An enterprise risk early warning system and method based on intelligent evolution map

Abstract

The application discloses an enterprise risk early warning system and method based on intelligent evolution map, relates to the technical field of computer information processing, extracts time nodes and characteristic words in enterprise penalty cases, respectively forms internal characteristic vectors and external characteristic vectors, integrates into a fusion characteristic vector to construct a risk evolution sequence, carries out time normalization processing on the characteristic sequences of multiple cases, clusters to form a risk evolution map, collects internal characteristics and external characteristics of a target enterprise, compares the fusion target vector of the target enterprise with the risk evolution map, positions a similar classification cluster, calculates a risk weight to generate a characteristic word push sequence, and provides quantitative basis for enterprise risk assessment through deep mining and multi-feature fusion of enterprise penalty cases, so that management personnel can timely master enterprise risk dynamics and take measures to cope with potential risks in advance.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, specifically to an enterprise risk early warning system and method based on intelligent evolutionary graphs. Background Technology

[0002] With the increasing demand for corporate credit risk early warning, risk assessment methods based on multi-dimensional data and intelligent analysis have gradually become a research hotspot. However, existing technologies still have shortcomings in integrating multi-source data from industry and commerce, taxation, and credit, as well as external characteristics such as public opinion, advertising placement, and business cooperation. This limits their ability to comprehensively reflect the risk evolution process and dynamic trends of enterprises, affecting the accuracy and timeliness of early warning results. Although the accumulation of multi-source data provides a foundation for risk assessment, the correlations and dynamic evolution patterns between data have not been fully explored, thus limiting the predictive ability for potential corporate risks.

[0003] Current risk assessment technologies mostly focus on the analysis of single dimensions or local features. For example, when processing corporate penalty records, they often only focus on static time-point information, ignoring the impact of external features on risk evolution. Furthermore, existing technologies employ fragmented implementations for time series alignment, feature fusion, and classification cluster construction, thus weakening their adaptability to complex corporate risk scenarios. Simultaneously, the heterogeneity between different data sources and the non-standardization of the time dimension also pose challenges to data analysis and model building, further impacting the actual effectiveness of risk warnings. Summary of the Invention

[0004] The purpose of this invention is to provide an enterprise risk early warning system and method based on intelligent evolutionary graphs to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for enterprise risk early warning based on intelligent evolutionary graphs, the method comprising:

[0006] Step S100: Obtain the enterprise's penalty case records and extract the time node sequence and node features to form a risk evolution sequence;

[0007] Step S200: Obtain the external features of the enterprise at each time point, arrange them in the order of time points to form an external feature evolution sequence, extract features from the external features at each time point, and obtain the external feature vector of each time point;

[0008] Step S300: Fuse the internal and external feature vectors of each time point to form a fused feature vector, and arrange the fused feature vectors to obtain the actual feature sequence of the penalty cases;

[0009] Step S400: Perform time normalization on the actual feature sequence to form a normalized feature sequence. Perform cluster analysis on the fused feature vectors of the normalized feature sequence to divide it into multiple classification clusters. Combine the directed connections between the fused feature vectors to obtain the risk evolution map.

[0010] Step S500: Compare the feature vector of the target enterprise with the classification clusters in the risk evolution map, select the target classification cluster, determine all the subsequent level classification clusters of the target classification cluster, and count the number of directed connections from the target classification cluster to each subsequent level classification cluster.

[0011] Step S600: Obtain the next-level category of the target category cluster, form a target feature word set by combining the internal feature words of the next-level category cluster, calculate the ranking weight of each feature word in the target feature word set, and sort the internal feature words according to the ranking weight to form a push sequence.

[0012] Furthermore, step S100 includes:

[0013] Step S101: Obtain a penalty case of a certain enterprise. The penalty case includes textual information from business registration, taxation and credit records. Extract the time nodes of each penalty case to form a time node sequence.

[0014] Step S102: Record the feature words in the penalty cases as internal feature words, extract the internal feature words from each penalty case as node features of the time node, and arrange the node features in the order of the time nodes to form a risk evolution sequence.

