Logistics industry green innovation performance evaluation method based on multi-modal data fusion

Abstract

The invention discloses a logistics industry green innovation performance evaluation method based on multi-modal data fusion, and the method comprises the steps: integrating enterprise operation data, social media public opinion and government policy texts based on a TOE framework, and constructing a multi-modal database; normalizing the numeric data, the text data and the image data respectively; constructing an index system: constructing a VIF-FA double screening method to screen indexes, and establishing a TOE three-dimensional evaluation system; aHP subjective weight and CRITIC method objective weight are fused based on the game theory, performance is calculated by constructing a DEA-TOPSIS two-stage model, and the first stage DEA calculates an enterprise comprehensive efficiency value; in the second stage, the DEA efficiency value serves as a newly-added index to extend a TOPSIS decision matrix, and a performance value is output in combination with the optimal combination weight. According to the method, the problem of singleness of traditional evaluation data is improved, subjective deviation and index redundancy are avoided, the index weighting scientificity is improved, and the quantitative logistics industry green innovation performance accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of logistics performance evaluation technology, and in particular to a method for evaluating the green innovation performance of the logistics industry based on multimodal data fusion. Background Technology

[0002] Current assessments of green innovation performance in the logistics industry are primarily limited to corporate financial data and government statistical annual reports, lacking multimodal data such as media sentiment and policy documents. This incomplete data coverage leads to a somewhat biased assessment perspective. Traditional indicator systems rely heavily on expert experience, easily introducing subjective bias. Multicollinearity and low-information indicators have not been systematically eliminated, resulting in reduced model interpretability. Existing weighting methods neglect the interplay between subjective and objective weights. Considering the issues of insufficient data sources, unsystematic indicator selection, and weighting deviations from reality, assessing the green innovation performance of the logistics industry based on multimodal data has become a research hotspot. A method for quantifying, screening, and weighting multimodal data is urgently needed for evaluating the green innovation performance of the logistics industry.

[0003] With the continuous development of the logistics industry and the increasing diversity of data formats, traditional performance evaluation methods can no longer meet current needs. A new, more efficient, and accurate evaluation method is emerging, highlighting its advantages. This green innovation performance evaluation method for the logistics industry, based on multimodal data fusion, ensures comprehensive data collection, accurate selection, scientific weighting, and reasonable and reliable evaluation, from data collection to indicator selection, game-theoretic weighting, and finally, performance evaluation completion. Summary of the Invention

[0004] This invention provides a method for evaluating the green innovation performance of the logistics industry based on multimodal data fusion. It rationally quantifies multimodal data, scientifically selects indicators and constructs an evaluation indicator system, and can accurately measure the green innovation performance of the logistics industry under multimodal data.

[0005] The technical solution adopted in this invention is: a performance evaluation method for green innovation in the logistics industry based on multimodal data fusion, comprising the following steps:

[0006] (1) Multimodal data collection and preprocessing: Based on the technology-organization-environment (TOE) framework, data is collected from the enterprise, social media and government policy aspects, and the data is preprocessed to build a multimodal database;

[0007] (2) Multimodal data standardization: Normalization processing is performed on numerical data, text data and image data;

[0008] (3) Constructing an evaluation index system: Based on the VIF-FA dual mechanism, key indicators are screened and a three-dimensional TOE evaluation system is constructed.

[0009] (4) Combination weighting and performance measurement: The game theory-based combination weighting method integrates subjective and objective weights, and calculates green innovation performance based on the weighting results through the DEA-TOPSIS combination model.

[0010] Based on the above scheme, as a preferred embodiment, step (2) includes the following steps:

[0011] (21) Numerical data are normalized using Min-Max and Max-Min methods;

[0012] (22) Text data is first vectorized using TF-IDF and then normalized using L2.

[0013] (23) Image data is first processed in grayscale, and the semantics of green elements are extracted from the grayscale image. Then, the range is mapped to [0, 1].

[0014] Based on the above scheme, as a preferred embodiment, step (3) includes the following steps:

[0015] (31) Based on the multicollinearity test, the index set I is screened. First, the coefficient of determination is calculated. Calculate the variance inflation factor (VIF) of index variable i. i ;

[0016] (32) Factor analysis was performed on the index set I after the VIF test to obtain the final key index set;

[0017] (33) After the indicators have been screened, they are further classified into three dimensions: technology, organization and environment, based on the three categories of government, society and enterprise, and a three-dimensional TOE evaluation system is constructed.

