Research on the Utilization Pattern Mining and Impact Mechanism of Open Government Data Based on Deep Learning Algorithms

Input the temporal feature E_t (parsed unique heat encoding) of the moment to be predicted and compute the importance score for each time slice at the time layer: (5) α=softmax(tanh(Yl+sWY+EtWE+b1)) \[\alpha =soft\max \left( \tanh \left( {{Y}_{l+s}}{{W}_{Y}}+{{E}_{t}}{{W}_{E}}+{{b}_{1}} \right) \right)\]

Where αϵR^(l+s), the learnable parameter dimension is denoted as follows: (6) WY∈R(l+s)×(N×dout)×1 \[{{W}_{Y}}\in {{R}^{\left( l+s \right)\times \left( N\times {{d}_{out}} \right)\times 1}}\] (7) WE∈Rde×(l+s) \[{{W}_{E}}\in {{R}^{{{d}_{e}}\times (l+s)}}\] (8) b1∈Rl+s \[{{b}_{1}}\in {{R}^{l+s}}\]

The above equation notation is to illustrate the dimension size of each parameter. d_out denotes the output dimension of the output convolutional features of the gated graph convolution module, W_Y denotes the weight parameter learnt for the input hidden layer vector Y_l+s (fused data), W_E denotes the weight parameter learnt for the moment to be predicted, and b₁ denotes the bias parameter learnt by the neural network.

The output of the temporal layer attention mechanism is accumulated from the time score results of each step, which are calculated as follows: (9) Yα=∑m=1l+sαmym∈RN×dout \[{{Y}_{\alpha }}=\sum\limits_{m=1}^{l+s}{{{\alpha }^{m}}}{{y}_{m}}\in {{R}^{N\times {{d}_{out}}}}\]

Where y_m is the result of slicing Y_l+s in the time dimension, Y_l+s=(y₁,…y_l+s), y_mϵY_l+s, and m refer to the indexes in the time dimension, and summation from 1 to l+s gives the result Y_α.

In the result Y_α obtained from the temporal layer attention mechanism, the attention mechanism is continued at the feature layer, analogous to the temporal attention approach, and is computed as follows: (10) β=softmax(tanh(YαWYF+EtWEF+b2)) \[\beta =soft\max \left( \tanh \left( {{Y}_{\alpha }}{{W}_{{{Y}_{F}}}}+{{E}_{t}}{{W}_{{{E}_{F}}}}+{{b}_{2}} \right) \right)\]

Where β denotes the feature layer attention mechanism score, βϵR^d_out, W_EF denotes the weight parameter learnt at the feature layer for input Y_α, W_EF denotes the weight parameter learnt at the feature layer for input E_t, and b₂ denotes the bias, and the dimensions of each parameter are set as follows: (11) WYF∈Rdout×N×1 \[{{W}_{{{Y}_{F}}}}\in {{R}^{{{d}_{out}}\times N\times 1}}\] (12) WEF∈Rde×dout \[{{W}_{{{E}_{F}}}}\in {{R}^{{{d}_{e}}\times {{d}_{out}}}}\] (13) b2∈Rdout \[{{b}_{2}}\in {{R}^{{{d}_{out}}}}\]

The preliminary results obtained from the final feature layer are shown in Eq. (14), where F_h denotes the vector of Y_α in the spatial feature dimension above, Y_α is a N×d_out vector that can be expressed as Y_α=(F₁,…F_dout), h refers to the index in the feature dimension, and h sums from 1 to d_out. Then: (14) Y=∑h=1aoutβhFh∈RN \[Y=\sum\limits_{h=1}^{{{a}_{out}}}{{{\beta }^{h}}}{{F}_{h}}\in {{R}^{N}}\]

Due to the large number of accident risk values in the sample that are zero, normal outputs will tend to predict all-zero values in order to reduce the error, a phenomenon known as the zero-inflation problem. In order to solve this problem, the scale reduction module is designed in this paper to combine the above results with the output.

The datasets used in this paper include three: traffic flow data, traffic accident data and weather data (rainfall and visibility). It is collected from open government data from June to December 2022 of a city. The processed raw dataset is split into training data, validation data and test data according to the ratio of 7:1:2, which is used for model training and validation.