[0015] Step S103: Obtain all internal feature words at the i-th time node in the risk evolution sequence, obtain the word vectors of the internal feature words, and perform a weighted summation of the word vectors of the internal feature words to obtain the internal feature vector at the i-th time node.

[0016] By collecting and analyzing the internal and external characteristics of corporate penalty cases and converting them into vector form for quantitative analysis, we can more comprehensively identify the risk factors faced by enterprises, improve the accuracy of risk identification, and avoid risk omissions due to insufficient or one-sided information.

[0017] Furthermore, step S200 includes:

[0018] Step S201: Obtain the external characteristics of the enterprise at each time point. The external characteristics include public opinion data, advertising data, and text information of business cooperation.

[0019] Step S202: Record the feature words in the external features as external feature words, and arrange the external feature words in chronological order to form an external feature evolution sequence;

[0020] Step S203: Obtain all external feature words at the i-th time node in the external feature evolution sequence, obtain the word vectors of the external feature words, and perform a weighted summation of the word vectors of the external feature words to obtain the external feature vector at the i-th time node.

[0021] By constructing a risk evolution map and normalizing and clustering the fused feature vectors at different time points, the evolution path and development trend of enterprise risks can be intuitively displayed. This helps managers clearly understand how risks develop and change over time, thereby better grasping risk dynamics and making early warning and response strategies.

[0022] Furthermore, step S300 includes:

[0023] Step S301: Perform feature fusion on the internal feature vector and external feature vector of the i-th time node to obtain the fused feature vector Vi of the i-th time node, Vi=α·Fi+β·Ei, where Fi represents the internal feature vector of the i-th time node, Ei represents the internal feature vector of the i-th time node, and satisfies the condition α+β=1;

[0024] Step S302: Arrange all the fused feature vectors of a certain penalty case in chronological order to obtain the actual feature sequence of the penalty case.

[0025] Step S303: Obtain the time length from the first time node to the last time node in the actual feature sequence, and record it as the actual time interval of a certain penalty case.

[0026] Furthermore, step S400 includes:

[0027] Step S401: Collect several penalty cases, obtain the actual feature sequence and actual time interval of each penalty case, obtain the time length of the actual time interval, perform time normalization processing on the actual feature sequence, and map the time nodes in the actual feature sequence to the normalized time nodes in the normalized feature sequence according to the normalized time.

[0028] Step S402: Arrange the fusion feature vectors of each penalty case according to the order of normalization nodes to obtain the normalization feature sequence of each penalty case. Take two adjacent fusion feature vectors in the same normalization feature sequence as a feature evolution pair. Establish a directed connection between the fusion feature vectors in the feature evolution pair. The direction of the directed connection is from the previous normalization time node to the next normalization time node.

[0029] Step S403: Perform cluster analysis on the fused feature vectors of several normalized feature sequences to divide them into several classification clusters. Combine all classification clusters and directed connections to obtain the risk evolution map.

[0030] Furthermore, step S500 includes:

[0031] Step S501: Take the current enterprise as the target enterprise, obtain the current internal feature words and external feature words of the target enterprise, establish the internal target vector corresponding to the internal feature words and the external target vector corresponding to the external feature words, and fuse the internal target vector and the external target vector to obtain the fused target vector;

[0032] Step S502: Compare the fusion target vector with the classification clusters in the risk evolution map, and denote the classification cluster with the highest similarity to the fusion target vector in the risk evolution map as the target classification cluster, and denote the similarity between the fusion target vector and the target classification cluster as w;

[0033] Step S503: The category clusters that have a directed connection with the target category cluster and are pointed to by the directed connection are recorded as the next-level category clusters of the target category cluster. All the next-level category clusters of the target category cluster are obtained, and the number of directed connections from the target category cluster to all the next-level category clusters is recorded as Q0.

[0034] Furthermore, step S600 includes:

[0035] Step S601: Obtain the j-th next-level classification cluster of the target classification cluster, denoted as classification cluster Gj; obtain the number of directed connections from the target classification cluster to classification cluster Gj, denoted as qj; calculate the risk weight λj of classification cluster Gj, λj = qj / Q0;

[0036] Step S602: Obtain the internal feature words corresponding to all fused feature vectors in the classification cluster Gj to form the j-th target feature word set, and use the product of risk weight and similarity w as the feature word weight of each internal feature word in the target feature word set;

[0037] Step S603: Obtain the target feature word set of each subsequent level of the target classification cluster. In all target feature word sets, sum the feature word weights of the same internal feature words to obtain the ranking weight of each internal feature word.

