A modeling method and device for identifying enterprise management risks

Abstract

The embodiment of the present invention discloses a modeling method and device for enterprise management risk identification, which relates to the field of computer technology and can effectively improve the accuracy of model prediction. The method includes: obtaining target data from an enterprise risk data source, wherein the target data includes character data and / or knowledge graph data; encoding the target data according to the adjacent relationship between the target data to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold; according to the generation time of the target data, the target feature code is time-series grouped, and the time evolution model between each group of target feature codes is trained; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is a time series feature code. The present invention is applicable to enterprise risk prediction.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a modeling method and device for identifying enterprise operation and management risks. Background Art

[0002] The operational and management risks of an enterprise are contained in a variety of data related to the enterprise. How to identify the risks that the enterprise may face from these huge amounts of data is of great significance to the enterprise itself and entities related to the enterprise.

[0003] Since enterprise risk-related data has the characteristics of complex data sources and diverse data formats, it is difficult to extract effective features when using mathematical models to identify enterprise risks, and it is even more difficult to use the extracted features for model training. Summary of the invention

[0004] In view of this, the embodiments of the present invention provide a modeling method, device, electronic device and storage medium for enterprise management risk identification, which can greatly reduce the workload and blindness of feature extraction and effectively improve the accuracy of model prediction.

[0005] In a first aspect, an embodiment of the present invention provides a modeling method for enterprise management risk identification, comprising: obtaining target data from an enterprise risk data source, the target data including character data and / or knowledge graph data; encoding the target data according to the adjacent relationship between the target data to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold; according to the generation time of the target data, the target feature codes are time-series grouped, and a time evolution model between each group of target feature codes is trained; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature code; using the features of the time series feature code as input and the probability of enterprise risk occurrence as output for model training to obtain a model for enterprise management risk identification.

[0006] Optionally, obtaining target data from the enterprise risk data source includes: performing word segmentation and key information extraction on the unstructured data in the enterprise risk data source so that each key information forms a character-type data; and / or converting the structured data in the enterprise risk data source into knowledge graph-type data.

[0007] Optionally, the target data includes character data; encoding the target data according to the adjacent relationship between the target data to obtain the target feature code includes: constructing the adjacent relationship between each character data according to a preset rule; and vectorizing the character data according to the adjacent relationship to form the target feature code.

[0008] Optionally, constructing the adjacent relationship between each character data according to preset rules includes: constructing the adjacent relationship between people in each type of position in each enterprise according to preset rules; vectorizing the character data according to the adjacent relationship to form the target feature code includes: vectorizing the identity identifiers of people in each type of position and calculating the average value to obtain the average identity feature code of such position; and concatenating the average identity feature codes of each type of position to form the target feature code.

[0009] Optionally, the target data includes knowledge graph data; encoding the target data according to the adjacent relationship between the target data to obtain the target feature code includes: based on the relationship between enterprises in the knowledge graph data, using the pagerank algorithm to determine the influence feature code of each enterprise; based on the relationship between enterprises in the knowledge graph data, using the label propagation algorithm to determine the risk feature code of each enterprise; based on the relationship between enterprises in the knowledge graph data, using the deepwalk algorithm to determine the feature code related to people in each enterprise; splicing the influence feature code, the risk feature code and the feature code related to people to form the target feature code.

[0010] Optionally, the target feature codes are grouped in time series according to the generation time of the target data, and the time evolution model between each group of target feature codes is trained, including: grouping the target feature codes corresponding to the target data generated within the same preset time period into one group; using N-1 groups of target feature codes corresponding to the first N-1 preset time periods as input and the target feature code corresponding to the Nth preset time period as output, to train the time evolution model, wherein N is an integer greater than 1.

[0011] Optionally, the target data includes character data and knowledge graph data; encoding the target data according to the adjacent relationship between the target data to obtain a target feature code includes: vectorizing the character data and the knowledge graph data according to the adjacent relationship between the target data, respectively, to form a character feature code and a graph feature code; and splicing the character feature code and the graph feature code to form the target feature code.

[0012] In a second aspect, an embodiment of the present invention further provides a modeling device for identifying enterprise management risks, comprising: an acquisition unit, for acquiring target data from an enterprise risk data source, wherein the target data includes character data and / or knowledge graph data; an encoding unit, for encoding the target data according to the adjacent relationship between the target data to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold; a first training unit, for time-series grouping the target feature codes according to the generation time of the target data, and training a time evolution model between each group of target feature codes; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature code; a second training unit, for performing model training with the features of the time series feature code as input and the probability of enterprise risk occurrence as output to obtain a model for identifying enterprise management risks.

[0013] Optionally, the acquisition unit includes: a first acquisition module, used to perform word segmentation and key information extraction on unstructured data in the enterprise risk data source, so that each key information forms a character-type data; and / or a second acquisition module, used to convert structured data in the enterprise risk data source into knowledge graph-type data.