In the experiments of this paper, the error functions of mean absolute error (MAE), mean squared error (MSE) and mean relative error (MRE) are used as the evaluation indexes of the algorithm. The smaller the value of the above three evaluation indexes, the more accurate the prediction results will be. This experiment predicts the risk of traffic accidents in the grid area, and the range of the predicted value is located between 0 and 1. When the predicted value is closer to 1, it means that the probability of traffic accidents in the area in the next time interval is greater.

This paper employs Linear Regression (LinR), Logistic Regression (LogR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Stack Noise Reduction Autocoding (SDAE to compare various algorithms.

The comparison of the experimental results of different models in the case where the input source is only traffic flow is shown in Figure 3. It can be found that (1) the prediction accuracy of deep learning models is overall higher than that of machine learning models. The reason for this is that the deep learning model is better at learning high-dimensional features and has a stronger learning ability than the machine learning model. (2) The prediction error of the SAGCN algorithm proposed in this paper is smaller than that of the SDAE model, and the values of its MAE, MSE, and MRE are 0.082, 0.038, and 0.808, respectively.

Figure 3.

Comparison of experimental results of different models

The proposed SAGCN model is combined with dynamic factors such as traffic accident data and weather data to carry out large-scale experiments to compare the impact of different influencing factors on the prediction of traffic accident risk in a city urban area.

A comparison of the experimental results based on various factors is shown in Table 1, where F-flow rate, A-number of traffic accidents, R-rainfall, and V-visibility are shown. The errors of the combination of flow, the number of historical accidents, rainfall and visibility are not the smallest, and the combination model of flow, the number of historical accidents, and rainfall has the smallest overall error, with MAE, MSE, and MRE of 0.075, 0.035, and 0.756, respectively, while the combination model of flow, visibility and rainfall has the largest error, with MAE, MSE, and MRE of 0.085, 0.044, and 0.851.

Considering single factors other than traffic flow, factors such as the number of historical accidents and rainfall are beneficial in reducing the error of the experimental results, while the visibility factor leads to an increase in the error of the experimental results. Therefore, it is not the case that the more combinations of factors are considered, the more favourable it is to reduce the prediction error of the model, and there are differences in the results of different combinations of factors. In this paper, rainfall and the number of accidents are favorable factors for reducing the prediction error of the risk of traffic accidents in the experiments.

Since different causal factors have different impacts in different scenarios, this section conducts model training for different scenarios, such as morning peak and flat peak, weekdays and holidays, and sunny and rainy days, in that order.

The comparison of the experimental results (MRE) for the morning peak and the flat peak is shown in Figure 4. The MRE of the prediction results for the morning peak is between 0.734 and 0.798, which are all lower than that of the flat peak.The prediction errors of various factors on the risk of traffic accidents are generally in line with those in the utility analysis mentioned earlier. There are more vehicles on the road during the morning peak.If a road traffic accident occurs, traffic congestion will be more serious than in the flat peak. Therefore, it is particularly important to accurately predict the risk of traffic accidents during the morning peak. The prediction errors of this paper’s model on the risk of traffic accidents in the morning peak are lower than those in the flat peak, which can help the personnel of the traffic management department to deploy the police to the high-risk areas of accidents in the morning peak in advance, and deal with the accidents in a timely manner, and can also help the drivers to avoid high-risk areas.

Figure 4.

The experimental comparison results of high peak and flat peak(MRE)

The effect of each different input on the model prediction error for weekdays and holidays was compared, and the comparison of the experimental results (MRE) for weekdays and holidays is shown in Figure 5. The model prediction MRE values for weekdays and holidays are 0.736~0.813 and 0.769~0.839, respectively. There are more vehicles on weekdays, and traffic congestion caused by road traffic accidents is more serious than that on ordinary holidays, so effective prediction of the risk of traffic accidents on weekdays is particularly important, and the prediction errors of this paper’s model on weekdays are lower than that of holidays, which can help the public to plan their routes to avoid high-risk areas. Citizens travelling on weekdays should plan their routes to avoid high-risk areas.

Figure 5.

