CN115526391A

CN115526391A - Method, device and storage medium for predicting enterprise risk

Info

Publication number: CN115526391A
Application number: CN202211143138.8A
Authority: CN
Inventors: 马宁亚; 林廷懋; 付博
Original assignee: CCB Finetech Co Ltd
Current assignee: CCB Finetech Co Ltd
Priority date: 2022-09-20
Filing date: 2022-09-20
Publication date: 2022-12-27

Abstract

The application discloses a method, a device and a storage medium for predicting enterprise risk. The method comprises the following steps: acquiring an initial affair map and an initial knowledge map of a risk prediction scene; embedding knowledge in the initial knowledge graph and projecting the knowledge to a hyperbolic space; embedding an event in the initial event map and projecting the event to a hyperbolic space; in a hyperbolic space, joint embedding is obtained according to knowledge embedding and event embedding; the joint embedding is transformed into Euclidean space and input into a predictor to predict enterprise risk. The method projects the joint embedding of the knowledge and the events to the hyperbolic space, so that the influence of the specific events on the enterprise is reflected comprehensively, the knowledge embedding is better processed by utilizing the characteristic of the hyperbolic space, and the prediction model of the enterprise risk has better effect and higher generalization capability.

Description

Method, device and storage medium for predicting enterprise risk

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method, an apparatus, and a storage medium for predicting enterprise risk.

Background

With the development of market economy, event-driven risk prediction is required for financial events that are continuously occurring in the market to evaluate the impact of the events. In the prior art, enterprise risk prediction is generally performed by extracting entities and relations of risk events from news, inputting a trained event prediction model to obtain event types and elements thereof, updating a knowledge graph, and obtaining an event risk value based on a better knowledge graph and the event prediction model. The core of the method is as follows: 1) Extracting entities of events and relations among the entities by using a BERT-QA mode; 2) Training an event risk model according to the risk event type and the risk event elements; 3) And weighting the risk of the related enterprises through the event risk model. Such methods have the following drawbacks in prediction: 1) The risk prediction for a risk event is too coarse, only a risk value is given, without further revealing which risks are; 2) Only triggering from the main body of the risk event, only considering the background knowledge of the event, but not considering the context knowledge of the event; 3) Embedding events also stays in the european space, which does not adequately respond to the hierarchical knowledge of the event itself. Thus, existing event-driven predictive models also have the following pain points: 1) The lack of completeness of event information, or the lack of background knowledge of the event, or the lack of contextual knowledge of the event, results in a bias in prediction, particularly for structurally similar events; 2) The traditional event embedding or entity embedding based on the BERT model is limited by the characteristics of the European space, the hierarchy of knowledge cannot be more accurately reserved and reflected, the knowledge is excessively compressed, and the effect is poor. Therefore, the prediction model for enterprise risk in the prior art has poor effect and low generalization capability.

Disclosure of Invention

An object of the embodiments of the present application is to provide a method, an apparatus, and a storage medium for predicting an enterprise risk, which are used to solve the problems in the prior art that a prediction model for an enterprise risk is poor in effect and low in generalization capability.

To achieve the above object, a first aspect of the present application provides a method for predicting risk of an enterprise, the method comprising:

acquiring an initial affair map and an initial knowledge map of a risk prediction scene;

embedding knowledge in the initial knowledge graph and projecting the knowledge to a hyperbolic space;

embedding an event in the initial event map and projecting the event to a hyperbolic space;

in a hyperbolic space, joint embedding is obtained according to knowledge embedding and event embedding;

the joint embedding is transformed into Euclidean space and input into a predictor to predict enterprise risk.

In the embodiment of the present application, obtaining an initial affairs graph and an initial knowledge graph of a risk prediction scenario includes:

extracting entities, risk events and risk event elements from the original financial corpus;

extracting causal relationships of the risk events and relationships between the entities to determine relationships between the risk events and relationships between the entities;

constructing an initial affair map according to the risk events, the risk event elements and the relationship among the risk events;

and constructing an initial knowledge graph according to the entities and the relationship among the entities.