[0014] Optionally, the target data includes character data; the encoding unit includes: a construction module, used to construct the adjacent relationship between each character data according to a preset rule; and a vectorization module, used to vectorize the character data according to the adjacent relationship to form the target feature code.

[0015] Optionally, the construction module is specifically used to construct the adjacent relationship between people in each type of position in each enterprise according to preset rules; the vectorization module is specifically used to: vectorize the identity identifiers of people in each type of position and calculate the average value to obtain the average identity feature code of such position; splice the average identity feature codes of each type of position together to form the target feature code.

[0016] Optionally, the target data includes knowledge graph data; the encoding unit includes: a first encoding module, used to determine the influence feature coding of each enterprise by using the pagerank algorithm based on the relationship between enterprises in the knowledge graph data; a second encoding module, used to determine the risk feature coding of each enterprise by using the label propagation algorithm based on the relationship between enterprises in the knowledge graph data; a third encoding module, used to determine the feature coding related to people in each enterprise by using the deepwalk algorithm based on the relationship between enterprises and people in the knowledge graph data; a splicing module, used to splice the influence feature coding, the risk feature coding and the feature coding related to people to form the target feature coding.

[0017] Optionally, the first training unit is specifically used to: group the target feature codes corresponding to the target data generated within the same preset time period into a group; use the N-1 groups of target feature codes corresponding to the first N-1 preset time periods as input and the target feature code corresponding to the Nth preset time period as output to train the time evolution model, wherein N is an integer greater than 1.

[0018] Optionally, the target data includes character data and knowledge graph data; the encoding unit is specifically used to: vectorize the character data and the knowledge graph data respectively according to the adjacent relationship between the target data to form character feature codes and graph feature codes; and splice the character feature codes and the graph feature codes to form the target feature codes.

[0019] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising: a housing, a processor, a memory, a circuit board and a power supply circuit, wherein the circuit board is placed inside the space enclosed by the housing, and the processor and the memory are arranged on the circuit board; a power supply circuit is used to supply power to various circuits or devices of the above-mentioned electronic device; the memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute any modeling method for enterprise management risk identification provided by the embodiments of the present invention.

[0020] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement any modeling method for enterprise management risk identification provided by an embodiment of the present invention.

[0021] The modeling method and device for enterprise management risk identification provided by the embodiments of the present invention can obtain target data from the enterprise risk data source, encode the target data according to the adjacent relationship between the target data, and obtain the target feature code. Since the target feature code can represent each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than the preset threshold, character data and / or knowledge graph data can be uniformly represented by a set of dense multidimensional vectors, which is conducive to feature extraction and model training. Then, according to the generation time of the target data, the target feature codes are time-series grouped, and the time evolution model between each group of target feature codes is trained to obtain the time-series feature code of the target data, thereby further introducing the time attribute of the enterprise risk into the model system, thereby effectively improving the accuracy of the model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0023] Figure 1 A flow chart of a modeling method for enterprise operation and management risk identification provided by an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of feature splicing in a modeling method for enterprise management risk identification provided by an embodiment of the present invention;

[0025] Figure 3 is a schematic diagram of an encoding method in an embodiment of the present invention;

[0026] Figure 4 Another feature splicing schematic diagram in a modeling method for enterprise management risk identification provided by an embodiment of the present invention;

[0027] Figure 5 A schematic diagram of a structure of a modeling device for enterprise operation and management risk identification provided by an embodiment of the present invention;

[0028] Figure 6 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] It should be clear that the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] On the first aspect, an embodiment of the present invention provides a modeling method for enterprise management risk identification, which can encode features related to enterprise risks and use the encoded features for model training, thereby greatly reducing the workload and blindness of feature extraction and effectively improving the accuracy of model prediction.

[0032] like Figure 1 As shown, the modeling method for enterprise operation and management risk identification provided by the embodiment of the present invention may specifically include:

[0033] S11, obtaining target data from the enterprise risk data source, wherein the target data includes character data and / or knowledge graph data.

[0034] Among them, the enterprise risk data source may include any information related to the enterprise, such as basic information of industrial and commercial enterprises, financial information of listed companies, operating status information of small and medium-sized enterprises nationwide, court judgment information, tax rating information, court announcement information, court execution information, trademark data, patent information, software / work copyright information, enterprise bidding information, and media reports involving enterprises, etc. The carriers of this information may include various databases, announcement documents, journals, etc., and the formats of the information are also very different. In this step, the required information can be integrated from these enterprise risk data sources to obtain the target data.

[0035] Optionally, in one embodiment of the present invention, the data type of the target data may include character data, knowledge graph data, or a combination of the two. When the data in the enterprise risk data source is not character data or knowledge graph data, the enterprise risk source data may be first converted into one of the two types of data before further processing.

