An index selection method of coupling boruta algorithm and correlation analysis

Abstract

This invention discloses a method for selecting indicators that couples the Boruta algorithm with correlation analysis. The method includes: S1, preprocessing existing data to generate a dataset containing a dependent variable and multiple independent variables to be selected; S2, ranking the n indicators in the dataset by Boruta importance and deleting unimportant indicators; S3, performing correlation analysis on the remaining indicators and setting multiple thresholds r for correlation coefficients; S4, for each threshold r, starting with the most important indicator, deleting all other indicators whose correlation coefficient is greater than the selected threshold; selecting the second most important indicator that was not deleted, and repeating this step until the last indicator is selected, with the deleted indicators forming the indicator set corresponding to the threshold r; S5, using the selected indicator set corresponding to each threshold as parameters to build a model; S6, selecting the final indicator set based on the model accuracy. This method has the advantages of selecting important indicators and avoiding multicollinearity problems.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to an index selection method that couples the Boruta algorithm with correlation analysis. Background Technology

[0002] When building regression models using machine learning methods (such as artificial neural networks and random forests), it is necessary to select an appropriate indicator system. Two common problems exist in indicator selection: first, the issue of indicator importance, meaning that important indicators need to be selected from the candidate indicators as model input; second, avoiding multicollinearity, meaning that if some indicators are too correlated, it may lead to overfitting. As a result, although the model has high training accuracy, its error on the test set may be too large. Therefore, it is necessary to remove some indicators based on their correlation with the candidate indicators.

[0003] However, there is no clear method for determining the importance of indicators, or for deciding at what level of correlation between indicators should some be removed, which leads to significant human influence in the indicator selection process. Summary of the Invention

[0004] The purpose of this invention is to provide an indicator selection method that couples the Boruta algorithm with correlation analysis, thereby solving the technical problems of which method to use to determine the importance of indicators during the indicator selection process, and at what level the correlation coefficient between indicators should be high enough to delete some indicators.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following solution:

[0006] A method for selecting indicators that couples the Boruta algorithm with correlation analysis includes the following steps:

[0007] Step S1: Preprocess the existing data of the research target to generate a variable y containing the dependent variable y and multiple independent variables x (x1, x2, ... x3) to be selected. n The dataset;

[0008] Step S2: Sort the n indicators in the independent variable dataset x by Boruta importance and delete the unimportant indicators obtained by the Boruta algorithm.

[0009] Step S3: Perform correlation analysis on the remaining indicators and set thresholds r for multiple correlation coefficients;

[0010] Step S4: For each correlation coefficient threshold r, starting with the most important indicator, delete all other indicators whose correlation coefficient is greater than the selected threshold; then select the second most important indicator that has not been deleted, and continue to delete other indicators according to the threshold; filter in this way until the last indicator is filtered, and the deleted indicators constitute the indicator set corresponding to the threshold r.

[0011] Step S5: Repeat step S4, generating an indicator set for each threshold; summarize all indicator sets into an indicator library;

[0012] Step S6: Use the set of indicators selected for each threshold as parameters to build a model;

[0013] Step S7: Train the model using the data collected in step S1, and select the final set of indicators based on the model's accuracy.

[0014] Further optimization involves using the Boruta index screening method to sort the n indicators in step S2, specifically including the following steps:

[0015] Step S2.1: The feature values ​​of each indicator in the dataset of independent variable x constitute the original features. Randomly arrange all the feature values ​​of the indicators to form the corresponding shadow features. Concatenate the original features and shadow features to form a new feature matrix.

[0016] Step S2.2: Use the new feature matrix as input to train and output a model of indicator importance;

[0017] Step S2.3: Calculate the Z-score of each index in the original feature and the corresponding shadow feature;

[0018] Step S2.4: Find the largest Z-score among the shadow features and denote it as Zmax;

[0019] Step S2.5: Mark the indicators with Z-score greater than Zmax as "important", mark the indicators with Z-score significantly less than Zmax as "not important", and remove them from the feature set;

[0020] Step S2.6: Delete all shadow features;

[0021] Step S2.7: Repeat steps S2.1 to S2.6 until all indicator features are marked as "important" or "not important".

