Mechanism Analysis and Response of Digital Financial Inclusion to Labor Economy based on ANN and Contribution Analysis

2.1 DFI

DFI, through big data, blockchain, and other technologies to break the boundaries of space and time, provides convenient, efficient, and reliable financial services to people in non-developed areas and rural areas. It specifically encompasses two concepts, one is “inclusion” and the other is “digital” (Ozili, 2018). To further clarify the connotation and meaning of DFI, this study understands it from the following two dimensions:

First, “Inclusion” is the core purpose of the financial system. It can be divided into two characters in Chinese which respectively mean “universal” and “benefit.” That is to say, the financial system not only emphasizes the universality and wide coverage of the audience but also emphasizes its essence of benefiting the people (Zhang et al., 2022). The concept of inclusive finance was first introduced by the United Nations in 2005 and is defined as “a financial system that effectively and comprehensively serves all segments and groups of society, especially poor, low-income people” (Yang & Zhang, 2020). In the early days, inclusive finance was mainly in the form of microfinance, providing affordable capital lending to the general public, especially the disadvantaged, and also providing channels for start-ups to raise funds. And with its gradual development, both internationally and in China, inclusive finance has gradually moved from a focus on microfinance to cover a wide range of business areas such as payments, deposits, loans, insurance, credit services, and securities (Li et al., 2022). However, while China has made significant progress in financial inclusion, digital financial services are still crucial to fill the gaps in financial inclusion (Ji et al., 2021).

Therefore, digital is an effective way to promote the vision of “inclusive.” Unlike traditional inclusive finance, DFI combines “digital” with “inclusive,” using digital technologies such as big data and cloud computing to increase the scope and coverage of financial services, lower the threshold for financial services, and provide lower cost, faster and wider coverage of financial services to a wider range of people (Gabor & Brooks, 2017). It also further reduces the financing constraints on entrepreneurial activities due to information asymmetry in financial transactions and thus improves the efficiency of financing (Fan & Zhang, 2017; Shen et al., 2022). In addition, DFI can contribute to urban innovation through improved access to credit, consumption, and industrial upgrading (Li & Li, 2022). Studies have shown that financial inclusion in practice is now strongly correlated with innovative digital finance, and it is evident that new digital economic models have become important drivers and sources of growth for sustainable financial inclusion (Chen et al., 2023).

In summary, DFI is a financial system that combines digital technology and finance to achieve inclusive purposes, and the addition of digital technology has opened up a new and viable path for the sustainable development of inclusive finance. To understand the current status of DFI development in each region and further achieve the goal of promoting inter-regional inclusive finance development, a set of indicators to measure the level of DFI development in different regions is essential (Chen et al., 2020). In 2011, a team of researchers from Peking University’s Digital Finance Research Centre and Ant Group Research Institute used Ant Group’s massive data on DFI to compile a “Peking University Digital Inclusion Index,” which provides a set of instrumental data reflecting the current status and evolution of DFI development in China. Based on clarifying the connotation and meaning of DFI, this study introduces and explains the first two dimensions of the index system, and explores the construction of a DFI -related index system in China, thus providing a reference for the construction of an international digital financial inclusion index system.

The Peking University Digital Inclusive Finance Index of China (PKU-DFIIC) meets all the principles in the construction process, conforms to the connotation of DFI, and provides a reliable and authoritative indicator system for measuring the development status of this financial system (Arner et al., 2020), which can provide a reference for the construction of an international digital financial inclusion index system. In this study, the PKU-DFIIC is compiled as follows (Table 1).