[0018] Based on the above scheme, as a preferred option, the specific steps of step (32) are as follows:

[0019] Calculate the correlation coefficient matrix R, and the elements r in the correlation coefficient matrix R. pq (p,q=1,2,…,m), representing index i p With index i q The correlation coefficient between them is calculated using the following formula:

[0020]

[0021] In the formula, N is the sample size, x ps For the index i in the s-th sample p The value, For index i p The sample mean

[0022] Solving for eigenvalues ​​and eigenvectors: Solving for the characteristic equation of the correlation coefficient matrix R: |R-λE m If |=0, then m eigenvalues ​​are obtained: λ1≥λ2≥…≥λ m ≥0 and the corresponding eigenvectors u1, u2, ..., u m E m It is an m-order identity matrix.

[0023] Determine the number of common factors: Calculate the contribution rate C of each eigenvalue. j and cumulative contribution rate AC j The formulas are as follows:

[0024]

[0025] Set a cumulative contribution rate threshold T AC Choose one that makes AC j ≥T AC Let the minimum value of j be denoted as k, then the number of common factors is k.

[0026] The factor loading matrix A is calculated using the following formula:

[0027]

[0028] In the formula, u jp It is the eigenvector u j The p-th component,

[0029] The initial factor loading matrix A is rotated using an orthogonal rotation method to obtain the rotated factor loading matrix A. * Let the rotation matrix be T, satisfying A * =AT, the objective function for maximum variance rotation is:

[0030]

[0031] Solve for the rotation matrix T such that the objective function Q reaches its maximum value.

[0032] For each common factor F j Set the factor loading threshold T A Select factors whose absolute value of factor loadings on the common factor is greater than T. A We select the key metrics, retain these metrics, and remove the others to obtain the final set of key metrics.

[0033] Based on the above scheme, as a preferred embodiment, step (4) includes the following steps:

[0034] (41) Subjective weights ω are generated using the AHP method. A The objective weights ω are generated using the CRITIC method. C ;

[0035] (42) The weight ω of the index combination expressed by the linear combination is:

[0036] ω=λ1ω A +λ2ω C

[0037] Based on the principles of game theory, an objective function is established, combining weights ω and ω A and ω C The objective is to minimize the sum of deviations between the coefficients, and to find the optimal linear combination coefficients. and At this point, the combined weight is the optimal combined weight ω. * The objective function and constraints are as follows:

[0038] min(||ω-ω A || 2 +||ω-ω C || 2 )

[0039] =min(||λ1ω) A +λ2ω C -ω A || 2 +||λ1ω A +λ2ω C -ω C || 2 )

[0040]

[0041] The system of equations is as follows:

[0042]

[0043] The obtained linear combination coefficients λ1 and λ2 are standardized:

[0044]

[0045] Finally, the optimal combination weight ω of the evaluation indicators is obtained. * for:

[0046]

[0047] (43) First-stage efficiency assessment:

[0048] The set of input variables X and the set of output variables Y are determined based on the selected indicators;

[0049] The formula for calculating the overall efficiency value of the kth enterprise is as follows:

[0050] max θ k

[0051]

[0052] Output overall efficiency value θ k ∈[0,1]

[0053] In the formula, θ k Y represents the overall efficiency value of the k-th firm, where m is the total number of decision-making units (i.e., the number of sample firms). rl Let X be the actual value of the l-th firm on the r-th output indicator. ol Y represents the actual value of the l-th enterprise on the o-th input indicator. rk Let X be the actual value of the k-th firm on the r-th output indicator. ok Let μ be the actual value of the k-th enterprise on the o-th input indicator. b Let be the weight coefficient of the b-th enterprise.

[0054] The second stage involves constructing the TOPSIS decision matrix and setting the DEA efficiency value θ. k As a new indicator, it together with the original TOE evaluation indicators forms an extended decision matrix.

[0055]

[0056] Calculate the weighting matrix:

[0057]

[0058] In the formula, It is the optimal combination weight of the j-th indicator, element a ij (1 << j << p) represents the value of index j of the i-th enterprise in the locality.

[0059] Determine the positive ideal solution R : and negative ideal solution R ； Then calculate the Euclidean distance.

[0060] Distance to the ideal solution:

[0061] This indicates the degree of deviation of the firm from the optimal state.

[0062] Distance to the negative ideal solution:

[0063] This indicates the degree of deviation of the firm from its worst-case scenario.