[0036] Among them, character data may include, for example, ID data. ID data is a type of category data. Compared with traditional category data, ID features have a huge value space, such as enterprise ID, corporate legal person ID, word ID after word segmentation, etc. Optionally, the value space of enterprise ID can be equal to the number of enterprises, and the value space of corporate legal person ID is equal to the total number of people; after a piece of news is segmented, the space composed of word IDs is even larger. Knowledge graph data may refer to data that reflects the internal structure of an enterprise or the external relations of an enterprise through a graph constructed through the triple structure of a knowledge graph. Given that character data and knowledge graph data are generally larger and more complex, in order to facilitate feature extraction and use the extracted features for model training, the embodiments of the present invention further process these two types of data as follows.

[0037] S12, encoding the target data according to the adjacent relationship between the target data to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold.

[0038] In this step, the target data can be encoded according to the adjacent relationship between the target data to obtain the target feature encoding. Through the above encoding, each specific feature in the target data can be converted into a feature represented by a vector, which is convenient for model training and prediction. In one embodiment of the present invention, the dimension of the vector can be less than a preset threshold, such as less than 200 or less than 500, so that the features can be represented more concentratedly, effectively improving the efficiency of the model. As for how each target data is specifically represented by a vector, it can be trained according to demand.

[0039] S13, according to the generation time of the target data, the target feature codes are grouped in time series, and a time evolution model between each group of target feature codes is trained; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature code;

[0040] The above steps encode the target data by only considering the content or structure of the target data itself. If the time element is added to the target data, it is equivalent to encoding the data at a fixed time point. However, the embodiments of the present invention are not limited to this. In other embodiments of the present invention, based on the above coding features, the law of changes in the business data of the enterprise over time can be taken into account to construct features with time-varying laws, so that the risk identification of the enterprise is more accurate and objective.

[0041] To this end, in one embodiment of the present invention, the target data generated within a period of time can be grouped together, for example, the business data of January is grouped into group 1, the business data of February is grouped into group 2, the business data of March is grouped into group 3, and the business data of April is grouped into group 4, etc. The time evolution model of the feature is trained using the evolution law between multiple groups of data. For example, the change law can be found between the data in group 1, group 2, group 3, and group 4 to obtain the time evolution model. The coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature encoding.

[0042] S14, taking the features of the time series feature encoding as input and the probability of enterprise risk occurrence as output to perform model training, so as to obtain a model for enterprise operation and management risk identification.

[0043] In this step, the target data can be re-encoded into a set of features with time parameters, i.e., features of time series feature encoding. The features of time series feature encoding are used as input, and the probability of enterprise risk occurrence is used as output for model training, so that a model for enterprise management risk identification can be obtained.

[0044] The modeling method for enterprise management risk identification provided by the embodiment of the present invention can obtain target data from the enterprise risk data source, encode the target data according to the adjacent relationship between the target data, and obtain the target feature code. Since the target feature code can represent each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than the preset threshold, character data and / or knowledge graph data can be uniformly represented by a set of dense multidimensional vectors, which is conducive to feature extraction and model training. Then, according to the generation time of the target data, the target feature codes are time-series grouped, and the time evolution model between each group of target feature codes is trained to obtain the time-series feature code of the target data, thereby further introducing the time attribute of the enterprise risk into the model system, thereby effectively improving the accuracy of the model prediction.

[0045] Optionally, in step S11, the processing method for obtaining target data from the enterprise risk data source may be different depending on the data type. For example, for the unstructured data in the enterprise risk data source, word segmentation and key information extraction may be performed so that each key information forms a character data; for the structured data in the enterprise risk data source, the structured data in the enterprise risk data source may be converted into knowledge graph data.

[0046] For example, for unstructured data such as a media report, we can segment the report and extract key information from it. Each key information forms a character data, such as "World Horticultural Exposition", "Opening", etc. For another example, for structured data such as a class's high school entrance examination results, we can form knowledge graph data according to student ID, name, subject, score, etc.

[0047] After obtaining the target data of character type or knowledge graph type, the target data can be encoded according to the adjacent relationship between the target data in step S12. The encoding of character type data and the encoding of knowledge graph type data are respectively described below.

[0048] Optionally, in one embodiment of the present invention, the target data obtained from the enterprise risk data source includes character data; then in step S12, encoding the target data according to the adjacent relationship between the target data to obtain the target feature code may include:

[0049] According to the preset rules, construct the adjacent relationship between each character data;

[0050] According to the adjacent relationship, the character data is vectorized to form the target feature code.

[0051] For example, in one embodiment of the present invention, there are two IDs, A and B. If there is a certain relationship between A and B, a directed edge can be connected between A and B, so that A and B are adjacent. For example, assuming that A and B are both legal persons of a certain enterprise X, and A becomes the legal person of the enterprise before B, then there is a relationship between A and B that "A and B are both legal persons of enterprise X and A is before B". Then, it can be defined according to this relationship whether an edge can be connected between other IDs. If an edge is connected between two IDs, the two IDs have an adjacent relationship. Of course, in other embodiments of the present invention, the adjacent relationship between each character data can also be defined by other connections, and the embodiments of the present invention do not limit this.

