An enterprise user portrait construction method and system based on big data mining

Abstract

The application provides an enterprise user portrait construction method and system based on big data mining, and relates to the technical field of data processing.The method comprises the following steps: extracting asset size values and semantic fluctuation features from a feature correlation graph, respectively converting the asset size values and the semantic fluctuation features into state inertia constraint parameters and market environment disturbance parameters, and constructing a migration correction vector; superimposing a state speed vector, a state acceleration vector, and the migration correction vector to obtain a pre-calibration coordinate, iteratively calibrating the pre-calibration coordinate by using the migration correction vector, and obtaining a state evolution sequence; dividing the state evolution sequence to obtain a distribution result; extracting core features according to the distribution result to obtain dynamic labels; and fusing the dynamic labels with basic attribute labels in an initial feature set to obtain an enterprise user portrait.The application improves the accuracy of the enterprise user portrait.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for constructing enterprise user profiles based on big data mining. Background Technology

[0002] In supply chain finance and corporate credit risk management by commercial banks, credit assessment of upstream and downstream SMEs of core enterprises is crucial. Existing enterprise user profiling technologies largely rely on static labeling systems or linear weighted scoring models. These methods typically treat various operational indicators (such as accounts receivable turnover, net cash flow, and inventory turnover days) as relatively independent feature inputs, calculating a comprehensive credit score through pre-set fixed weights or simple statistical regression. However, in actual operations, different operational indicators often exhibit non-linear coupling relationships. For example, some enterprises may experience a positive synergy between inventory backlog and short-term borrowing increases due to seasonal stockpiling; others may experience passive backlogs due to poor sales, where the increase in inventory and short-term borrowing reflects a tight cash flow, presenting a more complex asynchronous change. Due to a lack of quantitative analysis capabilities for such non-linear coupling relationships, existing technologies often treat each feature dimension as approximately orthogonal or independent variables, resulting in a sometimes discrete distribution of the constructed feature space vector, making it difficult to fully reconstruct the inherent logical structure of the enterprise's operating status. The resulting enterprise user profiles, when faced with data noise or normal fluctuations in indicators, sometimes struggle to accurately distinguish between substantive operational risks and regular cyclical fluctuations. This, to some extent, affects the reliability of credit risk assessment and limits the accuracy of credit decision support. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and system for constructing enterprise user profiles based on big data mining, which improves the accuracy of enterprise user profiles.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] Firstly, a method for constructing enterprise user profiles based on big data mining, the method comprising:

[0006] Step 1: Obtain multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set;

[0007] Step 2: Construct graph nodes based on the initial feature set, calculate the feature cosine similarity between graph nodes to obtain positive weights, and calculate the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights; synthesize the positive and negative weights into a feature association adjacency matrix, and perform graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain the feature association graph.

[0008] Step 3: Construct the business state migration trajectory based on the feature association map; perform differential processing on the business state migration trajectory to obtain the state rate vector and the state acceleration vector; extract the asset size value and semantic fluctuation features from the feature association map and convert them into state inertia constraint parameters and market environment disturbance parameters respectively, and construct the migration correction vector; superimpose the state rate vector, the state acceleration vector and the migration correction vector to obtain the pre-calibrated coordinates, and use the migration correction vector to iteratively calibrate the pre-calibrated coordinates to obtain the state evolution sequence;

[0009] Step 4: Divide the state evolution sequence to obtain the distribution results; extract core features based on the distribution results to obtain dynamic labels; fuse the dynamic labels with the basic attribute labels in the initial feature set to obtain the enterprise user profile.

[0010] Secondly, an enterprise user profile building system based on big data mining includes:

[0011] The processing module is used to acquire multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set.

[0012] The calculation module is used to construct graph nodes based on the initial feature set, calculate the feature cosine similarity between graph nodes to obtain positive weights, calculate the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights, synthesize the positive and negative weights into a matrix to obtain a feature association adjacency matrix, and perform graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain the feature association graph.

[0013] The calibration module is used to construct the business state migration trajectory based on the feature association map; perform differential processing on the business state migration trajectory to obtain the state rate vector and the state acceleration vector; extract the asset size value and semantic fluctuation features from the feature association map and convert them into state inertia constraint parameters and market environment disturbance parameters respectively, and construct the migration correction vector; superimpose the state rate vector, the state acceleration vector and the migration correction vector to obtain the pre-calibration coordinates; use the migration correction vector to iteratively calibrate the pre-calibration coordinates to obtain the state evolution sequence.

[0014] The output module is used to divide the state evolution sequence to obtain the distribution results; extract core features based on the distribution results to obtain dynamic labels; and fuse the dynamic labels with the basic attribute labels in the initial feature set to obtain the enterprise user profile.

[0015] Thirdly, a computing device includes:

[0016] One or more processors;

[0017] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0018] The above-described solution of the present invention has at least the following beneficial effects:

[0019] By synthesizing matrices of positive and negative weights, a feature association adjacency matrix is ​​constructed and graph topology reconstruction is completed. This can effectively quantify the nonlinear coupling relationships between different operational indicators, including positive collaboration, negative mutual exclusion, and complex asynchronous relationships. It helps avoid the problem of discrete distribution of feature space vectors, enabling the constructed feature association graph to more accurately reflect the internal logical structure of the enterprise's operating status. The reasonable reorganization of feature dimensions through graph topology reconstruction helps avoid interference from invalid features. Through multi-dimensional label fusion and structured coding, it helps improve the standardization and interpretability of the profile, enabling the constructed enterprise user profile to maintain high stability and reliability even when facing complex situations such as missing data and indicator fluctuations. To a certain extent, it solves the problems of weak anti-interference ability and poor interpretability of the profile. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a method for constructing enterprise user profiles based on big data mining, provided by an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of an enterprise user profile construction system based on big data mining, provided by an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, an embodiment of the present invention proposes a method for constructing enterprise user profiles based on big data mining, the method comprising the following steps:

[0024] Step 1: Obtain multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set;

[0025] Step 2: Construct graph nodes based on the initial feature set, calculate the feature cosine similarity between graph nodes to obtain positive weights, and calculate the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights; synthesize the positive and negative weights into a feature association adjacency matrix, and perform graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain the feature association graph.

[0026] Step 3: Construct the business state migration trajectory based on the feature association map; perform differential processing on the business state migration trajectory to obtain the state rate vector and the state acceleration vector; extract the asset size value and semantic fluctuation features from the feature association map and convert them into state inertia constraint parameters and market environment disturbance parameters respectively, and construct the migration correction vector; superimpose the state rate vector, the state acceleration vector and the migration correction vector to obtain the pre-calibrated coordinates, and use the migration correction vector to iteratively calibrate the pre-calibrated coordinates to obtain the state evolution sequence;

[0027] Step 4: Divide the state evolution sequence to obtain the distribution results; extract core features based on the distribution results to obtain dynamic labels; fuse the dynamic labels with the basic attribute labels in the initial feature set to obtain the enterprise user profile.

[0028] In this embodiment of the invention, by synthesizing positive and negative weight matrices, a feature association adjacency matrix is ​​constructed and graph topology reconstruction is completed. This can effectively quantify the nonlinear coupling relationships between different operational indicators, including positive collaboration, negative mutual exclusion, and complex asynchronous relationships. It helps avoid the problem of discrete distribution of feature space vectors, enabling the constructed feature association graph to more accurately reflect the internal logical structure of the enterprise's operating status. The reasonable reorganization of feature dimensions through graph topology reconstruction helps avoid interference from invalid features. Through multi-dimensional label fusion and structured coding, it helps improve the standardization and interpretability of the profile, enabling the constructed enterprise user profile to maintain high stability and reliability even when facing complex situations such as data loss and indicator fluctuations. To a certain extent, this solves the problems of weak anti-interference ability and poor interpretability of the profile.

[0029] In a preferred embodiment of the present invention, step 1 involves acquiring multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocessing the multi-source heterogeneous operational data, extracting key feature fields, and mapping the key feature fields to a high-dimensional feature space to obtain an initial feature set, which may include:

[0030] Step 101: Obtain multi-source heterogeneous operational data of the target enterprise within a preset time window. This multi-source heterogeneous operational data includes structured financial indicator data, semi-structured transaction flow data, and unstructured text data. Specifically, this involves acquiring multi-source heterogeneous operational data of the target enterprise within a preset time window. This specifically involves combining the actual needs of supply chain finance and commercial bank corporate credit risk management, targeting the credit assessment scenarios of upstream and downstream SMEs of the core enterprise, and collecting various operational-related data of the target enterprise within a pre-set time range. The preset time window is typically set to 12 months to ensure coverage of a complete business cycle and comprehensively reflect the enterprise's operational fluctuations. This multi-source heterogeneous operational data specifically covers three categories. The data comprises structured financial indicator data, semi-structured transaction flow data, and unstructured text data. Structured financial indicator data includes directly quantifiable financial data such as accounts receivable turnover, net cash flow, inventory turnover days, debt-to-equity ratio, and gross profit margin. Semi-structured transaction flow data includes various transaction records, fund transfer details, and invoice issuance records from the company's daily operations, which have a certain format but are not fully structured. Unstructured text data includes textual information that cannot be directly quantified, such as company annual reports, cooperation agreements, customer reviews, and industry research reports related to the company. These three types of data are collected through multiple channels, including the company's financial system, third-party data platforms, and interfaces with partner institutions.