In the embodiment of the present application, embedding knowledge in the initial knowledge-graph and projecting the knowledge-graph to the hyperbolic space includes:

obtaining initial embedding in an initial knowledge graph through a TransE algorithm;

projecting the initial embedding to a hyperbolic space through a hyperbolic transform;

an initial embedding is input to the graph attention model of the two-tier attention mechanism to derive an entity representation of each node of the knowledge-graph.

In the embodiment of the present application, embedding an event in an initial event graph and projecting the event to a hyperbolic space includes:

searching the target risk event in the initial event map to obtain a related subgraph of the target risk event;

re-inputting the relevant sub-graph into the pre-training model to obtain an initial event representation of the relevant sub-graph;

mapping an initial event representation of the associated subgraph into a hyperbolic space;

and (4) transferring and aggregating the evidence information of the initial event representation through a directed hyperbolic graph attention layer to obtain a target event representation.

In the embodiment of the present application, searching for a target risk event in a case graph to obtain a relevant subgraph of the target risk event includes:

determining the similarity between any event in the initial event map and the target risk event through a BM25 algorithm to obtain an anchor point event;

determining an evidence event of an anchor event in an initial affair map through a breadth-first algorithm;

connecting the target risk event, the anchor event and the evidence event according to a causal relationship to obtain a relevant subgraph of the target risk event;

wherein the target risk event comprises a target precondition event and a target guess event.

In this embodiment of the present application, re-inputting the relevant sub-graph into the pre-training model to obtain the initial event representation of the relevant sub-graph includes:

encoding event nodes in the relevant subgraphs to obtain event sequences;

the sequence of events is input to a pre-trained model to obtain an initial event representation of the associated sub-graph.

In this embodiment of the present application, the transferring and aggregating the evidence information of the initial event representation through the directed hyperbolic curve attention layer to obtain the target event representation includes:

the hyperbolic vectors represented by any two initial events are projected back to the Euclidean space, and the approximation of the attention value is calculated;

after the evidence information represented by any two initial events is propagated and aggregated, transforming any two updated event representations back into a hyperbolic space;

and determining any two updated event representations as target event representations.

In the embodiment of the present application, joint embedding satisfies formula (1):

Η＝αu _k +(1-α)u _e ； (1)

wherein H is combined insertion, u _k For knowledge embedding, u _e For event embedding, α is a hyper-parameter.

In an embodiment of the present application, transforming the joint embedding into Euclidean space and inputting a predictor to predict enterprise risk includes:

transforming the joint embedding into a Euclidean space to obtain a joint embedding vector representation;

the joint embedding vector representation is input into a predictor to determine the prediction probability of the causal relationship formed by the target risk event and the corresponding related event;

determining a score for the target risk event based on the predicted probability.

A second aspect of the present application provides an apparatus for predicting risk of an enterprise, comprising:

a memory configured to store instructions; and

a processor configured to call instructions from memory and upon execution of the instructions is capable of implementing the above-described method for predicting risk of an enterprise.

A third aspect of the present application provides a machine-readable storage medium having stored thereon instructions that, when executed by a processor, cause the processor to be configured to perform the above-described method for predicting risk of an enterprise.

A fourth aspect of the present application provides a computer program product comprising a computer program which, when executed by a processor, implements a method for predicting risk of an enterprise according to the above.

According to the technical scheme, the initial affair map and the initial knowledge map of the risk prediction scene are obtained, then knowledge embedding is carried out in the initial knowledge map and is projected to the hyperbolic space, event embedding is carried out in the initial affair map and is projected to the hyperbolic space, then joint embedding is obtained in the hyperbolic space, and finally the joint embedding is transformed to the Euclidean space and is input into a predictor so as to predict enterprise risks. The method and the system integrate the background knowledge in the knowledge graph and the logical relationship between the events in the case graph, thereby more comprehensively reflecting the influence of specific events on enterprises; the knowledge and event combined embedding is projected to a hyperbolic space, and knowledge embedding can be better processed by utilizing the good properties of knowledge layering and large parameter capacity of the hyperbolic space. In conclusion, the enterprise risk prediction model has a good effect and high generalization capability.