[0064] Calculate the green innovation performance value: C i A higher value indicates better green innovation performance of the logistics company.

[0065] Compared with the prior art, the present invention has the following technical advantages:

[0066] 1. In view of the problem that traditional methods rely on a single data source, this invention is based on the TOE framework and collects data from multiple dimensions, including enterprises, society, and government. By integrating multimodal data, a more comprehensive multimodal database is constructed, avoiding performance evaluation bias caused by single data sources.

[0067] 2. By analyzing traditional performance evaluation methods, it was found that indicator selection relies heavily on expert experience, which is highly subjective and does not systematically address the issues of multicollinearity and information redundancy among indicators. In order to select indicators more accurately, this invention constructs a VIF-FA dual indicator selection mechanism to eliminate multicollinearity and retain indicators with high core information content, thereby making the indicator system more hierarchical.

[0068] 3. This invention introduces game theory into the subjective and objective weighting process, avoiding subjective bias and neglect of actual conditions, thus making the subjective and objective weighting process more balanced with experience and data. Attached Figure Description

[0069] Figure 1 This is a flowchart of the present invention;

[0070] Figure 2 This is a flowchart of the data acquisition and processing process;

[0071] Figure 3 This is a diagram illustrating the overall framework for logistics performance evaluation. Detailed Implementation

[0072] To better understand this invention, a more detailed description of a performance evaluation method for green innovation in the logistics industry based on multimodal data fusion is provided below with reference to the accompanying drawings in the embodiments of this invention. Figure 1 It can be seen that, as Figure 1-3 As shown, the specific steps of the present invention are as follows:

[0073] Step 1: Multimodal data acquisition and preprocessing:

[0074] Based on the TOE (Technology-Organization-Environment) framework, data on logistics operations, social feedback on logistics, and government policies are collected from enterprise, social media, and government policy perspectives. Data preprocessing, including data cleaning, missing value imputation, and outlier removal, is then performed to construct a multimodal database.

[0075] like Figure 2 As shown, the diagram is a flowchart of data acquisition and processing. The main steps include multimodal data acquisition, preprocessing, and normalization of various data types.

[0076] Step 2: Multimodal data standardization processing:

[0077] Normalize numerical, textual, and image data.

[0078] (1) Numerical data are normalized using Min-Max and Max-Min methods. The normalization formula for positive indicators is:

[0079]

[0080] The formula for normalizing negative indices is:

[0081]

[0082] In the formula, X represents the actual value; max(X) represents the maximum value; and min(X) represents the minimum value.

[0083] (2) Text data is first vectorized using TF-IDF, and the formula is:

[0084]

[0085] TF-IDF = TF*IDF

[0086] In the formula, TF represents word frequency; WN represents the number of times a word appears in the article; TWN represents the total number of words in the article; IDF represents inverse document frequency; TAN represents the total number of articles; and TWAN represents the number of articles containing the word.

[0087] Then, the vectorized TF-IDF is subjected to L2 normalization, the formula of which is:

[0088]

[0089] In the formula, v i This represents the TF-IDF value of the i-th word.

[0090] (3) Image data is first processed using grayscale, and the formula is as follows:

[0091] Gray=0.299*R+0.587*G+0.114*B

[0092] In the formula, R, G, and B represent the pixel values ​​of the red, green, and blue channels, respectively.

[0093] Then, semantic extraction of green logistics is performed: first, a target detection model is used to identify green elements in the grayscale image, such as photovoltaic panels, and then the identification results are statistically analyzed to obtain: photovoltaic pixel area S and total image pixel area S. pic ,

[0094] Finally, the index is calculated: the proportion of the identified portion to the entire image is calculated, resulting in a dimensionless index in the 0-1 interval.

[0095] In the formula, S is the area of ​​the photovoltaic pixel identified in the image, S pic The total pixel area of ​​the image

[0096] like Figure 3 As shown,

[0097] Step 3: Construct an evaluation index system:

[0098] Based on the VIF-FA dual mechanism, key indicators are screened, and a three-dimensional evaluation system for TOE is constructed.

