CN120235634A

CN120235634A - Enterprise credit assessment method based on knowledge graph

Info

Publication number: CN120235634A
Application number: CN202510428052.7A
Authority: CN
Inventors: 倪晔波; 孔杰
Original assignee: Suzhou Huixi Zhirong Information Technology Co ltd
Current assignee: Suzhou Huixi Zhirong Information Technology Co ltd
Priority date: 2025-04-07
Filing date: 2025-04-07
Publication date: 2025-07-01

Abstract

The invention belongs to the technical field of enterprise credit assessment, and discloses an enterprise credit assessment method based on a knowledge graph, which comprises the steps of obtaining enterprise financial radar data, business ecological chain data and market pulsation data and constructing an enterprise holographic knowledge graph; carrying out credit period analysis based on a knowledge graph to obtain rating time anchor point data, carrying out time sequence fluctuation detection to obtain a financial fluctuation curve, carrying out behavior anomaly detection to identify a periodic camouflage behavior mode to obtain a camouflage behavior characteristic spectrum, constructing a credit distortion correction model to obtain an index correction matrix, carrying out credit authenticity reduction to obtain an enterprise intrinsic credit image, constructing a multidimensional credit quantization index to obtain a credit scoring result, carrying out risk level assessment to obtain an enterprise credit risk level, generating a credit early warning signal to obtain a risk early warning report, optimizing the camouflage behavior characteristic spectrum and the credit distortion correction model, and effectively improving the accuracy and reliability of enterprise credit assessment.

Description

Knowledge graph-based enterprise credit assessment method

Technical Field

The invention relates to the technical field of enterprise credit assessment, in particular to an enterprise credit assessment method based on a knowledge graph.

Background

With the rapid development of financial science and technology and big data technology, enterprise credit assessment has evolved from traditional financial index analysis to a comprehensive assessment system combining multi-source data, however, there are still some special and difficult technical challenges in the field of credit assessment.

At present, the enterprise credit assessment has a special problem that periodic camouflage behavior identification and real credit recovery of the enterprise are caused by the fact that part of the enterprise is specially subjected to short-term behavior adjustment aiming at a rating period, such as temporary optimization of financial indexes before financial report release, short-term improvement of operating conditions before rating inspection, and bad asset clearing by assault before financing, the periodic camouflage behavior causes significant deviation in a traditional assessment method, the real credit condition of the enterprise is difficult to reflect, the periodic abnormal mode in data is difficult to identify when the conventional credit assessment method is used for solving the problem, analysis capability of correlation between the enterprise behavior and the rating period is lacked, the real credit condition of the enterprise cannot be effectively restored through historical data, an assessment result is easily interfered by intentional operation of the enterprise, and further assessment accuracy is influenced, and underestimation of financial risks and mismatching resources are caused.

In view of the above, the present invention proposes an enterprise credit assessment method based on knowledge graph to solve the above-mentioned problems.

Disclosure of Invention

The application provides an enterprise credit assessment method based on a knowledge graph, which is used for effectively identifying periodic camouflage behaviors of enterprise credits and restoring real credit conditions, capturing periodic abnormal patterns in data, revealing hidden relations between enterprise behaviors and rating periods, realizing accurate identification of camouflage behaviors and objective assessment of real credit conditions, effectively solving the limitation that the traditional assessment method is easy to be operated by enterprises intentionally, reducing financial risks and optimizing resource allocation.

In a first aspect, the present application provides a knowledge-graph-based enterprise credit assessment method, where the knowledge-graph-based enterprise credit assessment method includes:

Step S1, acquiring enterprise financial radar data, business ecological chain data and market pulsation data, constructing an enterprise financial space-time trajectory graph based on the enterprise financial radar data, performing feature analysis on the enterprise financial space-time trajectory graph to obtain financial dynamic features, constructing a correlation entity mapping network based on the business ecological chain data and the market pulsation data, performing correlation analysis on the correlation entity mapping network to obtain entity relationship data, and constructing an enterprise holographic knowledge graph according to the financial dynamic features and the entity relationship data;

step S2, carrying out credit period analysis based on the enterprise holographic knowledge graph to obtain rating time anchor point data, carrying out time sequence fluctuation detection on financial indexes in the enterprise holographic knowledge graph to obtain a financial fluctuation curve, carrying out behavior anomaly detection based on the rating time anchor point data and the financial fluctuation curve, and identifying a periodic camouflage behavior mode to obtain a camouflage behavior characteristic spectrum;

Step S3, constructing a credit distortion correction model according to the camouflage behavior feature spectrum to obtain an index correction matrix, and applying the index correction matrix to enterprise historical data to restore credit authenticity to obtain an enterprise true credit image;

And S4, carrying out risk level assessment based on the credit scoring result to obtain an enterprise credit risk level, generating a credit early warning signal according to the enterprise credit risk level to obtain a risk early warning report, carrying out feedback acquisition on a credit assessment application effect to obtain assessment effect data, and optimizing the camouflage behavior characteristic spectrum and the credit distortion correction model based on the assessment effect data to realize spiral ascending optimization of enterprise credit assessment.

In a second aspect, the application provides an enterprise credit assessment device based on a knowledge graph, which comprises a memory and at least one processor, wherein the memory is stored with instructions, and the at least one processor calls the instructions in the memory to enable the enterprise credit assessment device based on the knowledge graph to execute the enterprise credit assessment method based on the knowledge graph.