[0022] To further optimize, in step S3, correlation analysis is performed on the remaining indicators. Specifically, Pearson regression or Spearman method is used to calculate the correlation coefficient of any two remaining indicators and form a correlation coefficient (r) matrix.

[0023] To further optimize, in step S3, a total of M correlation thresholds K are set. i K i =K i-1 +G, i∈[1,M], where G is a constant.

[0024] To further optimize this process, in step S6, a model is constructed using random forests and / or artificial neural networks.

[0025] Further optimization involves step S8, where the model is trained based on the data collected in step S1 to obtain predicted values. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are then calculated using the predicted values ​​and the actual collected data. 2 Then, cross-validation is used to evaluate the accuracy of the model.

[0026] To further optimize this process, in step S8, during cross-validation, the total data is divided into a training set and a validation set according to a certain ratio. Because the selected data has randomness, the calculation can be repeated multiple times to obtain the average value.

[0027] Chinese patent application CN115376011 A discloses a remote sensing method for forest aboveground biomass measurement based on an improved Boruta algorithm. The key to this patent application is replacing the random forest model iteration in the Boruta algorithm with XGboost iteration. The difference between this invention and the aforementioned prior application is that it first uses the Boruta algorithm to screen indicators, then combines correlation analysis with the selected indicators, and summarizes the screened indicators into an indicator library based on different thresholds. Then, it uses a model (such as ANN, RF, or others) to back-calculate the indicator library under each threshold, and selects the best indicator scheme.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The index selection method for coupling the Boruta algorithm and correlation analysis described in this invention has the advantages of selecting important indicators and avoiding multicollinearity problems. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the process for selecting indicators and simulating soil thickness in Example 1, which involves coupling the Boruta algorithm with correlation analysis.

[0031] Figure 2 This is a three-dimensional structure and soil thickness distribution map of the D5KH1 hillside in Example 1;

[0032] Figure 3 This is a schematic diagram illustrating the ranking of indicator importance using the Boruta algorithm in Example 1.

[0033] Figure 4 This is a comparison chart of the evaluation results of the two models under three indicator selection methods; among them, Figure 4 (a) Comparison of the results for the coefficient of determination R². Figure 4 (b) Comparison of the results of root mean square error (RMSE) Figure 4 (c) Comparison of the results of mean absolute error (MAE). Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] like Figure 1 As shown, this invention employs three index selection methods to simulate soil thickness. The study area is located in the Duodigou watershed in the northern suburbs of Lhasa. The specific research target is a source-type zero-order watershed within this watershed, named D5KH1. Figure 2 As shown, the area of ​​D5KH1 is 0.15 km². 2 The elevation ranges from 4870 to 5276 meters, with an average slope of 30.9°. The slopes are gentler in the lower valleys, but steeper in other areas. The vegetation type within D5KH1 is alpine meadow. A DJI Phantom 4 Pro drone was used for terrain surveying. Images were stitched together using Pix4dmapper software to generate orthophotos and a high-precision digital elevation model (DEM). The initial DEM resolution was 0.07 m, and Kriging interpolation was used to generate DEMs at six resolutions: 0.25, 0.50, 0.75, 1.00, 2.00, and 5.00 m. Soil thickness was measured using a steel rod. Samples were taken 2-3 times at each point, and the average value was used as the soil thickness value for that point. A total of 277 soil thickness samples were collected. Figure 2 The red dots indicate test points.

[0036] For each DEM resolution, 16 topographic parameters are extracted, including: elevation (m), slope (radian), and horizontal curvature (m). -1 ), Profile curvature (Profilec,m) -1 ), General curvature (Generalc,m) -1 Aspect (radian), convergent index (CI), and upslope attributing area (UAA, m²)2 ), specific catchment area (SCA, m²) 2 The topographic indicators include: topography wetness index (TWI), stream power index (SPI), slope-length factor (LS-factor, LS), topographic position index (TPI), valley depth (VD), terrain ruggedness index (TRI), and vertical distance to channel network (VDCN). In this embodiment, saga-9.1.1 software was used to extract the specific values ​​of these 16 topographic indicators.