[0099] Table 1. Initial TOE Three-Dimensional Evaluation Index System (Example)

[0100]

[0101] (1) Based on the multicollinearity test, the initial indicator set I after multimodal data standardization is screened. The initial indicator set I is a set of evaluation indicators obtained in step 2, which can be represented as I = {i1, i2, i3, i...} n First, calculate the coefficient of determination. The smaller the value, the more different the information reflected by indicator i is from other indicators, and indicator i should be retained; similarly, The larger the value, the more readily the information reflected by indicator i can be replaced by other indicators, thus indicator i can be deleted. The formula is:

[0102]

[0103] Finally, the variance inflation factor (VIF) of the indicator variable i is calculated. i This is used to determine whether indicator i exhibits multicollinearity with other indicators, thus eliminating information redundancy between indicators. If VIF i If the value is greater than 10, it indicates that indicator i exhibits multicollinearity with other indicators, and indicator i should be deleted. The formula is:

[0104]

[0105] In the formula: The coefficient of determination for index i; The mean of index i; VIF is the estimated value of the index of object i for object j. i is the variance inflation factor of index i.

[0106] The VIF results obtained after calculation are as follows:

[0107] Table 2. VIF Results Chart

[0108] <![CDATA[i n ]]> Evaluation indicators VIF <![CDATA[i1]]> Number of new energy vehicles 4.26 <![CDATA[i2]]> Green patent grants 4.81 <![CDATA[i3]]> fuel consumption 13.43 <![CDATA[i4]]> carbon dioxide emissions 16.06 <![CDATA[i5]]> The proportion of smart park entities 3.48 <![CDATA[i6]]> R&D personnel ratio 7.94 <![CDATA[i7]]> R&D investment as a percentage of total assets 8.70 <![CDATA[i8]]> Enterprise ESG maturity 2.73 <![CDATA[i9]]> Green logistics posts received many likes 5.82 <![CDATA[i 10 ]]> Green logistics posts 10.30 <![CDATA[i 11 ]]> Green Logistics Post Shares 7.42 <![CDATA[i 12 ]]> Green logistics policy efforts 2.24

[0109] From Table 2, we can see that i3, i4, i 10 The VIF values ​​were all greater than 10, so these three indicators were removed.

[0110] (2) For the index set I after the VIF test, {I=i1,i2,i5,i6,i7,i8,i9,i 11 i 12 The specific steps for factor analysis are as follows:

[0111] First, calculate the correlation coefficient matrix R: the elements r in the correlation coefficient matrix R pq (p,q=1,2,…,m), representing index i p With index i q The correlation coefficient between them is calculated using the following formula:

[0112]

[0113] In the formula, N is the sample size, x ps For the index i in the s-th sample p The value, For index i p The sample mean, x qs For the index i in the s-th sample q The value, For index i q The sample mean.

[0114] Table 2. Correlation coefficient matrix R

[0115] <![CDATA[i1]]> <![CDATA[i2]]> <![CDATA[i5]]> <![CDATA[i6]]> <![CDATA[i7]]> <![CDATA[i8]]> <![CDATA[i9]]> <![CDATA[i 11 ]]> <![CDATA[i 12 ]]> <![CDATA[i1]]> 1.00 <![CDATA[i2]]> -0.48 1.00 <![CDATA[i5]]> 0.89 -0.50 1.00 <![CDATA[i6]]> -0.84 0.28 -0.86 1.00 <![CDATA[i7]]> -0.26 0.08 -0.34 0.17 1.00 <![CDATA[i8]]> -0.15 -0.04 -0.15 0.21 -0.48 1.00 <![CDATA[i9]]> -0.05 -0.34 -0.05 0.14 0.25 -0.37 1.00 <![CDATA[i 11 ]]> -0.16 0.21 -0.30 0.16 0.57 -0.30 0.57 1.00 <![CDATA[i 12 ]]> -0.05 0.16 -0.07 0.23 0.13 0.20 -0.00 0.31 1.00

[0116] Then, solve for the eigenvalues ​​and eigenvectors: solve for the characteristic equation of the correlation coefficient matrix R, |R-λE|. m If |=0, then m eigenvalues ​​are obtained: λ1≥λ2≥…≥λ m ≥0 and the corresponding eigenvectors u1, u2, ..., u m E m It is an m-order identity matrix.

[0117] Then, determine the number of common factors: calculate the contribution rate C of each eigenvalue. j and cumulative contribution rate AC j The formulas are as follows:

[0118]

[0119] Set a cumulative contribution rate threshold T AC Choose one that makes ACj ≥T AC Let the minimum value of j be denoted as k, then the number of common factors is k.