According to the technical scheme provided by the application, the enterprise holographic knowledge graph is constructed, so that multidimensional integration and analysis of enterprise financial data, business relationship and market feedback are realized, the limitation that the traditional credit assessment only depends on a financial statement is broken through, and the comprehensiveness and depth of the enterprise credit assessment are improved. By utilizing the time sequence slicing processing and skeleton extraction algorithm, accurate capturing of the financial dynamic characteristics of the enterprise is realized, and the grasping capability of the time evolution law of the financial condition of the enterprise is enhanced. Through the association entity mapping network and spectral cluster analysis, the implicit association and influence factors of enterprises in the business ecological system are revealed, and the dimension of credit assessment is enriched. Based on time sequence alignment analysis of the rating time anchor point data and the financial fluctuation curve, accurate identification and quantification of periodic camouflage behaviors of enterprises are achieved, and the problem of evaluation distortion caused by enterprise information beautification in traditional credit evaluation is effectively solved. By disguising the behavior feature spectrum and the credit distortion correction model, the authenticity restoration of the enterprise financial data is realized, the credit is evaluated to return to essence, and the reliability and fairness of the evaluation result are improved. The construction of the multidimensional credit quantitative index system enables the evaluation result to be more three-dimensional and finer, and is convenient for a decision maker to comprehensively know the credit condition of an enterprise. By adopting contextualized risk level adjustment and a multi-stage early warning mechanism, dynamic monitoring and early warning of enterprise credit risk are realized, and the practicability and the foresight of an evaluation result are enhanced. Through the evaluation effect feedback acquisition and the closed loop optimization mechanism, the evaluation system has self-learning and iterative lifting capabilities, and the continuous effectiveness and adaptability of the evaluation method are ensured. The knowledge graph technology and the credit assessment depth are fused, a complete technical path from data acquisition, feature extraction and camouflage identification to credit restoration is established, and a more objective and accurate enterprise credit assessment tool is provided for financial institutions, investors and partners.

Drawings

FIG. 1 is a schematic diagram of an enterprise credit assessment method based on a knowledge graph of the present invention;

Fig. 2 is a schematic diagram of an enterprise credit evaluation system based on a knowledge graph according to the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Example 1;

Referring to fig. 1, the knowledge graph-based enterprise credit assessment method includes the steps of S1, acquiring enterprise financial radar data, business ecological chain data and market pulsation data, constructing an enterprise financial space-time trajectory graph based on the enterprise financial radar data, performing feature analysis on the enterprise financial space-time trajectory graph to obtain financial dynamic features, constructing a correlation entity mapping network based on the business ecological chain data and the market pulsation data, performing correlation analysis on the correlation entity mapping network to obtain entity relationship data, and constructing an enterprise holographic knowledge graph according to the financial dynamic features and the entity relationship data;

The process of performing step S1 may specifically include the steps of:

Step S11, collecting enterprise financial report data, audit report data and tax record data to form enterprise financial radar data, collecting provider data, customer data, stock right structure data and high management information data to form business ecological chain data, and obtaining media evaluation data, industry dynamic data and market transaction data to form market pulsation data;

Step S12, performing time sequence slicing processing on the enterprise financial radar data to obtain a financial time slice set, performing index extraction on the financial time slice set to obtain index sequence data, and constructing an enterprise financial space-time trajectory graph based on the index sequence data;

Step S13, entity identification and relation extraction are carried out on the business ecological chain data to obtain an entity relation initial network, emotion analysis and event extraction are carried out on the market pulsation data to obtain market reaction data, and an associated entity mapping network is constructed based on the entity relation initial network and the market reaction data;

S14, performing feature extraction on the enterprise financial space-time trajectory graph by using a skeleton extraction algorithm to obtain financial dynamic features;

s15, carrying out community discovery on the associated entity mapping network by utilizing a spectral clustering algorithm to obtain an associated entity community;

S16, constructing a knowledge graph body frame, and mapping the financial dynamic characteristics and entity relation data to the knowledge graph body frame to form an initial knowledge graph;

and S17, adding time attributes to the initial knowledge graph by using a time dimension coding technology to obtain a time sequence knowledge graph, and carrying out knowledge reasoning expansion on the time sequence knowledge graph to obtain an enterprise holographic knowledge graph.

Specifically, firstly, multi-source data acquisition is carried out, and a basic data set for enterprise credit evaluation is constructed. For enterprise financial radar data, acquiring financial report data from public channels such as enterprise annual report, quarter report and middle report, acquiring audit report data including audit opinion, key audit matters and abnormal adjustment items from an audit mechanism, and acquiring enterprise tax record data including tax payment grade, tax payment condition and tax penalty record from a tax department. These data together constitute corporate financial radar data, providing a comprehensive view for subsequent financial analysis. For business ecological chain data, main supplier information of enterprises is obtained from a supply chain management system, wherein the main supplier information comprises the number, concentration and cooperation stability of suppliers, client data is obtained from a client relationship management system, wherein the client data comprises the client scale, industry distribution and repayment conditions, stock right structure data is obtained from an industrial and commercial registration system, the stock right structure data comprises the stock right composition, actual control person and stock right change history, and high management information data is obtained from public materials, wherein the high management information data comprises high management background, job history and change conditions. Together, these data form business ecochain data reflecting business network relationships of the enterprise. For market pulsation data, media rating data including public opinion indexes, reputation ratings and crisis events are obtained from news media, social platforms and industry reports, industry dynamic data including industry scenic indexes, policy changes and technical innovations are obtained from industry associations and research institutions, and market transaction data including stock price fluctuations, bond transactions and financing costs are obtained from exchanges and transaction platforms. Together, these data constitute market pulsation data reflecting the external environment and market performance of the enterprise. And performing time sequence slicing processing on the enterprise financial radar data to obtain a financial time slice set. The time sequence slicing process aims at dividing continuous financial data according to time dimension to form a series of time slices, so that the time evolution characteristics of financial indexes can be analyzed conveniently. And carrying out time window division on the enterprise financial radar data according to the quarterly, semi-annual and annual time granularity to obtain a multi-granularity time window set. And carrying out normalization processing on the financial data in the multi-granularity time window set, and eliminating the dimension difference among different indexes to obtain standardized financial data. The formula of the normalization process is as follows: Wherein, the method comprises the steps of, For the normalized financial data to be of the order of magnitude,As a result of the original financial data,As the mean value of the index,Is the standard deviation of the index. And (3) carrying out homonymy and cyclic ratio calculation on the standardized financial data to obtain a change rate index set and reflect the dynamic change of the financial condition of the enterprise. The formula of the calculation of the same ratio is as follows:

the calculation formula of the ring ratio is as follows: Wherein, the method comprises the steps of, Is the rate of change of the same ratio,In order to achieve a rate of change of the ring ratio,As the index value of the current period,For the value of the index of the last year synchronization,Is the upper-period index value. And organizing the standardized financial data and the change rate index set according to time sequence to form a financial time slice set, and laying a foundation for the subsequent construction of a financial space-time trajectory graph. And (5) extracting indexes from the financial time slice set, and constructing index sequence data. Index extraction aims to extract key indexes reflecting financial conditions of enterprises from financial time slice sets, including repayment capability indexes (such as flow rate, snap rate, asset liability rate), profitability indexes (such as net asset profitability, sales gross rate, net profit rate), operational capability indexes (such as receivability turnover, inventory turnover, total asset turnover), and development capability indexes (such as business income increase rate, net profit increase rate, total asset increase rate). Based on the index sequence data, an enterprise financial space-time trajectory graph is constructed, the graph takes time as a horizontal axis and financial dimension as a vertical axis, the change trajectory of the financial condition of the enterprise along with the time is depicted, and the financial dynamic characteristics of the enterprise are intuitively displayed. And carrying out entity identification and relation extraction on the commercial ecological chain data to obtain an entity relation initial network. In the entity identification process, a named entity identification (NER) technology is adopted to identify key entities in the commercial ecological chain data, including enterprises, personnel, institutions and the like. In the relation extraction process, dependency syntactic analysis, pattern matching and other methods are adopted to extract the relation between entities from the text, such as 'supply', 'stock', 'cooperation', and the like. And constructing an entity relationship initial network by entity identification and relationship extraction, wherein the network takes an entity as a node and takes a relationship as an edge to display the association relationship between an enterprise and other business subjects. And carrying out emotion analysis and event extraction on the market pulsation data to obtain market response data. In the emotion analysis process, a dictionary method and a machine learning method are adopted to calculate emotion polarity and emotion intensity in market pulsation data, and emotion attitude of the market to enterprises is quantized. In the event extraction process, a time sequence analysis and anomaly detection method is adopted to extract important events related to enterprises from market pulsation data and evaluate the influence degree of the important events. Based on the initial network of entity relation and market reaction data, an associated entity mapping network is constructed, and the network not only contains static relation among entities, but also integrates the market reaction data, and shows dynamic evolution of business relation of enterprises. And extracting features of the enterprise financial space-time trajectory graph by using a skeleton extraction algorithm to obtain financial dynamic features. The skeleton extraction algorithm aims at extracting core structures and key features from complex financial space-time trajectory diagrams, simplifying data representation and improving analysis efficiency. Common skeleton extraction algorithms include morphological refinement and distance transformation. Taking morphological refinement as an example, the basic idea is to preserve the topology of the image by iteratively deleting boundary points until it cannot be deleted any more. The formula of the algorithm is as follows: Wherein, the method comprises the steps of, As a result of the extraction of the skeleton,As the original image is to be taken,In the case of a morphological etching operation,,,...,Is a structural element. And obtaining a core structure of the enterprise financial space-time trajectory graph through skeleton extraction, and generating financial dynamic characteristics, wherein the characteristics comprise key evolution modes and periodic variation rules of the enterprise financial conditions. And carrying out community discovery on the associated entity mapping network by utilizing a spectral clustering algorithm to obtain an associated entity community. The spectral clustering algorithm is a graph-based clustering method and is suitable for community structure discovery of complex networks such as a related entity mapping network. The basic steps include constructing a similarity matrixCalculating a Laplace matrix(WhereinDegree matrix), solve forSelecting the eigenvectors corresponding to K minimum non-zero eigenvalues to form an eigenvector, and carrying out K-means clustering on the row vectors of the eigenvector to obtain a community division result. Through spectral clustering, the associated entity mapping network is divided into a plurality of associated entity communities, and each community comprises a closely related entity set. Based on the association entity community, relationship intensity quantification is carried out, and the intensity index of the relationship among the entities is calculated, wherein the intensity index comprises relationship frequency, relationship persistence, relationship reciprocity and the like, so that entity relationship data is obtained. And constructing a knowledge graph ontology framework, and providing a structured template for knowledge organization in the enterprise credit evaluation field. The ontology framework includes a concept layer (defining core concepts of enterprises, financial indicators, associated entities, etc.), a relationship layer (defining relationship types between concepts), and an attribute layer (defining attribute characteristics of the concepts). And mapping the financial dynamic characteristics and the entity relation data to a knowledge graph ontology framework to form an initial knowledge graph. In the mapping process, the financial dynamic characteristics are converted into node attributes and relationship characteristics in the knowledge graph, and the entity relationship data are converted into nodes and edges in the knowledge graph to form a complete knowledge network structure. And adding time attributes to the initial knowledge graph by using a time dimension coding technology, and constructing a time sequence knowledge graph. The time dimension coding technology aims at integrating time information into a knowledge graph and expressing time sequence change characteristics of knowledge. Common time coding methods include time stamp coding, time interval coding, and periodic coding. Through time dimension coding, each relation in the knowledge graph is marked with time information, and the equal knowledge of the enterprise M having the financial index value v1 at time t1 can be expressed. And carrying out knowledge reasoning expansion on the time sequence knowledge graph, and expanding the knowledge coverage of the time sequence knowledge graph through logic reasoning and rule deduction. Common knowledge reasoning methods include rule-based reasoning, path-based reasoning and embedded reasoning. Through knowledge reasoning, a new knowledge triplet is generated, semantic content of the knowledge graph is enriched, and finally an enterprise holographic knowledge graph is formed, so that a comprehensive and three-dimensional knowledge base is provided for subsequent credit evaluation.

The process of performing step S2 may specifically include the steps of:

S21, acquiring an enterprise history rating date, a financing event date and a significant financial report release date from an external authority data source, and constructing a rating event time sequence;

S22, constructing a rating period model based on a rating event time sequence and an enterprise holographic knowledge graph to obtain rating time anchor point data;

s23, extracting a financial index sequence from the enterprise holographic knowledge graph, and constructing a financial index time sequence network to obtain index fluctuation original data;

S24, carrying out seasonal decomposition and trend analysis on the index fluctuation original data, and identifying a periodic fluctuation mode to obtain a financial fluctuation curve;

Step S25, aligning the time sequence of the rating time anchor point data with that of the financial fluctuation curve, and calculating the index change amplitude before and after rating to obtain a time correlation index;

step S26, constructing an abnormal behavior detector based on the time correlation index, and identifying index abnormality highly correlated with the rating period to obtain a periodic abnormal mode;

And S27, extracting and classifying the characteristics of the periodic abnormal patterns, and identifying a typical camouflage behavior pattern to obtain a camouflage behavior characteristic spectrum.