The experimental comparison results of weekdays and holidays(MRE)

The comparison of experimental results (MRE) between sunny and rainy days is shown in Figure 6. The model will show slightly better prediction results on rainy days than on sunny days, and the MRE for the prediction of traffic accident risk on rainy days is 0.745~0.781, and the MRE on sunny days is 0.772~0.793.

Figure 6.

The experimental comparison results of sunny and rainy(MRE)

In summary, open government data such as traffic flow, weather data and traffic accident data can be used to predict the risk of traffic accidents using deep learning models to help the relevant personnel make decisions and plans to reduce the occurrence of traffic accident incidents.

Let the linear regression model of random variable y and general variable x₁,x₂,⋯x_p be: (19) y=β0+β1x1+β2x2+⋯+βpxp+ε \[y={{\beta }_{0}}+{{\beta }_{1}}{{x}_{1}}+{{\beta }_{2}}{{x}_{2}}+\cdots +{{\beta }_{p}}{{x}_{p}}+\varepsilon \]

Where β₀,β₁⋯,β_p is the p+1 unknown parameter, β₀ is the regression constant, and β₁⋯,β_p is the regression coefficient. Y are called the explanatory variables (dependent variables), and x₁,x₂,⋯x_p are p general variables that can be precisely measured and controlled, called explanatory variables (independent variables).

For a practical problem, if n a set of observations (x_i1,x_i2,⋯x_ip;y)(i=1,2⋯n) is obtained, the linear regression model can be expressed as: (20) y=Xβ+ε \[y=X\beta +\varepsilon \]

Where y=[ y1y2⋮y3 ]$y=\left[ \begin{array}{*{35}{l}} {{y}_{1}} \\ {{y}_{2}} \\ \vdots \\ {{y}_{3}} \\ \end{array} \right]$, X=[ 1x11⋯x1p1x21⋯x2p⋮⋮⋮⋮1xn1⋯xnp ]$X=\left[ \begin{matrix} 1 & {{x}_{11}} & \cdots & {{x}_{1p}} \\ 1 & {{x}_{21}} & \cdots & {{x}_{2p}} \\ \vdots & \vdots & \vdots & \vdots \\ 1 & {{x}_{n1}} & \cdots & {{x}_{np}} \\ \end{matrix} \right]$, β=[ β1β2⋮βp ]$\beta =\left[ \begin{matrix} {{\beta }_{1}} \\ {{\beta }_{2}} \\ \vdots \\ {{\beta }_{p}} \\ \end{matrix} \right]$, and ε=[ ε1ε2⋮εn ]$\varepsilon =\left[ \begin{matrix} {{\varepsilon }_{1}} \\ {{\varepsilon }_{2}} \\ \vdots \\ {{\varepsilon }_{n}} \\ \end{matrix} \right]$ are random errors.

Goodness of fit is used to test the fit of the regression equation to the sample observations, as reflected in the value of the sample coefficient of determination R2=SSRSST${{R}^{2}}=\frac{SSR}{SST}$.

The sum of squares decomposition formula is: (21) ∑i=1n(yi−y¯)2=∑i=1n(yi−y^i)2+∑i=1n(y^i−y¯)2 \[\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-\bar{y} \right)}^{2}}}=\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-{{{\hat{y}}}_{i}} \right)}^{2}}}+\sum\limits_{i=1}^{n}{{{\left( {{{\hat{y}}}_{i}}-\bar{y} \right)}^{2}}}\]

The regression sum of squares, SSR=∑i=1n(y^i−y¯)2$SSR=\sum\limits_{i=1}^{n}{{{\left( {{{\hat{y}}}_{i}}-\bar{y} \right)}^{2}}}$, reflects the magnitude of fluctuations in the n estimates of ŷ₁,ŷ₂,…,ŷ_n which is due to the fact that there is indeed a linear relationship between Y and the independent variable x₁,x₂,…,x_m and through the variation of x₁,x₂,…,x_m.

The residual sum of squares ∑j=1n(yi−y^i)2$\sum\limits_{j=1}^{n}{{{\left( {{y}_{i}}-{{{\hat{y}}}_{i}} \right)}^{2}}}$, which is caused by everything other than the linear relationship between x₁,x₂,…x_n and Y (including the non-linear relationship between x₁,x₂,…x_m and Y and random error).