[0031] Step 102 involves format unification, missing value imputation, and outlier filtering of multi-source heterogeneous operational data to obtain a preprocessed data stream. Based on the preprocessed data stream, structured indicator parsing, semi-structured transaction sequence splitting, and unstructured text entity recognition are performed to extract key feature fields. Specifically, this includes: unifying the three types of collected multi-source heterogeneous operational data by performing format unification operations to adjust data from different sources and formats to a unified CSV format, avoiding data processing deviations caused by format differences; performing missing value imputation by using corresponding imputation methods based on the enterprise's operating scenario and data type, such as using the average value of the indicator over the past three months for missing financial indicators, using the median of similar transactions for missing transaction data, and using the text description corresponding to the industry average for missing text data, to avoid incomplete feature extraction due to missing values, which would affect the accuracy of subsequent profiling; and performing outlier filtering using the 3σ principle. The outlier identification and filtering process involves calculating the mean and standard deviation of each feature field in each type of operational data. The mean is calculated by dividing the sum of all valid data for that feature field by the number of valid data points. The standard deviation is calculated by taking the square root of the sum of the squares of the differences between all valid data points and the mean for that feature field, divided by the number of valid data points. Then, the outlier judgment range is set according to the 3σ principle, meaning that the range of normal data is from the mean minus 3 times the standard deviation to the mean plus 3 times the standard deviation. Values ​​exceeding this range are considered outliers. Factors such as the company's operating cycle and industry averages are further considered to verify whether the outliers are reasonable fluctuations, such as short-term inventory surges due to seasonal stockpiling or abnormal cash flow due to periodic payments. Values ​​verified as reasonable fluctuations are retained. Unreasonable outliers without reasonable cause or deviating from normal operating logic are filtered out to reduce data noise interference with subsequent processing. After these three steps, a preprocessed data stream is obtained.

[0032] Based on the preprocessed data stream, corresponding feature extraction operations are performed for the three types of data. For structured financial indicator data, structured indicator parsing is performed to filter out core financial indicators that reflect the company's operating status and credit level, while irrelevant and redundant financial data is removed. For semi-structured transaction data, semi-structured transaction sequence splitting is performed, breaking down the transaction flow according to dimensions such as transaction time, transaction type, and transaction amount to extract key information reflecting the company's cash flow and transaction activity. For unstructured text data, unstructured text entity recognition is performed. Specifically, a dictionary of business-related entities is first constructed, containing core entity categories such as company partners, business operations, risk warnings, and development plans. Each category includes high-frequency entity words from the corresponding industry, while also supplementing the target company's own specific entity words, thus completing the construction and updating of the entity dictionary. Then, text keyword matching is performed, splitting the preprocessed unstructured text into sentences, scanning the text content sentence by sentence, and comparing it with the words in the entity dictionary. Precise matching is performed to identify fully matched basic entities. Semantic association analysis is then conducted. For text content that is semantically related but does not directly match the entity dictionary, the sentences are first segmented and stop words are removed to extract core semantic words. Then, the semantic similarity between the core semantic words and entity words of each category in the entity dictionary is calculated. The semantic similarity is calculated by dividing the number of common semantic features between the core semantic words and entity words by the total number of semantic features of both. A semantic similarity threshold of 0.7 is set, and core semantic words with a semantic similarity greater than or equal to 0.7 are classified into the corresponding entity category and additionally marked as related entities. Finally, entity dictionary comparison and verification are performed. All entities marked by keyword matching and semantic association analysis are compared one by one with the entity dictionary. Mislabeled entities with large semantic deviations or irrelevant to the company's operations are removed, and accurate core entities are integrated and screened, including information related to the company's partners, business operations, risk warnings, and development plans. Through the feature extraction operations of the above three types of data, the key feature fields are finally integrated.

[0033] Step 103: Perform dimension normalization transformation on the key feature fields to obtain standardized feature fields; vectorize the standardized feature fields according to data type to obtain basic dimension feature vectors; input the basic dimension feature vectors into a preset dimension expansion matrix for linear weighted transformation, and stretch the spatial features through a nonlinear mapping function to obtain high-dimensional projection feature vectors; concatenate the feature dimensions and align the high-dimensional projection feature vectors temporally to map them to a unified high-dimensional feature space to obtain an initial feature set. Specifically, this includes: performing dimension normalization transformation on the key feature fields extracted in Step 102, using the min-max normalization method to adjust all key feature fields to a numerical range of 0 to 1 to eliminate the influence of dimension differences. The specific calculation process is to subtract the minimum value of all values ​​in each key feature field from its value, and then divide by the maximum value of all values ​​in the field. The difference between the minimum values ​​yields the standardized feature fields. These standardized feature fields are then vectorized and encoded according to their data types. The three data types each use their corresponding encoding methods. The specific calculation process is as follows: Structured financial indicator data uses one-hot encoding. The category attributes of all structured financial indicators are identified, and the discrete value range of each financial indicator is clarified. For example, the debt-to-asset ratio is divided into three categories: low, medium, and high; and the accounts receivable turnover ratio is divided into four categories: excellent, good, average, and poor. A one-hot encoding vector is constructed for each category attribute. The vector length is equal to the number of categories for that indicator. Each category corresponds to a position in the vector. If the value of the financial indicator is in a certain category, the value at that position is set to 1, and the values ​​at other positions are set to 0. For example, if the debt-to-asset ratio is medium, the encoding vector is [0,1,0]. This method converts the category information of structured financial indicators into a computable vector form.

[0034] Semi-structured transaction log data employs a bag-of-words (BOB) encoding method. Key information within the transaction log data is segmented to extract core terms such as transaction type, transaction object, and transaction scenario, constructing a dedicated vocabulary for each transaction log. This vocabulary contains all occurrences of these core terms. The frequency of each term in a single transaction log record is calculated, with the order of the vocabulary as the vector dimension. Each transaction log record corresponds to a vector, and the value at each position in the vector represents the frequency of the corresponding term in that record. If a term does not appear, the value is set to 0. For example, if the vocabulary is [purchase, sales, payment collection, payment], and a transaction is a purchase transaction, the encoded vector would be [1,0,0,0], thus achieving vectorization of the semi-structured data. Unstructured text data uses TF-IDF encoding. The specific calculation process involves calculating word frequencies. The first step is to calculate the term frequency (TF) by counting the number of times each core entity word appears in a single text and dividing it by the total number of words in that text. The second step is to calculate the inverse document frequency (IDF). This is done by counting the number of texts containing the word, dividing it by the total number of texts, taking the logarithm, and adding 1 (to avoid a logarithm of 0). The IDF is calculated as log(total number of texts ÷ number of texts containing the word) + 1. Finally, the TF and IDF of each word are multiplied to obtain the TF-IDF value. Using all core entity words as vector dimensions, each text corresponds to a vector, and the value at each position in the vector is the TF-IDF value of the corresponding word. This method quantifies the feature information of unstructured text. Through these three encoding methods, the standardized feature fields corresponding to the three types of data are converted into computable and processable vector forms, resulting in basic dimension feature vectors. This allows various features to be analyzed through vector operations for correlation analysis. The basic dimension feature vectors are input into a preset dimension expansion matrix for linear weighted transformation. The preset dimension expansion matrix has dimensions of 1024×256, which is set based on the company's business characteristics and industry rules. Its specific expression is as follows: ;

[0035] in (i=1,2,...,1024; j=1,2,...,256) represents the element in the i-th row and j-th column of the dimension expansion matrix, with values ​​ranging from 0.001 to 0.01, obtained by calibration using enterprise operating characteristic weights and industry benchmark parameters; the specific expression for the bias vector is: ,in The value of the i-th element of the bias vector ranges from 0.01 to 0.05; the specific calculation process of the linear weighted transformation is based on the feature vector of the fundamental dimension. With dimension expansion matrix Multiply, and add the bias vector The calculation formula is: ,in Based on the feature vector of the basic dimension, The feature vector is obtained after linear weighting transformation. This operation initially expands the feature dimension and uncovers potential correlations between features. Spatial feature stretching is performed using the Sigmoid nonlinear mapping function. The specific implementation of the Sigmoid function is as follows: each element of the linearly weighted feature vector is substituted into the function Sigmoid(x) = ... ,in The function is a natural constant. It maps the values ​​of the feature vectors to the range of 0 to 1, further expanding the feature dimension, breaking the limitations of the traditional linear model, and fully exploring the nonlinear correlation between different features to obtain high-dimensional projected feature vectors. All high-dimensional projected feature vectors are concatenated and time-series aligned, integrating high-dimensional projected feature vectors of different types and dimensions together. At the same time, the feature vectors are aligned according to the temporal order of a preset time window to ensure the temporal consistency of the feature data. The integrated and aligned high-dimensional projected feature vectors are mapped to a unified high-dimensional feature space to finally obtain the initial feature set.

[0036] This embodiment eliminates data noise, compensates for data defects, and standardizes data format through preprocessing and feature extraction operations, ensuring the integrity and accuracy of key feature fields. It maps key feature fields to a high-dimensional feature space, breaking the limitations of traditional linear feature processing, fully exploring the potential correlation between features, and solving the problems of single data, inaccurate feature extraction, and ignoring nonlinear correlations.

[0037] In a preferred embodiment of the present invention, step 2 involves constructing graph nodes based on an initial feature set, calculating the feature cosine similarity between graph nodes to obtain positive weights, and calculating the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights; synthesizing the positive and negative weights into a matrix to obtain a feature association adjacency matrix; and performing graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain a feature association graph, which may include:

[0038] Step 201: Map each feature dimension in the initial feature set to a graph node, extract the feature data sequence corresponding to each graph node, and construct a node attribute set. Calculate the cosine of the angle between the feature data sequences of any two graph nodes based on the node attribute set to obtain the feature cosine similarity. Specifically, this includes: combining the initial feature set obtained in step 103, mapping each feature dimension in the initial feature set to a graph node, with each graph node corresponding to a business operation feature dimension, ensuring that each core feature is reflected in the graph. Then, extract the feature data sequence corresponding to each graph node. This feature data sequence is all time-series data of that feature dimension within a preset time window, arranged in chronological order. Integrate the feature data sequences of all graph nodes to construct a node attribute set. The node attribute set completely contains the feature information and temporal changes of all graph nodes. Based on the constructed node attribute set, calculate the cosine of the angle between the feature data sequences of any two graph nodes. The specific calculation process is as follows: first, obtain the feature data sequences corresponding to each of the two graph nodes, and then use the two feature data sequences as vectors respectively. sum vector ,in The feature data sequence vector representing the first graph node. The feature data sequence vector representing the second graph node; calculate the dot product of the two vectors. ,in Representative vector with vector Calculate the dot product of the two vectors; then calculate the magnitudes of the two vectors respectively. and ,in Representative vector The length of the mold, Representative vector The modulus of the two vectors is calculated; finally, the dot product of the two vectors is divided by the product of their moduli, and the result is the feature cosine similarity between the two graph nodes. The calculation formula is: = ,in Representative vector with vector The included angle.