Additional features and advantages of embodiments of the present application will be described in detail in the detailed description which follows.

Drawings

The accompanying drawings, which are included to provide a further understanding of the embodiments of the disclosure and are incorporated in and constitute a part of this specification, illustrate embodiments of the disclosure and together with the description serve to explain the embodiments of the disclosure, but are not intended to limit the embodiments of the disclosure. In the drawings:

FIG. 1 schematically illustrates a flow chart of a method for predicting risk of an enterprise according to an embodiment of the application;

FIG. 2 schematically illustrates a flow chart for predicting risk of an enterprise according to a particular embodiment of the present application;

FIG. 3 is a block diagram schematically illustrating an apparatus for predicting risk of an enterprise according to an embodiment of the present application;

fig. 4 schematically shows an internal structure diagram of a computer device according to an embodiment of the present application.

Detailed Description

To make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it should be understood that the specific embodiments described herein are only used for illustrating and explaining the embodiments of the present application and are not used for limiting the embodiments of the present application. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

It should be noted that, if directional indications (such as up, down, left, right, front, and back … …) are involved in the embodiment of the present application, the directional indications are only used to explain the relative position relationship between the components, the motion situation, and the like in a specific posture (as shown in the drawing), and if the specific posture is changed, the directional indications are changed accordingly.

In addition, if there is a description of "first", "second", etc. in the embodiments of the present application, the description of "first", "second", etc. is for descriptive purposes only and is not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In addition, technical solutions between various embodiments may be combined with each other, but must be realized by a person skilled in the art, and when the technical solutions are contradictory or cannot be realized, such a combination should not be considered to exist, and is not within the protection scope of the present application.

In the embodiment of the present application, the hyperbolic space is a two-dimensional curved space having a uniform negative curvature everywhere, and unlike a limited spherical surface, the hyperbolic space has an infinite area. The hyperbolic space projection is a hyperbolic space matrix that transforms a representation matrix X of events or entities in euclidean space of d dimensions into a d' dimension. Specifically, a linear transformation is adopted for the event representation matrix X to map the event representation matrix to a d-dimensional Euclidean space, and then an exponential transformation is applied to project the transformed event representation matrix to a d' -dimensional hyperbolic space.

FIG. 1 schematically shows a flow chart of a method for predicting risk of an enterprise according to an embodiment of the application. As shown in fig. 1, in an embodiment of the present application, a method for predicting risk of an enterprise is provided, which may include the following steps.

And 102, acquiring an initial affair graph and an initial knowledge graph of the risk prediction scene.

In the embodiment of the application, the case diagram is a case logic knowledge base, describes an evolution rule and a mode between events, is structurally a directed cyclic diagram, nodes represent the events, and directed edges represent relations between the events, such as sequential bearing, cause and effect, upper and lower positions and the like. The knowledge graph is a series of different graphs for displaying the relation between the knowledge development process and the structure, and the knowledge resources and the carriers thereof are described, mined, analyzed, constructed, drawn and displayed by using a visualization technology, and the mutual relation among the knowledge resources and the carriers is realized. The risk prediction scene is a scene based on the risks of enterprise financial events which continuously occur in the market. Such as enterprise loss events, enterprise financial fraud events, enterprise asset devaluation events, and the like.

The initial event map of the embodiment of the application is an event map established according to risk events and risk event elements extracted from original financial corpora and the relationship between the risk events. In the initial event graph, nodes are risk events which are combined with a certain generalization of a specific scene and events related to the risk events, edges are association relations between the nodes, such as causal relations of the events, attributes are strength of the causal relations of the events, and calculation is carried out according to semantic strength between the events. The initial event map may be updated based on the continually extracted financial risk events and the relationships between the events. The initial knowledge graph of the embodiment of the application is a knowledge graph established according to entities extracted from original financial corpora and the relationship among the entities. In the initial knowledge graph, nodes are extracted entities such as names of people, names of places and the like, and edges are associated relations between nodes such as relations between companies, between companies and persons and between persons. The initial knowledge-graph may also be continuously updated based on the continuously extracted entities and relationships between the entities.