[0038] Step S604: Sort the internal feature words according to their ranking weight from largest to smallest to obtain the internal feature word push sequence, and send the internal feature word push sequence to the relevant management personnel.

[0039] To better implement the above methods, an enterprise risk early warning system based on intelligent evolution graph is also proposed. The system includes: a feature word management module, a feature fusion module, a risk evolution graph management module, a feature comparison module, and an information push module.

[0040] The feature word management module is used to collect enterprise risk evolution records, including penalty cases and external feature data. The feature fusion module is used to fuse node features and external features at time points to obtain actual feature sequences. The risk evolution map management module is used to perform cluster analysis on the normalized actual feature sequences and manage the risk evolution map. The feature comparison module is used to compare the fused target vector with the classification clusters in the risk evolution map to obtain the target classification cluster. The information push module is used to obtain internal feature words, arrange them according to the ranking weight, and push them to relevant management personnel.

[0041] Furthermore, the feature word management module includes: a time node management unit, a feature word management unit, and a word vector management unit. The time node management unit is used to manage the time nodes in the penalty cases, the feature word management unit is used to manage internal and external feature words, and the word vector management unit is used to manage the word vectors corresponding to the internal and external feature words.

[0042] Furthermore, the feature fusion module includes: a feature fusion unit, an actual feature sequence management unit, and a duration management unit. The feature fusion unit is used to fuse internal feature vectors and external feature vectors, the actual feature sequence management unit is used to manage the actual feature sequence, and the duration management unit is used to manage the duration of the actual time interval of the penalty case.

[0043] Furthermore, the risk evolution graph management module includes: a time normalization unit, a node mapping unit, a feature evolution pair management unit, and a graph management unit. The time normalization unit is used to perform time normalization processing on the actual feature sequence, the node mapping unit is used to map the time nodes in the actual feature sequence to the normalized time nodes in the normalized feature sequence, the feature evolution pair management unit is used to manage feature evolution pairs and directed connections, and the graph management unit is used to aggregate classification clusters and directed connections to manage the risk evolution graph.

[0044] Furthermore, the feature comparison module includes a target vector management unit and a similarity calculation unit. The target vector management unit is used to obtain the current internal and external feature words of the target enterprise and establish a fused target vector. The similarity calculation unit is used to calculate the similarity between the fused target vector and the classification clusters in the risk evolution map.

[0045] Furthermore, the information push module includes: a target feature word set management unit, a weight calculation unit, and a sequence management unit. The target feature word set management unit is used to collect internal feature words to obtain a target feature word set, the weight calculation unit is used to calculate the ranking weight of internal feature words, and the sequence management unit is used to sort the internal feature words to obtain an internal feature word push sequence.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1. This invention integrates multi-source data and external features to construct a dynamic risk evolution map, enabling in-depth analysis of the evolutionary patterns of enterprise risks. By fusing features at specific time points and performing time normalization, it addresses the issues of heterogeneity in multi-source data and non-standardization of the time dimension.

[0048] 2. The construction of the risk relationship graph combines the association probabilities between classification clusters, which can intuitively reflect the dynamic changing trends of enterprise risks. Simultaneously, this invention, through feature matching and similarity calculation, can quickly identify potential risks and provide targeted early warning information. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the structure of an enterprise risk early warning system based on intelligent evolution graph according to the present invention;

[0050] Figure 2 This is a flowchart illustrating an enterprise risk early warning method based on intelligent evolutionary graphs according to the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example: Figures 1-2 As shown, this invention provides a technical solution: an enterprise risk early warning system and method based on intelligent evolutionary graphs, the method comprising:

[0053] Step S100: Obtain the enterprise's penalty case records and extract the time node sequence and node features to form a risk evolution sequence;

[0054] Step S100 includes:

[0055] Step S101: Obtain a penalty case of a certain enterprise. The penalty case includes textual information from business registration, taxation and credit records. Extract the time nodes of each penalty case to form a time node sequence.

