Enterprise digital transformation maturity evaluation method and system

Abstract

The invention relates to the field of enterprise transformation evaluation, in particular to an enterprise digital transformation maturity evaluation method and system, and the method comprises the following steps: S1, data input: receiving and storing original data of an enterprise; s2, constructing an evaluation index system and performing evaluation: receiving the data in the step S1, constructing the evaluation index system, performing grade classification on the digital transformation maturity of the enterprise, performing attribution explanation on a classification result, and generating an evaluation report; s3, data comparison: comparing the evaluation report generated in the step S2 with a historical evaluation record, and analyzing the progress, stagnation or backward trend of the enterprise in the transformation process; s4, generating improvement suggestions: generating digital transformation suggestions for enterprises according to the maturity classification result generated in the step S2, attribution explanation and historical comparative analysis in the step S3 in combination with a business knowledge base; and S5, carrying out analysis data visualization: carrying out digital visualization presentation on the result of the previous step.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of enterprise transformation evaluation, and provides an enterprise digital transformation maturity evaluation method and system. BACKGROUND

[0002] The index system construction in the existing enterprise transformation evaluation method is static, lagging and subjective, and the evaluation index depends on the policy documents and reports obtained from official information sources on a regular basis. The index system constructed by this method is essentially static and lagging, and cannot reflect the fast-paced changes in the market and enterprise practice in real time. The effectiveness of the index has not been verified by the enterprise real data. It assumes that the index in the official document is universal and effective, but does not screen and verify which index can really distinguish and measure the digital transformation level of different enterprises through data analysis, which is easy to be out of touch with the actual business scenario of the enterprise. At the same time, it excessively relies on the blockchain, and describes the smart contract and data encryption scheme based on the blockchain in a large number of pages, which is undoubtedly beneficial to solving the data credibility problem. However, this scheme is technically complex and has high deployment cost, which may constitute an unbearable burden for enterprises with limited IT resources. More importantly, the blockchain only solves the problem of non-tamperability after the data is chained, but cannot guarantee the authenticity and accuracy before the data is chained. The evaluation result lacks depth explanation and operability, and the scheme finally outputs a "visual evaluation report" and a "transformation level", but this can only tell the enterprise which level it is in. It cannot answer the specific reasons and the key indicators that are lacking, so it lacks attribution analysis of the evaluation result, making it difficult for the enterprise to develop accurate and executable improvement measures even if it knows the level, so that the actual guidance value of the evaluation is limited. SUMMARY

[0003] The present application aims to provide an enterprise digital transformation maturity evaluation method and system, which can combine natural language processing, smart contract and blockchain technology, effectively solve the problems of inconsistent standards, low data credibility and complicated process in traditional evaluation methods, and also provide an improvement direction with actual guiding significance for enterprises.

[0004] The present application provides an enterprise digital transformation maturity evaluation method, which comprises the following steps: S1 data input: receiving and storing the basic information and multi-dimensional digital index original data of the enterprise; S2 constructing an evaluation index system and performing evaluation: receiving the data of step S1, automatically constructing an evaluation index system from the enterprise original data based on an unsupervised learning algorithm, classifying the maturity of the digital transformation of the enterprise based on a supervised learning algorithm, performing attribution explanation on the classification result, and generating an evaluation report; S3 Data Comparison: Compare the assessment report generated in step S2 with its historical assessment records to analyze the progress, stagnation, or regression trend of the enterprise in the transformation process; S4 Generate Improvement Suggestions: Based on the maturity classification results, attribution explanation generated in step S2, and historical comparison analysis in step S3, combined with the business knowledge base, generate digital transformation suggestions for the enterprise; S5 Data Visualization: Digitally visualize the maturity classification results in step S2, attribution explanation, historical comparison analysis in step S3, and digital transformation suggestions in step S4.

[0005] Preferably, the construction of the evaluation index system in step S2 includes the following steps: (1) Data Standardization: Use Z-Score standardization algorithm to process the original index data input in step S1, eliminate the influence of different index dimensions; (2) Natural grouping: Use K-Means clustering algorithm to automatically divide enterprises into K clustering groups based on their similarity in M standardized indicators without pre-set labels. These groups are considered to represent different natural modes or maturity stages of enterprise digital transformation; (3) Key indicator positioning: Use principal component analysis algorithm to interpret the clustering results in step (2); by analyzing the loading matrix of principal component analysis, identify the original indicators that contribute most to data variance as key driving factors for distinguishing different digital maturity modes.

[0006] Preferably, the step S2 of classifying the digital transformation maturity of the enterprise and attributing the classification results includes the following steps: (1) Model training: Use the extreme gradient boosting algorithm, use the optimized indicator data constructed in the first stage as features, and use a part of enterprise data labeled with maturity level as the training set to train a maturity classification model. Use genetic algorithm to globally optimize the hyperparameters of the extreme gradient boosting model; (2) Maturity prediction: Input new and unlabeled enterprise digital indicator data into the trained extreme gradient boosting model, and the model directly outputs the predicted maturity level of the enterprise; (3) Attribution analysis: Use the SHAP additive interpretation algorithm to explain the prediction results of the extreme gradient boosting model; calculate a SHAP value for each feature in the model, which accurately quantifies the contribution of the feature to the final maturity level prediction result, and based on this, generate a maturity report for individual enterprises.

[0007] Preferably, the data standardization in step (1) includes the following steps: S11 Input: raw data matrix X, with N rows of firms and M columns of indicators,

[0008] S12 Calculate the mean of each indicator: for each column j (i.e. each indicator):

[0009] : the mean of the jth indicator, representing the central position of all firms on this indicator; Calculate the standard deviation of each indicator: for each column j:

[0010] : the standard deviation of the jth indicator, measuring the size of data fluctuation of all firms on this indicator, The larger, the greater the difference between firms; S13 Standardize each data: for each element x in the matrix ij :

[0011] represent the standardized data, which is a dimensionless numerical value; S14 Output: form the standardized matrix Z from the data in step S13, where the mean of all data is 0 and the standard deviation is 1.