Specifically, key point-in-time data associated with an enterprise credit rating is first obtained from an external authoritative data source (e.g., ratings agency, stock exchange, enterprise bulletin platform). The data comprise historical rating dates of enterprises, past credit rating time of the enterprises, financing event dates, time of heavy financing activities such as enterprise bond release and bank loan, release dates of heavy financial reports, and release time of financial reports such as enterprise annual reports and quaternary reports. The time points are arranged according to the time sequence, a rating event time sequence is constructed, and a time reference is provided for the subsequent analysis of the behavior change of enterprises before and after rating. And constructing a rating period model based on the rating event time sequence and the enterprise holographic knowledge graph. The model aims at analyzing the periodic characteristics of enterprise credit rating activities and revealing the rules and patterns of rating activities. The construction of the rating period model comprises the following steps of time sequence analysis, analysis of time interval distribution of rating events, identification of statistical characteristics of rating periods, pattern identification, identification of correlation patterns of the rating events and enterprise behaviors, discovery of behavior change rules before and after rating, and period prediction, wherein the possible time window of the next rating is predicted based on historical data. And obtaining rating time anchor point data through a rating period model, wherein the data marks key time points of enterprise credit rating activities, and provides a reference standard for the follow-up analysis of the behavior change of the enterprise before and after rating. And (5) extracting a financial index sequence from the enterprise holographic knowledge graph, and constructing a financial index time sequence network. The financial index time sequence network takes time as a horizontal axis and takes different financial indexes as a vertical axis to form a multi-dimensional time sequence network structure. The network not only contains time series data of each financial index, but also characterizes interrelation and influence mechanisms among indexes. And obtaining index fluctuation original data by analyzing the financial index time sequence network, wherein the data records the change condition of each financial index of the enterprise along with time. And carrying out seasonal decomposition and trend analysis on the index fluctuation original data, and identifying a periodic fluctuation mode. Seasonal decomposition is intended to decompose the time series into trend components, seasonal components, and residual components, revealing the inherent structure of the time series. Common seasonal decomposition methods include classical decomposition and STL decomposition. Taking a classical decomposition method as an example, assuming that the time sequence Y can be decomposed into a trend component T, a seasonal component S and a random component R, the trend analysis aims at identifying the long-term change trend of the time sequence and revealing the overall development direction of the financial condition of an enterprise. Common trend analysis methods include moving average and exponential smoothing. And identifying the periodic fluctuation mode of the financial index through seasonal decomposition and trend analysis to obtain a financial fluctuation curve, wherein the fluctuation rule and the change characteristic of the financial index of the enterprise are reflected by the curve. And (5) carrying out time sequence alignment on the rating time anchor point data and the financial fluctuation curve, and calculating the index change amplitude before and after rating. The aim of time sequence alignment is to accurately match the rating time with the fluctuation time of the financial index, so that the relevance of the rating activity and the financial index change can be conveniently analyzed. After alignment, calculating the variation amplitude of the financial index in a certain time window (such as the first 6 months and the last 6 months) before and after each rating time point, and quantifying the influence degree of the rating activity on the financial index. The calculation formula of the variation amplitude is as follows: Wherein, the method comprises the steps of, In order to vary the amplitude of the variation,For the value of the index before the rating,Is the rated index value. According to the calculation result, a time correlation index is generated, the index reflects the correlation degree of the change of the financial index and the rating time, and a basis is provided for identifying abnormal behaviors related to the rating. Based on the time correlation index, an abnormal behavior detector is constructed. The abnormal behavior detector is used for identifying abnormal changes in the enterprise financial indexes, which are highly related to the rating period, and finding potential rating manipulation behaviors. The construction of the detector comprises the following steps of feature engineering, model training, training of an abnormality detection model by adopting an abnormality detection algorithm (such as isolated forest, single-class SVM and self-encoder), threshold setting, threshold of abnormality score setting and judging samples exceeding the threshold as abnormality. The basic idea of the isolated forest algorithm is that outliers are more easily isolated in random feature space. The anomaly score calculation formula is as follows:

Wherein, the method comprises the steps of, As the anomaly score, a score of the anomaly,Representing the data of the sample,In order to have an average path length that is,For a sample size ofIs used for the average path length of the binary search tree. And identifying index anomalies highly related to the rating period through an anomaly behavior detector to obtain periodic anomaly modes, wherein the modes reflect the abnormal financial behaviors of the enterprise before and after rating. And extracting and classifying the characteristics of the periodic abnormal patterns, and identifying typical camouflage behavior patterns. In the feature extraction process, representative features are extracted from the abnormal modes, including occurrence time, duration, influence range, severity and the like of the abnormality. In the classifying process, the abnormal patterns are classified into different types of camouflage behaviors based on the extracted features. And performing density cluster analysis on the periodic abnormal patterns, and identifying high-frequency abnormal patterns to obtain an abnormal pattern cluster. The basic idea of the DBSCAN density clustering algorithm is to divide the data into different clusters based on the density of the samples. The core parameters of the algorithm include the neighborhood radius epsilon and the minimum number of samples MinPts. And constructing a camouflage behavior classification system which comprises the classes of financial data beautification, asset reorganization camouflage, associated transaction control type, public opinion management type and the like. And labeling camouflage behavior types for the abnormal pattern clusters based on the expert knowledge base and the historical case base to obtain camouflage behavior type data. And calculating characteristic statistics of different camouflage behavior types, including occurrence frequency, severity and duration, to obtain a camouflage behavior quantization index. And constructing an enterprise camouflage tendency image based on the camouflage behavior quantification index to form a camouflage behavior characteristic spectrum. The feature spectrum comprehensively describes camouflage behavior features of enterprises in the credit rating process, and provides basis for follow-up credit distortion correction.