[0056] Step S102: Record the feature words in the penalty cases as internal feature words, extract the internal feature words from each penalty case as node features of the time node, and arrange the node features in the order of the time nodes to form a risk evolution sequence.

[0057] Step S103: Obtain all internal feature words at the i-th time node in the risk evolution sequence, obtain the word vectors of the internal feature words, and perform a weighted summation of the word vectors of the internal feature words to obtain the internal feature vector at the i-th time node.

[0058] In this embodiment, a corresponding time interval can be set for each time node, and the text information of the penalty cases and the text information of the external features within the time interval can be collected as the corresponding features of the node.

[0059] For the i-th time node, the collected internal feature words include a1, a2, a3, ..., an. Examples of internal feature words include: tax evasion, false invoices, false advertising, price fraud, substandard product quality, security vulnerabilities, and failure to pass environmental protection inspection.

[0060] Feature words are converted into word vectors using a pre-defined dictionary or a public language model such as Word2Vec, and are represented as ωi1, ωi2, ωi3, ..., ωin. Identical feature words are merged, and the weight of the feature words is determined by statistically analyzing the frequency of occurrence of the same feature word or by the importance of the feature word in the domain.

[0061] Calculate the internal feature vector Fi up to the i-th time node. Where rik represents the weight of the k-th internal feature word, ωik represents the word vector corresponding to the k-th feature word, and K represents the type of internal feature word.

[0062] Step S200: Obtain the external features of the enterprise at each time point, arrange them in the order of time points to form an external feature evolution sequence, extract features from the external features at each time point, and obtain the external feature vector of each time point;

[0063] Step S200 includes:

[0064] Step S201: Obtain the external characteristics of the enterprise at each time point. The external characteristics include public opinion data, advertising data, and text information of business cooperation.

[0065] Step S202: Record the feature words in the external features as external feature words, and arrange the external feature words in chronological order to form an external feature evolution sequence;

[0066] Step S203: Obtain all external feature words at the i-th time node in the external feature evolution sequence, obtain the word vectors of the external feature words, and perform a weighted summation of the word vectors of the external feature words to obtain the external feature vector at the i-th time node.

[0067] In the embodiments, external keywords include: consumer complaints, media exposure, stock price crash, trust crisis, partner default, contract disputes, supply chain disruption, loss of important customers, etc.

[0068] For the i-th time node, the collected external feature words include b1, b2, b3, ..., bm. The external feature words are converted into corresponding word vectors, which are denoted as ei1, ei2, ei3, ..., ein respectively. Identical feature words are merged, and the weight of the feature words is determined by statistically analyzing the frequency of occurrence of the same feature word or by the importance of the feature word in the domain.

[0069] Calculate the external feature vector Ei up to the i-th time node. Where pid represents the weight corresponding to the d-th feature word, eid represents the word vector corresponding to the d-th feature word, and D represents the number of types of external feature words.

[0070] Step S300: Fuse the internal and external feature vectors of each time point to form a fused feature vector, and arrange the fused feature vectors to obtain the actual feature sequence of the penalty cases;

[0071] Step S300 includes:

[0072] Step S301: Perform feature fusion on the internal feature vector and external feature vector of the i-th time node to obtain the fused feature vector Vi of the i-th time node, Vi=α·Fi+β·Ei, where Fi represents the internal feature vector of the i-th time node, Ei represents the internal feature vector of the i-th time node, and satisfies the condition α+β=1;

[0073] Step S302: Arrange all the fused feature vectors of a certain penalty case in chronological order to obtain the actual feature sequence of the penalty case.

[0074] Step S303: Obtain the time length from the first time node to the last time node in the actual feature sequence, and record it as the actual time interval of a certain penalty case.

[0075] Step S400: Perform time normalization on the actual feature sequence to form a normalized feature sequence. Perform cluster analysis on the fused feature vectors of the normalized feature sequence to divide it into multiple classification clusters. Combine the directed connections between the fused feature vectors to obtain the risk evolution map.

[0076] Step S400 includes:

[0077] Step S401: Collect several penalty cases, obtain the actual feature sequence and actual time interval of each penalty case, obtain the time length of the actual time interval, perform time normalization processing on the actual feature sequence, and map the time nodes in the actual feature sequence to the normalized time nodes in the normalized feature sequence according to the normalized time.