[0037] Example 1:

[0038] This embodiment uses a method of selecting indicators by coupling the Boruta algorithm and correlation analysis to simulate soil thickness. The specific steps are as follows:

[0039] Step S1: Collect soil thickness and corresponding topographic index data at 277 points; that is, at each resolution, 16 topographic indicators are used as independent variables and soil thickness is used as dependent variable.

[0040] Step S2: The importance of the 16 indicators is ranked by combining the Boruta index screening and correlation analysis methods, and is named "RF+COR" in the following text.

[0041] In this embodiment, taking 0.5m resolution as an example, 16 terrain indicators are extracted and used to form the original features. All indicators are randomly arranged to form the corresponding shadow features. The original features and shadow features are concatenated to form a new feature matrix. The new feature matrix is ​​used as input to train and output a model of indicator importance. The Z-score of each indicator in the original features and the corresponding shadow features is calculated. Figure 3 The indicators marked in blue according to their importance are shadow feature indicators, namely ShadowMin, ShadowMean, and ShadowMax, with the highest Z-score corresponding to ShadowMax. Indicators with less importance than ShadowMax are marked as "unimportant," such as... Figure 3 The three indicators marked in red for medium importance—Profile curvature, Planc, and Generalc—should be deleted, leaving the remaining 13 topographic indicators.

[0042] Step S3: Using the Pearson correlation method, the pairwise correlation coefficients of the 13 topographic indicators were calculated, and the results are shown in Table 1. The Pearson correlation coefficient between two indicators is equal to the product of their covariance and their respective standard deviations. The coefficient value always ranges from -1.0 to 1.0. Variables close to 0 are considered uncorrelated, while those close to 1 or -1 are considered strongly correlated. The Pearson correlation method is an existing technique and will not be elaborated further.

[0043] Table 1. Pairwise correlation coefficients of 13 topographic indicators

[0044] Elevation 1.00 -0.59 -0.35 0.21 -0.24 -0.22 -0.09 -0.33 -0.31 0.41 -0.57 -0.63 0.91 Slope -0.59 1.00 0.40 -0.29 0.21 0.21 -0.03 0.40 0.17 -0.31 0.97 0.26 -0.67 Aspect -0.35 0.40 1.00 -0.29 0.22 0.22 0.10 0.31 0.09 -0.16 0.36 0.21 -0.29 CI 0.21 -0.29 -0.29 1.00 -0.74 -0.73 -0.67 -0.74 -0.56 0.65 -0.25 -0.41 0.24 UAA -0.24 0.21 0.22 -0.74 1.00 1.00 0.96 0.98 0.85 -0.81 0.18 0.39 -0.35 SCA -0.22 0.21 0.22 -0.73 1.00 1.00 0.97 0.98 0.85 -0.81 0.17 0.38 -0.34 TWI -0.09 -0.03 0.10 -0.67 0.96 0.97 1.00 0.89 0.83 -0.76 -0.05 0.34 -0.19 SPI -0.33 0.40 0.31 -0.74 0.98 0.98 0.89 1.00 0.82 -0.81 0.36 0.40 -0.46 LS -0.31 0.17 0.09 -0.56 0.85 0.85 0.83 0.82 1.00 -0.71 0.16 0.40 -0.37 TPI 0.41 -0.31 -0.16 0.65 -0.81 -0.81 -0.76 -0.81 -0.71 1.00 -0.29 -0.64 0.51 TRI -0.57 0.97 0.36 -0.25 0.18 0.17 -0.05 0.36 0.16 -0.29 1.00 0.25 -0.64 VD -0.63 0.26 0.21 -0.41 0.39 0.38 0.34 0.40 0.40 -0.64 0.25 1.00 -0.50 VDCN 0.91 -0.67 -0.29 0.24 -0.35 -0.34 -0.19 -0.46 -0.37 0.51 -0.64 -0.50 1.00

[0045] Step S4: In this embodiment, a total of 15 correlation coefficient thresholds r are set, which are calculated every 0.05 from 0.2 to 0.9, r = 0.20, 0.25, ..., 0.85, 0.90.