The process of performing step S3 may specifically include the steps of:

s31, constructing an index torsion quantification model based on a camouflage behavior characteristic spectrum, and calculating the torsion degree of each financial index to obtain an index torsion coefficient;

s32, constructing a credit distortion correction model by using a reverse derivation algorithm, and converting the index distortion coefficient into an index correction matrix;

S33, extracting historical financial data from the enterprise holographic knowledge graph, and performing data reduction by applying an index correction matrix to obtain corrected financial data;

Step S34, recalculating the enterprise credit evaluation key index based on the corrected financial data to obtain a restoration credit index set;

Step S35, carrying out rationality verification on the restored credit index set by combining industry standard data and enterprise historical performance to obtain verified credit indexes;

Step S36, constructing an enterprise true credit portrait based on the verified credit index, and displaying the enterprise true credit status;

and S37, applying a credit scoring model to carry out multidimensional credit scoring on the enterprise true credit image to obtain a credit scoring result.

Specifically, an index torsion degree quantization model is constructed based on camouflage behavior feature spectrums. The model aims at quantifying the distortion degree of each financial index of the enterprise and provides a basis for the follow-up credit distortion correction. The construction of the index distortion quantization model comprises the following steps of feature selection, weight distribution, model design and distortion calculation method, wherein features related to financial index distortion are selected from camouflage behavior feature spectrums, weight distribution is carried out on each feature according to the importance of the features, and the features and the weights are converted into distortion quantization indexes. The formula for the torsion resistance calculation can be expressed as: Wherein, the method comprises the steps of, Is the firstThe degree of torsion of the individual financial indicators,Is the firstThe weight of the individual features is determined,Is the firstThe values of the individual features. And calculating the distortion degree of each financial index of the enterprise through the index distortion degree quantization model to obtain an index distortion coefficient, wherein the coefficient reflects the false degree of each financial index and provides a quantization reference for the follow-up credit distortion correction. And constructing a credit distortion correction model by using a reverse derivation algorithm. The model aims to deduce the true value of the financial index according to the index distortion coefficient, and correct the distortion data. The basic idea of the reverse derivation algorithm is to derive the true value of the financial index during rating by analyzing the normal financial performance of the enterprise during non-ratings. The method comprises the specific steps of establishing a benchmark, establishing a benchmark model of normal financial performance of an enterprise, analyzing deviation of a financial index of the enterprise and the benchmark model in a rating period, calculating correction values according to the deviation and a torsion coefficient. And converting the index distortion coefficient into an index correction matrix, wherein the matrix represents the correction mode and amplitude of each financial index, and provides an operation guide for subsequent data reduction. And extracting historical financial data from the enterprise holographic knowledge graph, and carrying out data reduction by applying an index correction matrix. The data recovery process comprises the steps of data extraction, extraction of historical financial data of enterprises from a knowledge graph, correction application, consistency check, and verification of internal consistency of corrected data, wherein the correction application corrects the financial data according to an index correction matrix, and the rationality of correction results is ensured. The formula for data correction can be expressed as: Wherein, the method comprises the steps of, In order to correct the true value after the correction,In order to observe the value of the distortion,Is the correction coefficient of the index. And restoring the historical financial data of the enterprise by applying the index correction matrix to obtain corrected financial data, wherein the data more truly reflects the financial condition of the enterprise. And recalculating the enterprise credit assessment key index based on the corrected financial data. The key indicators are calculated to include repayment capability indicators such as flow rate, snap action rate, and asset liability rate, profitability indicators such as net asset return rate, sales gross rate, and net profit rate, operational capability indicators such as receivability turnover rate, inventory turnover rate, and total asset turnover rate, and development capability indicators such as business income increase rate, net profit increase rate, and total asset increase rate. By recalculating these key metrics, a restored credit index set is obtained that reflects the actual credit status of the enterprise. And carrying out rationality verification on the restored credit index set by combining industry standard data and enterprise historical performance. The purpose of rationality verification is to ensure that the restored credit index accords with industry rules and the development logic of enterprises, and unreasonable results are avoided in the correction process. The verification method comprises the steps of industry standard comparing, judging whether the reduction index is in a reasonable range or not, historical comparing, judging whether the reduction index accords with the development trend of an enterprise or not by comparing the reduction index with the historical performance of the enterprise, and expert review, inviting an industry expert to review the reduction index to provide specialized comments and suggestions. Through rationality verification, verified credit indexes are obtained, and the indexes are subjected to multiple tests, so that the real credit condition of an enterprise is reflected more reliably. And constructing the enterprise true credit portrait based on the verified credit index. The enterprise true credit portrait is a comprehensive display of the enterprise true credit status, including aspects of credit advantage, credit disadvantage, credit risk, credit trend and the like of the enterprise. A multidimensional credit rating index system is constructed, which covers a repayment capacity dimension, a profitability dimension, an operational capacity dimension and a development capacity dimension. And normalizing the verified credit indexes to eliminate dimensional differences among different indexes, so that comprehensive evaluation and comparison are facilitated. The formula of normalization processing is:

Wherein, the method comprises the steps of, For the normalized index value of the index value,As the original index value of the index value,AndThe minimum and maximum values of the index, respectively. Mapping the normalized index to a multidimensional credit evaluation index system to obtain a standardized credit index set. And determining the weight of each dimension and each index by using an analytic hierarchy process. The analytic hierarchy process is a systematic analysis method, and the weights of all factors are obtained by constructing a judgment matrix and calculating a feature vector. The construction of the judgment matrix is based on expert judgment, and the factors are compared pairwise to represent the relative importance of the factors. The weight calculation step comprises the steps of constructing a judgment matrix, calculating the maximum eigenvalue and the corresponding eigenvector of the judgment matrix, and carrying out consistency test to obtain the final weight. And constructing a weighted score model, and multiplying the standardized index by the weight to obtain a weighted credit score. The weighted credit score is converted to a credit rating using a nonlinear mapping function. Common nonlinear mapping functions include Sigmoid functions and piecewise functions. Through the steps, the true credit portrait of the enterprise is generated, and the credit status of the enterprise is truly reflected. And applying a credit scoring model to carry out multidimensional credit scoring on the enterprise true credit image. The credit score model is an algorithm model for converting the multidimensional credit index of the enterprise into a single score, and is convenient for quantitative management and decision support of credit risk. Common credit scoring models include Logistic regression models, decision tree models, neural network models, and the like. Taking a Logistic regression model as an example, the formula is as follows:

Wherein, the method comprises the steps of, For the probability of an enterprise breach,,,...,As an index of the credit-on-demand (credit-on-demand) of the credit card,,,,...,Is a model parameter. And carrying out multidimensional scoring on the true credit image of the enterprise through a credit scoring model to obtain a credit scoring result, wherein the credit scoring result comprehensively reflects the credit condition and the risk level of the enterprise and provides a quantification basis for subsequent risk assessment.