[0012] Preferably, the step (2) natural grouping, comprising the following steps: S21 Input: input the standardized data matrix Z and the preset number of clusters K; S22 Initialization: randomly select K data points as the initial cluster centers, denoted as: C = {c1, c2,..., ck}, where each ck is an M-dimensional vector; S23 Iteration until convergence, i.e. the cluster centers no longer change significantly: Assign data points to the nearest cluster center: for each firm i, calculate its Euclidean distance to all K cluster centers, the Euclidean distance formula is:

[0013] where z ij is the value of firm i on the jth standardized indicator; c kj is the coordinate of the kth cluster center on the jth dimension, this distance measures the overall similarity between firm i and the centroid k, the smaller the distance, the more similar; firm i is assigned to the cluster where the nearest cluster center is located, all firms will be divided into K clusters, and the member set of cluster k is denoted as Sk ; S24 Recalculate cluster center: for each cluster k, calculate the average value of all data points in each dimension as the new cluster center, the new cluster center formula is:

[0014] where |S k | : the number of enterprises contained in cluster k; S25 Output: the cluster label y i ∈ {1, 2,..., K} of each enterprise i and the final cluster center set C.

[0015] Preferably, the step (3) key indicator positioning comprises the following steps: S31 Input: standardized data matrix Z; S32 Calculate the covariance matrix: the covariance matrix Σ can be simplified as:

[0016] Where the element of the pth row and the qth column in the matrix Σ represents the covariance between the pth indicator and the qth indicator, reflecting their linear correlation degree; calculate the eigenvalue and eigenvector of the covariance matrix: solve the characteristic equation:

[0017] Where λ is the eigenvalue, v is the eigenvector, which is an M-dimensional vector; S33 Select principal component: sort the eigenvalues from large to small: λ1≥λ2≥...≥λ M , the corresponding eigenvectors are denoted as v1, v2,..., v M , as the direction of M principal components, select the first P principal components so that the cumulative variance contribution rate exceeds the threshold: ; S34 Analyze the load matrix: Load: the load l of the original indicator k on the principal component j kj The calculation formula is:

[0018] Where, is the kth component of the eigenvector vj; S35 Output: output the indicator list screened out by S34 analysis in the order of load from large to small.

[0019] Preferably, the model in the step (1) model training is:

[0020] wherein, is a function that maps input features x i to a predicted output value , function is not a simple linear function, but is composed of the prediction results of K decision trees (DTs) , where is the predicted value for enterprise i, K is the total number of trees, is an independent decision tree, is the function space composed of all possible decision trees.

[0021] Preferably, the step (1) model training comprises the following steps: (1-1) Initialize the extreme gradient boosting model:

[0022] (1-2) Perform extreme gradient boosting model iteration: for each iteration t, it is equivalent to constructing the tth tree: (1-2-1) Calculate the negative gradient of the current model, i.e. the approximate residual:

[0023] wherein, L is the loss function, represents the prediction error degree of the current model on the i th sample, is the true maturity level label of enterprise i, used to measure the gap between the predicted value and the true situation, is the predicted value of the maturity level of enterprise i, is the predicted value of the maturity level of enterprise i after t iterations.

[0024] (1-2-2) Fit a new decision tree ftto predict these residuals ; (1-2-3) Find the optimal leaf weight w j of this tree by optimizing the objective function, so that after adding this tree, the loss of the total model is reduced the most; (1-2-4) Update the extreme gradient boosting model:

[0025] wherein, is the learning rate, is the feature data vector of the enterprise, which is the input of the model, containing the standardized data of the enterprise on all key indicators, used to support the prediction of the maturity level.

[0026] Preferably, the objective function of step (1-2-3) is:

[0027] where L is the loss term, measuring the gap between the predicted value and the true value, Ω is the regularization term, controlling the complexity of the model, preventing overfitting, and n represents the total number of samples (i.e., enterprises) in the training data set.

[0028]

[0029] where T is the number of leaves in a tree; w j : the weight (predicted value) of the jth leaf, is a complexity control parameter, and only when the loss reduction value is greater than can the node be split, is the L2 regularization weight, which penalizes the weight to avoid excessive weight.

[0030] Preferably, the step (1) model training adopts a genetic algorithm to globally optimize the hyperparameters of the extreme gradient boosting model, including the following steps: (1-2-4-1) Define the hyperparameter search space: define the extreme gradient boosting hyperparameters to be optimized and their value ranges to form a hyperparameter search space , which includes at least the maximum depth of the tree max_depth, the learning rate learning_rate, the number of sub-trees n_estimators, the row sampling ratio subsample, the column sampling ratio colsample_bytree, the L1 regularization weight reg_alpha, and the L2 regularization weight reg_lambda. The search space is defined as the Cartesian product of the value ranges of each hyperparameter: where the value range of each hyperparameter is set as a discrete value set; (1-2-4-2) Population initialization: set the population size N and use a random sampling method to randomly generate N combinations of hyperparameters from the hyperparameter search space to form the initial population P0:

[0031] (1-2-4-3) Fitness evaluation: define a fitness function to evaluate the pros and cons of individuals , the calculation method of the fitness function uses the average macro F1 score of K-fold cross-validation as the fitness value of the individual, and calculates the fitness of each individual in the current population Pt , the calculation formula is as follows:

[0032] in, The macro F1 score is the result of the validation at the k-th fold. The macro F1 score is calculated by first calculating the F1 score for each category and then taking the arithmetic mean of the F1 scores for all categories. (1-2-4-4) Crossover operation: Based on the fitness calculated in step (1-2-4-3), select superior individuals from the current population Pt to form the parent population. Pair individuals in the parent population Pparents with each other, and crossover probability p. c Perform a single-point crossover operation to generate a child population. Randomly pair up the parent populations to obtain N / 2 parent pairs. For each parent pair ( a, b) Generate a random number r that is uniformly distributed in the interval [0, 1]. If r <p c Then perform crossover: randomly select a crossover point position pos, swap all hyperparameter values ​​of the two parent individuals after the crossover point pos, and generate two new child individuals. a′, b′; if r≥p c If the parent is not copied, the offspring will be directly copied without crossover.

