Disclosure of Invention
The application provides an enterprise risk assessment method and related equipment based on a dynamic graph neural network, which can solve the problem of low accuracy of enterprise risk assessment.
In a first aspect, the present application provides an enterprise risk assessment method based on a dynamic graph neural network, where the enterprise risk assessment method includes:
acquiring relationship data of a plurality of target enterprises, wherein the relationship data of each target enterprise is used for describing enterprise relationships between the target enterprise and other target enterprises, and the enterprise relationships belong to one of a plurality of relationships;
The method comprises the steps of establishing a multi-layer enterprise social network based on all relationship data, wherein the multi-layer enterprise social network corresponds to various relationships one by one, a plurality of nodes in the multi-layer enterprise social network correspond to a plurality of target enterprises one by one, edges between two nodes in the same layer are enterprise relationships between the two corresponding target enterprises, and edges between two nodes in different layers are influence relationships between the two corresponding target enterprises;
the method comprises the steps of carrying out feature extraction on a multi-layer enterprise social network by utilizing a dynamic graph neural network to obtain short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network, wherein the short-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a short time span, the long-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a long time span, and the topology feature embedding is used for describing topology structure information of the multi-layer enterprise social network;
obtaining a persistence feature based on a multi-layer enterprise social network, and fusing the persistence feature, short-term feature embedding, long-term feature embedding and topology feature embedding to obtain a fusion feature, wherein the persistence feature is used for describing the stability degree of each side in the multi-layer enterprise social network;
And carrying out risk assessment on each target enterprise according to the fusion characteristics and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise, wherein the risk assessment result is used for describing the risk degree of the target enterprise.
Optionally, constructing the multi-layer enterprise social network based on all the relationship data includes:
Respectively aiming at each relation, taking the target enterprises with the relation in the relation data of all target enterprises as hierarchical target enterprises of the relation, and generating a single-layer enterprise social network of the relation according to the relation data of all hierarchical target enterprises;
And calculating influence relations between every two different single-layer enterprise social networks, and generating edges between nodes in the different single-layer enterprise social networks based on all the influence relations to obtain the multi-layer enterprise social network.
Optionally, fusing the persistent feature, the short-term feature embedding, the long-term feature embedding, and the topology feature embedding to obtain a fused feature, including:
embedding short-term features and persistent features to perform feature fusion to obtain first features;
Embedding and fusing the first characteristic and the long-term characteristic to obtain a second characteristic;
And embedding and integrating the second characteristic and the topological characteristic to obtain a fusion characteristic.
Optionally, feature fusion is performed on the short-term feature embedding and the persistent feature to obtain a first feature, including:
by the formula:
Calculating a first characteristic;
Wherein,Representing the dynamic attention weight of the person,Represent the firstA first node characteristic of the individual nodes,Representing short term feature embedding inThe short-term characteristics of the individual nodes,The characteristic of the persistence is indicated by the fact that,A scoring function is represented as a function of the score,Representing short term feature embedding inThe short-term characteristics of the individual nodes,,Representing nodesIs used to determine the neighbor node of a node (a),Representing a collection of nodes in a multi-tier enterprise social network.
Optionally, the first feature and the long-term feature are embedded and fused to obtain a second feature, including:
by the formula:
calculating a second characteristic;
Wherein,Represent the firstA second node characteristic of the individual node,Representing the first in a multi-tier enterprise social networkThe individual nodeThe weight between the individual nodes is such that,Representing the first in a multi-tier enterprise social networkThe neighboring nodes of the individual nodes are referred to as,Representing the enhanced long-term feature embedding,Representing long term feature embedding inThe components corresponding to the individual nodes are selected,Representing long term feature embedding inComponents corresponding to the individual nodes;
Embedding and integrating the second feature and the topological feature to obtain a fusion feature, wherein the method comprises the following steps:
by the formula:
computing fusion features;
Wherein,A neural network of the graph is shown,The weight is represented by a weight that,Representing the characteristics of the enhanced topology,Representing the first of the topological featuresThe components corresponding to the individual nodes are selected,Representing the first of the topological featuresComponents corresponding to the individual nodes.
Optionally, performing risk assessment on each target enterprise according to the fusion feature and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise, including:
the following steps are respectively carried out for each target enterprise:
calculating the centrality score of the node corresponding to the target enterprise according to the multi-layer enterprise social network;
and performing risk assessment on the target enterprise based on the centrality score and the fusion characteristic to obtain a risk assessment result of the target enterprise.