[0078] Step S402: Arrange the fusion feature vectors of each penalty case according to the order of normalization nodes to obtain the normalization feature sequence of each penalty case. Take two adjacent fusion feature vectors in the same normalization feature sequence as a feature evolution pair. Establish a directed connection between the fusion feature vectors in the feature evolution pair. The direction of the directed connection is from the previous normalization time node to the next normalization time node.

[0079] Step S403: Perform cluster analysis on the fused feature vectors of several normalized feature sequences to divide them into several classification clusters. Combine all classification clusters and directed connections to obtain the risk evolution map.

[0080] Collect the actual time intervals of all penalty cases, obtain the time length of each actual time interval, and record the longest actual time interval among all penalty cases as tmax and the shortest actual time interval as tmin.

[0081] Retrieve the h-th penalty case from all penalty cases, and denote the actual time interval of the h-th penalty case as t. h Calculate the normalized time length T of the h-th penalty case. h , t j With T j The ratio is used as the mapping ratio to map each time node in the actual time interval to a normalized time node in the normalized time length.

[0082] For two fused feature vectors at adjacent normalized time nodes, a directed connection is established, with the direction pointing from the fused feature vector corresponding to the previous normalized time node to the fused feature vector corresponding to the next normalized time node. The fused feature vectors are then concatenated to form a normalized feature sequence.

[0083] Step S500: Compare the feature vector of the target enterprise with the classification clusters in the risk evolution map, select the target classification cluster, determine all the subsequent level classification clusters of the target classification cluster, and count the number of directed connections from the target classification cluster to each subsequent level classification cluster.

[0084] Step S500 includes:

[0085] Step S501: Take the current enterprise as the target enterprise, obtain the current internal feature words and external feature words of the target enterprise, establish the internal target vector corresponding to the internal feature words and the external target vector corresponding to the external feature words, and fuse the internal target vector and the external target vector to obtain the fused target vector;

[0086] Step S502: Compare the fusion target vector with the classification clusters in the risk evolution map, and denote the classification cluster with the highest similarity to the fusion target vector in the risk evolution map as the target classification cluster, and denote the similarity between the fusion target vector and the target classification cluster as w;

[0087] Step S503: The category clusters that have a directed connection with the target category cluster and are pointed to by the directed connection are recorded as the next-level category clusters of the target category cluster. All the next-level category clusters of the target category cluster are obtained, and the number of directed connections from the target category cluster to all the next-level category clusters is recorded as Q0.

[0088] When calculating the similarity between the fusion target vector and the classification cluster, the methods that can be used include: calculating the average similarity between the fusion target vector and all fusion feature vectors in the classification cluster, or first calculating the cluster core of the fusion feature vector in each classification cluster, and then obtaining the similarity between the fusion target vector and the classification cluster by calculating the similarity between the fusion target vector and the cluster core.

[0089] The direction pointed to by the directed connection is defined as the next level direction. For example, if classification cluster A includes a fused feature vector v1 and classification cluster B includes a fused feature vector v2, and there is a directed connection v1→v2 between v1 and v2, then classification cluster B is the next level classification cluster of classification cluster A.

[0090] Step S600: Obtain the next-level classification cluster of the target classification cluster, form a target feature word set by combining the internal feature words of the next-level classification cluster, calculate the ranking weight of each feature word in the target feature word set, and sort the internal feature words according to the ranking weight to form a push sequence;

[0091] Step S600 includes:

[0092] Step S601: Obtain the j-th next-level classification cluster of the target classification cluster, denoted as classification cluster Gj; obtain the number of directed connections from the target classification cluster to classification cluster Gj, denoted as qj; calculate the risk weight λj of classification cluster Gj, λj = qj / Q0;

[0093] Step S602: Obtain the internal feature words corresponding to all fused feature vectors in the classification cluster Gj to form the j-th target feature word set, and use the product of risk weight and similarity w as the feature word weight of each internal feature word in the target feature word set;

[0094] Step S603: Obtain the target feature word set of each subsequent level of the target classification cluster. In all target feature word sets, sum the feature word weights of the same internal feature words to obtain the ranking weight of each internal feature word.