[0033] (1-2-4-5) Mutation operation: For the offspring population produced after the crossover operation, the mutation probability p is... m Perform a random mutation operation on each individual in the offspring population. Iterate through each hyperparameter of the individual. [j] generates a random number r that is uniformly distributed in the interval [0,1]. j If r j <p m Then this hyperparameter [j] Mutation occurs: A new value is randomly selected from its predefined search space range to replace the original value. If r j ≥p m If so, the hyperparameter remains unchanged; (1-2-4-6) Select the E individuals with the highest fitness from the previous generation population Pt and keep them directly into the next generation. Merge the offspring population obtained after mutation in step (1-2-4-5) with the E elite individuals kept to form a new generation population Pt+1. If the number of individuals after merging exceeds the population size N, randomly delete the excess individuals; if it is insufficient, randomly generate new individuals to make up the difference. (1-2-4-7) Iterative evolution: Repeat steps (1-2-4-4) to (1-2-4-6) to perform T generations of evolution, where T is the preset maximum number of generations. In each generation of evolution, record the best fitness individual in the current population and its fitness value. (1-2-4-8) Optimal solution output: when evolution reaches the maximum number of generations T, select the individual with the highest fitness from the final generation population P T as the optimal hyperparameter combination output, as the selection of the final extreme gradient boosting hyperparameters.

[0034] Preferably, the step (2) of maturity prediction includes the following steps: (2-1) Traverse each decision tree to obtain the output of the tree: a weight w is stored on this leaf node k ; (2-2) Summarize the outputs of all trees: for a regression task that predicts a continuous value, the final prediction value is the weighted sum of all tree outputs, plus the initial prediction value:

[0035] where, is the learning rate, used to weaken the influence of each tree; for classification tasks, the Softmax function is used for calculation.

[0036] (2-3) Make the final prediction: compare the probabilities of all categories c , select the category with the highest probability as the final prediction of maturity level:

[0037] where, argmax is the parameter maximization function, used to find the variable value that makes the objective function maximum.

[0038] Preferably, the step (2-2) of classification task includes the following steps: (2-2-1) Calculate the original score of each category: when predicting each maturity level c, the model will calculate an original score s c

[0039] where, is the output score of the kth tree trained specifically for predicting category c; (2-2-2) Convert to probability by Softmax function: input these original scores into the Softmax function to calculate the probability of belonging to each category

[0040] where, is the exponential of the original score of category c; is the sum of the exponents of the original scores of all categories.

[0041] ​Preferably, the SHAP value calculation formula of the attribution analysis in step (3) is:

[0042] wherein, is the SHAP value of feature j, i.e. the contribution degree, is the model prediction value for feature subset S, S is all possible feature subsets without considering feature j, is the weight term, used for weighted average of all possible subset combinations, is the marginal contribution of feature j, i.e. the change of prediction value after adding feature j, and finally output the maturity report: for the prediction of enterprise i, each feature j has a SHAP value The sum of SHAP values of all features is equal to the difference between the prediction value and the average prediction value:

[0043] wherein, is the maturity level prediction value of the model for a single enterprise i, E[f(X)] is the mathematical expectation of the maturity prediction value of the model for all enterprise samples, X represents the feature data set of all enterprises, and E[] is the expectation operator.

[0044] A digital transformation maturity evaluation system for enterprises, comprising the following functional modules: An input module for receiving and storing basic information and multi-dimensional digital index raw data of an enterprise; An evaluation module, which is a core processing module connected to the input module, comprising: An evaluation index system construction unit for automatically constructing an evaluation index system from enterprise raw data based on an unsupervised learning algorithm; A maturity classification and attribution unit for classifying the digital transformation maturity of an enterprise into grades based on a supervised learning algorithm and explaining the classification results; A comparison module connected to the evaluation module for comparing the latest evaluation report of an enterprise with its historical evaluation records to analyze the progress, stagnation or regression trend of the enterprise in the transformation process; A suggestion generation module connected to the comparison module and the evaluation module for generating customized and operable digital transformation suggestions for an enterprise based on maturity classification results, attribution explanation and historical comparison analysis, combined with a business knowledge base; A visualization module connected to the evaluation module, the comparison module and the suggestion generation module for displaying the evaluation results, comparison analysis, attribution explanation and improvement suggestions in the form of charts and dashboards.

[0045] By adopting the above scheme, the enterprise digital transformation maturity evaluation method and system provided by the application has the following advantages and beneficial effects: the enterprise digital transformation maturity evaluation method and system provided by the application discards the mode of completely relying on expert subjective setting of indexes and scoring, automatically mines and constructs an index system from real enterprise data through unsupervised learning, and makes the evaluation standard more objective and data-supported. Through SHAP attribution analysis, not only the grade can be given, but also the contribution degree of each specific index can be accurately pointed out, the advantages and shortcomings of the enterprise are accurately positioned, the evaluation result has unprecedented interpretability and depth of insight. From index construction, model training, grade prediction to report generation, the whole process is highly automated. In particular, the genetic algorithm is used to optimize the hyperparameters, which reduces the tedious manual parameter adjustment work and greatly improves the model construction and evaluation efficiency. The data-driven index system can be dynamically updated as new data is included, better adapting to changes in technology and market. Unlike the general template type evaluation report, the report generated by the system for each enterprise is based on the attribution analysis and historical trend comparison of its own data, and the improvement suggestions provided are highly customized and operable. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the specific embodiments of the present application. It should be understood that the following drawings only show some specific embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0047] Figure 1 is the enterprise digital transformation maturity evaluation system framework of the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application.

[0049] In addition, the terms "first", "second", etc. are used only for descriptive purposes and are not to be construed as indicating or implying relative importance or an indicated number of features. Thus, features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0050] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting", "fixing" and the like should be understood broadly, for example, can be fixedly connected, can be detachably connected, or integrally connected; can be mechanically connected, or electrically connected; can be directly connected, or indirectly connected through an intermediate medium, or can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0051] The preferred embodiments of the present application are described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the scope of protection of the present application is more clearly defined.