Optionally, after the steps of obtaining the persistent feature based on the multi-layer enterprise social network, and fusing the persistent feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain the fused feature, the enterprise risk assessment method further includes:
and calculating a comprehensive opportunity score of each target enterprise based on the multi-layer enterprise social network and the persistence features for each target enterprise, wherein the comprehensive opportunity score is used for describing the probability that the target enterprise has a potential development opportunity.
In a second aspect, the present application provides an enterprise risk assessment apparatus based on a dynamic graph neural network, including:
The system comprises an acquisition module, a storage module and a storage module, wherein the acquisition module is used for acquiring relationship data of a plurality of target enterprises, and the relationship data of each target enterprise is used for describing enterprise relationships between the target enterprise and other target enterprises, and the enterprise relationships belong to one of a plurality of relationships;
the system comprises a construction module, a multi-layer enterprise social network, a plurality of target enterprises, a relation module and a relation module, wherein the construction module is used for constructing a multi-layer enterprise social network based on all relation data;
The system comprises a feature extraction module, a feature embedding module, a feature extraction module and a feature extraction module, wherein the feature extraction module is used for carrying out feature extraction on a multi-layer enterprise social network by utilizing a dynamic graph neural network to obtain short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network, wherein the short-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a short time span, the long-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a long time span, and the topology feature embedding is used for describing topology structure information of the multi-layer enterprise social network;
the fusion module is used for acquiring the durability characteristic based on the multi-layer enterprise social network, and fusing the durability characteristic, the short-term characteristic embedding, the long-term characteristic embedding and the topology characteristic embedding to obtain the fusion characteristic, wherein the durability characteristic is used for describing the stability degree of each side in the multi-layer enterprise social network;
The risk assessment module is used for carrying out risk assessment on each target enterprise according to the fusion characteristics and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise, and the risk assessment result is used for describing the risk degree of the target enterprise.
In a third aspect, an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, where the processor implements the method for evaluating an enterprise risk based on a dynamic graph neural network when the processor executes the computer program.
In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, where a computer program is stored, where the computer program is executed by a processor to implement the enterprise risk assessment method based on a dynamic graph neural network.
The scheme of the application has the following beneficial effects:
In the embodiment of the application, the multi-layer enterprise social network is constructed based on the relationship data of a plurality of target enterprises, the dynamic graph neural network is utilized to conduct feature extraction on the multi-layer enterprise social network, short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network are obtained, the persistent features are obtained based on the multi-layer enterprise social network, the persistent features, the short-term feature embedding, the long-term feature embedding and the topology feature embedding are fused to obtain fusion features, and finally risk assessment is conducted on each target enterprise according to the fusion features and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise. The method comprises the steps of constructing a multi-layer enterprise social network, fully expressing the association relation between target enterprises, realizing accurate modeling of the enterprise relation, fusing persistent features, short-term feature embedding, long-term feature embedding and topology feature embedding to obtain fusion features, improving the information richness and comprehensiveness of the fusion features, performing risk assessment based on the information rich fusion features and the accurate multi-layer enterprise social network, and effectively improving the accuracy of enterprise risk assessment.
Other advantageous effects of the present application will be described in detail in the detailed description section which follows.
Detailed Description
In the following description, for purposes of explanation and not limitation, specific details are set forth such as the particular system architecture, techniques, etc., in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.
It should be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.
It should also be understood that the term "and/or" as used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations.
As used in the present description and the appended claims, the term "if" may be interpreted as "when..once" or "in response to a determination" or "in response to detection" depending on the context. Similarly, the phrase "if a determination" or "if a [ described condition or event ] is detected" may be interpreted in the context of meaning "upon determination" or "in response to determination" or "upon detection of a [ described condition or event ]" or "in response to detection of a [ described condition or event ]".
Furthermore, the terms "first," "second," "third," and the like in the description of the present specification and in the appended claims, are used for distinguishing between descriptions and not necessarily for indicating or implying a relative importance.
Reference in the specification to "one embodiment" or "some embodiments" or the like means that a particular feature, structure, or characteristic described in connection with the embodiment is included in one or more embodiments of the application. Thus, appearances of the phrases "in one embodiment," "in some embodiments," "in other embodiments," and the like in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments" unless expressly specified otherwise. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless expressly specified otherwise.