And 104, embedding knowledge in the initial knowledge graph and projecting the knowledge to a hyperbolic space.

Knowledge embedding refers to a method of mapping the content of the knowledge-graph, including entities and relationships, to a continuous vector space. In the prior art, more knowledge embedding is projected to the Euclidean space, and the European space cannot reflect the hierarchy of the knowledge, so that in the embodiment of the application, the knowledge embedding can be projected to the hyperbolic space, and the knowledge embedding is better processed by utilizing the characteristic of the hyperbolic space. In one example, the processor may obtain an initial embedding, that is, an entity embedding, by using a hyperbolic space transform in an initial knowledge Graph, project the initial embedding to the hyperbolic space through a hyperbolic transform, and obtain an entity representation of each node of the knowledge Graph by using a Graph Attention model (HGAT) of a two-layer Attention mechanism. By considering heterogeneous graph problems of edges of various types of nodes existing in event background knowledge, information of the heterogeneous graph is processed by utilizing a layered graph attention machine mechanism, and the topological structure of the graph is better reserved.

And 106, embedding the event in the initial event map and projecting the event to a hyperbolic space.

Event embedding refers to projecting events into a particular embedding space, representing the events by a dense vector of a particular dimension in the space, while preserving invariance of the physical meaning and relationship between events in the embedding space. In the embodiment of the application, the processor may first search out a relevant subgraph of the input target risk event from the constructed initial event graph, then encode event nodes in the relevant subgraph, and encode the event in the text form into a dense vector, that is, an initial event representation of the relevant subgraph. And then mapping the initial event representation to a hyperbolic space, and transmitting and updating information by using a cause and effect reasoning machine to finally obtain an updated target event representation. In one example, a Directed hyperboloid Graph Attention Layer (DHGAT Layer) may be used as a causal reasoner to communicate and aggregate evidence information of the initial event representation to obtain a target event representation. According to the embodiment of the application, the causal logic relation among risk events is considered through event embedding, the interpretability is stronger, and the requirements of a financial application scene are better met.

And 108, in the hyperbolic space, obtaining joint embedding according to knowledge embedding and event embedding.

In the embodiment of the application, joint embedding refers to merging knowledge representations of different spaces in the same task. The processor may derive joint embedding from knowledge embedding and event embedding projected into the hyperbolic space. In one example, entity representations of nodes derived from an initial knowledge graph and target event representations derived from an initial event graph may be weighted to ultimately yield a jointly embedded vector representation containing knowledge and events. In another example, the entity representation and the target event representation of the node are updated by means of a gate mechanism.

And 110, transforming the joint embedding into Euclidean space and inputting a predictor to predict the enterprise risk.

In the embodiment of the present application, considering that the final joint embedded representation is a vector representation in a hyperbolic space, in order to facilitate performing a final prediction task, the joint embedded vector representation in the hyperbolic space may be transformed back to an euclidean space, and then the prediction task of the target risk event is performed to determine a prediction probability that the target risk event and the corresponding related event form a causal relationship, and determine a score of the target risk event according to the prediction probability. Knowledge and event information are comprehensively considered, and information related to risks is fully utilized.

According to the technical scheme, the initial affair map and the initial knowledge map of the risk prediction scene are obtained, then knowledge embedding is carried out in the initial knowledge map and is projected to the hyperbolic space, event embedding is carried out in the initial affair map and is projected to the hyperbolic space, then joint embedding is obtained in the hyperbolic space, and finally the joint embedding is transformed to the Euclidean space and is input into a predictor so as to predict enterprise risks. The method and the system integrate the background knowledge in the knowledge graph and the logical relationship between the events in the affair graph, so that the influence of specific events on enterprises is reflected more comprehensively; the joint embedding of knowledge and events is projected to the hyperbolic space, and the knowledge embedding can be better processed by utilizing the good properties of knowledge layering and large parameter capacity of the hyperbolic space. In conclusion, the enterprise risk prediction model has the advantages of being good in effect and high in generalization capability.