[0039] Step 202: Perform interval normalization on the feature cosine similarity to obtain positive weights; select highly correlated node pairs based on the positive weights, extract the feature data sequences corresponding to the highly correlated node pairs to construct a joint covariance matrix, perform eigenvalue decomposition on the joint covariance matrix to extract principal component eigenvalues, proportionally reduce the principal component eigenvalues ​​to obtain the variance contribution rate, and calculate the information overlap of the feature dimensions based on the variance contribution rate to obtain the covariance overlap. Specifically, this includes: normalizing the feature cosine similarity... Interval normalization is performed using the min-max normalization method to adjust all feature cosine similarity values ​​to a preset uniform range of 0.1 to 0.9, eliminating the magnitude differences between different similarity values. The specific calculation process is as follows: Let the original feature cosine similarity value be s, where s represents the feature cosine similarity calculated in step 201. The original values, after normalization, have an intermediate value of The final positive weight is The calculation formulas are as follows: , After this normalization process, the positive weights are obtained. Positive weights The value ranges from 0.1 to 0.9, with positive weights. The positive weights directly correspond to the strength of the positive correlation between the two feature dimensions. The larger the value, the more pronounced the positive synergistic relationship between the two feature dimensions; according to the positive weight... Filter highly correlated node pairs, set the positive weight threshold to 0.6, and then apply the positive weight... Node pairs with a correlation coefficient greater than 0.6 are considered highly correlated node pairs. The feature data sequences corresponding to these highly correlated node pairs are as follows: and (n is the number of time series data within the preset time window), where The feature data sequence representing the first highly correlated node. The feature data sequence representing the second highly correlated node; the feature data sequences of each pair of highly correlated nodes are integrated to construct a joint covariance matrix. ,in The joint covariance matrix T represents the highly correlated node pairs. The joint covariance matrix T has a dimension of 2×2, and its specific expression is as follows: ,in The autocovariance of the characteristic data sequence x and Representative feature data sequence and The cross-variance, Representative feature data sequence The autocovariance, the specific formula for calculating the covariance is: ( Sequences (mean), joint covariance matrix It can reflect the collaborative change relationship of feature data between highly correlated node pairs.

[0040] For the constructed joint covariance matrix Perform eigenvalue decomposition to extract the joint covariance matrix. Principal component eigenvalues and ,in This represents the eigenvalue of the first principal component. This represents the eigenvalue of the second principal component. and It can reflect the main changing patterns and information content of the feature data; the extracted principal component eigenvalues ​​are proportionally converted, and the proportion of each principal component eigenvalue to the sum of all principal component eigenvalues ​​is calculated. This proportion is the variance contribution rate. Let the variance contribution rate of the first principal component be... The variance contribution rate of the second principal component ,in The variance contribution rate of the eigenvalues ​​of the first principal component. The variance contribution rate representing the eigenvalues ​​of the second principal component is calculated using the following formulas: , Variance contribution rate and The values ​​range from 0 to 1. A higher variance contribution rate indicates greater information overlap between the two feature dimensions. The degree of information overlap of the feature dimensions is calculated based on the variance contribution rate. Specifically, the variance contribution rates of the two principal components corresponding to highly correlated nodes are added together, then divided by 2 to obtain the covariance overlap degree 'o'. The calculation formula is as follows: The covariance overlap value o ranges from 0 to 1.

[0041] Step 203: Analyze the data distribution offset of highly correlated node pairs based on the covariance overlap. Calculate the fluctuation asymmetry parameter based on the distribution offset and the fluctuation range of the feature data sequence. Weight the distribution offset, fluctuation asymmetry parameter, and positive weights to obtain the mutual exclusion index. Specifically, this includes: first, analyzing the data distribution offset of highly correlated node pairs based on the covariance overlap 'o' obtained in step 202. The data distribution offset refers to the degree of deviation of the feature data sequence corresponding to the highly correlated node pair from the distribution trend. Let the mean of the first feature data sequence be... The mean of the second feature data sequence is The data distribution offset is d, and the specific calculation formula is d= The data distribution offset d ranges from 0 to 1. A larger d indicates a more significant difference in the distributions of the two feature dimensions. Based on the calculated distribution offset d and the fluctuation range of the feature data sequence, a fluctuation asymmetry parameter is calculated. Let the fluctuation range of the first feature data sequence be... The fluctuation range of the second characteristic data sequence is ,in The range of fluctuations in the feature data sequence representing the first highly correlated node. The fluctuation range represents the feature data sequence of the second highly correlated node. The fluctuation range is the difference between the maximum and minimum values ​​in the feature data sequence; the fluctuation asymmetry parameter is k, and the two preset weights are respectively... and ,in The weights representing the distribution offset The weights representing the fluctuation range are summed to 1. The specific calculation process is as follows: The value range of the fluctuation asymmetry parameter k is from 0 to the maximum fluctuation range of the feature data sequence. The fluctuation asymmetry parameter k can reflect the degree of asymmetry of feature fluctuations between highly correlated node pairs.

[0042] The distribution offset d, the fluctuation asymmetry parameter k, and the positive weights are combined. The weighted fusion calculation is performed, with three preset weights as follows: =0.3、 =0.4 and =0.3, where This represents the weight of the distribution offset in the calculation of the mutual exclusion index. The weight of the fluctuation asymmetric parameter in the calculation of the mutual exclusion index is represented. This represents the weight of the positive weights in the calculation of the mutual exclusion index, and the sum of these weights is 1; the mutual exclusion index is m, and the specific calculation process is m = The results of the three calculations are added together to obtain the mutual exclusion index m, which ranges from 0.03 to 0.97.

[0043] Step 204: Extract the numerical extrema of the mutual exclusion exponents and set the polarity reversal reference interval. Based on the polarity reversal reference interval, establish numerical inversion and bias compensation rules, and perform sign flipping and bias compensation processing on the mutual exclusion exponents to obtain the initial reverse sequence. Specifically, this includes: extracting the numerical extrema of all mutual exclusion exponents m obtained in step 203, including the maximum value among the mutual exclusion exponents. and minimum value Based on these two extreme values, a polarity reversal reference interval is set, ranging from 0.3 to 0.7. This interval distinguishes the positive and negative meanings of the mutual exclusion index, ensuring the accuracy of subsequent sign reversal. According to the set polarity reversal reference interval, numerical inversion and bias compensation rules are established. The numerical inversion rule involves reversing the sign of the mutual exclusion index value within the polarity reversal reference interval, i.e., subtracting the mutual exclusion index value from 1. The bias compensation rule sets a corresponding bias compensation value based on the degree of deviation of the mutual exclusion index value from the polarity reversal reference interval. Let the bias compensation value be b, where b represents the bias compensation value of the mutual exclusion index. The further the deviation from the center of the reference interval (0.5), the smaller the bias compensation value b. The range of the bias compensation value b is 0.02 to 0.05. The specific calculation process is b = This is used to correct numerical deviations that may occur after sign flipping. Based on the established rules, sign flipping and bias compensation are performed on all mutual exclusion indices m. Let the initial reverse sequence value be t. The specific calculation process is as follows: when the mutual exclusion index m is within the baseline interval (0.3≤m≤0.7), t=(1-m)+b; when the mutual exclusion index m is outside the baseline interval (m<0.3 or m>0.7), t=m+b. Through the above processing, the mutual exclusion index that originally reflected the reverse correlation is converted into a data form that can correspond to the positive weight. At the same time, the numerical deviation caused by sign flipping is eliminated through bias compensation. After the above processing, the initial reverse sequence is obtained.

[0044] Step 205: Perform boundary limiting and truncation on the initial reverse sequence to obtain a limited reverse sequence; perform amplitude compression and normalization on the limited reverse sequence to obtain negative weights. Specifically, this includes: performing boundary limiting and truncation on the initial reverse sequence value t obtained in step 204, setting the numerical boundary range of the initial reverse sequence to 0.1 to 0.9, and trunculating values ​​exceeding this boundary range to the boundary value. That is, when t > 0.9, the truncated value is 0.9; when t < 0.1, the truncated value is 0.1; when 0.1 ≤ t ≤ 0.9, the value remains unchanged, and the value of the limited reverse sequence is set to... After boundary-limiting truncation, a limit-limited inverse sequence is obtained. The numerical values ​​of the limit-limited inverse sequence are then analyzed. Amplitude compression and normalization are performed using the same min-max normalization method as the positive weights. This adjusts the values ​​of the amplitude-limited reverse sequence to a uniform range of 0.1 to 0.9, ensuring that the positive and negative weights can be used for subsequent matrix synthesis and eliminating the difference in magnitude between them. Let the median value after normalization be... The negative weight is The specific calculation process is as follows: , After amplitude compression and normalization, negative weights are obtained. .

[0045] Step 206: Align the positive and negative weights according to the mapping relationship of their corresponding graph nodes, and perform matrix dimension alignment and weighted superposition to construct a weight synthesis matrix. Based on the weight synthesis matrix, perform diagonal zeroing and threshold filtering to obtain the feature association adjacency matrix. Specifically, this includes: aligning the positive and negative weights according to the mapping relationship of their corresponding graph nodes to ensure that the dimensions of the positive and negative weight matrices are consistent, both being N×N, where N represents the total number of graph nodes; the specific expression of the positive weight matrix is: ;

[0046] in The element in the i-th row and j-th column of the positive weight matrix represents the positive weight between the i-th graph node and the j-th graph node, where i and j are both positive integers from 1 to N. The specific expression for the negative weight matrix is: ;

[0047] in The element in the i-th row and j-th column of the negative weight matrix represents the negative weight between the i-th graph node and the j-th graph node, where i and j are both positive integers from 1 to N. Each position and The weight values ​​correspond to the same pair of graph nodes; the aligned positive weight matrix and negative weight matrix are weighted and superimposed, with the two preset weights set as follows: =0.6 and =0.4, the sum of the weights is 1, construct the weight composition matrix. ,in The weighting coefficients represent the positive weighting matrix. The weighting coefficients representing the negative weight matrix are expressed as follows: After expansion, the element calculation formula is: ,in The element in the i-th row and j-th column of the weighted composite matrix represents the element where i and j are positive integers from 1 to N. The weighted composite matrix constructed through weighted superposition fully contains the forward and reverse correlation weights between all graph nodes, with values ​​ranging from 0.04 to 0.86. Based on the constructed weighted composite matrix... Perform diagonal zeroing, that is, reduce the number of elements on the diagonal of the weight synthesis matrix to zero. Set all values ​​to 0 to obtain the zeroed matrix. ,in The element in the i-th row and i-th column of the weighted composite matrix represents the correlation within the same graph node. Zeroing this correlation avoids interference from subsequent analyses. The specific expression of the matrix after zeroing is: ;

[0048] Perform threshold filtering, setting the weight threshold to 0.2, and then zeroing the matrix. Elements with values ​​less than 0.2 are set to 0, and elements with values ​​greater than or equal to 0.2 are retained, resulting in the feature association adjacency matrix A, with the specific expression as follows: ,in The element in the i-th row and j-th column of the adjacency matrix represents the feature association, where i and j are both positive integers from 1 to N, and their values ​​follow the following rules: .