Aiming at the problem of low accuracy of the existing enterprise risk assessment, the embodiment of the application provides an enterprise risk assessment method based on a dynamic graph neural network, which comprises the steps of acquiring the relationship data of a plurality of target enterprises, constructing a multi-layer enterprise social network based on all the relationship data, extracting the characteristics of the multi-layer enterprise social network by using the dynamic graph neural network, obtaining short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network, obtaining persistent features based on the multi-layer enterprise social network, fusing the persistent features, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain fusion features, and finally performing risk assessment on each target enterprise according to the fusion features and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise. The method comprises the steps of constructing a multi-layer enterprise social network, fully expressing the association relation between target enterprises, realizing accurate modeling of the enterprise relation, fusing persistent features, short-term feature embedding, long-term feature embedding and topology feature embedding to obtain fusion features, improving the information richness and comprehensiveness of the fusion features, performing risk assessment based on the information rich fusion features and the accurate multi-layer enterprise social network, and effectively improving the accuracy of enterprise risk assessment.
The enterprise risk assessment method based on the dynamic graph neural network provided by the application is exemplified as follows.
As shown in fig. 1, the enterprise risk assessment method based on the dynamic graph neural network provided by the application comprises the following steps:
and step 11, acquiring relationship data of a plurality of target enterprises.
The relationship data of each target enterprise is used for describing an enterprise relationship (the relationship data comprises the enterprise relationship between the target enterprise and at least one other target enterprise) between the target enterprise and other target enterprises, the enterprise relationship (such as project cooperation, result sharing, market competition, supply chain relationship, son-mother company, same board, same stakeholder and the like) belongs to one of various relationships (such as cooperation relationship, supply relationship, competition relationship, affiliation and association relationship), such as project cooperation and result sharing, belongs to cooperation relationship, supply chain relationship belongs to supply relationship, market competition belongs to competition relationship, son-mother formula belongs to affiliation, same board and same stakeholder belongs to association relationship.
It should be noted that the relationship data further includes time information of the enterprise relationship between the target enterprises, and can describe the generation time and the ending time of the relationship between the enterprises, for example, for the enterprise a, the relationship data of the enterprise a is that the relationship between the enterprise B and the 2016-2017 is a cooperative relationship, and the relationship between the enterprise a and the enterprise B is an affiliated relationship since 2018.
In some embodiments of the present application, the relationship data of the target enterprise may be obtained by accessing a website that exposes enterprise relationship data, such as enterprise search, sky eye search, etc.
And step 12, constructing a multi-layer enterprise social network based on all the relationship data.
The multi-layer enterprise social network corresponds to the multiple relationships one by one, the multiple nodes in the multi-layer enterprise social network correspond to the multiple target enterprises one by one, edges between two nodes in the same layer are enterprise relationships between the corresponding two target enterprises, edges between two nodes in different layers are influence relationships between the corresponding two target enterprises (the influence relationships between the enterprise relationships of the nodes in the layer and the enterprise relationships of the nodes in the other layer are described, for example, the supply chain relationships tend to promote the cooperative relationships, and the competing relationships tend to block the cooperative relationships). Each node has a corresponding set of characteristics for the target business including basic characteristics (business name, industry type, number of employees, location), financial status (revenue, profit, total assets), operational status (recent financing, credit rating, market share), etc.
In an embodiment of the present application, the step of constructing the multi-layer social network for enterprises based on all the relationship data includes:
The first step is to take the target enterprises with the relation in the relation data of all target enterprises as the hierarchical target enterprises of the relation for each relation, and generate a single-layer enterprise social network of the relation according to the relation data of all hierarchical target enterprises.
Specifically, if an enterprise relationship corresponding to the relationship exists between the two hierarchical target enterprises, an edge is generated.
It should be noted that edges in a single-layer enterprise social network reflect direct relationships between enterprises at the same level.
For example, for the cooperative relationship, in the enterprises a, B, C and D, the enterprise a, B and C have the cooperative relationship, and the enterprise C and D have the cooperative relationship, so in the single-layer enterprise social network for the cooperative relationship, an edge is formed between the node corresponding to the enterprise a and the node corresponding to the enterprise B, and between the node corresponding to the enterprise C and the node corresponding to the enterprise D.
After the edges are generated, the attributes of each edge are attached with a timestamp in the corresponding relationship data so as to embody the time information. While the attributes of the edges include enterprise relationships and relationship strengths (e.g., collaboration times, dependencies, etc.) of the corresponding two target enterprises.
And secondly, calculating influence relations between every two different single-layer enterprise social networks, and generating edges between nodes in the different single-layer enterprise social networks based on all the influence relations to obtain the multi-layer enterprise social network.
It should be noted that edges between nodes in different single-tier enterprise social networks reflect the interleaving and interaction of relationships between different tiers, revealing that tier-to-tier interactions, such as supply chain tier dependencies, may affect the risk intensity of the cooperating tiers.