[0052] After constructing the adjacent relationship between each character data, each character data can be represented by a set of vectors according to this adjacent relationship. By representing character data with vectors, the amount of character data can be greatly reduced, effectively improving the efficiency of the model.

[0053] For example, in one embodiment of the present invention, each enterprise is taken as a unit, and a "paragraph" Pi = {C i1 , C i2 , C i3 ......C ini}.

[0054] Where Pi represents the "paragraph" constructed by the ith enterprise, and Cij represents the 1-hot code of the jth executive of the ith enterprise. Considering executives as a word, the 1-hot codes of the first n executives are taken as n input vectors, and the 1-hot code of the last executive is taken as the output vector. A deep learning model is constructed and trained with all the "paragraph" data composed of Pi. The coefficient matrix between the input layer and the first layer of the hidden layer of the trained model is used as the embedding parameter, and each C ini The 1-hot encoding of is embedded into a dense vector space.

[0055] Furthermore, in the above embodiment, the neighbor relationship constructed based on Pi only reflects the ID features related to people in the enterprise. In view of the fact that the business risk of an enterprise is based on the enterprise, in one embodiment of the present invention, the ID features related to people can also be spliced ​​into features of the enterprise dimension.

[0056] Specifically, in one embodiment of the present invention, for an enterprise, according to a preset rule, constructing the adjacent relationship between each character data may include:

[0057] According to the preset rules, construct the adjacent relationship between people in each position in each enterprise;

[0058] The vectorizing the character data according to the adjacent relationship to form the target feature code includes: vectorizing the identity identifiers of people in each type of position and calculating the average value to obtain the average identity feature code of such position;

[0059] The average identity characteristic codes of various positions are concatenated to form the target characteristic code.

[0060] For example, for any type of position (i.e., dimension k (such as executives, middle-level managers, or ordinary employees, etc.)), find the jth person (a total of J people) who meets the dimension, and record the characteristic code of the person in this type of position as:

[0061]

[0062] For each dimension k, a fixed-length feature vector is constructed. The lengths of feature vectors of different dimensions can be the same or different. Figure 2 As shown, by concatenating these feature vectors, we can obtain the ID-type features for the enterprise.

[0063] For knowledge graph data, in one embodiment of the present invention, the target data obtained from the enterprise risk data source includes knowledge graph data, and then encoding the target data according to the adjacent relationship between the target data to obtain the target feature encoding may specifically include:

[0064] Based on the relationship between enterprises in the knowledge graph data, the PageRank algorithm is used to determine the influence feature coding of each enterprise;

[0065] Based on the relationship between enterprises in the knowledge graph data, the label propagation algorithm is used to determine the risk feature coding of each enterprise;

[0066] Based on the relationship between enterprises and people in knowledge graph data, the deepwalk algorithm is used to determine the feature coding related to people in each enterprise;

[0067] The influence feature code, the risk feature code and the person-related feature code are concatenated to form the target feature code.

[0068] Specifically, in knowledge graph data, the relationship between enterprises can be simulated as the relationship between nodes in the knowledge graph. The influence of an enterprise can be determined by the number and / or quality of nodes connected to the node representing the enterprise.

[0069] In the embodiment of the present invention, the Pagerank algorithm simulates enterprise A as a node in the knowledge graph data, and the "outbound link" of this node is regarded as a "directed edge" pointing to other nodes, and the "inbound link" is a directed edge from other nodes to this node. The entire knowledge graph can form a directed graph.

[0070] Alternatively, the assessment of corporate impact can follow the following two principles:

[0071] Quantitative hypothesis: The larger the in-degree (i.e., the number of links) of a node, the higher the influence of the enterprise represented by the node; for example, if the in-degree of node A1 is 3 and the in-degree of node A2 is 6, the influence of the enterprise represented by node A2 is greater than that of the enterprise represented by node A1.

[0072] Quality assumption: The greater the influence of the source node of the in-degree of a node, the greater the influence of the enterprise represented by the node. For example, the influence of the source node B1 of the in-degree of node A3 is greater than the influence of the source node B2 of the in-degree of node A4, then the influence of the enterprise represented by node A3 is greater than the influence of the enterprise represented by node A4. Optionally, in one embodiment of the present invention, the influence of the source node of the in-degree can also be expressed as the in-degree of the source node of the in-degree, based on which, the influence of node B1 can be expressed as the in-degree of node B1.

[0073] The "random walk model" of random processes in mathematics can be used to walk in the directed graph to represent the influence of each company. The random walk model means that the current state of a thing is only related to its previous state, and has nothing to do with its previous state. Assuming that in the initial state, the probability of a user walking through all pages is equal, and after each walk, the user will walk to the page pointed to by the link with equal probability according to the link given on the page at this time. Through the probability analysis of this random walk, it can be obtained that the probability of the user staying on a certain node is greater. The greater the probability, the higher the influence.