[0049] Step 207: Extract the topological structure of connected paths between graph nodes based on the feature association adjacency matrix to obtain spatial mapping constraint rules; based on the spatial mapping constraint rules, reorganize the dimensions and rearrange the positions of the initial feature set to construct an initial node layout set. Specifically, this includes: extracting the topological structure of connected paths between graph nodes based on the feature association adjacency matrix obtained in step 206. This structure reflects the connection relationships and connected paths between graph nodes, clarifying the core associations and associated path distributions between nodes; based on the extracted connected path topological structure, setting specific spatial mapping constraint rules. The core of the rules is to clarify the quantitative correspondence between association weights and node spatial distances, while limiting the basic range and adjustment step size of node spatial coordinates; the initial value range of node spatial coordinates is uniformly set to 0 to 10, and the coordinate adjustment step size is fixed at 0.1 to ensure accurate and controllable layout adjustments; positive weights and node spatial distances... The spatial distance between nodes is negatively correlated; for every 0.1 increase in positive weight, the spatial distance between corresponding nodes decreases by 0.5 units, with the maximum positive weight of 0.9 corresponding to a minimum spatial distance of 5.5 units. Conversely, the spatial distance between nodes is positively correlated; for every 0.1 increase in negative weight, the spatial distance between corresponding nodes increases by 0.5 units, with the maximum negative weight of 0.9 corresponding to a maximum spatial distance of 9.5 units. Furthermore, a minimum spatial distance threshold of 2 units is set between nodes to avoid node overlap, and a maximum spatial distance threshold of 10 units is set to prevent overly dispersed node layouts. Based on this spatial mapping constraint rule, the initial feature set obtained in step 103 undergoes dimensional restructuring and position rearrangement. The feature dimensions in the initial feature set are recombined according to the rules, and the spatial coordinates of the corresponding graph nodes for each feature dimension are adjusted to ensure the node layout conforms to the feature association relationships. After dimensional restructuring and position rearrangement, an initial node layout set is constructed.

[0050] Step 208 involves performing a connection density balancing iterative process on the initial node layout set and the feature association adjacency matrix, dynamically converging the relative spatial distances between nodes and solidifying strongly associated connections, thus completing the graph topology reconstruction process and obtaining the feature association graph. Specifically, this includes: performing a connection density balancing iterative process on the initial node layout set obtained in step 207 and the feature association adjacency matrix obtained in step 206; the number of iterations for the connection density balancing iterative process is set to 50, the step size for each iteration is set to 0.1, and the preset threshold for the number of surrounding nodes of a node is set to 3 to 5. This threshold is calibrated based on the total number of nodes N in the graph to ensure a balanced node distribution; specifically, in each iteration, the number of surrounding nodes of each node is calculated. Within a 5-unit range, if there are more than 5 nodes, the distance between that node and its surrounding nodes is increased by 0.1 units each time; if there are fewer than 3 nodes, the distance between that node and its surrounding nodes is decreased by 0.1 units each time. This maintains a balanced connection density between nodes, preventing them from being too dense or too sparse. Simultaneously, the relative spatial distance between nodes is dynamically converged by combining the core association weights in the feature association adjacency matrix. Specifically, the target distance between nodes equals the initial distance minus the positive weight multiplied by 0.5, plus the negative weight multiplied by 0.5. This causes closely related nodes to move closer together and weakly related nodes to move further apart, gradually optimizing the node layout. During the iteration process, strong association edges are fixed, with a strong association weight threshold of 0.6. For strongly related node pairs with weights greater than 0.6 in the feature association adjacency matrix, their corresponding edges are fixed to ensure clear representation of strong associations. After 50 iterations, until the node layout is stable and the connection density is within the preset threshold range of 3 to 5 nodes, the graph topology reconstruction is completed, resulting in the final feature association graph.

[0051] This embodiment constructs a feature association adjacency matrix and reconstructs the graph topology to form an intuitive feature association graph, which clearly presents the association logic of various business features of an enterprise. It can accurately restore the internal structure of the enterprise's business status and avoid the problem of discrete feature space vector distribution.

[0052] In a preferred embodiment of the present invention, step 3 involves constructing an operational state migration trajectory based on a feature association graph; performing differential processing on the operational state migration trajectory to obtain a state rate vector and a state acceleration vector; extracting asset size values ​​and semantic fluctuation features from the feature association graph and converting them into state inertia constraint parameters and market environment disturbance parameters, respectively, and constructing a migration correction vector; superimposing the state rate vector, state acceleration vector, and migration correction vector to obtain pre-calibrated coordinates; and using the migration correction vector to iteratively calibrate the pre-calibrated coordinates to obtain a state evolution sequence, which may include:

[0053] Step 301: Generate time-series partitioning nodes based on a preset time window, and segment the feature data sequence corresponding to the feature association graph according to the time-series partitioning nodes to obtain a time-series graph snapshot set. Specifically, this includes: combining the feature association graph obtained in step 208, first determining the specific partitioning rules of the preset time window. The preset time window follows the previously set 12-month cycle, and the time-series is divided by month to generate 12 time-series partitioning nodes. Each time-series partitioning node corresponds to a monthly time node, ensuring that the time-series partitioning is uniform and can accurately reflect the monthly operating characteristics changes of the enterprise; extract the feature data sequence corresponding to all graph nodes in the feature association graph, and segment the feature data sequence for each feature association graph. The feature data sequences are all complete time-series data of the feature dimension within a preset 12-month time window, arranged in chronological order. According to the 12 generated time-series partitioning nodes, each feature data sequence is divided into segments, and each time-series partitioning node corresponds to a feature data subsequence. Then, the feature data subsequences of all graph nodes under the same time-series partitioning node are integrated to construct the feature association graph snapshot corresponding to that time node. Each snapshot completely retains the feature information, association relationship and spatial layout of all graph nodes under the corresponding time node. The snapshot construction of all time-series partitioning nodes is completed in sequence, and finally a time-series graph snapshot set containing 12 snapshots is obtained.

[0054] Step 302: Extract the spatial layout coordinates of all graph nodes within each snapshot in the time-series graph snapshot set, calculate the geometric center of each snapshot, and obtain the time-series center point sequence; perform spatial connection mapping on the time-series center point sequence according to the order of the time-series partitioning nodes to generate the initial migration path. Specifically, this includes: traversing each snapshot in the time-series graph snapshot set, extracting the spatial layout coordinates of all graph nodes within each snapshot, where the spatial layout coordinates of each graph node are the two-dimensional spatial coordinates determined in step 207 based on the spatial mapping constraint rules, and recording each... The x-axis and y-axis coordinates of all graph nodes within a snapshot are calculated. For each snapshot, its geometric center is calculated by adding the x-axis coordinates of all graph nodes within the snapshot, then dividing by the total number of graph nodes in the snapshot to obtain the x-axis geometric center coordinates. Similarly, the y-axis coordinates of all graph nodes within the snapshot are added together, then divided by the total number of graph nodes in the snapshot to obtain the y-axis geometric center coordinates. The combination of the x-axis and y-axis geometric center coordinates is the geometric center of the snapshot. The geometric centers of all snapshots are calculated sequentially. The nodes are divided according to time sequence, and all geometric centers are arranged in chronological order to obtain a time-series center point sequence. Each element in the time-series center point sequence corresponds to the geometric center of a snapshot at a time node. Adjacent geometric centers in the time-series center point sequence are connected spatially according to the time sequence, using straight lines to ensure that the lines reflect the spatial positional changes of the geometric centers at adjacent time nodes. After integrating all the lines, an initial migration path is generated, which can initially reflect the spatial migration trend of the enterprise's operating status over time.

[0055] Step 303: Perform curve smoothing fitting and abnormal offset point removal processing on the initial migration path to obtain the operational status migration trajectory. Specifically, this includes: performing abnormal offset point removal processing on the initial migration path obtained in step 302. First, calculate the spatial distance between each coordinate geometric center in the time-series center point sequence and the two adjacent coordinate geometric centers. The calculation method is to subtract the x-axis coordinate of the previous coordinate geometric center from the x-axis coordinate of the current coordinate geometric center and take the absolute value of the difference. Then, subtract the y-axis coordinate of the previous coordinate geometric center from the y-axis coordinate of the current coordinate geometric center and take the absolute value of the difference. Add the two absolute values ​​to obtain the spatial distance between the current coordinate and the previous coordinate. Similarly, calculate the spatial distance between the current coordinate and the next coordinate. Take the average of the two spatial distances as the average of the adjacent distances of the current coordinate.