For example, the influence relationship may be obtained by calculating indexes such as similarity (which may be calculated by a calculation formula such as cosine similarity) between the single-layer social networks of enterprises, and for analyzing the mutual influence between the single-layer social networks of enterprises in the cooperative relationship and the single-layer social networks of enterprises in the feed relationship, the similarity indexes (such as cosine similarity) of the two networks are calculated first. If the similarity is greater than the set threshold, then a significant impact relationship is considered to exist between the two networks. On the basis, cross-layer edges are further generated for related nodes in the two networks according to the relation strength, interaction frequency or time dependency. The weight of the cross-layer edge can be quantified according to the similarity value, the time attenuation factor or the interaction strength and the like so as to reflect the strength of the influence relationship between the two layers of networks. And generate cross-layer association matrixWhereinThe cross-layer influence relationship and the influence degree are quantized, and how the relationship of the target enterprises vi and vj in the supply layer l affects the relationship in the cooperation layer m is expressed.
It should be noted that, in the multi-layer enterprise social network, each target enterprise only corresponds to one node, and for target enterprises with various relationships (all existing in a plurality of single-layer enterprise social networks), after the multi-layer enterprise social network is formed, only one node is reserved, and the nodes at other layers are equivalent to the mapping of the node, so that the structure analysis and the display of the multi-layer enterprise social network are facilitated. If the enterprise a is in a cooperative relationship with two target enterprises and is in a competitive relationship with one target enterprise, in the multi-layer enterprise social network, the node corresponding to the enterprise a should have two edges with the attribute of the cooperative relationship and one edge with the attribute of the competitive relationship, instead of the node corresponding to the enterprise a in the cooperative layer having two edges with the attribute of the cooperative relationship, the other node corresponding to the enterprise a in the competitive layer has one edge with the attribute of the cooperative relationship.
Because the edges between the interlayer nodes are provided with the time stamps, when the multi-layer enterprise social network is analyzed, the time of the relationship in the multi-layer enterprise social network is dynamically updated according to the time stamps, and if the relationship of the target enterprise between 2015 and 2017 needs to be analyzed, the edges with the time stamps ending before 2015 and starting after 2017 are marked as invalid.
It is worth mentioning that the multi-layer enterprise social network is constructed, the association relation between target enterprises can be fully expressed, accurate modeling of the enterprise relation is achieved, and time-dependent analysis of the relation between enterprises is facilitated.
The multi-tiered social network is illustrated below in connection with a specific example.
The multi-layer enterprise social network is shown in fig. 2, and comprises five layers of networks, namely a cooperation layer, a supply layer, a competition layer, an affiliated layer and an association layer, A, B, C is the number of the nodes, a solid line represents edges between the nodes in the layers, and a dotted line represents edges between the nodes in the layers.
And 13, performing feature extraction on the multi-layer enterprise social network by using the dynamic graph neural network to obtain short-term feature embedding, long-term feature embedding and topological feature embedding of the multi-layer enterprise social network.
The short-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a short time span, the long-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a long time span, and the topology feature embedding is used for describing topology structure information of the multi-layer enterprise social network. The short time span and the long time span are set according to the actual period requirement of the enterprise risk assessment, and the short time span is smaller than the long time span, for example, the short time span is a quarter, and the long time span can be one year.
The dynamic graph neural network is composed of a time sequence graph sub-network (TGN, temporal Graph Networks), a double graph attention sub-network (DGAT, dual Graph Attention Network) and a graph rolling neural sub-network (GCN, graph Convolutional Network) which are connected in sequence, wherein the input end of the time sequence graph sub-network is the input end of the dynamic graph neural network, the output end of the time sequence graph sub-network outputs short-term characteristic embedding, the output end of the double graph attention sub-network outputs long-term characteristic embedding, and the output end of the graph rolling neural sub-network outputs topological characteristic embedding.
Specifically, the multi-layer enterprise social network is input into the dynamic graph neural network, and short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network are obtained.
And 14, obtaining persistent features based on the multi-layer enterprise social network, and fusing the persistent features, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain fusion features.
The persistence features described above are used to describe the degree of stability of each edge in a multi-layer enterprise social network.
In some embodiments of the present application, the steps of obtaining the persistent feature based on the multi-layer social network of the enterprise, and fusing the persistent feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain the fused feature include:
first, a persistence feature is obtained based on a multi-layer enterprise social network.