2.2 Labor Economics

The theory of labor economics mainly focuses on labor relations and summarizes the laws of development of labor economy. It analyzes how to create the highest economic efficiency with the lowest labor cost (Wu et al., 2022). With the development of labor economics, it was labeled as “personal economics” in 1980 (Yuan & Lin, 2011). However, this label ignored the impact of intrinsic management on labor economics. There are two important assets in labor economics, namely, fixed assets and human assets, among which there is a distinction between individual human assets and team human assets (Arif, 2021). Also, human capital is an indirect market influenced by salary designation mechanism and labor market wages (Chen & Ji, 2020). Thus, it can be concluded that labor economics is an emerging doctrine with a special character because it takes “labor force” as its object of study (Bobba et al., 2021). The core of this doctrine is to study the factors affecting labor supply and labor demand in the labor market and the interaction between them (Shi, 2021; Villamil et al., 2020). Since labor has special properties that make it different from ordinary commodities, classical theories that do not distinguish it cannot fully take into account the problem of unsystematic risks and residuals (Cui et al., 2018; Singh, 2023). In summary, labor economics is used here as a primary indicator, while secondary indicators are established by considering the factors that affect labor supply and labor demand from a microeconomics perspective and the factors that cause interactions from a macroeconomics perspective. The labor economic indicators and related meanings are shown in Table 2.

2.2.1 Analysis of the Impact Mechanism of DFI

DFI facilitates the action of financial inclusion through digital technology, which contains mobile payment and microfinance to serve the majority of micro and small enterprises as well as residents (Sun et al., 2020). The labor economy contains various factors such as labor supply and demand allocation, macro and micro management of labor, labor efficiency, and labor reproduction, and the specific influence mechanism is shown in Figure 2.

Figure 2

The transmission mechanism of DFI on labor economy.

From the connotation of both, DFI has a coupling effect with the benign development of the labor economy. First of all, a sound financial system also has very good support for financial products and services for the lack of credit history, non-high-income groups, and micro and small enterprises (Ren et al., 2023; Wang & Chen, 2022). A complete credit system for micro and small enterprises brought by DFI can, to a certain extent, alleviate the financing constraints of micro and small enterprises (Xu et al., 2023) and facilitate the expansion of business scale (Dai et al., 2023). In addition, the development of DFI can stimulate the demand side of business and improve the unemployment problem to some extent and raise the hourly wage rate of workers (Du et al., 2023; Zhu, 2023).

In addition, the development of DFI can further enhance the productive capacity and technology level of firms by alleviating the level of financing constraints in the financial market (Li et al., 2024; Sun & Tang, 2022), which leads to an increase in the marginal output per unit of labor in firms, which in turn increases the demand for labor or the ability of firms to pay wages, and this transmission mechanism promotes labor productivity and employment levels (Wu & Chen, 2023). As a result, this study proposes hypothesis H1.

H1: DFI can alleviate the financing constraint of business operation and entrepreneurship process and thus promote the labor economy.

DFI contributes to the mobility and convenience of services, and on this basis, it has both breadth and depth of service coverage, and digital inclusive means represented by mobile payment contribution to the development of the platform economy (Liu & Xu, 2022). The development of platform economy contributes to the refinement of the demand side of labor, the increase in labor flexibility, the diversification of employment forms, and the significant increase in the number of payments and insurance holdings per capita of labor (Li & Wan, 2022). As a result, this study proposes hypothesis H2.

H2: DFI can promote the development of labor economy by promoting the mobility, convenience, and coverage of various services and thus the development of labor economy.

Second, due to the existence of market segmentation, factors of production cannot flow freely, or it is difficult to promote the flow of factors to more efficient areas due to information asymmetry, which will lead to factor mismatch (Zhang & Wang, 2020). DFI can improve information asymmetry and alleviate the problem of weakened free mobility of factors for the purpose of protecting local industries, thus improving labor mismatch and macro and micro management of labor, but there is some heterogeneity in this mechanism, so DFI has a moderating effect on the benign development of labor economy. As a result, this study proposes hypothesis H3.

H3: DFI can improve labor allocation and macro and micro management and thus promote the development of labor economy.