[0095] Step S604: Sort the internal feature words according to their ranking weight from largest to smallest to obtain the internal feature word push sequence, and send the internal feature word push sequence to the relevant management personnel.

[0096] In this embodiment, seven directed connections are drawn from category cluster A1 to the next level, of which three directed connections point to category cluster B1 and four directed connections point to category cluster B2.

[0097] Therefore, the risk weight of B1 is λ1 = 0.429, and the risk weight of B2 is λ2 = 0.571;

[0098] The internal feature words of classification clusters B1 and B2 are collected respectively. The backtracking path of the internal feature words is: fused feature vector → internal feature vector → word vector corresponding to internal feature word → internal feature word;

[0099] According to step S603, the feature words in classification cluster B1 and classification cluster B2 are aggregated, the cumulative risk weight of the same feature words is calculated and sorted to obtain the internal feature word push sequence.

[0100] In another embodiment, the fused target vector is compared with the classification clusters in the risk evolution map, and the similarity between the fused target vector and multiple classification clusters is calculated, denoted as w1, w2, ... wf, respectively;

[0101] Calculate the risk weight corresponding to each category cluster and multiply it by the similarity of the corresponding category cluster to obtain the ranking weight of the internal feature words. For example, for the j-th next-level category cluster in the first target category cluster, the corresponding feature word weight is w1×λj.

[0102] The system includes: a feature word management module, a feature fusion module, a risk evolution graph management module, a feature comparison module, and an information push module;

[0103] The feature word management module is used to collect the risk evolution records of enterprises, including penalty cases and external feature data. The feature word management module includes: a time node management unit, a feature word management unit, and a word vector management unit. The time node management unit is used to manage the time nodes in penalty cases, the feature word management unit is used to manage internal and external feature words, and the word vector management unit is used to manage the word vectors corresponding to internal and external feature words.

[0104] The feature fusion module is used to fuse the node features and external features of time nodes to obtain the actual feature sequence. The feature fusion module includes a feature fusion unit, an actual feature sequence management unit, and a duration management unit. The feature fusion unit is used to fuse internal feature vectors and external feature vectors. The actual feature sequence management unit is used to manage the actual feature sequence. The duration management unit is used to manage the duration of the actual time interval of the penalty case.

[0105] The risk evolution graph management module is used to perform cluster analysis on the normalized actual feature sequences and manage the risk evolution graph. The risk evolution graph management module includes: a time normalization unit, a node mapping unit, a feature evolution pair management unit, and a graph management unit. The time normalization unit is used to perform time normalization processing on the actual feature sequences. The node mapping unit is used to map the time nodes in the actual feature sequences to the normalized time nodes in the normalized feature sequences. The feature evolution pair management unit is used to manage feature evolution pairs and directed connections. The graph management unit is used to aggregate classification clusters and directed connections and manage the risk evolution graph.

[0106] The feature comparison module is used to compare the fused target vector with the classification clusters in the risk evolution graph to obtain the target classification cluster. The feature comparison module includes a target vector management unit and a similarity calculation unit. The target vector management unit is used to obtain the current internal and external feature words of the target enterprise and establish the fused target vector. The similarity calculation unit is used to calculate the similarity between the fused target vector and the classification clusters in the risk evolution graph.

[0107] The information push module is used to acquire internal feature words, arrange them according to their ranking weights, and push them to relevant management personnel. The information push module includes a target feature word set management unit, a weight calculation unit, and a sequence management unit. The target feature word set management unit is used to collect internal feature words to obtain a target feature word set, the weight calculation unit is used to calculate the ranking weights of the internal feature words, and the sequence management unit is used to sort the internal feature words to obtain an internal feature word push sequence.