By way of example, a persistent coherent method may be used to perform topology analysis on a multi-layer enterprise social network to obtain a persistent barcode, and then an embedding algorithm Embedding may be used to convert the persistent barcode into a persistent feature, so that the persistent barcode and other features may be embedded in the same feature space to perform fusion, and the persistent feature is ensured to embody its stability in the fusion.
The continuous coherent method extracts topological structure characteristics, such as connected branches and rings, in the multi-layer enterprise social network by constructing Rips complex shapes (used for describing the topological structure of connection between nodes in the multi-layer enterprise social network) under different thresholds so as to reveal potential risks and opportunities in the network. The selection of the distance threshold epsilon is typically based on the characteristics of the data and the analysis targets, typically determined by node characteristics. For example, in a multi-tier enterprise social network analysis, ε may be set to the average strength of interactions between nodes, and 75% or 90% quantiles of the average strength may be taken as initial thresholds to ensure that the primary structural features are captured, and experiments are performed using different thresholds ε to find the threshold ranges under which the topological features are most stable. At the threshold epsilon a Rips complex of the network is generated. This step is repeated at different epsilon values, capturing the generation and disappearance of each topology (e.g., ring, hole, etc.) at different thresholds. While quantifying the duration of the topology by the persistent barcode. The length of the persistent bar code reflects the stability of the topological feature, the bar code with high persistence represents a long-term relationship, and the bar code with low persistence points to a short-term or unstable relationship and is used for analyzing a key relationship mode in the multi-layer enterprise social network.
For each topology that is a hole (holes may be considered missing connections or structures in the network, their presence may indicate a potential risk or opportunity, e.g., holes may represent potential points of supply relationship break in a supply relationship; holes may represent splits or disparities in a community in a social network), their birth and death times at different thresholds ε are recorded, which may be automatically calculated by a persistent coherent software package (e.g., gudhi, dionysus, etc.). The persistence measure is the difference between the death time and the birth time of the hole, expressed asWhere di is the death time of the hole and bi is the birth time of the hole.
The persistence metric may be used to risk and opportunistic identify holes, e.g., set a risk thresholdA persistence metric value of 1.0, above which is a high risk hole, an opportunity threshold0.5, Below which is a potential opportunity hole.
And secondly, embedding short-term features and fusing persistent features to obtain first features.
Specifically, the formula is as follows:
Calculating a first characteristic;
Wherein,Representing the dynamic attention weight of the person,Represent the firstA first node characteristic of the individual nodes,Representing short term feature embedding inThe short-term characteristics of the individual nodes,The characteristic of the persistence is indicated by the fact that,A scoring function is represented as a function of the score,Representing short term feature embedding inThe short-term characteristics of the individual nodes,,Representing nodesIs used to determine the neighbor node of a node (a),Representing a collection of nodes in a multi-tier enterprise social network.
It should be noted that a attention weight is defined in the above formulaThe balance of time dynamic information and network structure stability is captured by calculating the attention weights of short term feature embedding and persistent features in each time step. Short term feature embedding is used to capture short term dynamics of relationships, while persistent features reveal more stable long term relationships in the network. At each time step, according to the dynamic attention weight, a scoring function of short-term feature embedding and persistence features is calculatedAnd the fusion weight is adaptively adjusted, so that the fusion of the short-term and long-term characteristics is realized, and the dynamic change of the network is dealt with.
Before this step is performed, it is necessary to perform time-step alignment and preprocessing on the persistent feature, the short-term feature embedding, the long-term feature embedding, and the topology feature embedding, which specifically includes:
(1) The method comprises the steps of time step alignment, namely firstly, arranging a persistence feature, a short-term feature embedding, a long-term feature embedding and a topology feature embedding according to time stamps to determine a unified time scale, secondly, aiming at different features, adopting the same time scale to ensure that all the features are aligned on the same time step so as to capture potential risks in a dynamic network, and finally, using linear interpolation to process missing values to ensure that effective data exists in each time step.
(2) And selecting an adaptive time window, namely performing adaptive adjustment on the time window of each characteristic according to the dynamic change condition of the enterprise relationship. For example, for short term feature embedding, a shorter time window (1-3 days) is selected to capture rapid changes, and for topological feature embedding and persistent features, a longer time window (7-30 days) is selected to reflect more stable structural features.
(3) Embedding normalization three feature embedding was normalized separately first using Z-score normalization. So as to ensure that the characteristics are on the same scale and eliminate the influence caused by the output difference of different models. The formula for Z-score normalization is:
Wherein,The data after the normalization is represented and,The data before the normalization is represented by the data,The standard deviation is indicated as such,Representing the mean.