In this embodiment of the present application, the step 102 of obtaining the initial case map and the initial knowledge map of the risk prediction scenario may include:

Specifically, the initial event map is a event map constructed according to the original financial corpus (i.e., the historical financial corpus), and the processor may establish the event map according to the risk event and the risk event element extracted from the original financial corpus and the relationship between the risk events. When the original financial corpus is updated, the initial affair atlas is also updated. The initial event map can be constructed by firstly extracting the risk events and the risk event elements and the relationship between the risk events and then constructing the initial event map according to the risk events, the risk event elements and the relationship between the risk events.

First, the processor may extract from the original financial corpus the entities of interest, the risk events and the subjects of the risk events, the time of interest, etc. Wherein, the risk events can include but are not limited to loss events, financial fraud events, dong Gaojian member abnormal events, asset devaluation events, rating deterioration events, and other risk events; entities may include, but are not limited to, person names, place names, organization names, and the like. In one example, the processor may extract the risk events and risk event elements through a Named Entity Recognition (NER) model.

Second, the processor may extract causal relationships of the risk events to determine relationships between the risk events. Similarly, the processor may also extract relationships through the NER model. In addition, the processor needs to calculate the causal strength of the extracted events as the attributes of the event relationship.

Finally, after extracting the risk event, the risk event elements and the relationship among the risk events, the processor can establish a case map of the default prediction scene based on the causal relationship among the risk events, integrate the data into triples in the form of points, edges and attributes, and import the triples into a map database for storage. A triplet is referred to herein as a "start-edge-end" form. The mechanism of the initial case graph includes nodes and edges. The nodes are risk events extracted by the processor and combined with a specific scene to be generalized, and events related to the risk events are extracted from the risk events according to a affair map accumulated by a company. The edges are the association between entities, and the general case graph includes the relationships of cause, effect, condition, order, up and down, etc. The causal relationship between the nodes (events) is mainly researched through the relationship of the event graph in the embodiment of the application, the attribute is the strength of the causal relationship between the events, and the calculation is carried out according to the semantic strength between the events.

The initial knowledge graph of the embodiment of the application is a knowledge graph established according to entities extracted from original financial corpora and the relationship among the entities. In the initial knowledge graph, nodes are extracted entities such as names of people, names of places and the like, and edges are association relations between nodes such as relations between companies, between companies and persons and between persons. The initial knowledge-graph may also be updated on a continuous basis based on the continuously extracted entities and the relationships between the entities. When the original financial corpus is updated, the initial knowledge map is also updated. The initial knowledge graph can be constructed by firstly extracting the entities and the relations among the entities and then constructing the initial knowledge graph according to the relations among the entities.

It should be noted that, if new data is added based on the constructed event graph, the initial event graph and the initial knowledge graph are updated according to the method and the structure for constructing the event graph, so as to process the continuously extracted financial risk events, the relationship between the events, and the relationship between the entities.

In this embodiment of the present application, the embedding and projecting knowledge into the hyperbolic space in the initial knowledge-graph at step 104 may include:

obtaining initial embedding in the initial knowledge graph through a TransE algorithm;

Specifically, the processor may derive an initial embedding, i.e., entity i embedding, in the initial knowledge-graph using hyperbolic space TransE. The initial embedding satisfies formula (2):

wherein, the first and the second end of the pipe are connected with each other,

in order to be initially embedded in the optical fiber,

and

is a head entity, and is characterized in that,

and

is a tail entity, l is a relationship,

and

for the positive example of the existence of the relationship,

and

indicating that there is no negative example of the relationship of/,

the predicted and actual errors.