The total sum of squared deviations SST=∑i=1n(yi−y¯)2$SST=\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-\bar{y} \right)}^{2}}}$ reflects the magnitude of the total fluctuations in Y the observations y₁,y₂,…y_n.

From the significance of the regression sum of squares and the residual sum of squares, it can be known that the greater the proportion of the regression sum of squares in the total sum of squares of deviations, the better the linear regression, which indicates that the regression straight line fits the sample observations better, and if the proportion of the residual sum of squares is large, the regression straight line does not fit the sample observations satisfactorily. The ratio of the sum of squared regressions SSR to the sum of squared total deviations SST is defined as the coefficient of determination: (22) (coefficient of determination)——R2=∑i=1n(y^i−y¯)2∑i=1n(yi−y¯)2 \[\text{ }\left( coefficient\text{ }of\text{ }\det er\min ation \right){{R}^{2}}=\frac{\sum\limits_{i=1}^{n}{{{\left( {{{\hat{y}}}_{i}}-\bar{y} \right)}^{2}}}}{\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-\bar{y} \right)}^{2}}}}\]

The coefficient of determination R² is a relative indicator of the goodness of fit of a regression line to the sample observations, reflecting the proportion of the variation in the dependent variable that the independent variable can explain. If the coefficient of determination R² is close to 1, it means that the regression equation can explain the vast majority of the uncertainty in the dependent variable, and the regression equation is a good fit; conversely, if R² is not too large, it means that the regression equation is not working well, and should be modified.

There exists complete multicollinearity, i.e., for the column vectors of the design matrix X there exists a set of numbers C₀,C₁,⋯C_p that are not all zero such that: (24) C0+C1xi1+⋯Cpxip+0,i=1,2,⋯n \[{{C}_{0}}+{{C}_{1}}{{x}_{i1}}+\cdots {{C}_{p}}{{x}_{ip}}+0,i=1,2,\cdots n\]

The rank rank(X)<p+1 of the design matrix X, at which point X′X|=0$X\prime X\left| =0 \right.$, the solution to the system of regular equations X′Xβ=X′y; is not unique, (X′X)⁻¹ does not exist, and the expression for the least squares estimate of the regression parameters β^=(X′X)−1X′y$\hat{\beta }={{\left( X\prime X \right)}^{-1}}X\prime y$ does not hold.

AMO theory is the ‘ability-motivation-opportunity’ theory, which believes that an individual’s ability, motivation and opportunity together determine their behavioural tendencies, while TPB theory (Theory of Planned Behaviour) believes that perceived behavioural control, behavioural attitudes and subjective norms together influence the emergence of behaviours. Combining AMO-TPB theory, this paper constructs the theoretical framework of AMO-TPB.

The research model of open government data utilisation is shown in Figure 7. The capability dimension contains perceived behavioural control and data literacy, i.e., the degree of control felt by the public when using government data and the public’s data awareness and capability. The motivation dimension contains behavioural attitudes and subjective norms, i.e., the public’s evaluation of the behaviour of utilising government data and the social pressure felt by the public to perform the utilisation behaviour. The opportunity dimension includes facility readiness, data quality and platform quality, i.e., the readiness of the infrastructure that the public has to support the utilisation behaviour, and the availability and ease of use of government data and open government data platforms. In addition, individual characteristics such as gender and age affect the consistency between individual willingness and actual behavior. Risk attitude is the individual’s tendency to choose between risky and safe options. In this paper, risk attitude refers to the public’s weighing and choosing between risky and safe options when facing whether to make use of government data or not, and it is also an important factor that affects the relationship between individual willingness and actual behaviour. Therefore, this paper argues that gender, age, and risk-taking attitude also have an impact.

Figure 7.

Open government data utilization model

The independent variables in this paper are 10 variables that fall under the dimensions of ability, motivation, opportunity, and individual characteristics, and the dependent variable is the utilization of open government data. Considering that the research object is the general public, the questionnaire needs to be widely applicable, so the questions are set in accordance with the principle of simplicity and comprehensibility. The total number of questionnaires returned was 525, with 483 of them being valid, with a validity rate of 92%.The survey sample is representative of different provinces, age groups, and genders in the PRD region.The collected data passed the reliability and validity tests.