And thirdly, embedding and fusing the first characteristic and the long-term characteristic to obtain a second characteristic.
Specifically, the formula is as follows:
calculating a second characteristic。
Wherein,Represent the firstA second node characteristic of the individual node,Representing the first in a multi-tier enterprise social networkThe individual nodeThe weight between the individual nodes is such that,Representing the first in a multi-tier enterprise social networkThe neighboring nodes of the individual nodes are referred to as,Representing the enhanced long-term feature embedding,Representing long term feature embedding inThe components corresponding to the individual nodes are selected,Representing long term feature embedding inComponents corresponding to the individual nodes.
It should be noted that, based on the first feature, long-term feature embedding is used to further focus on relationships and inter-layer interactions between nodes. By weightThe first feature and the long-term feature embedding are weighted and updated to ensure that the features associated with the layers are properly focused. For example, if there is a strong association between the provisioning layer and the collaboration layer, the associated cross-layer information is preferentially focused by the attention mechanism in the long-term feature embedding.
And fourthly, embedding and integrating the second characteristic and the topological characteristic to obtain a fusion characteristic.
Specifically, the formula is as follows:
computing fusion features。
Wherein,A neural network of the graph is shown,The weight is represented by a weight that,Representing the characteristics of the enhanced topology,Representing the first of the topological featuresThe components corresponding to the individual nodes are selected,Representing the first of the topological featuresComponents corresponding to the individual nodes.
It should be noted that, in actual use, the dynamic change of the persistence feature is continuously monitored, and the following formula is adopted:
Updating the persistence features, and when the enterprise risk assessment is carried out by using the method of the application at the future time, using the persistence features updated in real time to participate in the fusion of the steps to obtain fusion features.
Wherein,Representing the updated persistence feature(s),Representing the persistence characteristic of the time step t,Is a dynamic weight.
It is worth mentioning that by updating the persistence feature in real time, the timeliness of the persistence feature can be improved, and further the timeliness of the fusion feature is improved, and the persistence feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding are fused to obtain the fusion feature, so that the information richness and the comprehensiveness of the fusion feature are improved.
And 15, performing risk assessment on each target enterprise according to the fusion characteristics and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise.
The risk assessment result is used for describing the risk degree of a target enterprise, and can be used for carrying out risk analysis on the enterprise or other enterprises of the user, focusing on the enterprise with high risk degree and carrying out manual analysis, and analyzing whether the enterprise currently has risks in aspects of supply chains, technical cooperation, capital association, market competition, legal compliance, market operation, systematic risk and the like, so that corresponding measures are adopted or cooperation with the enterprise with high risk degree is avoided.
In some embodiments of the present application, the step of performing risk assessment on each target enterprise according to the fusion feature and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise includes:
the following steps are respectively carried out for each target enterprise:
First, calculating the centrality score of a node corresponding to a target enterprise according to a multi-layer enterprise social network.
For example, the centrality score may be a betting centrality or a degree centrality of the node in the multi-tier enterprise social network.
And secondly, performing risk assessment on the target enterprise based on the centrality score and the fusion characteristic to obtain a risk assessment result of the target enterprise.
The comprehensive risk score of the target enterprise can be calculated through the centrality score and the fusion characteristic, and the risk degree of the target enterprise is determined according to the comprehensive risk score and a preset threshold. If a first preset threshold is setFor the limit between low risk and medium risk, a second preset threshold valueTo limit risk of stroke to high risk, if the risk score is integrated<Defined as low risk ifDefined as stroke risk, ifThen it is defined as high risk.
Specifically, the formula can be used:
calculate the firstComprehensive risk scoring for individual target enterprises。
Wherein,、、As a parameter of the weight-bearing element,Represent the firstThe centrality score of the individual nodes,Representing risk conduction intensity extracted from cross-layer incidence matrix of multi-layer enterprise social network, representing the firstThe risk of interaction of individual nodes in a multi-tier enterprise social network,Representing the continuous coherent computationThe persistence metric of the high risk bar code corresponding to the individual node represents the systematic risk in the topology.
It should be noted that, the method of the present application can also identify potential opportunities of target enterprises, identify target enterprises having higher cooperative will or supply chain potential as opportunity enterprises, specifically calculate, for each target enterprise, a comprehensive opportunity score of each target enterprise based on a multi-layer enterprise social network and persistence features, where the comprehensive opportunity score is used to describe probability that the target enterprise has potential development opportunities, and the higher the comprehensive opportunity score is, the greater the probability that the target enterprise has potential development opportunities is, so that a user can pay attention to the enterprise having high comprehensive opportunity score, and consider developing cooperative relationships or supply relationships with the enterprise.