[0052] The accompanying drawings are referred to in the description of the present application Figure 1 An enterprise digital transformation maturity assessment system includes the following functional modules: Input module: for receiving and storing basic information and multi-dimensional digital index raw data of the enterprise.

[0053] Evaluation module: the core processing module, connected to the input module, including: Evaluation index system construction unit: based on unsupervised learning algorithm to automatically construct evaluation index system from enterprise raw data.

[0054] Maturity classification and attribution unit: based on supervised learning algorithm to classify the digital transformation maturity of the enterprise, and to attribute the classification results.

[0055] Comparison module: connected to the evaluation module, for comparing the latest evaluation report of the enterprise with its historical evaluation records, analyzing the progress, stagnation or regression trend of the enterprise in the transformation process.

[0056] Suggestion generation module: connected to the comparison module and evaluation module, generating customized and operable digital transformation suggestions for the enterprise according to the maturity classification results, attribution explanation and historical comparison analysis, combined with the business knowledge base.

[0057] Visualization module: connected with the evaluation module, comparison module and suggestion generation module, for displaying the evaluation results, comparison analysis, attribution interpretation and improvement suggestions in the form of charts, dashboards, etc.

[0058] The working method of the evaluation index system construction unit in the evaluation module includes the following steps: S11, data standardization: using Z-Score standardization algorithm to process the original index data of N enterprises collected by the input module, eliminating the influence of different index dimensions.

[0059] S12, natural grouping: using K-Means clustering algorithm, without pre-set label, based on the similarity of enterprises in M standardized indexes, automatically divide the enterprises into K clustering groups; these groups are regarded as representing different natural modes or maturity stages of enterprise digital transformation.

[0060] S13, key index positioning: using principal component analysis (PCA) algorithm to interpret the clustering results of step S12; by analyzing the loading matrix of PCA, the original indexes that contribute most to the variance of data, i.e. distinguish different clustering groups, are identified, these high-loading indexes are the key driving factors that distinguish different digital maturity modes.

[0061] The working method of the maturity classification and attribution unit in the evaluation module includes the following steps: S21, model training: using extreme gradient boosting (XGBoost) algorithm, taking the optimized index data constructed in the first stage as features, and taking a part of enterprise data labeled with maturity level labels by experts as training set, training a maturity classification model. To further improve the model performance, genetic algorithm (GA) can be used to globally optimize the hyperparameters of XGBoost model.

[0062] S22, maturity prediction: input the new, unlabeled enterprise digital index data into the trained XGBoost model, the model directly outputs the predicted maturity level of the enterprise.

[0063] S23, attribution analysis: using SHAP (Shapley Additive Explanations) algorithm to interpret the prediction results of XGBoost model; for each feature in the model, a SHAP value is calculated, which accurately quantifies the contribution of the feature to the final maturity level prediction result, based on which, an interpretable report is generated for individual enterprises.

[0064] The S11 data standardization includes the following steps: 1. Input: original data matrix X, N rows (enterprises), M columns (indexes).

[0065]

[0066] 2. Calculate the mean of each metric: for each column j (i.e. each metric):

[0067] : the mean of the jth metric. It represents the "center position" of all firms on this metric.

[0068] Calculate the standard deviation of each metric: for each column j:

[0069] : the standard deviation of the jth metric. It measures the size of the fluctuation of data of all firms on this metric. The larger, the greater the difference between firms.

[0070] 3. Standardize each data: for each element x in the matrix ij :

[0071] ( ) : calculate how much the original value deviates from the mean.

[0072] Divide by : express this deviation in terms of the number of standard deviations. The result is a dimensionless value.

[0073] 4. Output: the standardized matrix Z, where all data has mean 0 and standard deviation 1.

[0074] The S12 natural grouping, comprising the following steps: 1. Input: the standardized data matrix Z, and the preset number of clusters K.

[0075] 2. Initialization: randomly select K data points as the initial cluster centers (also called centroids). Denoted as: C = {c1, c2,..., cK}, where each c k is an M-dimensional vector. k

[0076] 3. Iterate until convergence (centroids no longer change significantly): Assign data points to the nearest centroid: for each firm i (corresponding to data point z i ): calculate the Euclidean distance from it to all K centroids. The Euclidean distance formula is:

[0077] where z​ij Vi,j: the value of enterprise i on the jth standardized indicator; c kj Ck,j: the coordinate of the kth centroid on the jth dimension. This distance measures the overall similarity between enterprise i and centroid k. The smaller the distance, the more similar.

[0078] Assign enterprise i to the cluster where the nearest centroid lies. In this way, all enterprises will be divided into K clusters. Denote the member set of cluster k as S k .

[0079] 4. Recalculate centroids: for each cluster k, calculate the average value of all data points within this cluster on each dimension as the new centroid. The formula for the new centroid is:

[0080] where |S k |: the number of enterprises included in cluster k.

[0081] 5. Output: the cluster label y i i of each enterprise i and the final centroid set C.

[0082] The S13 key indicator positioning includes the following steps: The purpose of this step is to find out which original indicators contribute the most to distinguishing different clusters (i.e., different maturity patterns).

[0083] 1. Input: the standardized data matrix Z.

[0084] 2. Calculate the covariance matrix: since the data has been standardized (mean is 0), the covariance matrix Σ can be simplified as:

[0085] The meaning of covariance: the element in the pth row and qth column of matrix Σ represents the covariance between the pth indicator and the qth indicator, reflecting their linear correlation degree.

[0086] Calculate the eigenvalues and eigenvectors of the covariance matrix: solve the characteristic equation:

[0087] where λ is the eigenvalue, which is a scalar, and its size represents the amount of information (variance) carried by its corresponding principal component; v is the eigenvector, which is an M-dimensional vector, defining the direction of the principal component.

[0088] 3. Select principal components: Sort the eigenvalues from large to small: λ1≥ λ2≥... ≥ λ M .