The process of performing step S4 may specifically include the steps of:

Step S41, constructing a risk grade dividing standard based on a credit grading result, and grading the credit risk of the enterprise to obtain the credit risk grade of the enterprise;

step S42, combining the industry risk sensitivity index and the macro economic environment index, and carrying out situation adjustment on the enterprise credit risk level to obtain a contextualized risk level;

Step S43, designing a risk early warning threshold according to the contextualized risk level, constructing a multi-level early warning mechanism and generating a risk early warning signal;

Step S44, compiling enterprise credit risk early warning reports based on the risk early warning signals, wherein the enterprise credit risk early warning reports comprise risk grades, risk causes and risk evolution trends;

Step S45, acquiring application feedback and decision effect data of a user on a credit evaluation report to form an evaluation effect database;

Step S46, performing model effect analysis based on the evaluation effect database, and identifying a model optimization direction to obtain model tuning parameters;

And S47, optimizing camouflage behavior characteristic spectrum and credit distortion correction model by using model tuning parameters to realize closed-loop optimization of the evaluation system.

Specifically, a risk ranking criterion is constructed based on the credit scoring results. The risk ranking criteria are intended to translate continuous credit scores into discrete risk rankings, facilitating risk management and decision support. The risk classification method comprises the steps of equally dividing a credit scoring interval into a plurality of levels, dividing the credit scoring interval into quantiles, classifying the risk levels according to the quantiles of the scores, clustering, and classifying the scores into natural groups through a clustering algorithm. A common risk level hierarchy includes AAA, AA, A, BBB, BB, B, CCC, CC, C levels, each corresponding to a different risk level. And grading the credit risk of the enterprise, determining the credit risk grade of the enterprise according to the credit score and the risk grade grading standard of the enterprise, and providing a standardized risk measure for risk management. And carrying out situation adjustment on the credit risk level of the enterprise by combining the industry risk sensitivity index and the macro economic environment index. The situation adjustment aims at considering the influence of industry characteristics and macroscopic environment on the credit risk of enterprises, so that the risk assessment is more comprehensive and accurate. Industry risk sensitive indicators include industry concentration, industry periodicity, industry policy sensitivity, etc., which reflect the impact of industry characteristics on enterprise credit risk. Macroeconomic environmental indicators include GDP growth rate, CPI growth rate, monetary supply, etc., which reflect the impact of the macroeconomic environment on the credit risk of the enterprise. The situation adjustment method comprises the steps of weighting risks, weighting and adjusting basic risk levels according to industrial risks and macroscopic risks, simulating situations, evaluating risk performances of enterprises under different situations by simulating different industries and macroscopic situations, and adjusting the risk levels based on expert experience and judgment by expert adjustment. By means of context adjustment, a contextualized risk level is obtained, the level comprehensively considers the influence of industry characteristics and macroscopic environments, and the credit risk condition of an enterprise is more comprehensively reflected. And designing a risk early warning threshold according to the contextualized risk level, and constructing a multi-stage early warning mechanism. The risk early warning threshold is a critical value for triggering risk early warning, and the multi-stage early warning mechanism is used for setting a plurality of early warning levels according to different risk degrees so as to realize hierarchical management of risks. The design method of the early warning threshold comprises a historical analysis method, an expert judgment method and a statistical method, wherein the threshold is set according to characteristics of risk events in historical data, the expert judgment method is based on expert experience and judgment, and the statistical method is used for setting the threshold according to statistical characteristics (such as mean value and standard deviation) of the data. The multi-level early warning mechanism generally comprises three levels, namely a prompt level which indicates that potential risks exist and need to be concerned, a warning level which indicates that the risks are obviously increased and need to be vigilant, and an emergency level which indicates that the risks are serious and needs to take measures immediately. And generating a risk early warning signal by setting a risk early warning threshold value and constructing a multi-stage early warning mechanism, and providing timely early warning support for risk management. And compiling an enterprise credit risk early warning report based on the risk early warning signal. The early warning report is a system display of risk early warning results and provides information support of risk management for decision makers. The early warning report comprises a risk level representing the current credit risk level of an enterprise, a risk cause, a risk evolution trend, and a future development direction and change trend of the risk. The method for compiling the early warning report comprises the steps of data visualization, visual display of risk characteristics and change trends through charts, text analysis, detailed explanation of risk causes and management suggestions through characters, situation analysis and situation simulation display of risk performance under different situations. By compiling the enterprise credit risk early warning report, the system displays the credit risk condition of the enterprise and provides valuable risk management information for a decision maker. And collecting application feedback and decision effect data of the user on the credit evaluation report to form an evaluation effect database. The purpose of feedback acquisition is to know the user experience and decision effect, and provide basis for model optimization. And designing an evaluation effect feedback acquisition table comprising accuracy evaluation, timeliness evaluation and validity evaluation. accuracy evaluation focuses on the coincidence degree of the evaluation result and the actual situation, timeliness evaluation focuses on the timeliness of the evaluation result, and effectiveness evaluation focuses on the supporting value of the evaluation result on the decision. And tracking the consistency of the follow-up actual performance of the enterprise and the credit evaluation result, and calculating the prediction deviation to obtain accuracy data. The calculation formula of the prediction bias is: Wherein, the method comprises the steps of, In order to predict the deviation, the deviation is calculated,In order to be able to predict the value,Is an actual value. And collecting credit decision application scene data, and analyzing the application effect of the evaluation result in credit decisions, investment decisions and cooperation decisions to obtain decision value evaluation data. Integrating accuracy data and decision value evaluation data, constructing an evaluation effect comprehensive index, forming an evaluation effect database, and providing data support for model optimization. And carrying out model effect analysis based on the evaluation effect database, and identifying the model optimization direction. The purpose of model effect analysis is to evaluate the performance of the current model, finding potential problems and optimizing space. The method for analyzing the effect comprises the steps of carrying out statistical analysis, evaluating the overall performance of the model through statistical indexes, carrying out classification analysis aiming at the performance difference of different types of sample analysis models, carrying out deep analysis on model error cases through error analysis, and finding out error reasons. And (3) identifying the optimization direction of the model, such as feature engineering optimization, algorithm optimization, threshold adjustment and the like, and generating specific model tuning parameters to provide guidance for subsequent model optimization through model effect analysis. And optimizing camouflage behavior characteristic spectrum and credit distortion correction model by using model tuning parameters. The model optimization aims at improving the accuracy, stability and generalization capability of the model and realizing continuous improvement of an evaluation system. The optimization method comprises the steps of parameter adjustment, feature optimization, feature selection and combination optimization, feature expression improvement, new algorithm introduction or existing algorithm improvement, and model performance improvement. Through continuous model optimization, spiral ascending optimization of the enterprise credit evaluation system is realized, and the accuracy and effectiveness of credit evaluation are continuously improved. By taking a specific enterprise as an example, an enterprise credit assessment method based on a knowledge graph is used, financial reports, business relations and market evaluation data of the enterprise are firstly collected to construct an enterprise holographic knowledge graph, then correlation between financial fluctuation and rating period of the enterprise is analyzed to identify potential camouflage behaviors, then a correction model is constructed according to camouflage behavior characteristics to restore real credit conditions of the enterprise, finally risk assessment and early warning are carried out based on restored credit indexes, and the assessment model is continuously optimized according to actual effects. Through the whole flow, the comprehensive, objective and dynamic assessment of the credit condition of the enterprise is realized, and reliable basis is provided for decisions such as credit, investment and the like.