[0108] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for enterprise risk early warning based on intelligent evolution map, characterized in that: The method comprises: Step S100: obtaining the penalty case records of an enterprise and extracting the time node sequence and node features to form a risk evolution sequence; Step S200: obtaining the external features of the enterprise at each time node, arranging the external features in the order of time nodes to form an external feature evolution sequence, extracting the external features at each time node to obtain the external feature vectors of the time nodes; Step S300: fusing the internal feature vectors and the external feature vectors of each time node to form a fused feature vector, and arranging the fused feature vector to obtain the actual feature sequence of the penalty case; Step S400: performing time normalization processing on the actual feature sequence to form a normalized feature sequence, and performing clustering analysis on the fused feature vectors of the normalized feature sequence to divide them into a plurality of classification clusters, and obtaining a risk evolution graph by combining the directed connections between the fused feature vectors; Step S400 comprises: Step S401: collecting a plurality of penalty cases, obtaining the actual feature sequence and the actual time interval of each penalty case, obtaining the time length of the actual time interval, and performing time normalization processing on the actual feature sequence to map the time nodes in the actual feature sequence to the normalized time nodes in the normalized feature sequence according to the normalized time; Step S402: arranging the fused feature vectors of each penalty case in the order of normalized nodes to obtain the normalized feature sequence of each penalty case, and taking the two adjacent fused feature vectors in the same normalized feature sequence as a feature evolution pair, and establishing a directed connection between the fused feature vectors in the feature evolution pair, wherein the direction of the directed connection is from the previous normalized time node to the next normalized time node; Step S403: performing clustering analysis on the fused feature vectors of the plurality of normalized feature sequences to divide them into a plurality of classification clusters, and collecting all classification clusters and directed connections to obtain a risk evolution graph; Step S500: comparing the feature vectors of the target enterprise with the classification clusters in the risk evolution graph, screening out the target classification cluster, determining all next-level classification clusters of the target classification cluster, and counting the number of directed connections of the target classification cluster to each next-level classification cluster; Step S600: obtaining the next-level classification clusters of the target classification cluster, grouping the internal feature words of the next-level classification clusters into a target feature word set, calculating the ranking weight of each feature word in the target feature word set, ranking the internal feature words according to the ranking weight, and forming a push sequence.

2. The enterprise risk early warning method based on intelligent evolution map according to claim 1, characterized in that: Step S100 comprises: Step S101: obtaining a penalty case of an enterprise, wherein the penalty case comprises text information of business, tax and credit records, and extracting the time nodes of each penalty case to form a time node sequence; Step S102: taking the feature words in the penalty case as internal feature words, extracting the internal feature words from each penalty case as the node features of the time nodes, and arranging the node features in the order of time nodes to form a risk evolution sequence; Step S103: obtain all internal feature words of the i-th time node in the risk evolution sequence, obtain the word vector of the internal feature words, and perform weighted summation on the word vector of the internal feature words to obtain the internal feature vector of the i-th time node.

3. The enterprise risk early warning method based on intelligent evolution map according to claim 2, characterized in that: Step S200 includes: Step S201: obtain the external features of the enterprise at each time node, including public opinion data, advertising data, and commercial cooperation text information; Step S202: record the feature words in the external features as external feature words, arrange the external feature words in time node order to form an external feature evolution sequence; Step S203: obtain all external feature words of the i-th time node in the external feature evolution sequence, obtain the word vector of the external feature words, and perform weighted summation on the word vector of the external feature words to obtain the external feature vector of the i-th time node.

4. The enterprise risk early warning method based on intelligent evolution map according to claim 3, characterized in that: Step S300 includes: Step S301: perform feature fusion on the internal feature vector and the external feature vector of the i-th time node to obtain the fusion feature vector Vi of the i-th time node, Vi = a Fi + b Ei, wherein Fi represents the internal feature vector of the i-th time node, Ei represents the external feature vector of the i-th time node, and a + b = 1 is satisfied; Step S302: arrange all fusion feature vectors of the certain penalty case in time node order to obtain the actual feature sequence of the certain penalty case, Step S303: obtain the time length from the first time node to the last time node in the actual feature sequence, and record it as the actual time interval of the certain penalty case.