[0089] The corresponding feature vectors are denoted as v1, v2, ..., v M This is the direction of the M principal components.

[0090] Typically, the top P principal components are selected such that their cumulative variance contribution rate exceeds a threshold (e.g., 85%).

[0091]

[0092] 4. Analyze the load matrix: Loading: The loading l of the original index k on the principal component j kj The calculation formula is:

[0093] in, It is the feature vector v j The kth component.

[0094] Interpretation of loadings: Essentially, they are the correlation coefficients between the original index and the principal components. The larger the absolute value of the loading, the greater the contribution of the original index to the composition of this principal component. Focus on the original indices with large absolute values ​​of loadings on the first few important principal components (such as PC1, PC2); these are the key driving factors.

[0095] 5. Output: A list of selected key metrics.

[0096] The purpose of training the S21 model is to learn a function F(X) to predict the maturity level Y.

[0097] XGBoost model format:

[0098] in, It is a function that takes input feature x i Mapping to a predicted output value .function It is not a simple linear function, but rather a function composed of K decision trees ( It is composed of the sum of the prediction results. Let K be the predicted value for firm i; K is the total number of trees. It is an independent decision tree; It is the function space consisting of all possible decision trees.

[0099] Training process: Initialize the model (usually the initial prediction is the average of all sample labels).

[0100]

[0101] For each iteration t (construct the tth tree): 1. Calculate the negative gradient (approximate residual) of the current model:

[0102] where L is the loss function; represents the prediction error degree of the current model on the ith sample.

[0103] 2. Fit a new decision tree f t to predict these residuals .

[0104] 3. Find the optimal leaf weight w j of this tree (by optimizing the objective function) so that the total model loss is reduced the most after adding this tree.

[0105] 4. Update the model:

[0106] where is the learning rate, a very small number (such as 0.1), which controls the influence of each tree and prevents the step size from being too large, helping to prevent overfitting.

[0107] Objective function of XGBoost:

[0108] where ∑L is the loss term, measuring the difference between the predicted value and the true value. ∑Ω is the regularization term, controlling the complexity of the model to prevent overfitting, and n represents the total number of samples (i.e. enterprises) in the training data set.

[0109]

[0110] where T is the number of leaves in a tree; w j : the weight (predicted value) of the jth leaf; is the complexity control parameter, and only when the loss reduction value is greater than will the node be split; is the L2 regularization weight, which penalizes the leaf weight to avoid excessive weight.

[0111] In the training process, in order to make the trained model better, GA is used to optimize the hyperparameters of XGBoost, the specific process is as follows: 1. Define the hyperparameter search space: define the XGBoost hyperparameters to be optimized and their value ranges to form a hyperparameter search space .

[0112] The hyperparameters include at least: maximum tree depth (max_depth), learning rate (learning_rate), number of subtrees (n_estimators), row sampling ratio (subsample), column sampling ratio (colsample_bytree), L1 regularization weights (reg_alpha), and L2 regularization weights (reg_lambda), where L1 regularization (reg_alpha) applies an "absolute value penalty" to the leaf weights wj of the decision tree in the model, and the penalty term is... (α is reg_alpha, wj is the weight of the j-th leaf of the decision tree, and T is the number of leaves in a single tree). This means that for each decision tree, the sum of the absolute values ​​of all leaf weights is calculated, multiplied by the hyperparameter α as a penalty term, and added to the model's total loss. Its core function is feature sparsity—it reduces the weights of some unimportant features to 0, indirectly achieving feature selection and simplifying the model structure. L2 regularization (reg_lambda): applies a "squared penalty" to the leaf weights wj of the decision trees in the model. The penalty term is... (λ is reg_lambda, and the definitions of wj and T are consistent with L1). This means that for all leaf weights of each decision tree, the sum of squares is calculated, multiplied by the hyperparameter λ as a penalty term, and added to the model's total loss. Its core function is to smooth the weights—preventing weights from becoming zero, but rather reducing excessively large weights, avoiding the overly significant impact of a single feature on the prediction result, and improving the model's generalization ability. The search space... Defined as the Cartesian product of the ranges of each hyperparameter:

[0113] The range of values ​​for each hyperparameter is set as a discrete set based on prior knowledge or experience.

[0114] 2. Population Initialization: Set the population size N. Use a random sampling method to search the hyperparameter search space. N hyperparameter combinations are randomly generated to form the initial population P0:

[0115] 3. Fitness Assessment: Define a fitness function F(θ) to assess an individual. The fitness function has advantages and disadvantages. The calculation method uses the average macro F1 score obtained through K-fold cross-validation as the fitness value of the individual. The principle of K-fold cross-validation is to divide the training data into 5 equal parts (assuming K=5), use 4 parts to train the model and 1 part to validate the model each time, repeating this process 5 times, using a different part for validation each time. For the current population P... t Each individual in Compute its fitness , the formula is as follows:

[0116] wherein, is the macro F1 score of the k-fold validation. The macro F1 score is to first calculate the F1 score of each classification, and then take the arithmetic mean of the F1 scores of all classifications.

[0117] 4. Cross operation: based on the fitness calculated in step S3, select excellent individuals (individuals with high fitness) from the current population Pt to form a parent population. Pair the individuals in the parent population P parents with each other to perform single-point crossover operation with a crossover probability p c , and generate a child population. The parent population is randomly paired with each other to obtain N / 2 parent pairs. For each parent pair a , b , a random number r uniformly distributed in the interval [0, 1] is generated. If r c , perform crossover: randomly select a crossover position pos (1≤pos<L, L is the number of hyperparameters), exchange all hyperparameter values of the two parent individuals after the crossover position pos to generate two new child individuals a ′, b ′. If r c , no crossover is performed, and the parent individual is directly copied as the child individual.

[0118] 5. Mutation operation: perform random mutation operation on the child population generated after the crossover operation with a mutation probability p m . For each individual in the child population, traverse each hyperparameter [j] of the individual, and generate a random number r j uniformly distributed in the interval [0, 1]. If r j <p m , the hyperparameter [j] is mutated: a new value is randomly selected from its predefined search space value range to replace the original value. If r j ≥p m , the hyperparameter remains unchanged.