3.1 ANN Method

The ANN system, a highly advanced computational model inspired by the information processing mechanism of nerve cells in the human brain, simulates the process of receiving and transmitting stimuli to nerve cells in the form of linear weighted summaries and functional mappings, which replicate to a large extent the basic principles of the work of the human brain. Unlike traditional linear regression, which is single-threaded, unidirectional, and difficult to deal with the defects of complex problems, ANN is possible to adjust the weights in the network structure by means of optimized learning algorithms, thus achieving astonishing accuracy and efficiency in the learning and information processing process. We highlight and use the ability of the ANN system to self-organize, to automatically discover patterns and regularities in the input data without explicit instructions, and to adjust its internal structure accordingly. Associated with this is its self-learning ability, where the system is able to optimize and improve its performance on its own, meaning that with training, the ANN system can become more accurate and progressively improve its problem-solving capabilities.

As a means of statistical modelling, ANNs have the following characteristics: (1) They have the ability to perform better for complex non-linear relationships. (2) No pre-specified fitting functions are required, with learning and highly adaptive recognition patterns, i.e., they do not require a pre-determined mathematical relationship for mapping between input and output, and can reproduce complex non-linear input-output relationships and be applied in the sequence training process. ANNs as deep learning methods are highly adaptive and better able to interactively respond to increasingly complex real-world economic systems than predictions from traditional linear fitting models or traditional machine learning methods

The network consists of at least three layers of nodes forming a synaptic hierarchical framework: the basic components of an ANN are an input layer, a hidden layer, and an output layer, combined by weights, biases, and activation functions (Pua et al., 2022). The general topology of the ANN is shown in Figure 3. The dots represent the nodes of the neural network and are often referred to as neurons. The lines connecting the neurons are called synapses, which can be understood as connections between neurons, and serve to allow signals from the previous layer to be received by the next layer of neurons. Figure 4 illustrates the working mechanism of neurons in a typical ANN structure.

Figure 3

ANN training architecture.

Figure 4

Hidden layer cell activation function.

During ANN training, each neuron in the input layer receives data, multiplies the values by the weights and passes the results to the hidden layer. The neuron in the hidden layer generates the output signal by means of an activation function on the activation signal transmitted by the previous layer. The activation signal is a weighted sum of all signals entering the neuron, as shown in equation (1).

x i is the neuron in the input layer and ω ij is the first neuron of the input layer neuron i of the hidden layer and x j is the weight vector between the first neuron in the input layer and the first neuron in the hidden layer, and x j is a neuron j is the activation signal received by the neuron in the hidden layer. The hidden layer neuron generates the output signal through the activation function. Each neural unit of the input layer directly corresponds to the original data and then provides information to the hidden layer, and each neural unit of the hidden layer has different weights for different neural units of the input layer, thus biasing it to be sensitive to a certain recognition pattern. There are various forms of activation functions for the hidden layer, usually Sigmoid and RELU functions, which produce a non-linear dependency between the input signal and the output result, and therefore a non-linear fitting algorithm must be used to determine the model parameters (Ravnik et al., 2021). The general form of one of these Sigmoid functions is as follows.

3.2 Sensitivity Analysis

Sensitivity analysis is a very important class of methods for quantitatively describing the degree of importance of a model’s input variables to its output variables by varying each of the model’s attributes over a range of possible variations, and examining and predicting the degree to which changes in these attributes affect the output values and quantifying them as sensitivity coefficients.

Garson’s algorithm is a method for weight analysis based on the connection weights between layers of a neural network (Wong et al., 2011). It is a typical algorithm for sensitivity analysis based on connection weights. The basic principle is to calculate the contribution of the input variables to the output variables by the product of the connection weights, and the Garson method takes into account the effect of the interaction of multiple variables on the output compared to other methods (Zheng et al., 2018). The Garson algorithm uses the connection weights between each input signal in the network as a “bridge” and calculates the relative contribution of each input parameter to the output parameter using a matrix of weights (Cao et al., 2022). In overall, the algorithm determines the contribution of each input value by allocating the connection weights between the hidden layer and the output layer of each hidden layer neuron to the connection weights of each related input signal (Bai et al., 2020). The formula for the degree of influence of the algorithm’s input variables on the output variables is

where n , l , and q are the number of neurons in the input layer, the hidden layer, and the output layer, respectively, and w is the connection weight of the input layer to the hidden layer, and v is the connection weight of the hidden layer to the output layer. Sensitivity analyses based on Garson’s algorithm will attribute interpretability to the input features or samples, making them easier to understand. The results will be output together as part of interpreting the neural network learning results.