Specifically, the formula can be used:
calculate the firstComprehensive opportunity scoring for individual target enterprises。
Wherein,The composite opportunity score for node i is given,Is the firstThe connectivity of individual nodes, representing their potential strength of cooperation in different tier networks,Is the firstCross-layer similarity (e.g., industry similarity, technology similarity) of individual nodes, by cross-layer correlation matrix calculation,Calculated for continuous coherentA persistence metric for the low risk bar code corresponding to the individual node, representing the node's stable relationship potential,、、Is a weight parameter that adjusts the contribution of connectivity, similarity, and topological features to the opportunity score.
It is worth mentioning that constructing the multi-layer enterprise social network can fully express the association relation between target enterprises, realizes the accurate modeling of the enterprise relation, fuses the persistent feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain the fusion feature, improves the information richness and the comprehensiveness of the fusion feature, performs risk assessment based on the information rich fusion feature and the accurate multi-layer enterprise social network, and can effectively improve the accuracy of enterprise risk assessment.
Furthermore, the main advantages of the method of the application include:
(1) And combining the dynamic graph neural network with continuous coherence, namely fusing the persistence characteristic with the embedding of the dynamic graph neural network, improving the recognition capability of complex relations in the multi-layer enterprise social network, and effectively capturing the characteristics of time variation and topological structure.
(2) And the self-adaptive characteristic alignment adopts the self-adaptive time window selection and characteristic standardization technology to ensure the consistency of the characteristics in the time dimension and enhance the adaptability and the accuracy.
(3) And the dynamic attention mechanism is used for carrying out weighted fusion on the characteristics of different layers, enhancing the sensitivity to key nodes and relations and improving the recognition capability to potential risks and opportunities.
(4) The real-time dynamic monitoring is provided with real-time updating capability, can quickly respond to the dynamic change of the enterprise relationship, and provides scientific basis for enterprise decision-making.
The method mainly evaluates potential risks of enterprises in aspects of supply chains, technical cooperation, capital association, market competition, legal compliance, market operation, systematic risks and the like, namely the risks can influence the services of the enterprises or damage the assets and reputation of the enterprises due to the relations of the supply chains, cooperation, competition and the like, and the opportunities are identified to bring positive influence and value to the enterprises and identify and evaluate conditions and conditions which bring potential growth, profit or strategic advantages to the enterprises so as to keep leading in the competitive market. The method breaks through the technical bottlenecks of the traditional enterprise information system in the aspects of dynamic modeling, multi-layer network characterization, topology feature extraction and the like. By combining the time sequence perceived graph neural network with the persistence coherent theory in the computation topology, the accurate modeling of the enterprise relation network in two dimensions of time and structure is realized. The innovative technical scheme not only can capture the dynamic evolution mode of the enterprise relationship, but also can reveal the potential risks contained in the network structure through the topological feature analysis, thereby providing a more systematic and deep analysis framework for enterprise risk management.
The enterprise risk assessment device based on the dynamic graph neural network provided by the application is exemplified below.
As shown in fig. 3, an embodiment of the present application provides an enterprise risk assessment apparatus based on a dynamic graph neural network, where the enterprise risk assessment apparatus 300 based on the dynamic graph neural network includes:
The system comprises an acquisition module 301, a storage module and a storage module, wherein the acquisition module is used for acquiring relationship data of a plurality of target enterprises, and the relationship data of each target enterprise is used for describing enterprise relationships between the target enterprise and other target enterprises, and the enterprise relationships belong to one of a plurality of relationships;
The construction module 302 is configured to construct a multi-layer enterprise social network based on all the relationship data, wherein the multi-layer enterprise social network corresponds to a plurality of relationships one by one, a plurality of nodes in the multi-layer enterprise social network correspond to a plurality of target enterprises one by one, edges between two nodes in the same layer are enterprise relationships between two corresponding target enterprises, and edges between two nodes in different layers are influence relationships between two corresponding target enterprises;
The feature extraction module 303 is configured to perform feature extraction on the multi-layer enterprise social network by using the dynamic graph neural network to obtain short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network, where the short-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a short time span, the long-term feature embedding is used for describing enterprise relationship information in the multi-layer enterprise social network over a long time span, and the topology feature embedding is used for describing topology structure information of the multi-layer enterprise social network;
The fusion module 304 is configured to obtain a persistent feature based on the multi-layer social network of the enterprise, and fuse the persistent feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain a fusion feature;
The risk assessment module 305 is configured to perform risk assessment on each target enterprise according to the fusion feature and the multi-layer enterprise social network, so as to obtain a risk assessment result of each target enterprise, where the risk assessment result is used to describe the risk degree of the target enterprise.