[0119] 6. Select the E individuals with the highest fitness from the last generation population P t , and directly retain them to the next generation. Combine the child population obtained after step 5 mutation with the retained E elite individuals to form a new generation population Pt+1 If the number of individuals after merging exceeds the population size N, randomly delete the extra individuals; if not, randomly generate new individuals to make up for it.

[0120] 7. Iterative evolution: repeat steps 4-6 for T generations of evolution, where T is the pre-set maximum number of generations. In each generation of evolution, record the best individual in the current population and its fitness value.

[0121] 8. Optimal solution output: when the evolution reaches the maximum number of generations T, select the individual with the highest fitness value from the final population P T T as the optimal hyperparameter combination output, as the final selection of XGBoost hyperparameters.

[0122] The S22 maturity prediction includes the following steps: 1. Traverse each decision tree For each tree f k (k from 1 to K), perform the following operations: Start from the root node: input the enterprise's data x new into the root node of this tree.

[0123] Traverse down according to the rules: at each node, there is a split rule, for example, if cloud platform usage <0.65. Check the value of the enterprise x new on the feature "cloud platform usage". If its value <0.65, go to the left child node; otherwise, go to the right child node.

[0124] Repeat step 2: at the new child node, apply the split rule of this node again and continue to walk down. Traverse down until you reach a leaf node that has no child nodes.

[0125] Get the output of this tree: this leaf node stores a weight (or score) w k . The "view" or "vote" of this tree f k for this enterprise is this score w k .

[0126] 2. Summarize the outputs of all trees Now we have the outputs w1, w2,..., w k of K trees.

[0127] For a regression task (predicting a continuous value), the final prediction value is the weighted sum of all tree outputs plus the initial prediction value:

[0128] where,​ is the learning rate, used to attenuate the influence of each tree.

[0129] For classification tasks (maturity assessment is a multi-classification), use the Softmax function, the process is as follows: Calculate the raw score for each class: for each possible maturity level c (such as L1, L2, L3, L4, L5). When making a prediction, the model will calculate a raw score s c for each class c.

[0130]

[0131] Here, is the output score of the kth tree trained specifically for predicting class c.

[0132] Convert to probability by Softmax function: input these raw scores into the Softmax function to calculate the probability of belonging to each class.

[0133]

[0134] where, is the exponential of the raw score for class c; is the sum of the exponentials of the raw scores for all classes. Divide the two, and you get a probability value between 0 and 1, and the sum of the probabilities of all classes is 1.

[0135] 3. Make the final prediction: compare the probabilities of all classes c , select the class with the highest probability as the final predicted maturity level.

[0136]

[0137] The S23 attribution analysis aims to fairly allocate the contribution of each feature to a single prediction result. The theoretical basis of attribution analysis is the Shapley value of cooperative game theory.

[0138] SHAP value calculation formula:

[0139] where, is the SHAP value of feature j, i.e. its contribution. is the model prediction value for feature subset S. S is all possible feature subsets excluding feature j. is the weight term, used to weight average all possible subset combinations. is the marginal contribution of feature j, i.e. the change in prediction value after adding feature j.

[0140] The SHAP value is a weighted average of the marginal contribution that feature j makes across all possible combinations of features. It fairly satisfies the axioms of validity, symmetry, and additivity.

[0141] Final output: For each prediction for enterprise i, there is a SHAP value for each feature j The sum of the SHAP values for all features equals the difference between the prediction and the average prediction: .

[0142] The above is only a preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution that belongs to the concept of the present application falls within the protection scope of the present application.

Claims

1. A method for assessing the maturity of enterprise digital transformation, characterized in that, Includes the following steps: S1 Data Input: Receives and stores raw data on the company's basic information and multi-dimensional digital indicators; S2: Construct and evaluate the assessment index system: Receive the data from step S1, and automatically construct the assessment index system from the enterprise's original data based on the unsupervised learning algorithm. Also, classify the enterprise's digital transformation maturity based on the supervised learning algorithm, and interpret the classification results to generate an evaluation report. S3 Data Comparison: Compare the evaluation report generated in step S2 with its historical evaluation records to analyze the company's progress, stagnation, or regression trends during the transformation process; S4 generates improvement suggestions: Based on the maturity classification results and attribution explanations generated in step S2, and the historical comparative analysis in step S3, combined with the business knowledge base, suggestions for digital transformation are generated for enterprises. S5 involves data visualization analysis: presenting the maturity classification results and attribution explanations from step S2, the historical comparative analysis from step S3, and the digital transformation recommendations from step S4 in a digitally visualized manner.

2. The enterprise digital transformation maturity assessment method according to claim 1, characterized in that, The construction of the evaluation index system in step S2 includes the following steps: (1) Data standardization: The Z-Score standardization algorithm is used to process the original index data input in step S1 to eliminate the influence of different index units; (2) Natural grouping: Using the K-Means clustering algorithm, without preset labels, enterprises are automatically divided into K clusters based on the similarity of enterprises on M standardized indicators. These groups are considered to represent different natural patterns or maturity stages of enterprise digital transformation. (3) Key indicator positioning: The clustering results of step (2) are interpreted using the principal component analysis algorithm; by analyzing the loading matrix of the principal component analysis, the original indicators that constitute the largest variance of the data are identified as key driving factors to distinguish different digital maturity models.

3. The enterprise digital transformation maturity assessment method according to claim 1, characterized in that, Step S2, which involves classifying enterprises' digital transformation maturity levels and attributing the classification results, specifically includes the following steps: (1) Model training: The extreme gradient boosting algorithm is adopted, using the optimized index data constructed in the first stage as features, and a portion of enterprise data labeled with maturity level as training set to train a maturity classification model. The genetic algorithm is used to globally optimize the hyperparameters of the extreme gradient boosting model. (2) Maturity prediction: Input new and unlabeled enterprise digital indicator data into the trained extreme gradient boosting model, and the model directly outputs the predicted maturity level of the enterprise; (3) Attribution analysis: The Shapley additive interpretation algorithm is used to interpret the prediction results of the extreme gradient boosting model; a SHAP value is calculated for each feature in the model, which accurately quantifies the contribution of the feature to the final maturity level prediction result, and a maturity report is generated for each enterprise based on this.

