journal.pone.0307721.g004

Ablation experiment

To measure the effectiveness of the PBQ index proposed in this paper, we conducted ablation experiments on this index. The BRP-Net without the PBQ index reduced from eleven input variables to ten. The results of the ablation qualitative and quantitative experiments are presented in Table 6 and Fig 5. These results clearly demonstrate that excluding the PBQ as an input variable for prefecture-level population birth rate prediction leads to a certain degree of reduction in prediction accuracy. This also indicates that incorporating local policies into the influencing factors of prefecture-level population birth rates, as proposed in the BRP-Net, aligns with the actual situation in China, thus achieving more accurate prediction accuracy under the same model.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Quantitative ablation experiments to measure the effectiveness of the PBQ index.

Bold is best.

https://doi.org/10.1371/journal.pone.0307721.t006

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Visualization of comparative results for ablation experiments targeting PBQ metrics.

https://doi.org/10.1371/journal.pone.0307721.g005

Correlation analysis and contribution weight analysis of factors affecting birth rate

We also conducted correlation analysis on the attribute features of the input multivariate data to provide support for prefecture-level city governments to comprehensively control local development direction and improve birth rates. The heatmap represents the size of the correlation coefficient, and the depth of color represents the strength of the correlation. In Fig 6, the relationship between color and correlation strength corresponds to the color bar on the right side of the figure. The more blue it is, the stronger the positive correlation, and the more orange it is, the stronger the negative correlation. By analyzing the correlation results of the heatmap, it can be found that the overall linear correlation between variables is not strong, which is a key premise to ensure that the proposed model can fully explore the time series feature information of different variables.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Heatmap of correlation among multi-dimensional input variables and importance assessment plot.

https://doi.org/10.1371/journal.pone.0307721.g006

At the same time, we evaluated the feature importance of the predicted results of the proposed BRP-Net through SHapley Additive exPlanations (SHAP) [56]. This method mainly obtains the feature importance evaluation by iteratively calculating feature interactions, differential feature calculation, and weighted average calculation of Shapley values for all feature subsets. This theory provides a global explanation method for model prediction, which can help understand the degree of influence of each input feature on the prediction results in neural networks. The experimental results are also shown in Fig 6. Similarly, the thickness of the line represents the size of the correlation coefficient, and the depth of color represents the strength of the correlation. The more blue it is, the stronger the positive correlation, and the more orange it is, the stronger the negative correlation. We can see that GDP and PBQ have the most significant impact on birth rate in prefecture-level cities, with the former showing a negative correlation connection and the latter showing a positive correlation connection. In addition, social security and employment expenditure, percentage of women over age 6 with a college degree or above, and old-age dependency ratio also have a significant impact on birth rate in prefecture-level cities.