[0046] Step S5: For each correlation coefficient threshold r, starting with the most important indicator, the topographic position index (TPI), delete other indicators with correlation coefficients greater than the selected threshold. Then select the second most important indicator that was not deleted, and continue to delete other indicators according to the threshold. Repeat this process until the last indicator is selected. The deleted indicators constitute the indicator set corresponding to the threshold r. Repeat this step 15 times. Each threshold corresponds to an indicator set. Summarize the 15 indicator sets into an indicator library, as shown in Table 2. The larger the threshold, the more indicators are selected.

[0047] Table 2. Index library consisting of the index sets corresponding to the 15 thresholds.

[0048]

[0049] Due to space limitations, this section will only use a threshold r = 0.8 to illustrate the specific selection process. From Figure 3 The most important indicator is TPI. According to Table 1, indicators with a correlation coefficient greater than 0.8 in the TPI row / column will be deleted, namely UAA, SCA, and SPI. Next, since the second most important indicator LS was not deleted in the previous selection, we search for the row / column with LS and delete indicators with a correlation coefficient greater than 0.8, namely TWI. We continue this process until the last indicator is selected.

[0050] The remaining indicators are: VD TPI, LS, CI, Aspect, Slope, and Elevation. These seven indicators constitute the indicator set corresponding to the threshold of 0.8.

[0051] Step S6: Use the selected index set corresponding to each threshold in Table 2 as parameters to construct a soil thickness prediction model. In this embodiment, for each index set, two soil thickness prediction models based on Random Forest (RF) and Artificial Neural Network (ANN) are constructed at each resolution.

[0052] The basic unit of a random forest is a decision tree. Multiple decision trees are built and merged, and the prediction results of each tree in the forest for a given sample are statistically analyzed. Then, a voting method is used to select the final result from these predictions. This invention uses the "randomForest" package in R to write a random forest model, and the tuneRF function is used to automatically optimize the mtry parameter (the number of variables randomly sampled when building the decision tree branches).

[0053] Artificial neural network (ANN) models are inspired by the learning process of the human brain and contain many processing neurons, which are typically classified into different layers: input layer, output layer, and hidden layer. Neurons connect to neurons in adjacent layers but not to neurons in the same layer. ANNs can effectively solve nonlinear problems. This invention uses the "neuralnet" package in the R language to run the neural network model. The number of neurons in the hidden layer is crucial in the model settings because it involves 16 variables. In this embodiment, neural network models with 1 to 16 hidden layer neurons were constructed, and the results were run at six different resolutions. It was found that the model performed best when the number of hidden layers was 2. Therefore, the number of neurons in the ANN model used in this embodiment is uniformly set to 2.

[0054] Step S7: After the model is built, 90% of the 277 points in step S1 are used as the training set and 10% as the validation set to train the model, and the final metric set is selected based on the model accuracy.

[0055] Since 90% of the data is randomly sampled each time (there may be discrepancies between two calculations), each model is run 1000 times and the average is taken as the final calculation result. This invention uses the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 To evaluate the model prediction performance of DEMs at different resolutions.

[0056]

[0057]

[0058]

[0059] Among them, X samp It is the actual soil thickness measured in the field, X pred N is the soil thickness predicted by the model, and N is the number of selected points.

[0060] Due to space limitations, this embodiment only shows the root mean square error (RMSE) of the running results of the RF and ANN models. The calculation results are combined for six resolutions from 0.25 to 5.00. Table 3 shows the RMSE of the running results of the RF model, and Table 4 shows the RMSE of the running results of the ANN model. It can be seen that the model performs best when the threshold r = 0.90.

[0061] Table 3 shows the root mean square error results of the RF model.

[0062] 0.20 28.28 28.19 28.56 28.64 28.70 30.47 0.25 28.18 28.32 28.10 28.67 27.18 30.71 0.30 29.15 28.15 28.16 28.42 26.96 30.48 0.35 27.69 28.20 25.63 24.98 27.02 29.06 0.40 27.65 27.61 25.86 25.66 26.71 27.76 0.45 26.71 26.89 25.72 25.58 26.74 27.82 0.50 26.94 26.91 25.64 25.77 27.19 27.89 0.55 27.80 26.75 26.36 25.64 27.38 27.93 0.60 27.60 27.19 26.20 26.21 27.51 27.81 0.65 27.63 26.89 25.67 26.02 27.44 27.93 0.70 26.41 26.65 25.62 25.80 26.73 27.37 0.75 26.20 25.28 26.02 25.75 26.66 27.11 0.80 26.31 25.27 25.55 25.93 26.59 27.05 0.85 26.12 25.34 25.76 25.76 26.45 27.05 0.90 26.20 25.20 25.61 25.82 26.26 26.04

[0063] Table 4 shows the root mean square error results of the ANN model.