4. The enterprise digital transformation maturity assessment method according to claim 2, characterized in that, The data standardization in step (1) specifically includes the following steps: Input S11: Original data matrix X, with N rows of companies and M columns of indicators. S12 calculates the average of each indicator: for each column j: The average value of the j-th indicator represents the central position of all enterprises on this indicator. Calculate the standard deviation for each indicator: For each column j: The standard deviation of the j-th indicator measures the magnitude of data fluctuation across all companies on this indicator. The larger the value, the greater the differences between companies; S13 standardizes each data point: for each element x in the matrix ij : It represents standard data, specifically a dimensionless numerical value; S14 Output: Form a standardized matrix Z from the data in step S13, where the mean of all data is 0 and the standard deviation is 1.

5. The enterprise digital transformation maturity assessment method according to claim 4, characterized in that, The natural grouping in step (2) includes the following steps: S21 Input: Input the standardized data matrix Z and the preset number of clusters K; S22 Initialization: Randomly select K data points as the initial cluster centers, denoted as: C={c1,c2,...,ck}, where each ck is an M-dimensional vector; S23 Iteration continues until convergence, i.e., the cluster center no longer changes significantly: Assign data points to the nearest cluster centers: For each enterprise i, calculate its Euclidean distance to all K cluster centers. The Euclidean distance formula is: Where z ij Let c be the value of enterprise i on the j-th standardized index; kj Let S be the coordinates of the k-th cluster center in the j-th dimension. This distance measures the overall similarity between enterprise i and the centroid k; the smaller the distance, the more similar they are. Enterprise i is assigned to the cluster containing the nearest cluster center. All enterprises will be divided into K clusters, and the membership set of cluster k is denoted as S. k ; S24 Recalculate the cluster center: For each cluster k: Calculate the average of all data points within the cluster in each dimension, and use this average as the new cluster center. The formula for the new cluster center is: Among them |S k |: The number of firms contained in cluster k; S25 Output: Cluster label y for each enterprise i i ∈{1,2,...,K} and the final set of cluster centers C.

6. The enterprise digital transformation maturity assessment method according to claim 5, characterized in that, The key indicator positioning in step (3) is as follows: Includes the following steps: S31 Input: Standardized data matrix Z; S32 Calculate the covariance matrix: The covariance matrix Σ can be simplified to: The element in the p-th row and q-th column of matrix Σ represents the covariance between the p-th and q-th indicators, reflecting their linear correlation. Calculate the eigenvalues ​​and eigenvectors of the covariance matrix; solve the characteristic equation. Where λ is the eigenvalue and v is the eigenvector, which is an M-dimensional vector; S33 Principal Component Selection: Sort the eigenvalues ​​from largest to smallest: λ1≥λ2≥...≥λ M The corresponding feature vectors are denoted as v1, v2, ..., v M Assuming the directions of the M principal components, select the top P principal components such that their cumulative variance contribution rate exceeds a threshold: ； S34 Analysis of Load Matrix: Loading: The loading l of the original index k on the principal component j kj The calculation formula is: in, It is the k-th component of the feature vector vj; S35 Output: Outputs a list of indicators in descending order of load. S34 Analysis: Filters and selects indicators.

7. The enterprise digital transformation maturity assessment method according to claim 3, characterized in that, The model used in step (1) of model training is: in, To input feature x i Mapping to a predicted output value The function, Let K be the predicted value for firm i, and K be the total number of trees. As an independent decision tree, It is the function space consisting of all possible decision trees.

8. The enterprise digital transformation maturity assessment method according to claim 7, characterized in that, The model training in step (1) includes the following steps: (1-1) Initialize the extreme gradient boosting model: (1-2) Iterate the extreme gradient boosting model: For each iteration t, it is equivalent to constructing the t-th tree: (1-2-1) Calculate the negative gradient of the current model, i.e., the approximate residual: Where L is the loss function, This indicates the degree of prediction error of the current model on the i-th sample. This serves as a true maturity level label for company i, used to measure the gap between the predicted value and the actual situation. This is the predicted value for the maturity level of enterprise i. For after t The predicted maturity level of firm i after one iteration; (1-2-2) Fit a new decision tree ft to predict these residuals. ; (1-2-3) Find the optimal leaf weight w of this tree by optimizing the objective function. j This results in the greatest reduction in the overall model loss after adding this tree; (1-2-4) Update the extreme gradient boosting model: in, For learning rate, The feature data vector of the enterprise is the input to the model, containing standardized data of the enterprise on all key indicators, used to support the prediction of maturity level.

9. The enterprise digital transformation maturity assessment method according to claim 8, characterized in that, The objective function for step (1-2-3) is: Where L is the loss term, which measures the difference between the predicted value and the true value; Ω is the regularization term, which controls the complexity of the model and prevents overfitting; and n represents the total number of samples in the training dataset. Where T is the number of leaves in a tree; w j The weight of the j-th leaf. As a complexity control parameter, when splitting nodes, only when the loss reduction value is greater than It can only split when The L2 regularized weights are used to penalize the leaf weights and prevent them from becoming too large.