By constructing the enterprise holographic knowledge graph, the embodiment realizes multidimensional integration and analysis of enterprise financial data, business relations and market feedback, breaks through the limitation that the traditional credit assessment only depends on financial reports, and improves the comprehensiveness and depth of enterprise credit assessment. By utilizing the time sequence slicing processing and skeleton extraction algorithm, accurate capturing of the financial dynamic characteristics of the enterprise is realized, and the grasping capability of the time evolution law of the financial condition of the enterprise is enhanced. Through the association entity mapping network and spectral cluster analysis, the implicit association and influence factors of enterprises in the business ecological system are revealed, and the dimension of credit assessment is enriched. Based on time sequence alignment analysis of the rating time anchor point data and the financial fluctuation curve, accurate identification and quantification of periodic camouflage behaviors of enterprises are achieved, and the problem of evaluation distortion caused by enterprise information beautification in traditional credit evaluation is effectively solved. By disguising the behavior feature spectrum and the credit distortion correction model, the authenticity restoration of the enterprise financial data is realized, the credit is evaluated to return to essence, and the reliability and fairness of the evaluation result are improved. The construction of the multidimensional credit quantitative index system enables the evaluation result to be more three-dimensional and finer, and is convenient for a decision maker to comprehensively know the credit condition of an enterprise. By adopting contextualized risk level adjustment and a multi-stage early warning mechanism, dynamic monitoring and early warning of enterprise credit risk are realized, and the practicability and the foresight of an evaluation result are enhanced. Through the evaluation effect feedback acquisition and the closed loop optimization mechanism, the evaluation system has self-learning and iterative lifting capabilities, and the continuous effectiveness and adaptability of the evaluation method are ensured. The knowledge graph technology and the credit assessment depth are fused, a complete technical path from data acquisition, feature extraction and camouflage identification to credit restoration is established, and a more objective and accurate enterprise credit assessment tool is provided for financial institutions, investors and partners.

Example 2;

referring to fig. 2, the detailed description of the embodiment 1 is omitted, and the enterprise credit evaluation system based on knowledge graph includes:

The map construction module is used for acquiring enterprise financial radar data, business ecological chain data and market pulsation data, constructing an enterprise financial space-time trajectory graph based on the enterprise financial radar data, carrying out feature analysis on the enterprise financial space-time trajectory graph to obtain financial dynamic features, constructing a related entity mapping network based on the business ecological chain data and the market pulsation data, carrying out association analysis on the related entity mapping network to obtain entity relation data, and constructing an enterprise holographic knowledge map according to the financial dynamic features and the entity relation data;

The characteristic detection module is used for carrying out credit period analysis based on the enterprise holographic knowledge graph to obtain rating time anchor point data, carrying out time sequence fluctuation detection on financial indexes in the enterprise holographic knowledge graph to obtain a financial fluctuation curve, carrying out behavior anomaly detection based on the rating time anchor point data and the financial fluctuation curve, and identifying a periodic camouflage behavior mode to obtain a camouflage behavior characteristic spectrum;

the portrait construction module is used for constructing a credit distortion correction model according to the camouflage behavior characteristic spectrum to obtain an index correction matrix, carrying out credit authenticity restoration on enterprise historical data by applying the index correction matrix to obtain an enterprise true credit portrait;

The evaluation optimization module is used for evaluating the risk level based on the credit grading result to obtain an enterprise credit risk grade, generating a credit early warning signal according to the enterprise credit risk grade to obtain a risk early warning report, carrying out feedback acquisition on a credit evaluation application effect to obtain evaluation effect data, optimizing the camouflage behavior characteristic spectrum and the credit distortion correction model based on the evaluation effect data, and realizing spiral ascending optimization of enterprise credit evaluation.

The application also provides an enterprise credit assessment device based on the knowledge graph, which comprises a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions, when executed by the processor, cause the processor to execute the steps of the enterprise credit assessment device method based on the knowledge graph.

The above formulas are all formulas with dimensionality removed and numerical calculation, the formulas are formulas with the latest real situation obtained by software simulation through collecting a large amount of data, and preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to those skilled in the art without departing from the principles of the present invention are intended to be comprehended within the scope of the present invention.