5. The enterprise risk early warning method based on intelligent evolution map according to claim 4, characterized in that: Step S500 includes: Step S501: take a current enterprise as a target enterprise, obtain the internal feature words and external feature words of the target enterprise, establish an internal target vector corresponding to the internal feature words and an external target vector corresponding to the external feature words, and perform feature fusion on the internal target vector and the external target vector to obtain a fusion target vector; Step S502: compare the fusion target vector with the classification clusters in the risk evolution graph, record the classification cluster with the greatest similarity to the fusion target vector in the risk evolution graph as a target classification cluster, and record the similarity between the fusion target vector and the target classification cluster as w; Step S503: record the classification cluster that is connected to the target classification cluster and is pointed to by the target classification cluster as a next-level classification cluster of the target classification cluster, obtain all next-level classification clusters of the target classification cluster, and record the number of directed connections from the target classification cluster to all next-level classification clusters as Q0.

6. The enterprise risk early warning method based on intelligent evolution map according to claim 5, characterized in that: Step S600 includes: Step S601: record the j-th next-level classification cluster of the target classification cluster as a classification cluster Gj, record the number of directed connections from the target classification cluster to the classification cluster Gj as qj, calculate the risk weight of the classification cluster Gj, and Step S602: obtain all fusion feature vectors in the classification cluster Gj to form a j-th target feature word set composed of internal feature words, and take the product of the risk weight and the similarity w as the feature word weight of each internal feature word in the target feature word set. Step S603: respectively acquiring a target feature word set of each next-level classification cluster of the target classification cluster, and performing feature word weight summation on the same internal feature word in all target feature word sets to respectively obtain a ranking weight of each internal feature word; Step S604: ranking the internal feature words according to the ranking weights from large to small to obtain an internal feature word pushing sequence, and sending the internal feature word pushing sequence to a relevant management personnel.

7. An enterprise risk early warning system based on intelligent evolution map, used for executing the enterprise risk early warning method based on intelligent evolution map in any one of claims 1-6. The system comprises: a feature word management module, a feature fusion module, a risk evolution graph management module, a feature comparison module, and an information pushing module; The feature word management module is used for collecting risk evolution records of an enterprise, including penalty cases and external feature data, the feature fusion module is used for fusing node features of time nodes and external features to obtain actual feature sequences, the risk evolution graph management module is used for performing cluster analysis on normalized actual feature sequences to manage a risk evolution graph, the feature comparison module is used for comparing a fusion target vector with classification clusters in the risk evolution graph to obtain a target classification cluster, and the information pushing module is used for acquiring internal feature words and arranging the internal feature words according to ranking weights to push the internal feature words to a relevant management personnel.

8. The enterprise risk early warning system based on an intelligent evolution graph according to claim 7, characterized in that: The feature word management module comprises a time node management unit, a feature word management unit, and a word vector management unit, wherein the time node management unit is used for managing time nodes in penalty cases, the feature word management unit is used for managing internal feature words and external feature words, and the word vector management unit is used for managing word vectors corresponding to the internal feature words and the external feature words; The feature fusion module comprises a feature fusion unit, an actual feature sequence management unit, and a time length management unit, wherein the feature fusion unit is used for fusing internal feature vectors and external feature vectors, the actual feature sequence management unit is used for managing actual feature sequences, and the time length management unit is used for managing a time length of an actual time interval of a penalty case; The risk evolution graph management module comprises a time normalization unit, a node mapping unit, a feature evolution pair management unit, and a graph management unit, wherein the time normalization unit is used for performing time normalization processing on actual feature sequences, the node mapping unit is used for mapping time nodes in actual feature sequences into normalized time nodes in normalized feature sequences, the feature evolution pair management unit is used for managing feature evolution pairs and directed connections, and the graph management unit is used for collecting classification clusters and directed connections to manage a risk evolution graph.

9. The enterprise risk early warning system based on an intelligent evolution graph according to claim 7, characterized in that: The feature comparison module comprises a target vector management unit and a similarity calculation unit, wherein the target vector management unit is used for acquiring internal feature words and external feature words of a target enterprise at present to establish a fusion target vector, and the similarity calculation unit is used for calculating a similarity between the fusion target vector and classification clusters in the risk evolution graph. The information pushing module comprises a target feature word set management unit, a weight calculation unit and a sequence management unit, wherein the target feature word set management unit is used to collect internal feature words to obtain a target feature word set, the weight calculation unit is used to calculate the sorting weight of the internal feature words, and the sequence management unit is used to sort the internal feature words to obtain an internal feature word pushing sequence.

Images