[0120] The calculation results are shown in the table below:

[0121] Table 3. Eigenvalues ​​and Contribution Rates

[0122]

[0123]

[0124] Factors with eigenvalues ​​greater than 1 are retained. The cumulative contribution rate of the extracted factors is greater than 85%, so only the first four factors need to be retained.

[0125] Then, calculate the factor loading matrix A: the element a in factor loading matrix A pj (p = 1, 2, ..., m; j = 1, 2, ..., k) represents i p The loading on the j-th common factor is given by the formula:

[0126]

[0127] In the formula, u jp It is the eigenvector u j The p-th component.

[0128] Table 4. Calculated factor loading matrix A:

[0129] index Factor 1 Factor 2 Factor 3 Factor 4 <![CDATA[i1]]> 0.85 -0.12 0.31 0.05 <![CDATA[i2]]> -0.43 0.76 0.08 -0.13 <![CDATA[i5]]> 0.82 -0.21 0.42 0.11 <![CDATA[i6]]> -0.79 0.31 0.14 0.08 <![CDATA[i7]]> -0.25 0.12 0.31 0.57 <![CDATA[i8]]> -0.14 -0.04 0.88 0.05 <![CDATA[i9]]> -0.05 -0.34 0.05 0.91 <![CDATA[i 11 ]]> -0.16 0.21 0.09 0.82 <![CDATA[i 12 ]]> -0.05 0.16 0.20 0.31

[0130] Then, the initial factor loading matrix A is rotated using an orthogonal rotation method to obtain the rotated factor loading matrix A. * Let the rotation matrix be T, satisfying A * =AT. The goal of rotation is to polarize the loadings on each factor as much as possible towards 0 and 1. The objective function of variance-maximizing rotation is:

[0131]

[0132] Solve for the rotation matrix T such that the objective function Q reaches its maximum value.

[0133] The loadings of index i on each factor were calculated, and the results are shown in the table below:

[0134] Table 5. Rotated factor loading matrix

[0135] index Factor 1 Factor 2 Factor 3 Factor 4 Maximum load <![CDATA[i1]]> 0.91 0.03 0.12 0.04 0.91 <![CDATA[i5]]> 0.87 -0.08 0.21 0.09 0.87 <![CDATA[i6]]> -0.11 0.76 0.13 0.08 0.76 <![CDATA[i7]]> 0.09 0.69 -0.14 0.23 0.69 <![CDATA[i2]]> 0.05 0.83 -0.04 0.11 0.83 <![CDATA[i8]]> 0.14 -0.03 0.88 0.05 0.88 <![CDATA[i 12 ]]> -0.01 0.13 0.78 -0.04 0.78 <![CDATA[i9]]> 0.07 0.11 0.05 0.91 0.91 <![CDATA[i 11 ]]> -0.04 0.08 0.13 0.86 0.86

[0136] Finally, for each common factor F j Set the factor loading threshold TA The factor loading is set to 0.5, and the absolute value of the factor loadings on this common factor is greater than T. A We select the key metrics, retain these metrics, and remove the others to obtain the final set of key metrics.

[0137] (3) After the indicators have been screened, they are further classified into three dimensions: technology, organization, and environment, based on the three categories of government, society, and enterprises, and a three-dimensional TOE evaluation indicator system is constructed. See Table 6 below for an example. The table below shows the evaluation indicators after VIF-FA screening (example).

[0138] Table 6. Evaluation Indicators After Screening (Example)

[0139]

[0140]

[0141] Step 4: Combined weighting and performance measurement:

[0142] The game theory-based combinatorial weighting method integrates subjective and objective weights, and calculates green innovation performance based on the weighting results using the DEA-TOPSIS combinatorial model.

[0143] (1) First, the subjective weights of the indicators are calculated using the Analytic Hierarchy Process (AHP). Based on the characteristics of AHP, the weights derived from expert experience are used to characterize the actual role of the indicators in actual production. Then, the objective weights of the indicators are calculated using the CRITIC method. The CRITIC method takes into account the magnitude of indicator variability while also considering the correlation between indicators. It comprehensively measures the objective weights of the indicators based on the comparative strength of the evaluation indicators and the conflict between them to characterize the objectivity of the indicators in production activities.