3.3 Algorithm Test Index Setting

The number of nodes in the hidden layer is determined by training the network, and in order to construct the optimal performance network, when testing the trained ANN model with test samples, some test metrics need to be selected to better guide the evaluation of the results of the trained model. In this work, two common tests are selected, the mean absolute percentage error (MAPE) and the coefficient of determination ( R 2 ). The MAPE error evaluation metric is calculated in equation (4), where y i − y i ^ is the true value on the test set. MAPE is commonly used as a measure of accuracy, reflecting the degree to which the model is good or bad, and is a holistic approach for evaluating the model from an “average” perspective (Li & Peng, 2018). It is an overall assessment of the model from an “average” perspective. The closer the MAPE is to zero, the better the model is (Zhi et al., 2022). MAPE is a measure of relative error. MAPE is a relative error measure that uses absolute values to avoid positive and negative errors cancelling each other out.

R 2 is the range of values of [ 0 , 1 ] , the formula for the calculation process is given in equation (5), where the numerator part represents the sum of squared differences between the true and predicted values; the denominator part represents the sum of squared differences between the true and mean values. If R 2 is larger and closer to 1, the higher the correlation, the closer the forecast value is to the actual observed value, and the model provides a good fit. If R 2 is close to 0, the model is a poor fit. R 2 is an important indicator for determining the goodness of fit of a model (Zhang, 2015).

4.1 Selection of the Study Data

The Northeast China region, which comprises the three provincial administrative divisions of Heilongjiang, Jilin, and Liaoning provinces (from 118°53′E in the west to 135°05′E in the east, from 38°43′N in the south to 53°33′N in the north), was an extremely important center of heavy industry in the early years of China’s history, based on its geographical location, natural resources, and history. The data from the Northeast region of China is chosen because its major economic sectors are more closely aligned with the labor economy, and the data from the Northeast region is more representative of the impact of financial inclusion on the labor economy than the more advanced regions, which is of greater importance to the empirical evidence in this study (Figure 5).

Figure 5

Geographical location of the North East Region. Source: Ministry of Natural Resources of PRC.

As the origin of the rapid industrial development of new China, the Northeast rapidly developed an industrial and agricultural economic production model in the 1950s, characterized by planned production and Soviet model management. By the end of the 1950s, the GDP of the Northeast had exceeded 15% of China’s total GDP, reaching 19% at one point, and has remained at around 13% since then.

However, as China’s modernization process has increased the need for productive forces and relations of production, the main contradictory features of the old industrial bases in the Northeast, such as a single industrial structure, a rigid management system, and a lack of innovation, have gradually come to light, leading to a slowdown in economic growth and, with it, a gradual loss of labor. According to the data disclosed in the China Labor Statistics Yearbook by year, the number of laid-off workers in the Northeast Region has been increasing rapidly year on year, from a total of 60,500 in 1992, accounting for 3.7% of the region’s working population, to 225,400 in 2000, accounting for 10.95% of the region’s working population, which was among the highest in China during the economic transformation and reform of the 1990s. Between 1998 and 2000, the number of registered urban unemployed nationwide reached 5.71, 5.75, and 5.95 million, respectively, while the number of unemployed workers in the Northeast Region reached 833,000, 854,000, and 895,000. In the twenty-first century, in the context of the current new industrialization path of high quality and light pollution, the focus on innovation-driven, transformational development and digitization has led to the integration of information technology into the industrial model, led the Northeast to finally face the serious challenges posed by the resource-based industrial layout for sustainable urban development after decades of resource consumption and environmental pollution. With a total GDP of RMB 5.57 trillion by 2021, the economic growth rate of the Northeast has been lower than China’s average growth rate of 8.1%.