[0056] Calculate the mean of adjacent distances between all geometric centers of the time series. Then, calculate the mean and standard deviation of all adjacent distances. The mean is calculated by summing the means of all adjacent distances and dividing by the length of the time series center point sequence. The standard deviation is calculated by summing the squares of the differences between the means of all adjacent distances and the overall mean, dividing by the length of the time series center point sequence, and then taking the square root. An anomaly threshold is set as the overall mean plus twice the standard deviation. Geometric centers with adjacent distances exceeding this threshold are identified as anomalous offset points. These anomalous offset points correspond to abnormal fluctuations in the company's operating status and do not conform to normal operating patterns; therefore, they are removed from the time series center point sequence, resulting in a center point sequence after anomaly removal. Curve smoothing fitting is then performed on the initial migration path corresponding to the anomaly-removed center point sequence. The fitting process involves using the x-axis coordinates of the anomaly-removed center point sequence as the independent variable and the y-axis coordinates as the dependent variable to construct a polynomial equation. The polynomial equation expression is as follows: Where x represents the x-axis coordinate of the centroid sequence after outlier removal, and y represents the corresponding y-axis coordinate. , , , These represent the constant term and coefficients of each degree of the polynomial, respectively. By solving the polynomial coefficients, a smooth fit is achieved, ensuring that the fitted curve accurately matches the spatial variation trend of the center point sequence while avoiding overfitting. The specific process of solving the polynomial coefficients is as follows: first, assume the coefficients of the polynomial equation, calculate the sum of squares of the differences between the coordinates of each center point and the corresponding coordinates of the assumed polynomial equation, and continuously adjust the polynomial coefficients until the sum of squares reaches its minimum value. The polynomial coefficients obtained at this time are the fitting coefficients, and the corresponding polynomial curve is the smoothed business state migration trajectory.

[0057] Step 304: Extract continuous coordinate points on the operational status migration trajectory, calculate the spatial position difference between adjacent coordinate points to obtain a trajectory coordinate difference sequence; perform first-order difference processing based on the time interval between the trajectory coordinate difference sequence and the corresponding time-series partitioning node to obtain the state rate vector. Specifically, this includes: uniformly extracting continuous coordinate points from the operational status migration trajectory in chronological order, with the extraction interval consistent with the time interval of the time-series partitioning node, i.e., extracting one coordinate point per month to ensure that the temporal sequence of the coordinate points corresponds to the previous time-series partitioning node, and ensuring that the number of extracted continuous coordinate points is consistent with the number of center point sequences after removing anomalies; calculating the spatial position difference between two adjacent coordinate points by subtracting the x-axis coordinate of the previous coordinate point from the x-axis coordinate of the subsequent coordinate point to obtain the x-axis position difference; and subtracting the y-axis coordinate of the previous coordinate point from the y-axis coordinate of the subsequent coordinate point to obtain the y-axis position difference. Position difference; combine the x-axis position difference and y-axis position difference to obtain the spatial position difference between adjacent coordinate points. Calculate the spatial position difference of all adjacent coordinate points in sequence and arrange them in chronological order to obtain the trajectory coordinate difference sequence. Since the time sequence division nodes are divided by month, and the time interval is 1 month, convert the time interval to a unified time unit (month), i.e., the value of each time interval is 1. Perform first-order difference processing based on the trajectory coordinate difference sequence and the time interval. The first-order difference processing process is to divide each spatial position difference in the trajectory coordinate difference sequence by the corresponding time interval to obtain the state rate within each time interval. Each state rate contains two components: x-axis rate and y-axis rate. Arrange all state rates in chronological order to obtain the state rate vector. The state rate vector can quantify the migration speed and direction of the enterprise's operating state within each time interval.

[0058] Step 305: Calculate the velocity difference between adjacent time points based on the state velocity vector to obtain a velocity difference sequence. Perform second-order difference processing on the velocity difference sequence to obtain the state acceleration vector. Specifically, this includes: extracting the magnitude of each state velocity from the state velocity vector obtained in step 304 by performing a square root operation on the sum of the squares of the x-axis and y-axis velocities of each state velocity; calculating the velocity difference between two adjacent time points by subtracting the velocity value of the previous time point from the velocity value of the later time point, and calculating the velocity difference for all adjacent time points sequentially. The rate difference sequence is obtained by arranging the values ​​sequentially. A second-order difference process is then performed on this sequence, which involves dividing each rate difference by the corresponding time interval (still one month) to obtain the state acceleration for each time period. Each state acceleration corresponds to a numerical value, the sign of which reflects the direction of acceleration: a positive value indicates increasing rate, a negative value indicates decreasing rate, and the magnitude of the value reflects the magnitude of the acceleration. All state accelerations are arranged chronologically to obtain a state acceleration vector. This vector quantifies the changing trend of the enterprise's operational state transition speed, further improving the quantitative analysis of the temporal changes in the enterprise's operational state.

[0059] Step 306: Analyze the node attributes of the feature association graph to extract the asset size value and semantic fluctuation features; perform a magnitude mapping transformation on the asset size value to obtain the state inertia constraint parameters. Specifically, this includes: analyzing the attribute information of all graph nodes in the feature association graph obtained in step 208. The node attribute information includes the specific numerical value and semantic description of the enterprise's operating characteristics corresponding to each graph node. From this, graph nodes corresponding to the asset size feature are selected, and the asset size value corresponding to that node is extracted. The asset size value is the average total assets of the enterprise within a preset time window, ensuring that the value can reflect the actual asset size of the enterprise; at the same time, graph nodes corresponding to the enterprise's operating fluctuation features are selected, and the semantic fluctuation features corresponding to that node are extracted. The semantic fluctuation features include the direction, amplitude, duration, and other semantic aspects of the enterprise's operating fluctuations. The descriptions, such as stable operation, slight fluctuations, and significant decline, are quantified and transformed into calculable numerical forms. For example, stable operation corresponds to 0.1, slight fluctuations to 0.3, significant fluctuations to 0.5, and significant declines to 0.7, ensuring that semantic fluctuation features can participate in subsequent calculations. The extracted asset size values ​​are then subjected to a magnitude mapping transformation. The mapping transformation rules are set in conjunction with the industry benchmark asset size. First, the ratio of the target company's asset size value to the industry average asset size value is calculated. The target company's asset size value is divided by the industry average asset size value to obtain the asset size ratio. Then, the asset size ratio is mapped to a value range of 0.1 to 0.9. The mapping process is to multiply the asset size ratio by 0.8 and add 0.1 to obtain the state inertia constraint parameters.

[0060] Step 307: Based on the state inertia constraint parameters, set the amplitude calculation benchmark. Use the amplitude calculation benchmark to perform range normalization on the semantic fluctuation features to obtain market environment disturbance parameters. Align the state inertia constraint parameters with the market environment disturbance parameters in terms of vector dimensions to construct a migration correction vector. Specifically, this includes: setting the amplitude calculation benchmark based on the state inertia constraint parameters obtained in Step 306. The process of setting the amplitude calculation benchmark is to multiply the state inertia constraint parameters by 0.5 and add 0.2 to obtain the amplitude calculation benchmark. The value range of the amplitude calculation benchmark is 0.25 to 0.65 to ensure that the benchmark value can adapt to the market environment disturbance quantification needs of enterprises with different asset sizes. Use the amplitude calculation benchmark to perform range normalization on the semantic fluctuation features obtained in Step 306. The range normalization process is to first subtract the minimum value of the semantic fluctuation feature from the maximum value of the semantic fluctuation feature to obtain the range of the semantic fluctuation feature, and then subtract the minimum value of the semantic fluctuation feature from the value of each semantic fluctuation feature to obtain the difference. Dividing by the range of the semantic fluctuation feature yields the normalized semantic fluctuation feature. Multiplying this normalized feature by the amplitude calculation benchmark yields the market environment disturbance parameter. This parameter quantifies the disturbance effect of market environment changes on the migration of a company's operating state; a larger value indicates a stronger disturbance. The state inertia constraint parameter and the market environment disturbance parameter are then aligned in vector dimension, both converted to two-dimensional vectors. When the state inertia constraint parameter is converted to a two-dimensional vector, both the x-axis and y-axis components are the values ​​of the state inertia constraint parameter, and vice versa, ensuring consistent dimensions. Finally, the aligned state inertia constraint parameter vector and the market environment disturbance parameter vector are weighted and superimposed, with each weight set to 0.5. Multiplying the state inertia constraint parameter vector by 0.5 and adding the market environment disturbance parameter vector by 0.5 yields the migration correction vector.

[0061] Step 308: Spatial dimension alignment and component superposition of the state velocity vector and state acceleration vector are performed to obtain the base position coordinates; the base position coordinates and migration correction vector are superimposed to obtain the pre-calibration coordinates, specifically including: aligning the state velocity vector with the state acceleration vector and performing component superposition to obtain the pre-calibration coordinates. With the state acceleration vector Perform spatial dimension alignment, state-rate vector ,in The x-axis velocity component, The y-axis velocity component; the state acceleration vector is converted into a two-dimensional vector. ,in The x-axis acceleration component is... For the y-axis acceleration component, ensure the two vector spaces have the same dimension; perform component superposition on the aligned two vectors, using the superposition formula: ,in Based on the base position coordinates, Based on the x-axis component, Based on the basic y-axis component, the basic position coordinates can reflect the basic trend of the enterprise's operational status migration, combining the effects of migration speed and acceleration; the basic position coordinates are then compared with the migration correction vector obtained in step 307. The vector components are superimposed using the following formula: ,in For pre-calibration coordinates, To precalibrate the x-axis components, This is for pre-calibrating the y-axis component.

[0062] Step 309: Extract the spatial position difference between the pre-calibrated coordinates and the migration correction vector to construct a coordinate deviation sequence; set a convergence threshold based on the coordinate deviation sequence, and use the migration correction vector to perform direction compensation and step size adjustment on the pre-calibrated coordinates to obtain the coordinates for the first round of iterations. Specifically, this includes: calculating the pre-calibrated coordinates obtained in step 308. With migration correction vector The spatial difference between them is calculated using the following formula: ,in For the spatial difference between a single pre-calibrated coordinate and the migration correction vector, For the difference in position along the x-axis, To calculate the spatial position difference between the y-axis and all pre-calibrated coordinates and the migration correction vector, and arrange them in chronological order, we obtain the coordinate deviation sequence. (where 'a' represents the number of pre-calibrated coordinates), the coordinate deviation sequence reflects the degree of deviation between the pre-calibrated coordinates and the ideal coordinates; based on the coordinate deviation sequence, a convergence threshold is set, and the mean of all spatial position differences in the coordinate deviation sequence is first calculated using the following formula: ,in The mean of the deviation, , Let x and y be the x-axis and y-axis components of the i-th deviation vector, respectively, and then set a convergence threshold. The convergence threshold is dynamically adjusted based on the actual situation of the coordinate deviation sequence to ensure accurate determination of whether the coordinate deviation has converged. The migration correction vector is used to perform direction compensation and step size adjustment on the pre-calibrated coordinates. The direction compensation formula is as follows: ,in For sign function; step size adjustment coefficient The actual compensation step size is The final adjusted formula is as follows: , in These are the coordinates for the first iteration.