[0144] The calculated subjective weights ω of AHP are... A and CRITIC objective weight ω C As shown in the table below:

[0145] Table 7. Subjective and Objective Weighting Table

[0146]

[0147] (2) After obtaining the subjective weight ω A Objective weight ω C Then, by using a linear combination method, the indicators are combined together to generate a new weight ω, the formula of which is:

[0148] ω=λ1ω A +λ2ω C

[0149] In the formula, ωA To give the subjective weights of the indicators in the AHP method, ω C Here, λ1 and λ2 are the objective weights of the indicators given by the CRITIC method, and λ1 and λ2 are linear combination coefficients that satisfy λ1 + λ2 = 1.

[0150] (3) Based on the ideas of game theory, establish an objective function to combine weights ω and ω A and ω C The objective is to minimize the sum of deviations between the coefficients, and to find the optimal linear combination coefficients. and At this point, the combined weight is the optimal combined weight ω. * The objective function and constraints are as follows:

[0151] min(||ω-ω A || 2 +||ω-ω C || 2 )

[0152] =min(||λ1ω) A +λ2ω C -ω A || 2 +||λ1ω A +λ2ω C -ω C || 2 )

[0153]

[0154] The system of equations is as follows:

[0155]

[0156] The obtained linear combination coefficients λ1 and λ2 are standardized:

[0157]

[0158] Finally, the optimal combination weight ω of the evaluation indicators is obtained. * for:

[0159]

[0160] The optimal combination weight ω was obtained through calculation. * The results are shown in the table below:

[0161] Table 8. Optimal Combination Weight Table

[0162]

[0163] (4) First-stage efficiency evaluation:

[0164] The BCC model is used to calculate the green innovation performance value of logistics enterprises. According to the TOE evaluation indicators screened by the VIF-FA method in step 3, the indicators are determined as two sets: the input variable set X and the output variable set Y. Taking the logistics enterprise as the decision-making unit DMU, calculate the comprehensive performance value of the k-th decision-making unit DMU (logistics enterprise), and the comprehensive performance value θ k The formula is:

[0165] max θ k

[0166]

[0167] Output the comprehensive efficiency value θ k ∈[0, 1]

[0168] In the formula, θ k is the comprehensive efficiency value of the k-th enterprise, m is the total number of decision-making units (i.e., the number of sample enterprises), Y rl is the actual value of the l-th enterprise on the r-th output indicator, X ol is the actual value of the l-th enterprise on the o-th input indicator, Y rk is the actual value of the k-th enterprise on the r-th output indicator, X ok is the actual value of the k-th enterprise on the o-th input indicator, μ b is the weight coefficient of the b-th enterprise.

[0169] In the second stage, construct the TOPSIS decision matrix, and use the DEA efficiency value θ k as a new indicator, and jointly form an extended decision matrix with the original TOE evaluation indicators

[0170]

[0171] Calculate the weighted normalized matrix:

[0172]

[0173] In the formula, m represents the number of logistics enterprises; p represents the number of original TOE indicators; is the optimal combination weight of the j-th indicator, which is weighted by the above game subjective and objective weights; the element a ij (1 << j << p) represents the value of the i-th enterprise on the j-th indicator.

[0174] Determine the positive ideal solution R : and the negative ideal solution R ； .

[0175] Let the j-th attribute value of the positive ideal solution R : be Negative ideal solution R ； The value of the j-th attribute is but

[0176]

[0177] Then calculate the Euclidean distance to the ideal solution:

[0178] This indicates the degree of deviation of the firm from the optimal state. Distance to the negative ideal solution:

[0179] This indicates the degree to which a firm deviates from its worst-case scenario. Calculating the green innovation performance value:

[0180] C i A higher value indicates better green innovation performance of the logistics company.

Claims

1. A performance evaluation method for green innovation in the logistics industry based on multimodal data fusion, characterized in that, Includes the following steps: (1) Multimodal data collection and preprocessing: Based on the technology-organization-environment (TOE) framework, data is collected from the enterprise, social media and government policy aspects, and the data is preprocessed to build a multimodal database; (2) Multimodal data standardization: Normalization processing is performed on numerical data, text data and image data; (3) Constructing an evaluation index system: Based on the VIF-FA dual mechanism, key indicators are screened, and a three-dimensional evaluation system for TOE is constructed. (4) Combination weighting and performance measurement: The game theory-based combination weighting method integrates subjective and objective weights, and calculates green innovation performance based on the weighting results through the DEA-TOPSIS combination model.