It should be noted that, because the content of information interaction and execution process between the above devices/units is based on the same concept as the method embodiment of the present application, specific functions and technical effects thereof may be referred to in the method embodiment section, and will not be described herein.
It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-described division of the functional units and modules is illustrated, and in practical application, the above-described functional distribution may be performed by different functional units and modules according to needs, i.e. the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-described functions. The functional units and modules in the embodiment may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit, where the integrated units may be implemented in a form of hardware or a form of a software functional unit. In addition, the specific names of the functional units and modules are only for distinguishing from each other, and are not used for limiting the protection scope of the present application. The specific working process of the units and modules in the above system may refer to the corresponding process in the foregoing method embodiment, which is not described herein again.
As shown in fig. 4, an embodiment of the present application provides a terminal device D10 of the embodiment comprising at least one processor D100 (only one processor is shown in fig. 4), a memory D101 and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, the processor D100 implementing the steps of any of the respective method embodiments described above when executing the computer program D102.
Specifically, when the processor D100 executes the computer program D102, a multi-layer enterprise social network is constructed based on all relationship data by acquiring the relationship data of a plurality of target enterprises, then the feature extraction is performed on the multi-layer enterprise social network by using a dynamic graph neural network to obtain short-term feature embedding, long-term feature embedding and topology feature embedding of the multi-layer enterprise social network, then the persistent feature is acquired based on the multi-layer enterprise social network, the persistent feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding are fused to obtain fusion features, and finally risk assessment is performed on each target enterprise according to the fusion features and the multi-layer enterprise social network to obtain a risk assessment result of each target enterprise. The method comprises the steps of constructing a multi-layer enterprise social network, fully expressing the association relation between target enterprises, realizing accurate modeling of the enterprise relation, fusing the persistence feature, the short-term feature embedding, the long-term feature embedding and the topology feature embedding to obtain fusion features, improving the information richness and the comprehensiveness of the fusion features, performing risk assessment based on the information rich fusion features and the accurate multi-layer enterprise social network, and effectively improving the accuracy of enterprise risk assessment.
The Processor D100 may be a central processing unit (CPU, central Processing Unit), the Processor D100 may also be other general purpose processors, digital signal processors (DSP, digital Signal processors), application SPECIFIC INTEGRATED integrated circuits (ASICs), off-the-shelf Programmable gate arrays (FPGAs) or other Programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
The memory D101 may in some embodiments be an internal storage unit of the terminal device D10, for example a hard disk or a memory of the terminal device D10. The memory D101 may also be an external storage device of the terminal device D10 in other embodiments, for example, a plug-in hard disk, a smart memory card (SMC, smart Media Card), a Secure Digital (SD) card, a flash memory card (FLASH CARD) or the like, which are provided on the terminal device D10. Further, the memory D101 may also include both an internal storage unit and an external storage device of the terminal device D10. The memory D101 is used for storing an operating system, an application program, a boot loader (BootLoader), data, other programs, etc., such as program codes of the computer program. The memory D101 may also be used to temporarily store data that has been output or is to be output.
Embodiments of the present application also provide a computer readable storage medium storing a computer program which, when executed by a processor, implements steps for implementing the various method embodiments described above.
Embodiments of the present application provide a computer program product enabling a terminal device to carry out the steps of the method embodiments described above when the computer program product is run on the terminal device.
The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the present application may implement all or part of the flow of the method of the above embodiments, and may be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the computer program may implement the steps of each of the method embodiments described above. Wherein the computer program comprises computer program code which may be in source code form, object code form, executable file or some intermediate form etc. The computer readable medium can include at least any entity or device capable of carrying computer program code to a dynamic graph neural network based enterprise risk assessment method device/terminal apparatus, a recording medium, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, and a software distribution medium. Such as a U-disk, removable hard disk, magnetic or optical disk, etc.
In the foregoing embodiments, the descriptions of the embodiments are emphasized, and in part, not described or illustrated in any particular embodiment, reference is made to the related descriptions of other embodiments.
Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
While the foregoing is directed to the preferred embodiments of the present application, it will be appreciated by those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present application, and such modifications and adaptations are intended to be comprehended within the scope of the present application.