[0063] Step 310 involves recalculating the spatial position difference between the first-round iteration coordinates and the migration correction vector to update the second-round deviation sequence. Based on this second-round deviation sequence, direction compensation and step size adjustment are performed iteratively until the deviation sequence converges to the convergence threshold range, resulting in the final calibration coordinate set. Specifically, this includes calculating the first-round iteration coordinates obtained in step 309. With migration correction vector The spatial difference between them is calculated using the following formula: The spatial differences of all coordinates from the first iteration are calculated sequentially, arranged in chronological order, and the resulting second-round deviation sequence is updated. Determine the modulus of all spatial positional differences in the second-round deviation sequence. Are all values ​​less than or equal to the convergence threshold? If there exists a value greater than The deviation is then calculated based on the second-round deviation sequence, and the direction compensation and step size adjustment formula in step 309 is used to readjust the coordinates of the first-round iteration to obtain the coordinates of the second-round iteration. Then, the formula is applied... Recalculate the deviation, update the deviation sequence, and repeat the above process of adjusting and recalculating the deviation, performing iterative iterations. In each iteration, gradually reduce the coordinate deviation until all coordinates in the deviation sequence are correct. ( (where the iteration number is the number of iterations), then stop iterating. All the iterative coordinates obtained at this point constitute the final calibration coordinate set. ,in This determines the final number of calibration coordinates.

[0064] Step 311: Perform boundary scanning and spatial connectivity determination on the final calibration coordinate set to extract peripheral coordinate points and generate an initial boundary node set; perform curvature calculation and inflection point screening on the initial boundary node set, connect adjacent inflection points according to the curvature change gradient to obtain the state boundary contour line; perform orthogonal projection mapping on the final calibration coordinate set into the state boundary contour line to obtain the contour constraint coordinate set; reassemble the contour constraint coordinate set into a sequence to obtain the state evolution sequence, specifically including: performing boundary scanning and spatial connectivity determination on the final calibration coordinate set; performing boundary scanning and spatial connectivity determination on the final calibration coordinate set to obtain the state evolution sequence ... Boundary scan processing is performed, using a point-by-point scanning method to traverse all... Determine whether each coordinate is an outer coordinate point, and the determination criteria are as follows: and , or and Coordinates that meet this standard are identified as peripheral coordinate points. All peripheral coordinate points are extracted and arranged in chronological order to generate an initial set of boundary nodes. ( The initial boundary node set (number of peripheral coordinate points) can reflect the peripheral range of the enterprise's operational status migration. A shape contour extraction algorithm is introduced to process the initial boundary node set. The specific calculation process is to traverse each peripheral coordinate point. Calculate the coordinates of this point and its two adjacent coordinates. The angle between the lines is calculated by first calculating the vector. The formula for calculating the cosine of the included angle is: The corresponding included angle ,Will The coordinate points are identified as redundant noise points and removed, resulting in the denoised boundary node set: , where f is the number of boundary nodes after denoising. Number of peripheral coordinate points ; Calculate the number of adjacent nodes The coordinates of the midpoint are given by the formula: Using midpoint coordinates Connect adjacent nodes to form a continuous and smooth initial contour line. Complete the shape and contour extraction.

[0065] For the extracted initial contour line The Hu invariant moment calculation is performed by first calculating the geometric center of all coordinate points on the initial contour line. The calculation formula is:

[0066] ,by Establish a local coordinate system with the origin; then calculate the moments of the contour line, including the first moment. ;

[0067] Second moment , , ;

[0068] Third moment , , ;

[0069] The central moment is calculated based on the moments of each order, and the second-order central moment. , , ;

[0070] Third-order central moment , , , ;

[0071] Calculate the normalized central moments Finally, seven Hu invariant moments were constructed, namely: ; ; ; ; ; ;

[0072] By using Hu invariant moment calculations, the shape characteristics of the contour line are ensured to remain unchanged with translation, rotation, and scaling, thus improving the stability of the contour line; the initial boundary node set... Perform curvature calculation and inflection point filtering. The curvature calculation process involves taking three adjacent nodes. Calculate the radius of the arc formed by the three points. curvature The curvature of each of the three adjacent nodes is calculated sequentially to obtain the curvature sequence. The inflection point screening process involves setting a curvature threshold. ,in The number of peripheral coordinate points. For curvature; The nodes are identified as inflection points, and all inflection points are extracted to obtain the inflection point set. (j represents the number of inflection points), and these inflection points correspond to key change nodes in the migration of the enterprise's operating status.

[0073] Connect adjacent inflection points according to the curvature change gradient; curvature change gradient = ,according to Connect the elements sequentially from largest to smallest, using smooth curves during the connection process, and combining the initial contour lines obtained from the shape contour extraction algorithm. Based on the calculation results of Hu's invariant moments, the accuracy of inflection point connections was optimized. After all inflection points were connected, the state boundary contour line was obtained. The state boundary contour line can clearly define the scope and key change nodes of the enterprise's business state transition, and has invariance to translation, rotation, and scaling, thus exhibiting stronger stability; then the final calibration coordinate set is... Each coordinate in Towards the state boundary contour line Internally, orthogonal projection mapping is performed, and the projection mapping formula is: ,in Outline The normal vector at the intersection of the projections. for To the outline The vertical distance ensures the projected coordinates Located inside the state boundary contour line, extract all projected coordinates to obtain the contour constraint coordinate set. ,in The set of contour constraint coordinates, representing the number of coordinates required, ensures that the enterprise's operational trajectory better aligns with its actual business scope, preventing it from exceeding reasonable boundaries. Finally, the contour constraint coordinate set... The sequence is reassembled chronologically, supplementing the time node information and feature association information corresponding to each coordinate. Combined with the shape features reflected by Hu invariant moments, the state evolution sequence is obtained. ,in The total number of nodes in the state evolution sequence.

[0074] This embodiment combines the complete calculation of each order of moments, central moments, and normalized moments using the Hu invariant moment calculation algorithm, making the state boundary contour line more stable and accurate, possessing invariance to translation, rotation, and scaling, and clearly defining the reasonable range of enterprise business state transitions.

[0075] In a preferred embodiment of the present invention, step 4, dividing the state evolution sequence to obtain a distribution result; extracting core features based on the distribution result to obtain dynamic labels; and fusing the dynamic labels with the basic attribute labels in the initial feature set to obtain an enterprise user profile, may include:

[0076] Step 401: Perform time-series segmentation and density assessment on the state evolution sequence to obtain a feature density distribution sequence; set interval boundaries based on the feature density distribution sequence to divide the time segments corresponding to the stable period, growth period, fluctuation period, and decline period, and obtain the life cycle division result; map the state evolution sequence to the time segments corresponding to the life cycle division result, extract the spatial clustering pattern of each time segment, and obtain the distribution result. Specifically, this includes: obtaining the state evolution sequence obtained in step 311, which is a set of coordinates of the continuous change of the enterprise's operating status over time, containing the operating status coordinates and related feature information corresponding to each time node, which meets the needs of monitoring the time-series changes of enterprise operations in supply chain finance and corporate credit risk management of commercial banks; perform time-series segmentation processing on the state evolution sequence, and use the sliding time window method for segmentation. The length of the sliding time window is set to a monthly cycle consistent with the previous time-series division nodes, that is, each time window corresponds to a monthly time segment, and the sliding step size is set to 1 time unit to ensure that each time node can be included in the corresponding time segment. By sliding the window sequentially, the entire state evolution sequence is divided into several continuous and non-overlapping monthly time segments.

[0077] Density assessment is performed on the state evolution sequence coordinates within each time segment. The specific process involves counting the number of all state coordinates within each time segment and calculating the clustering degree of the coordinates. Specifically, this involves first counting the total number of coordinates within each time segment, then calculating the area of ​​the corresponding spatial region, and dividing the total number of coordinates by the area of ​​the spatial region to obtain the characteristic density value for each time segment. This process is repeated for all time segments, and the characteristic density values ​​are arranged chronologically to obtain the characteristic density distribution sequence. Based on this sequence, interval boundaries are set. This involves first calculating the maximum and minimum values ​​of all characteristic density values ​​in the sequence, then calculating the difference between the maximum and minimum values. This difference is then divided into four intervals, each corresponding to a stage in the enterprise life cycle. The interval with the highest and least fluctuating characteristic density value corresponds to the stable period, the interval with a continuously rising characteristic density value corresponds to the growth period, and the interval with large fluctuations in characteristic density value without significant increase corresponds to the growth period. The intervals with a downward trend correspond to the fluctuation period, and the intervals with a continuous downward trend in feature density values ​​correspond to the decay period. Based on the range of these four intervals, interval boundaries are set, and each feature density value in the feature density distribution sequence is mapped to the corresponding interval, thus dividing the time segments corresponding to the stable period, growth period, fluctuation period, and decay period. Each time segment corresponds to a life cycle stage, thereby obtaining the life cycle division result. The state evolution sequence is mapped to the time segments corresponding to the life cycle division result, that is, the coordinates and associated features of each time node in the state evolution sequence are mapped to the time segment of its life cycle stage. Then, the spatial clustering pattern of each time segment is extracted. The extraction process involves analyzing the spatial distribution of all state coordinates within each time segment, determining whether the coordinates are in a concentrated or dispersed state, counting the number of cluster centers and the size of the cluster range within each time segment, and integrating these cluster-related features to obtain the spatial clustering pattern of each time segment. After integrating the spatial clustering patterns of all time segments, the distribution result is obtained.