2. The method for evaluating the performance of green innovation in the logistics industry based on multimodal data fusion according to claim 1, characterized in that, Step (2) includes the following steps: (21) Numerical data are normalized using Min-Max and Max-Min methods; (22) Text data is first vectorized using TF-IDF and then normalized using L2. (23) Image data is first processed in grayscale, and the semantics of green elements are extracted from the grayscale image. Then, the range is mapped to [0, 1].

3. The method for evaluating the performance of green innovation in the logistics industry based on multimodal data fusion according to claim 1, characterized in that, Step (3) includes the following steps: (31) Based on the multicollinearity test, the index set I is screened. First, the coefficient of determination is calculated. Calculate the variance inflation factor (VIF) of index variable i. i ; (32) Factor analysis was performed on the index set I after the VIF test to obtain the final key index set; (33) After the indicators have been screened, they are further classified into three dimensions: technology, organization and environment, based on the three categories of government, society and enterprise, and a three-dimensional TOE evaluation system is constructed.

4. The method for evaluating the performance of green innovation in the logistics industry based on multimodal data fusion according to claim 3, characterized in that, The specific steps of step (32) are as follows: Calculate the correlation coefficient matrix R based on the sample size and the index i in the s-th sample. p Values, Indicators i p The sample mean, the index i in the s-th sample q Values ​​and index i q The sample mean is used to determine the elements r in the correlation coefficient matrix R. pq (p,q=1,2,…,m), representing index i p With index i q The correlation coefficient between them Solving for eigenvalues ​​and eigenvectors: Solving for the characteristic equation of the correlation coefficient matrix R: |R-λE m If |=0, then m eigenvalues ​​are obtained: λ1≥λ2≥…≥λ m ≥0 and the corresponding eigenvectors u1, u2, ..., u m E m It is an m-order identity matrix. Determine the number of common factors: Calculate the contribution rate C of each eigenvalue. j and cumulative contribution rate AC j The formulas are as follows: In the formula, λ j Let be the j-th eigenvalue, m be the total number of original indicators, and k be the number of common factors ultimately retained. Set the cumulative contribution rate threshold T AC Choose one that makes AC j ≥T AC Let the minimum value of j be denoted as k, then the number of common factors is k. The factor loading matrix A is calculated using the following formula: In the formula, a jp Let u be the loading of the j-th original index on the p-th common factor. jp It is the eigenvector u j The p-th component, The initial factor loading matrix A is rotated using an orthogonal rotation method to obtain the rotated factor loading matrix A. * Let the rotation matrix be T, satisfying A * =AT, the objective function for maximum variance rotation is: Solve for the rotation matrix T such that the objective function Q reaches its maximum value. For each common factor F j Set the factor loading threshold T A Select factors whose absolute value of factor loadings on the common factor is greater than T. A We select the key metrics, retain these metrics, and remove the others to obtain the final set of key metrics.

5. The method for evaluating the performance of green innovation in the logistics industry based on multimodal data fusion according to claim 1, characterized in that, Step (4) includes the following steps: (41) Subjective weights ω are generated using the AHP method. A The objective weights ω are generated using the CRITIC method. C ; (42) The weight ω of the index combination expressed by the linear combination is: ω=λ1ω A +λ2ω C Based on the principles of game theory, an objective function is established, using the combined weights ω. j With ω A and ω C The objective is to minimize the sum of deviations between the coefficients, and to find the optimal linear combination coefficients. and At this point, the combined weight is the optimal combined weight ω. * .

6. The method for evaluating the performance of green innovation in the logistics industry based on multimodal data fusion according to claim 1, characterized in that, Step (4) also includes (43) the first-stage efficiency assessment: The set of input variables X and the set of output variables Y are determined based on the selected indicators; Calculate the overall efficiency value θ of the kth firm. k , The second stage involves constructing the TOPSIS decision matrix and setting the DEA efficiency value θ. k As a new indicator, it together with the original TOE evaluation indicators forms an extended decision matrix. Calculate the weighting matrix: In the formula, r ij For the j-th indicator of the i-th enterprise, It is the optimal combination weight of the j-th indicator, element a ij (1 << j << p) represents the value of the i-th enterprise on the j-th index. Determine the positive ideal solution R : and negative ideal solution R ； Then calculate the Euclidean distance to the positive ideal solution: This indicates the degree of deviation of the firm from the optimal state. Distance to the negative ideal solution: This indicates the degree to which a firm deviates from its worst-case scenario. Calculate the green innovation performance value: C i A higher value indicates better green innovation performance of the logistics company.

Images