[0074] After determining the influence of the enterprise, in one embodiment of the present invention, the label propagation algorithm can be used to further determine the risk feature code of each enterprise. In this way, the greater the influence of the enterprise, once the enterprise has risk features, the greater the probability of the enterprise having risks, and the smaller the influence of the enterprise, when the risk features appear, the probability of the enterprise having risks is relatively small.

[0075] Specifically, label propagation is a semi-supervised learning algorithm, and its core idea is that similar data should have the same label. It can be considered that label propagation tends to use the label with the largest number of labels of a node's neighbor nodes as the label of the node itself. Optionally, these labels can be labels that represent different types of risks or labels that represent different degrees of risk. Based on this, if most of the nodes in all the neighbor nodes of an enterprise are labeled with the risk label R1, then the enterprise will also be labeled with the risk label R1.

[0076] In the knowledge graph of enterprise relationships, equity investment relationships and guarantee relationships between enterprises are more likely to cause enterprise risks. In one embodiment of the present invention, risk feature coding can be performed on enterprises with these two types of relationships. For example, if enterprise A guarantees enterprise B, a directed edge is connected between enterprise A and enterprise B. After determining the connection relationship, the risk coding value of each enterprise in the knowledge graph can be obtained by using pagerank and label propagation.

[0077] After determining the impact of the relationship between enterprises on enterprise risk through pagerank and label propagation algorithms, in an embodiment of the present invention, the feature coding related to people in the enterprise can also be determined based on the relationship between enterprises and people in the knowledge graph.

[0078] Based on the relationship between enterprises and people in knowledge graph data, the deepwalk algorithm is used to determine the feature coding related to people in each enterprise.

[0079] In the relationship between enterprises and people, enterprise risks can be characterized through human relationships, and the idea of ​​deepwalk is used for feature encoding. The Deepwalk algorithm is also similar to the word2vec method. Deepwalk evenly selects network nodes on the graph to generate a random walk sequence of fixed length, and compares it to a sequence in natural language, and uses the DNN model to obtain the feature encoding method.

[0080] like Figure 3 As shown, in one embodiment of the present invention, the company and executive relationships are used to form the network relationship between companies and people. Figure 3 The six nodes 1-6 in the left figure can represent companies or executives, and the lines between the nodes represent the connections between the nodes. According to the idea of ​​Deepwalk, we can randomly walk between the nodes along these lines to form several sequences (see Figure 3 The sequence relationship is analogous to the sequence in natural language. Given a node, a model is trained to predict nearby nodes (see Figure 3 The coefficient matrix between the input layer and the first layer nodes of the hidden layer of the model is the characteristic encoding of the enterprise in the relationship between enterprises and people.

[0081] After obtaining the above feature encoding, Figure 4 As shown, different feature codes based on pagerank, labelpropagation and Deepwalk can be "stitched" together to form a feature code that represents the enterprise knowledge graph.

[0082] The above embodiments respectively describe in detail the feature encoding of character data and knowledge graph data, but the present invention is not limited thereto. In one embodiment of the present invention, the target data may include character data and knowledge graph data, and then in step S12, the target data is encoded according to the adjacent relationship between the target data to obtain the target feature encoding, which may specifically include:

[0083] According to the adjacent relationship between the target data, the character data and the knowledge graph data are respectively vectorized to form character feature coding and graph feature coding;

[0084] The character feature code and the graph feature code are concatenated to form the target feature code.

[0085] That is to say, after obtaining the feature coding of character data and the feature coding of knowledge graph data, these two feature codings can be spliced ​​together to form the target feature coding.

[0086] Furthermore, the target feature code is a risk feature code for an enterprise at the same time node. When the influence of time factors on risk features is taken into consideration, the target feature code can be further processed. Specifically, in step S14, the target feature codes are grouped in time series according to the generation time of the target data, and the time evolution model between each group of target feature codes is trained, which may include:

[0087] Grouping the target feature codes corresponding to the target data generated within the same preset time period into one group;

[0088] The time evolution model is trained by taking N-1 groups of target feature codes corresponding to the first N-1 preset time periods as input and taking the target feature code corresponding to the Nth preset time period as output, wherein N is an integer greater than 1.

[0089] For example, in one embodiment of the present invention, the characteristic code of enterprise i at time j is x ij , take the time interval n, where n is a hyperparameter. Combined with the data change characteristics of enterprise risks, n is 1 week, which not only ensures that the enterprise data has certain changes in different periods, but also creates more training samples compared to one month.

[0090] The time series of enterprise i is cut into m-n+1 samples according to n:

[0091] (x i1 , x i2 , …, x in-1 )->(x in )

[0092] (x i2 , x i3 , …, x in )->(x in+1 )

[0093] …

[0094] (x im-n+1 , x i3 , …, x im-1 )->(x im )

[0095] The data of all enterprises are combined together to form a training sample, and a DNN model is trained.

[0096] After training, the shared weights of the input layer and the hidden layer are multiplied by the input layer as the final feature encoding input.