[0078] Step 402: Perform feature variance calculation on the distribution results to obtain a feature fluctuation amplitude sequence; perform numerical sorting processing on the feature fluctuation amplitude sequence to obtain a feature sorting sequence; extract risk indication features and value-driven features at preset sorting positions according to the feature sorting sequence to construct a core feature set; assign weight parameters for each time segment to the core feature set to obtain a stage weight coefficient matrix. Specifically, this includes: obtaining the distribution results obtained in step 401, which contains the spatial clustering pattern and corresponding state coordinate features of time segments in each life cycle stage; for each time segment in the distribution results, extracting key features of all state coordinates within that time segment. These key features include spatial location features and correlation features of the coordinates, etc. These features are all extracted from the previous feature correlation map and state evolution sequence, which are closely aligned with the enterprise's... The core monitoring dimension of operational status involves calculating the variance of key features within each time segment. The specific process is as follows: first, calculate the mean of each key feature across all coordinates within that time segment; then, calculate the difference between the feature value corresponding to each coordinate and the mean; square each difference; sum all squared differences; and divide the sum of squared differences by the total number of coordinates within that time segment to obtain the variance of the key feature within that time segment. This process is repeated for all key features across all time segments. These variance values ​​are then arranged chronologically and by feature type to obtain a feature fluctuation amplitude sequence. This sequence reflects the degree of fluctuation of each key feature across different time segments. Greater fluctuations indicate more drastic changes in the company's operational status along that feature dimension, potentially corresponding to higher operational risks.

[0079] Numerical sorting is performed based on the characteristic fluctuation amplitude sequence. The specific process involves first arranging all variance values ​​in the sequence from largest to smallest, preserving the feature type and time segment information corresponding to each variance value during the sorting process to avoid disconnection between the sorted features and the time segments. After sorting, a feature sorting sequence is obtained. The features corresponding to the highest-ranking variance values ​​in this sequence are the most volatile, while those at the bottom are the most stable. Risk indicator features and value-driving features are extracted from the feature sorting sequence at preset sorting positions. The preset sorting positions are set according to the needs of enterprise operational risk monitoring and value assessment. The first 30% of the features in the sorting sequence are used as risk indicator features, which exhibit drastic fluctuations and can promptly reflect abnormal changes in the enterprise's operational status, providing support for operational risk prediction. The middle 40% of the features in the sorting sequence are used as value-driving features, which exhibit relatively stable fluctuations and can reflect the enterprise's core operational value and development potential. These risk indicator features and value-driving features are extracted and integrated to construct a core feature set, which accurately covers the two core dimensions of enterprise operational risk and operational value.

[0080] Weight parameters for each time segment are assigned to the core feature set. The specific process involves setting a base weight based on the importance of each life cycle stage. The base weight ranges uniformly from 0 to 1, and the sum of the base weights for all life cycle stages is 1. Specifically, the base weight ranges from 0.3 to 0.4 for the stable period, 0.25 to 0.35 for the growth period, 0.15 to 0.25 for the fluctuation period, and 0.05 to 0.15 for the decline period. The base weights for the stable and growth periods are in a higher range, while those for the fluctuation and decline periods are in a relatively lower range. Then... The weights of each feature are adjusted based on its variance within the corresponding time segment. Features with larger variance values ​​have higher weights, and features with smaller variance values ​​have lower weights. Specifically, the weight is calculated by multiplying the base weight by the ratio of the feature's variance value to the sum of the variance values ​​of all features within that time segment. This yields the weight parameter for each core feature within each time segment. The weight parameter for each core feature within its corresponding time segment ranges from 0 to 1, and the sum of the weight parameters for all core features within each time segment is 1. The weight parameters of all core features within each time segment are then organized by feature type and time segment to construct a stage weight coefficient matrix. The specific expression for the stage weight coefficient matrix is ​​as follows: ;

[0081] in, The total number of time segments representing the lifecycle stage. Represents the total number of core features in the core feature set. Representing the Within the first time segment The weight parameters of each core feature, and satisfying The sum of all elements in the row vector corresponding to each time segment is 1.

[0082] Step 403: Perform semantic label mapping on the core feature set based on the stage weight coefficient matrix to obtain dynamic labels; extract enterprise registration information and industry classification dimensions from the initial feature set to obtain basic attribute labels; align the dynamic labels and basic attribute labels by dimension and hierarchical association to construct a multi-dimensional label fusion matrix; perform structured encoding processing on the multi-dimensional label fusion matrix to output the enterprise user profile. Specifically, this includes: obtaining the stage weight coefficient matrix and core feature set obtained in step 402; performing semantic label mapping on the core feature set based on the stage weight coefficient matrix; the specific process of semantic label mapping is to first construct a semantic label dictionary, which contains the correspondence between core features and semantic labels, combined with the needs of supply chain finance and commercial bank corporate credit risk management. The goal is to map risk indication features to risk-related semantic tags and value-driven features to value-related semantic tags. The strength of the tag for each core feature is determined by its weight parameter; the higher the weight parameter, the stronger the tag. The specific mapping process involves first finding the basic tag corresponding to each core feature in the semantic tag dictionary, and then multiplying the weight parameter of the core feature by the baseline strength of the basic tag. The baseline strength ranges from 0 to 10, resulting in the semantic tag and strength corresponding to the core feature. The tag strength ranges from 0 to 10. The semantic tags corresponding to all core features are integrated according to time segments to obtain dynamic tags. Dynamic tags can reflect the changes in the operating status of an enterprise at different life cycle stages, possessing temporal dynamism and conforming to the dynamic evolution characteristics of the enterprise's operating status.

[0083] Enterprise registration information and industry classification dimensions are extracted from the initial feature set. The initial feature set is a collection of basic enterprise information obtained through big data collection and integration. The extraction process involves filtering out enterprise registration information and industry classification-related information from the initial feature set. Enterprise registration information includes basic information such as enterprise registration time, registered address, registered capital, and core business personnel. The industry classification dimension includes classification information such as the enterprise's primary industry, secondary industry, and sub-sector. This extracted information is standardized to remove duplicate and redundant information and unify the information format, resulting in basic attribute tags. Basic attribute tags reflect the static basic characteristics of an enterprise and are a fundamental component of the enterprise user profile. Dynamic tags are then aligned and hierarchically associated with the basic attribute tags. The specific process of dimension alignment involves unifying the temporal dimension of the dynamic tags with the static dimension of the basic attribute tags to determine the common association dimension of the two tag sets, ensuring that the tags are related. The specific process of hierarchical association without dimensional conflicts involves using basic attribute tags as the bottom layer and dynamic tags as the upper layer. The bottom-layer basic attribute tags support the upper-layer dynamic tags, which reflect changes in the enterprise's operating status based on the basic attributes. The two tag sets are linked hierarchically and dimensionally to construct a multi-dimensional tag fusion matrix. This matrix contains information such as basic attribute tags, dynamic tags, tag strength, and association relationships, achieving the organic integration of static and dynamic tags. Structured encoding is then performed on the multi-dimensional tag fusion matrix. The specific process involves first determining the encoding rules, converting each tag and its association information in the multi-dimensional tag fusion matrix into a standardized encoding form, and preserving the hierarchical relationship and strength information of the tags during the encoding process to ensure that the encoded tags can be recognized and used by the subsequent system. After encoding, the encoding results are integrated and sorted to remove encoding errors and redundant information, forming a structured enterprise user profile.

[0084] This embodiment accurately divides the enterprise life cycle stages by temporal segmentation and density assessment of the state evolution sequence, clearly presenting the clustering characteristics of the enterprise's operating state at different stages.

[0085] like Figure 2 As shown, embodiments of the present invention also provide an enterprise user profile construction system based on big data mining, comprising:

[0086] The processing module is used to acquire multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set.

[0087] The calculation module is used to construct graph nodes based on the initial feature set, calculate the feature cosine similarity between graph nodes to obtain positive weights, calculate the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights, synthesize the positive and negative weights into a matrix to obtain a feature association adjacency matrix, and perform graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain the feature association graph.

[0088] The calibration module is used to construct the business state migration trajectory based on the feature association map; perform differential processing on the business state migration trajectory to obtain the state rate vector and the state acceleration vector; extract the asset size value and semantic fluctuation features from the feature association map and convert them into state inertia constraint parameters and market environment disturbance parameters respectively, and construct the migration correction vector; superimpose the state rate vector, the state acceleration vector and the migration correction vector to obtain the pre-calibration coordinates; use the migration correction vector to iteratively calibrate the pre-calibration coordinates to obtain the state evolution sequence.

[0089] The output module is used to divide the state evolution sequence to obtain the distribution results; extract core features based on the distribution results to obtain dynamic labels; and fuse the dynamic labels with the basic attribute labels in the initial feature set to obtain the enterprise user profile.

[0090] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0091] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0092] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing enterprise user profiles based on big data mining, characterized in that, The method includes: Step 1: Obtain multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set; Step 2: Construct graph nodes based on the initial feature set, calculate the feature cosine similarity between graph nodes to obtain positive weights, and calculate the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights; synthesize the positive and negative weights into a feature association adjacency matrix, and perform graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain the feature association graph. Step 3: Construct the business state migration trajectory based on the feature association map; perform differential processing on the business state migration trajectory to obtain the state rate vector and the state acceleration vector; extract the asset size value and semantic fluctuation features from the feature association map and convert them into state inertia constraint parameters and market environment disturbance parameters respectively, and construct the migration correction vector; superimpose the state rate vector, the state acceleration vector and the migration correction vector to obtain the pre-calibrated coordinates, and use the migration correction vector to iteratively calibrate the pre-calibrated coordinates to obtain the state evolution sequence; Step 4: Divide the state evolution sequence to obtain the distribution results; extract core features based on the distribution results to obtain dynamic labels; fuse the dynamic labels with the basic attribute labels in the initial feature set to obtain the enterprise user profile.