The Northeast, with its natural resources and other advantages, was once the lifeblood of China’s economic development as the first heavy industrial base built after the founding of New China. However, the problem of slow economic growth and overall economic decline caused by a series of factors such as difficulties in industrial transformation, severe environmental pollution problems, and a large loss of young people is a dilemma that the Northeast has always faced.

4.2 Data Collection and Pre-processing

In summary, the Northeast region of China is used as the objective of this study. Combined with the character of informatization and universality of DFI, the purpose of this study is to quantify the response mechanism and correlation degree of digital financial inclusion to the labor economy by taking into account the problems of an aging population, imbalance of labor supply and demand, and weak blood generation mechanism in the Northeast region, and to explore the solution of labor economy problems through the construction of digital financial inclusion, thus further promoting the modernization process. Given the small number of years for which laboratory data and instantaneous sampling data were available, and that some of these years had missing data, we first performed Multiple Imputation using IBM SPSS Statistics 26.0, using standardization for data categories with inconsistent exposure, and focused fitting on a few important samples. Tables 3 and 4 show the data of DFI and labor economy after descriptive statistical processing.

4.3 Multiple Imputation

Among the above data and indicators, three of the digital financial inclusion indicators, X ₆ (monetary funds), X ₇ (investment), and X ₉ (credit) are identified as invalid variables and removed from this study because of the large amount of missing data. The labor economic indicators of Y ₅ (birth rate), Y ₆ (mortality rate), Y ₇ (natural growth rate), Y ₁₀ (average wage index), and Y ₁₁ (employed persons) have missing data for individual provinces in individual years. Therefore, this study performs multiple interpolations of missing data for the scalar variables Y ₅, Y ₆, Y ₇, Y ₁₀, and Y ₁₁ by the fully conditional designation (Markov Chain Monte Carlo) method.

The multiple imputation process incorporates analytical weights into the regression models used to interpolate missing values, excluding cases with negative or zero analytical weights. The fully conditional specification (FCS) method fits a single dependent variable model using other variables in the model as predictors and then interpolates missing values for the fitted variables. This method is executed continuously until the maximum number of iterations is reached and the interpolated values are saved to the interpolated dataset.

Finally, the parameters, mean values, labelled differences, and confidence intervals of the initial data and multiple interpolated datasets are compared to select the most reasonable dataset.

4.4 ANN Training and Results

The article uses IBM SPSS Statistics 26.0 as the ANN training software. The ANN is trained using a neural network multiple perceptions, with a hyperbolic tangent function for the hidden layer activation function with a range of 10–30 hidden layer units on a single hidden layer, to measure the R ², and MAPE between the predicted and true values for different hidden layer units, as a criterion to evaluate the optimal model. It was found that the prediction deviation was the smallest for a single hidden layer of unit 14, and the model converges to an optimal state or achieves a high level of accuracy. In this study, a single hidden layer of unit 14 is chosen as the optimal parameter.

R² represents the proportion of variance explained by the model and is the Sum of the Squares Regression and the Sum of the Squares Total between the actual and predicted values (SSR/SST), which is the default metric for scikit-learn regression problems. R ² is used to measure the proportion of variance explained by the model to all variance on a relative basis. R ² tends to 1 representing the more variance that can be covered by the model predictions.

MAPE based on the mean absolute error (MAE), the mean value of the absolute error between the predicted and true values is determined as a percentage unit instead of an absolute value unit. MAPE is more sensitive to relative error and does not change due to the global scaling of the target variable, which makes it suitable for problems where there is a large gap between the target scales of the metric variables. The closer the MAPE index converges to 0%, the better the model is set up, while the MAPE index is greater than 100%, which means that the model is a poor-quality model (Figure 6).

Figure 6

The degree of prediction deviation under different hidden layer units (layers 10–20).