Claims

1. The enterprise credit assessment method based on the knowledge graph is characterized by comprising the following steps of:

2. The knowledge-graph-based enterprise credit assessment method according to claim 1, wherein step S1 comprises:

step S13, entity identification and relation extraction are carried out on the commercial ecological chain data to obtain an entity relation initial network, emotion analysis and event extraction are carried out on the market pulsation data to obtain market reaction data, and an associated entity mapping network is constructed based on the entity relation initial network and the market reaction data;

S14, extracting features of the enterprise financial space-time trajectory graph by using a skeleton extraction algorithm to obtain financial dynamic features;

Step 15, carrying out community discovery on the associated entity mapping network by utilizing a spectral clustering algorithm to obtain an associated entity community;

S16, constructing a knowledge graph body frame, and mapping the financial dynamic characteristics and the entity relation data to the knowledge graph body frame to form an initial knowledge graph;

3. The knowledge-graph-based enterprise credit assessment method according to claim 2, wherein the performing time-series slicing processing on the enterprise financial radar data to obtain a financial time slice set includes:

Dividing time windows of the enterprise financial radar data according to three time granularities of quarter, half year and year to obtain a multi-granularity time window set;

Normalizing the financial data in the multi-granularity time window set to obtain standardized financial data;

performing homonymy and cyclic ratio calculation on the standardized financial data to obtain a change rate index set;

the normalized financial data and the change rate index set are organized chronologically to form a set of financial time slices.

4. The knowledge-graph-based enterprise credit assessment method according to claim 1, wherein step S2 comprises:

s22, constructing a rating period model based on the rating event time sequence and the enterprise holographic knowledge graph to obtain rating time anchor point data;

S25, aligning the time sequence of the grading time anchor point data with the financial fluctuation curve, and calculating the index change amplitude before and after grading to obtain a time correlation index;

S26, constructing an abnormal behavior detector based on the time correlation index, and identifying index abnormality highly correlated with the rating period to obtain a periodic abnormal mode;

and step S27, carrying out feature extraction and classification on the periodic abnormal patterns, and identifying typical camouflage behavior patterns to obtain a camouflage behavior feature spectrum.

5. The knowledge-graph-based enterprise credit assessment method of claim 4, wherein the feature extraction and classification of the periodic anomaly patterns, and the identification of typical camouflage behavior patterns, and the obtaining of camouflage behavior feature spectrums, comprises:

Performing density cluster analysis on the periodic abnormal patterns, and identifying high-frequency abnormal patterns to obtain an abnormal pattern cluster;

Constructing a camouflage behavior classification system which comprises the categories of financial data beautification, asset reorganization camouflage, associated transaction control type, public opinion management type and the like;

labeling camouflage behavior types for the abnormal pattern clusters based on an expert knowledge base and a historical case base to obtain camouflage behavior type data;

calculating characteristic statistics of different camouflage behavior types, including occurrence frequency, severity and duration, to obtain a camouflage behavior quantization index;

and constructing an enterprise camouflage tendency image based on the camouflage behavior quantification index to form a camouflage behavior characteristic spectrum.

6. The knowledge-graph-based enterprise credit assessment method according to claim 1, wherein step S3 comprises:

S31, constructing an index torsion quantification model based on the camouflage behavior characteristic spectrum, and calculating the torsion degree of each financial index to obtain an index torsion coefficient;

s33, extracting historical financial data from the enterprise holographic knowledge graph, and carrying out data reduction by applying the index correction matrix to obtain corrected financial data;

step S34, recalculating enterprise credit evaluation key indexes based on the corrected financial data to obtain a restoration credit index set;

step S35, carrying out rationality verification on the restoration credit index set by combining industry standard data and enterprise historical performance to obtain verified credit indexes;

And S37, applying a credit scoring model to carry out multidimensional credit scoring on the enterprise true credit portrait to obtain a credit scoring result.

7. The knowledge-graph-based enterprise credit assessment method according to claim 6, wherein the constructing an enterprise true credit image based on the verified credit index, displaying the enterprise true credit status, comprises:

constructing a multidimensional credit evaluation index system which covers a repayment capacity dimension, a profitability dimension, an operation capacity dimension and a development capacity dimension;

Normalizing the verified credit index, mapping the normalized credit index to a multidimensional credit evaluation index system, and obtaining a standardized credit index set;

Determining the weight of each dimension and each index by using an analytic hierarchy process, and constructing a weighted scoring model to obtain a weighted credit score;

and converting the weighted credit score into a credit grade by using a nonlinear mapping function, and generating the enterprise true credit portrait.

8. The knowledge-graph-based enterprise credit assessment method according to claim 1, wherein step S4 comprises:

Step S41, constructing a risk classification standard based on the credit scoring result, and classifying the credit risk of the enterprise to obtain the credit risk level of the enterprise;

step S42, combining an industry risk sensitivity index and a macro economic environment index, and carrying out situation adjustment on the enterprise credit risk level to obtain a contextualized risk level;

Step S45, collecting application feedback and decision effect data of a user on a credit evaluation report to form an evaluation effect database;

and S47, optimizing camouflage behavior characteristic spectrum and credit distortion correction model by using the model tuning parameters to realize closed-loop optimization of the evaluation system.

9. The knowledge-graph-based enterprise credit assessment method according to claim 8, wherein step S45 comprises:

step S451, designing an evaluation effect feedback acquisition table, wherein the evaluation effect feedback acquisition table comprises accuracy evaluation, timeliness evaluation and validity evaluation;

step S452, tracking the consistency of the follow-up actual performance of the enterprise and the credit evaluation result, and calculating a prediction deviation to obtain accuracy data;

step S453, credit decision application scenario data are collected, and application effects of the evaluation results in credit decisions, investment decisions and cooperation decisions are analyzed to obtain decision value evaluation data;

And step 454, integrating the accuracy data and the decision value evaluation data, and constructing an evaluation effect comprehensive index to form an evaluation effect database.

10. The enterprise credit assessment equipment based on the knowledge graph is characterized by comprising a memory and at least one processor, wherein the memory stores instructions;

The at least one processor invoking the instructions in the memory to cause the knowledge-graph based enterprise credit assessment apparatus to perform the knowledge-graph based enterprise credit assessment method of any of claims 1-9.