[0097] To verify the effectiveness, in one embodiment of the present invention, a simple attempt is made on a one-dimensional time series data set constructed as follows, and the effectiveness of the method can be seen.

[0098] Let X n =0.5*X n-1 +0.2*X n-2

[0099] Here X n It can be regarded as a one-dimensional feature encoding of the enterprise, taking the time slice parameter as 2, that is, by constructing X n About X n-2 and X n-1 The deep learning model is used to learn the feature encoding method. In this simple time evolution formula, if the hidden layer features can learn the time evolution law, that is, the expression X n =0.5*X n-1 +0.2*X n-2 , it means that the hidden layer features can well represent the time evolution law. When n is large enough (100,000 data points), and the hidden layer feature is a neuron, we can see that the feature and X n-1 , X n-2 The link layer coefficients are 0.4899 and 0.1998, that is, X n =0.4899*X n-1 +0.1998*X n-2 , which fully demonstrates that the feature encoding can effectively learn the law of time evolution.

[0100] To verify the validity of the model, the embodiment of the present invention selected a total of 430,430 data of corporate loan applications in history, of which a total of 5,209 loans were eventually overdue, and the overall loan overdue rate was: 5,209 / 430,430=1.21%.

[0101] Two versions of the model were established using the above sample data. Both the Baseline-Model and the 4P-Model are based on the machine learning gbdt algorithm. The difference between the two is that the features of the Baseline-Model come from the summary of business experts, and this method heavily relies on the modeler's personal experience and understanding of the business scenario; while the 4P-Model, in addition to adding the experience features summarized by business experts, also uses a large number of embedded features generated by the deep learning-based business management risk coding technology.

[0102] Based on the above two versions of the model, the two versions of the model were applied to the experimental set data for offline verification. The comparison of the AUC, precision and recall rate of the final experimental results is shown in Table 1 and Table 2 respectively:

[0103] Table 1

[0104] Model version AUC KS LIFT Baseline 0.909 0.647 12.976 4PModel 0.846 0.55 13.895

[0105] Table 2

[0106]

[0107] It can be seen from the above two tables that by combining the embedded features generated by deep learning management risk coding technology in the GBDT algorithm, the performance indicators of the model can be significantly improved, thereby enabling the model to have a stronger ability to distinguish between the risks and debt repayment results of the enterprise.

[0108] Taking the experimental data in Table 2 as an example, if risk intervention is also carried out on the top 30% of corporate users with the highest risk - that is, the loan applications are rejected, the GBDT model based on traditional feature technology can intercept 77.1% of risky customers, while the 4P-Model, due to its stronger risk identification ability, can intercept 90.48% of risky customers. This means that when 30% of applications are intercepted, the experimental group can recall 13.4% compared with the control group - the experimental set involves nearly 700 defaulting customers. Based on the average loan of 500,000 yuan per small and micro customer, the model can intercept more than 300 million potential overdue loans for small and micro enterprises for banks.

[0109] In addition, the number of risky customers identified through the 4P-Model, among the top 20% of customers, is almost close to the number of risky customers recalled through the traditional model for the top 30% of customers. This means that with the risk indicators remaining almost unchanged, the 4P-Model can approve nearly 10% more customer loan applications (about 40,000 applications). Based on an average loan amount of 500,000 yuan per small and micro customer, the model can directly bring an additional loan size of about 20 billion yuan to the bank.

[0110] On the second aspect, an embodiment of the present invention also provides a modeling device for enterprise management risk identification, which is capable of encoding features related to enterprise risks and using the encoded features for model training, thereby greatly reducing the workload and blindness of feature extraction and effectively improving the accuracy of model prediction.

[0111] like Figure 5 As shown, the modeling device for enterprise operation and management risk identification provided by the embodiment of the present invention may specifically include:

[0112] An acquisition unit 31 is used to acquire target data from an enterprise risk data source, wherein the target data includes character data and / or knowledge graph data;

[0113] The encoding unit 32 is used to encode the target data according to the adjacent relationship between the target data to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold;

[0114] The first training unit 33 is used to group the target feature codes in time series according to the generation time of the target data, and train the time evolution model between each group of target feature codes; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature code;

[0115] The second training unit 34 is used to perform model training with the features of the time series feature encoding as input and the probability of enterprise risk occurrence as output, so as to obtain a model for enterprise operation and management risk identification.

[0116] The modeling device for enterprise management risk identification provided by the embodiment of the present invention can obtain target data from the enterprise risk data source, encode the target data according to the adjacent relationship between the target data, and obtain the target feature code. Since the target feature code can represent each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than the preset threshold, character data and / or knowledge graph data can be uniformly represented by a set of dense multidimensional vectors, which is conducive to feature extraction and model training. Then, according to the generation time of the target data, the target feature code is time-series grouped, and the time evolution model between each group of target feature codes is trained to obtain the time-series feature code of the target data, thereby further introducing the time attribute of the enterprise risk into the model system, thereby effectively improving the accuracy of the model prediction.