[0064] 0.20 28.89 28.86 28.44 28.26 28.82 28.87 0.25 28.87 28.78 28.18 28.31 28.42 28.78 0.30 28.27 28.84 28.29 28.40 28.33 28.91 0.35 28.85 28.85 28.37 28.12 28.13 28.62 0.40 28.45 28.82 28.35 28.38 28.37 28.58 0.45 26.87 26.69 28.44 27.85 28.26 28.58 0.50 26.40 26.91 28.50 27.64 28.13 28.77 0.55 27.22 26.87 28.36 27.74 28.41 28.78 0.60 27.41 26.05 28.54 28.06 28.38 28.79 0.65 27.33 26.27 28.30 28.48 28.56 28.69 0.70 27.41 26.34 28.15 28.34 29.05 29.00 0.75 27.16 26.37 26.34 26.39 28.99 29.00 0.80 27.27 26.44 26.23 26.57 28.96 29.10 0.85 26.76 26.39 26.13 26.66 28.98 29.09 0.90 26.74 26.20 26.39 26.29 26.02 26.11

[0065] Comparative Example 1:

[0066] In this comparative example, the difference from Example 1 is that the indicators are not deleted; they are referred to as "All" below, meaning all 16 indicators are used as parameters to construct the soil thickness prediction model. Two soil thickness prediction models are constructed for each resolution, one based on Random Forest (RF) and the other on Artificial Neural Network (ANN). The 277 points from step S1 are divided into a training set (90%) and a validation set (10%) to train the model. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used to evaluate the model. 2 To evaluate the model prediction performance of DEMs at different resolutions.

[0067] Comparative Example 2:

[0068] In this comparative example, unlike Example 1, a stepwise regression model was used to screen the indicators, hereinafter referred to as "Stepwise". During the screening, all 277 data points were used as the training set. SPSS software (version 22) was used to select 16 indicators using a stepwise regression model, automatically calculating the relationship between soil thickness and each selected indicator at various resolutions, as shown in Table 5. Using a stepwise regression model for screening is existing technology and will not be described in detail here.

[0069] The other steps are the same as in Example 1. Two soil thickness prediction models are constructed for each resolution, one based on Random Forest (RF) and the other on Artificial Neural Network (ANN). The 277 points from step S1 are divided into a training set (90%) and a validation set (10%). The models are trained using the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 To evaluate the model prediction performance of DEMs at different resolutions.

[0070] Table 5 shows the results of the indicator selection using the stepwise regression model.

[0071] 0.25 SD=-2.183×TPI+0.205×LS+4.497×SPI+0.613×CI+78.888 0.56 <0.001 0.50 SD=-2.164×TPI+0.204×LS+5.786×SPI+0.830×CI+72.923 0.57 <0.001 0.75 SD=-3.846×TPI+0.325×LS-1.807×VD-0.107×Ele-67.734×TRI+670.471 0.59 <0.001 1.00 SD=-3.479×TPI+0.366×LS-0.980×VD-0.118×VDCN+107.273 0.56 <0.001 2.00 SD=-1.940×TPI+8.294×SPI+104.117×Planc+70.258 0.55 <0.001 5.00 SD=37.119×SPI-1.444×TPI+0.189×VD+34.416×TWI-64.487×UAA+180.197 0.57 <0.001

[0072] In Table 5, SD refers to soil thickness, and R... 2 This represents the coefficient of certainty.

[0073] Results comparison and analysis:

[0074] A comparison of the results of the RF and ANN models under the three index selection methods involved in Example 1, Comparative Example 1, and Comparative Example 2. Figure 4 As shown; where, Figure 4 (a) represents the deterministic coefficient RF of the RF and ANN models under the three selection methods. 2 The comparison Figure 4 (b) Comparison of the root mean square error (RMSE) of the RF and ANN models under the three selection methods. Figure 4 (c) Comparison of the mean absolute error (MAE) of the RF and ANN models under the three selection methods.