10. The enterprise digital transformation maturity assessment method according to claim 9, characterized in that, The step (1) model training uses a genetic algorithm to globally optimize the hyperparameters of the extreme gradient boosting model, including the following steps: (1-2-4-1) Define the hyperparameter search space: Define the extreme gradient boosting hyperparameters to be optimized and their range of values, which constitute a hyperparameter search space. The hyperparameters include at least: the maximum tree depth (max_depth), the learning rate (learning_rate), the number of subtrees (n_estimators), the row sampling ratio (subsample), the column sampling ratio (colsample_bytree), the L1 regularization weight (reg_alpha), and the L2 regularization weight (reg_lambda), and the search space. Defined as the Cartesian product of the ranges of each hyperparameter: The range of values ​​for each hyperparameter is set as a discrete set of values; (1-2-4-2) Population initialization: Set the population size N and use a random sampling method to search the hyperparameter space. N hyperparameter combinations are randomly generated to form the initial population P0: (1-2-4-3) Fitness Evaluation: Define the fitness function Used to assess individuals The fitness function is superior or inferior to the fitness function. The calculation method uses the average macro F1 score obtained through K-fold cross-validation as the fitness value of the individual. This applies to each individual in the current population Pt. Calculate its fitness The calculation formula is as follows: in, The macro F1 score is the result of the validation at the k-th fold. The macro F1 score is calculated by first calculating the F1 score for each category and then taking the arithmetic mean of the F1 scores for all categories. (1-2-4-4) Crossover operation: Based on the fitness calculated in step (1-2-4-3), select superior individuals from the current population Pt to form the parent population. Pair individuals in the parent population Pparents with each other, and crossover probability p. c Perform a single-point crossover operation to generate a child population. Randomly pair up the parent populations to obtain N / 2 parent pairs. For each parent pair ( a, b) Generate a random number r that is uniformly distributed in the interval [0, 1]. If r <p c Then perform crossover: randomly select a crossover point position pos, swap all hyperparameter values ​​of the two parent individuals after the crossover point pos, and generate two new child individuals. a′, b′; if r≥p c If the parent generation is not copied, the offspring generation will be directly copied into the offspring generation. (1-2-4-5) Mutation operation: For the offspring population produced after the crossover operation, the mutation probability p is... m Perform a random mutation operation on each individual in the offspring population. Iterate through each hyperparameter of the individual. [j] generates a random number r that is uniformly distributed in the interval [0,1]. j If r j <p m Then this hyperparameter [j] Mutation occurs: A new value is randomly selected from its predefined search space range to replace the original value. If r j ≥p m If so, the hyperparameter remains unchanged; (1-2-4-6) Select the E individuals with the highest fitness from the previous generation population Pt and keep them directly into the next generation. Merge the offspring population obtained after mutation in step (1-2-4-5) with the E elite individuals kept to form a new generation population Pt+1. If the number of individuals after merging exceeds the population size N, randomly delete the excess individuals; if it is insufficient, randomly generate new individuals to make up the difference. (1-2-4-7) Iterative evolution: Repeat steps (1-2-4-4) to (1-2-4-6) to perform T generations of evolution, where T is the preset maximum number of generations. In each generation of evolution, record the best fitness individual in the current population and its fitness value. (1-2-4-8) Optimal solution output: When the evolution reaches the maximum number of generations T, start from the final generation population P T Select the individual with the highest fitness The output serves as the optimal combination of hyperparameters, and is used as the final extreme gradient boosting hyperparameter selection.

11. The enterprise digital transformation maturity assessment method according to claim 10, characterized in that, The maturity prediction in step (2) includes the following steps: (2-1) Traverse each decision tree and obtain the output of that tree: the leaf node stores a weight w. k ; (2-2) Summarizing the outputs of all trees, for a regression task predicting a continuous value, the final predicted value is the weighted sum of the outputs of all these trees, plus the initial predicted value: in, It is the learning rate, used to reduce the influence of each tree; for classification tasks, it is calculated using the Softmax function. (2-3) Make the final prediction: compare the probabilities of all categories c. The category with the highest probability is selected as the final predicted maturity level. Here, argmax is the parameter maximization function, used to find the variable value that maximizes the objective function.

12. The enterprise digital transformation maturity assessment method according to claim 11, characterized in that, The classification task in step (2-2) includes the following steps: (2-2-1) Calculate the raw score for each category: For each maturity level c, the model calculates a raw score s for each category c during prediction. c in It is the output score of the k-th tree, which is specifically trained to predict category c; (2-2-2) Convert to probability using the Softmax function: Input these raw scores into the Softmax function to calculate the probability of belonging to each category. in, An index for calculating the original score of category c; To calculate the sum of the raw score indices for all categories.

13. The enterprise digital transformation maturity assessment method according to claim 10, characterized in that, The formula for calculating the SHAP value in the attribution analysis in step (3) is as follows: in, The SHAP value of feature j, i.e., its contribution. Let S be the model prediction for a feature subset S, where S represents all possible feature subsets without considering feature j. This is the weighting term, used to calculate a weighted average over all possible subset combinations. The marginal contribution of feature j, i.e., the change in the predicted value after adding feature j, is used to ultimately output the maturity report: for the prediction of firm i, each feature j has a SHAP value. The sum of the SHAP values ​​of all features equals the difference between the predicted value and the average predicted value. in, Let f(X) be the model's predicted maturity level for a single firm i, and let E[f(X)] be the mathematical expectation of the model's predicted maturity levels for all firm samples. Let X represent the feature data set of all firms, and E[ ] be the expectation operator.

14. A system for assessing the maturity of enterprise digital transformation, characterized in that, Includes the following functional modules: Input module: Used to receive and store raw data of the company's basic information and multi-dimensional digital indicators; Evaluation module: This is the core processing module, connected to the input module, and includes: Evaluation index system construction unit: Automatically constructs an evaluation index system from the enterprise's original data based on unsupervised learning algorithms; Maturity Classification and Attribution Unit: Based on supervised learning algorithms, enterprises' digital transformation maturity is classified into levels, and the classification results are explained by attribution. Comparison Module: Connected to the assessment module, this module compares the company's latest assessment report with its historical assessment records to analyze the company's progress, stagnation, or regression trends during the transformation process. Suggestion generation module: Connects the comparison module and the assessment module, and generates customized and actionable digital transformation suggestions for enterprises based on maturity classification results, attribution explanations, and historical comparative analysis, combined with the business knowledge base; Visualization module: Connects the evaluation module, comparison module and suggestion generation module, and is used to display the evaluation results, comparative analysis, attribution explanation and improvement suggestions in an intuitive form such as charts and dashboards.

Images