[0117] Optionally, the acquisition unit 31 may include: a first acquisition module, used to perform word segmentation and key information extraction on the unstructured data in the enterprise risk data source, so that each of the key information forms a character-type data; and / or a second acquisition module, used to convert the structured data in the enterprise risk data source into knowledge graph-type data.

[0118] Optionally, the target data includes character data; the encoding unit 32 may include: a construction module, used to construct the adjacent relationship between each character data according to a preset rule; a vectorization module, used to vectorize the character data according to the adjacent relationship to form the target feature code.

[0119] Optionally, the construction module is specifically used to construct the adjacent relationship between people in each type of position in each enterprise according to preset rules; the vectorization module is specifically used to: vectorize the identity identifiers of people in each type of position and calculate the average value to obtain the average identity feature code of such position; splice the average identity feature codes of each type of position together to form the target feature code.

[0120] Optionally, the target data includes knowledge graph data; the encoding unit includes: a first encoding module, used to determine the influence feature coding of each enterprise by using the pagerank algorithm based on the relationship between enterprises in the knowledge graph data; a second encoding module, used to determine the risk feature coding of each enterprise by using the label propagation algorithm based on the relationship between enterprises in the knowledge graph data; a third encoding module, used to determine the feature coding related to people in each enterprise by using the deepwalk algorithm based on the relationship between enterprises and people in the knowledge graph data; a splicing module, used to splice the influence feature coding, the risk feature coding and the feature coding related to people to form the target feature coding.

[0121] Optionally, the first training unit is specifically used to: group the target feature codes corresponding to the target data generated within the same preset time period into a group; use the N-1 groups of target feature codes corresponding to the first N-1 preset time periods as input and the target feature code corresponding to the Nth preset time period as output to train the time evolution model, wherein N is an integer greater than 1.

[0122] Optionally, the target data includes character data and knowledge graph data; the encoding unit is specifically used to: vectorize the character data and the knowledge graph data respectively according to the adjacent relationship between the target data to form character feature codes and graph feature codes; and splice the character feature codes and the graph feature codes to form the target feature codes.

[0123] like Figure 6 As shown, the electronic device provided by an embodiment of the present invention may include: a shell 51, a processor 52, a memory 53, a circuit board 54 and a power circuit 55, wherein the circuit board 54 is arranged inside the space enclosed by the shell 51, and the processor 52 and the memory 53 are arranged on the circuit board 54; the power circuit 55 is used to supply power to various circuits or devices of the above-mentioned electronic device; the memory 53 is used to store executable program codes; the processor 52 runs the program corresponding to the executable program code by reading the executable program code stored in the memory 53, so as to execute the modeling method for enterprise management risk identification provided by any of the aforementioned embodiments.

[0124] The specific execution process of the above steps by the processor 52 and the steps further executed by the processor 52 by running the executable program code can be found in the description of the previous embodiment, which will not be repeated here.

[0125] The above electronic devices exist in many forms, including but not limited to:

[0126] (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (such as iPhone), multimedia phones, functional phones, and low-end phones.

[0127] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing capabilities, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0128] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0129] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0130] (5) Other electronic devices with data interaction functions.

[0131] Correspondingly, an embodiment of the present invention also provides a computer-readable storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement any one of the modeling methods for enterprise management risk identification provided by the aforementioned embodiments, thereby also being able to achieve the corresponding technical effects, which have been described in detail above and will not be repeated here.

[0132] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0133] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0134] In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0135] For the convenience of description, the above device is described by dividing the functions into various units / modules. Of course, when implementing the present invention, the functions of each unit / module can be implemented in the same or multiple software and / or hardware.

[0136] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0137] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A modeling method for identifying business management risks, characterized in that: include: Acquire target data from an enterprise risk data source, wherein the target data includes character data and / or knowledge graph data; According to the adjacent relationship between the target data, the target data is encoded to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold; According to the generation time of the target data, the target feature codes are grouped in time series, and a time evolution model between each group of target feature codes is trained; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature code; Taking the features of the time series feature encoding as input and the probability of enterprise risk occurrence as output, model training is performed to obtain a model for enterprise operation and management risk identification; The target data includes character data; encoding the target data according to the adjacent relationship between the target data to obtain the target feature code includes: constructing the adjacent relationship between each character data according to a preset rule; vectorizing the character data according to the adjacent relationship to form the target feature code; and / or The target data includes knowledge graph data; encoding the target data according to the adjacent relationship between the target data to obtain the target feature code includes: based on the relationship between enterprises in the knowledge graph data, using the pagerank algorithm to determine the influence feature code of each enterprise; based on the relationship between enterprises in the knowledge graph data, using the label propagation algorithm to determine the risk feature code of each enterprise; based on the relationship between enterprises in the knowledge graph data, using the deepwalk algorithm to determine the feature code related to people in each enterprise; splicing the influence feature code, the risk feature code and the feature code related to people to form the target feature code.