[0075] When the coefficient of determination R 2 The simulation results are best when the overall values ​​are maximized, and the overall values ​​of root mean square error (RMSE) and mean absolute error (MAE) are minimized. Therefore, from... Figure 4 It is clear from the results that the Random Forest (RF) model achieved the best simulation results. The order of merit for the three metric selection schemes under the Random Forest model is: RF + COR > All > Stepwise. Furthermore, the results from the Artificial Neural Network (ANN) model also show that the order of merit for the three metric selection schemes is: RF + COR > All > Stepwise.

[0076] In summary, the index selection method that couples the Boruta algorithm with correlation analysis achieves the best results.

[0077] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for selecting indicators that couples the Boruta algorithm with correlation analysis, characterized in that, The steps used to simulate soil thickness include the following: Step S1: Preprocess the existing data of the research target to generate a variable y containing the dependent variable y and multiple independent variables x (x1, x2, ... x3) to be selected. n The dataset contains n, representing the number of topographic indicators, with soil thickness as the dependent variable. Step S2: Sort the n indicators in the independent variable dataset x by Boruta importance and delete the unimportant indicators obtained by the Boruta algorithm. Step S3: Perform correlation analysis on the remaining terrain indicators and set thresholds r for multiple correlation coefficients; Step S4: For each correlation coefficient threshold r, starting with the most important indicator, the terrain location index, delete all other indicators whose correlation coefficient is greater than the selected threshold; then select the second most important indicator that has not been deleted, and continue to delete other indicators according to the threshold; filter in this way until the last indicator is filtered, and the deleted indicators constitute the indicator set corresponding to the threshold r. Step S5: Repeat step S4, generating an indicator set for each threshold; summarize all indicator sets into an indicator library; Step S6: Use the set of indicators selected for each threshold as parameters to construct a soil thickness prediction model; Step S7: Train the model using the data collected in step S1, and select the final set of indicators based on the model's accuracy.

2. The index selection method for coupling Boruta algorithm and correlation analysis according to claim 1, characterized in that: Step S2 uses the Boruta index screening method to sort the n indicators, specifically including the following steps: Step S2.1: Each indicator in the dataset of independent variable x constitutes the original feature. Randomly arrange all indicators to form the corresponding shadow feature. Concatenate the original feature and the shadow feature to form a new feature matrix. Step S2.2: Use the new feature matrix as input to train and output a model of indicator importance; Step S2.3: Calculate the Z-score of each index in the original feature and the corresponding shadow feature; Step S2.4: Find the largest Z-score among the shadow features and denote it as Zmax; Step S2.5: Mark the indicators with Z-score greater than Zmax as "important", mark the indicators with Z-score significantly less than Zmax as "not important", and remove them from the feature set; Step S2.6: Delete all shadow features; Step S2.7: Repeat steps S2.1 to S2.6 until all indicator features are marked as "important" or "not important".

3. The index selection method for coupling Boruta algorithm and correlation analysis according to claim 2, characterized in that: In step S3, correlation analysis is performed on the remaining indicators. Specifically, Pearson regression or Spearman method is used to calculate the correlation coefficient between any two remaining indicators, forming a correlation coefficient matrix.

4. The index selection method for coupling Boruta algorithm and correlation analysis according to claim 3, characterized in that: In step S3, a total of [number] settings are made. M Threshold of correlation K i , K i = K i-1 +G , i ∈[1, M ], G It is a constant value.

5. The index selection method for coupling Boruta algorithm and correlation analysis according to claim 4, characterized in that: In step S6, a model is constructed using random forest and / or artificial neural networks.

6. The index selection method for coupling Boruta algorithm and correlation analysis according to claim 4, characterized in that: In step S7, the model is trained based on the data collected in step S1 to obtain predicted values. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are then calculated using the predicted values ​​and the actual collected data. 2 Then, cross-validation is used to evaluate the accuracy of the model.

7. The index selection method for coupling Boruta algorithm and correlation analysis according to claim 4, characterized in that: In step S7, during cross-validation, the total data is divided into a training set and a validation set according to a certain ratio.

Images