2. The method for constructing enterprise user profiles based on big data mining according to claim 1, characterized in that, Acquire multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set, including: Acquire multi-source heterogeneous operational data of the target enterprise within a preset time window. The multi-source heterogeneous operational data includes structured financial indicator data, semi-structured transaction flow data, and unstructured text data. Multi-source heterogeneous operational data is formatted, missing values ​​are filled, and outliers are filtered to obtain a preprocessed data stream. Based on the preprocessed data stream, structured indicator parsing, semi-structured pipeline sequence splitting, and unstructured text entity recognition are performed to extract key feature fields. The key feature fields are normalized to obtain standardized feature fields. The standardized feature fields are then vectorized according to data type to obtain basic dimension feature vectors. The basic dimension feature vectors are input into a preset dimension expansion matrix for linear weighting transformation and spatial feature stretching is performed through a nonlinear mapping function to obtain high-dimensional projected feature vectors. The high-dimensional projected feature vectors are then concatenated by feature dimensions and aligned temporally to map to a unified high-dimensional feature space to obtain the initial feature set.

3. The method for constructing enterprise user profiles based on big data mining according to claim 2, characterized in that, Graph nodes are constructed based on the initial feature set. Positive weights are obtained by calculating the cosine similarity between graph nodes. Negative weights are obtained by calculating the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights, including: Each feature dimension in the initial feature set is mapped to a graph node, and the feature data sequence corresponding to each graph node is extracted to construct a node attribute set. Based on the node attribute set, the cosine value of the angle between the feature data sequences of any two graph nodes is calculated to obtain the feature cosine similarity. The cosine similarity of features is normalized to obtain positive weights. Highly correlated node pairs are selected based on the positive weights, and the feature data sequences corresponding to the highly correlated node pairs are extracted to construct a joint covariance matrix. The joint covariance matrix is ​​decomposed to extract principal component eigenvalues. The principal component eigenvalues ​​are proportionally converted to obtain the variance contribution rate. The information overlap of the feature dimensions is calculated based on the variance contribution rate to obtain the covariance overlap. The data distribution offset of highly correlated node pairs is analyzed based on the covariance overlap. The fluctuation asymmetry parameter is calculated based on the distribution offset and the fluctuation range of the feature data sequence. The distribution offset, fluctuation asymmetry parameter and positive weight are weighted and fused to obtain the mutual exclusion index. Extract the numerical extreme values ​​of the mutual exclusion index to set the polarity reversal reference interval. Based on the polarity reversal reference interval, establish numerical inversion and bias compensation rules, and perform sign flipping and bias compensation processing on the mutual exclusion index to obtain the initial reverse sequence. The initial inverse sequence is subjected to boundary limiting and truncation to obtain the limiting inverse sequence; the limiting inverse sequence is then subjected to amplitude compression and normalization to obtain negative weights.

4. The method for constructing enterprise user profiles based on big data mining according to claim 3, characterized in that, The positive and negative weights are combined to obtain a feature association adjacency matrix. Based on the feature association adjacency matrix, the initial feature set is reconstructed using graph topology to obtain a feature association graph, including: The positive and negative weights are aligned and weighted according to the mapping relationship of the corresponding graph nodes to construct a weight synthesis matrix; based on the weight synthesis matrix, diagonal zeroing and threshold filtering are performed to obtain the feature association adjacency matrix. The topology of connected paths between nodes in the graph is extracted based on the feature association adjacency matrix to obtain spatial mapping constraint rules; the initial feature set is reorganized in dimensions and rearranged in position based on the spatial mapping constraint rules to construct an initial node layout set. The initial node layout set and the feature association adjacency matrix are subjected to connection density balancing iterative processing. The relative spatial distance between nodes is dynamically converged and the strong association connection edges are solidified. The graph topology reconstruction process is completed to obtain the feature association graph.

5. The method for constructing enterprise user profiles based on big data mining according to claim 4, characterized in that, Constructing operational status migration trajectories based on feature association maps, including: Based on a preset time window, time-series partitioning nodes are generated, and the feature data sequence corresponding to the feature association map is segmented according to the time-series partitioning nodes to obtain a set of time-series map snapshots. Extract the spatial layout coordinates of all nodes in each snapshot of the time-series map snapshot set, calculate the geometric center of each snapshot, and obtain the time-series center point sequence; perform spatial connection mapping on the time-series center point sequence according to the order of the time-series node division to generate the initial migration path; The initial migration path is subjected to curve smoothing fitting and abnormal offset point removal processing to obtain the operational status migration trajectory.

6. The method for constructing enterprise user profiles based on big data mining according to claim 5, characterized in that, The state velocity vector and state acceleration vector are obtained by performing differential processing on the operational state transition trajectory; Asset size numerical values ​​and semantic fluctuation features are extracted from the feature association graph and transformed into state inertia constraint parameters and market environment disturbance parameters, respectively. A migration correction vector is then constructed, including: Extract continuous coordinate points on the operational state migration trajectory, calculate the spatial position difference between adjacent coordinate points to obtain a trajectory coordinate difference sequence; perform first-order difference processing based on the time interval between the trajectory coordinate difference sequence and the corresponding time sequence division node to obtain the state rate vector; The numerical difference in velocity between adjacent time points is calculated based on the state velocity vector to obtain a velocity difference sequence. The second-order difference processing is then performed on the velocity difference sequence to obtain the state acceleration vector. The node attributes of the feature association graph are analyzed to extract the asset size numerical value and semantic fluctuation features; the magnitude mapping transformation of the asset size numerical value is performed to obtain the state inertia constraint parameters. Based on the state inertia constraint parameters, an amplitude calculation benchmark is set, and the semantic fluctuation features are normalized using the amplitude calculation benchmark to obtain market environment disturbance parameters. The state inertia constraint parameters and market environment disturbance parameters are aligned in vector dimension to construct a migration correction vector.

7. The method for constructing enterprise user profiles based on big data mining according to claim 6, characterized in that, The pre-calibrated coordinates are obtained by superimposing the state velocity vector, state acceleration vector, and migration correction vector. The pre-calibrated coordinates are then iteratively calibrated using the migration correction vector to obtain the state evolution sequence, including: The state velocity vector and state acceleration vector are spatially aligned and their components are superimposed to obtain the base position coordinates; the base position coordinates are then superimposed with the migration correction vector to obtain the pre-calibration coordinates. Extract the spatial position difference between the pre-calibrated coordinates and the migration correction vector to construct a coordinate deviation sequence; set a convergence judgment threshold based on the coordinate deviation sequence; use the migration correction vector to perform direction compensation and step size adjustment on the pre-calibrated coordinates to obtain the coordinates of the first round of iteration. The spatial position difference between the first round of iteration coordinates and the migration correction vector is recalculated to update the second round of deviation sequence. Based on the second round of deviation sequence, direction compensation and step size adjustment are performed. The iteration is repeated until the deviation sequence converges to the convergence judgment threshold range to obtain the final calibration coordinate set. Boundary scanning and spatial connectivity determination are performed on the final calibration coordinate set to extract peripheral coordinate points and generate an initial boundary node set. Curvature calculation and inflection point screening are performed on the initial boundary node set, and adjacent inflection points are connected according to the curvature change gradient to obtain the state boundary contour line. The final calibration coordinate set is orthogonally projected onto the interior of the state boundary contour line to obtain the contour constraint coordinate set. The contour constraint coordinate set is then reassembled into a sequence to obtain the state evolution sequence.

8. The method for constructing enterprise user profiles based on big data mining according to claim 7, characterized in that, The state evolution sequence is divided to obtain the distribution results; core features are extracted based on the distribution results to obtain dynamic labels; By fusing dynamic tags with basic attribute tags from the initial feature set, a user profile for the enterprise is obtained, including: The state evolution sequence is segmented temporally and evaluated for density to obtain a characteristic density distribution sequence. Based on the characteristic density distribution sequence, interval boundaries are set to divide the time segments corresponding to the stable period, growth period, fluctuation period and decline period, and the life cycle division results are obtained. The state evolution sequence is mapped to the time segments corresponding to the life cycle division results, and the spatial clustering pattern of each time segment is extracted to obtain the distribution results. The characteristic variance is calculated on the distribution results to obtain the characteristic fluctuation amplitude sequence; numerical sorting is performed on the characteristic fluctuation amplitude sequence to obtain the characteristic sorting sequence; risk indication features and value driving features at preset sorting positions are extracted from the characteristic sorting sequence to construct a core feature set; weight parameters for each time segment are assigned to the core feature set to obtain the stage weight coefficient matrix. Semantic label mapping is performed on the core feature set based on the stage weight coefficient matrix to obtain dynamic labels; enterprise registration information and industry classification dimensions are extracted from the initial feature set to obtain basic attribute labels; the dynamic labels and basic attribute labels are dimensionally aligned and hierarchically associated to construct a multi-dimensional label fusion matrix; structured encoding processing is performed on the multi-dimensional label fusion matrix to output the enterprise user profile.

9. A system for constructing enterprise user profiles based on big data mining, wherein the system implements the method as described in any one of claims 1 to 8, characterized in that, include: The processing module is used to acquire multi-source heterogeneous operational data of the target enterprise within a preset time window, preprocess the multi-source heterogeneous operational data, extract key feature fields, and map the key feature fields to a high-dimensional feature space to obtain an initial feature set. The calculation module is used to construct graph nodes based on the initial feature set, calculate the feature cosine similarity between graph nodes to obtain positive weights, calculate the covariance overlap and mutual exclusion index between feature dimensions based on the positive weights to obtain negative weights, synthesize the positive and negative weights into a matrix to obtain a feature association adjacency matrix, and perform graph topology reconstruction processing on the initial feature set based on the feature association adjacency matrix to obtain the feature association graph. The calibration module is used to construct the business state migration trajectory based on the feature association map; and to perform differential processing on the business state migration trajectory to obtain the state velocity vector and the state acceleration vector. The asset size numerical value and semantic fluctuation features are extracted from the feature association map and converted into state inertia constraint parameters and market environment disturbance parameters, respectively, and a migration correction vector is constructed. The state rate vector, state acceleration vector and migration correction vector are superimposed to obtain the pre-calibrated coordinates. The pre-calibrated coordinates are iteratively calibrated using the migration correction vector to obtain the state evolution sequence. The output module is used to divide the state evolution sequence and obtain the distribution results; based on the distribution results, core features are extracted to obtain dynamic labels. By fusing dynamic tags with basic attribute tags from the initial feature set, a user profile for the enterprise is obtained.

10. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 8.

Images