2. The method according to claim 1, characterized in that The obtaining of target data from the enterprise risk data source includes: Perform word segmentation and key information extraction on the unstructured data in the enterprise risk data source, so that each key information forms a character data; and / or Convert structured data in enterprise risk data sources into knowledge graph data.

3. The method according to claim 1, characterized in that The step of constructing the adjacent relationship between the character data according to the preset rules includes: According to the preset rules, construct the adjacent relationship between people in each position in each enterprise; The vectorizing the character data according to the adjacent relationship to form the target feature code includes: The identity identifiers of people in each type of position are vectorized and averaged to obtain the average identity feature encoding of this type of position; The average identity characteristic codes of various positions are concatenated to form the target characteristic code.

4. The method according to claim 1, characterized in that: The step of grouping the target feature codes in time series according to the generation time of the target data and training the time evolution model between each group of target feature codes includes: Grouping the target feature codes corresponding to the target data generated within the same preset time period into one group; The time evolution model is trained by taking N-1 groups of target feature codes corresponding to the first N-1 preset time periods as input and taking the target feature code corresponding to the Nth preset time period as output, wherein N is an integer greater than 1.

5. The method according to any one of claims 1 to 4, characterized in that The target data includes character data and knowledge graph data; The step of encoding the target data according to the adjacent relationship between the target data to obtain the target feature code comprises: According to the adjacent relationship between the target data, the character data and the knowledge graph data are respectively vectorized to form character feature coding and graph feature coding; The character feature code and the graph feature code are concatenated to form the target feature code.

6. A modeling device for identifying business management risks, characterized in that: include: An acquisition unit, used to acquire target data from an enterprise risk data source, wherein the target data includes character data and / or knowledge graph data; An encoding unit, used for encoding the target data according to the adjacent relationship between the target data to obtain a target feature code; wherein the target feature code represents each target data with a multidimensional vector, and the number of dimensions of the multidimensional vector is less than a preset threshold; A first training unit is used to group the target feature codes in time series according to the generation time of the target data, and train a time evolution model between each group of target feature codes; wherein the coefficient matrix between the input layer nodes and the first layer nodes of the hidden layer of the time evolution model is the time series feature code; A second training unit is used to perform model training using the features of the time series feature encoding as input and the probability of enterprise risk occurrence as output, so as to obtain a model for enterprise operation and management risk identification; The target data includes character data; the encoding unit includes: a construction module, which is used to construct the adjacent relationship between each character data according to a preset rule; a vectorization module, which is used to vectorize the character data according to the adjacent relationship to form the target feature code; and / or The target data includes knowledge graph data; the encoding unit includes: a first encoding module, which is used to determine the influence feature coding of each enterprise by using the pagerank algorithm based on the relationship between enterprises in the knowledge graph data; a second encoding module, which is used to determine the risk feature coding of each enterprise by using the labelpropagation algorithm based on the relationship between enterprises in the knowledge graph data; a third encoding module, which is used to determine the feature coding related to people in each enterprise by using the deepwalk algorithm based on the relationship between enterprises and people in the knowledge graph data; a splicing module, which is used to splice the influence feature coding, the risk feature coding and the feature coding related to people to form the target feature coding.

7. The device according to claim 6, characterized in that The acquisition unit comprises: The first acquisition module is used to perform word segmentation and key information extraction on the unstructured data in the enterprise risk data source, so that each key information forms a character data; and / or The second acquisition module is used to convert the structured data in the enterprise risk data source into knowledge graph data.

8. The device according to claim 6, characterized in that The construction module is specifically used to construct the adjacent relationship between people of each type of position in each enterprise according to preset rules; The vectorization module is specifically used for: The identity identifiers of people in each type of position are vectorized and averaged to obtain the average identity feature encoding of this type of position; The average identity characteristic codes of various positions are concatenated to form the target characteristic code.

9. The device according to claim 6, characterized in that The first training unit is specifically used for: Grouping the target feature codes corresponding to the target data generated within the same preset time period into one group; The time evolution model is trained by taking N-1 groups of target feature codes corresponding to the first N-1 preset time periods as input and taking the target feature code corresponding to the Nth preset time period as output, wherein N is an integer greater than 1.

10. The device according to any one of claims 6 to 9, characterized in that The target data includes character data and knowledge graph data; The encoding unit is specifically used for: According to the adjacent relationship between the target data, the character data and the knowledge graph data are respectively vectorized to form character feature coding and graph feature coding; The character feature code and the graph feature code are concatenated to form the target feature code.

11. An electronic device, characterized in that: The electronic device comprises: a shell, a processor, a memory, a circuit board and a power supply circuit, wherein the circuit board is placed inside the space enclosed by the shell, and the processor and the memory are arranged on the circuit board; the power supply circuit is used to supply power to various circuits or devices of the above-mentioned electronic device; the memory is used to store executable program code; the processor runs the program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute the modeling method for enterprise management risk identification described in any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the modeling method for enterprise management risk identification described in any one of claims 1 to 5.

Images