The processor may then project the initial embedding into the hyperbolic space through a hyperbolic transform. The initial embedding after projection satisfies formula (3) and formula (4):

wherein the content of the first and second substances,

for the initial embedding after the projection,

is a head entity in a hyperbolic space,

is a tail entity in a hyperbolic space,

is a relationship in a hyperbolic space,

the method is a hyperbolic space, c is an Euclidean space, h is vector representation, and u is a vector.

Finally, the processor can perform GAT on the HGAT model initially embedded and input into the hyperbolic space at the node and the relationship level respectively, and perform neighbor aggregation on the neighborhood nodes to obtain an entity representation of each node i of the knowledge graph, wherein the entity representation of the node i satisfies the formula (5):

for the physical representation of event node i, K is the number of attention weights, δ is a function of δ,

is the kth node weight, w ^k Is a weight coefficient of k, x _ij And a vector of the node ij, vi is a point after the node i is aggregated, vj is any neighborhood node, and N (vi) is a set of neighborhood nodes.

α _ij Satisfy equation (6) for attention weight between nodes ij

Wherein the content of the first and second substances,

wherein alpha is _ij Is the attention weight between nodes ij, α ^T Is the target adjacency matrix A, wx _i Is x _i Weight coefficient of (Wx) _j Is x _i V (k) is a neighborhood node of node k, N (v) _i ) Is a collection of neighborhood nodes.

In the embodiment of the present application, the HGAT may use heterogeneous convolution to consider the heterogeneity of different types of information, capture the importance of different neighboring nodes (reduce the weight of noise) and the importance of different node (information) types to a specific node.

It should be noted that knowledge embedding may be performed not only from the perspective of the spatial domain by using HGAT but also from the perspective of the frequency domain by using a relational graph convolution. According to the embodiment of the application, heterogeneous graphs of various relations are considered in training and prediction of various relation information, so that heterogeneous graph information is processed by utilizing a layered graph attention machine, and the topological structure of the graph is better reserved.

In this embodiment of the present application, the step 106 of embedding and projecting the event in the initial event graph to the hyperbolic space may include:

mapping an initial event representation of the correlated subgraph into a hyperbolic space;

Specifically, the processor may first search out relevant sub-graphs of the input target risk event in the constructed initial event graph. And encoding event nodes in the related subgraphs, and encoding the events in the text form into dense vectors, namely initial event representation of the related subgraphs. In one example, the processor may encode event nodes in the relevant subgraph to obtain a sequence of events; the sequence of events is input to a pre-trained model to obtain an initial event representation of the associated sub-graph. And then mapping the initial event representation to a hyperbolic space, and transmitting and updating information by using a cause and effect reasoning machine to finally obtain an updated target event representation. In one example, evidence information of an initial event representation can be passed and aggregated to arrive at a target event representation using DHGAT as a cause and effect reasoner. According to the embodiment of the application, the causal logic relation among risk events is considered through event embedding, the interpretability is stronger, and the requirements of a financial application scene are better met.

In this embodiment of the present application, searching for the target risk event in the event graph to obtain a relevant subgraph of the target risk event may include:

Specifically, for a given target Premise (Premise) event and a target guess (Hypothesis) event, the processor may search the two events in a constructed case map, that is, an initial case map, respectively, to obtain one or more events similar to the Premise event and the guess event, and the specific search method may be through a BM25 algorithm. For convenience, the target precondition event and the target guess event are not represented as P and H without loss of generality. Taking the target precondition event P as an example, for any event e in all event sets in the target event graph, the calculation method of the BM25 score of the target precondition event P satisfies formula (7) and formula (8):

BM25 _score (P,e)＝∑ _t∈P w(t,e)； (7)

wherein, P is a target precondition, t is any event in the target precondition, e is any event in all event sets in the target event graph, w (t, e) is the relevance of word segmentation in the target precondition, qtf is the frequency of the precondition t in the target precondition, tf is the frequency of the precondition t in any event, d is a target text, l _d Length of target text, avg _l B is a constant, df is the frequency of the target text d contained in the target precondition event, N is the total number of texts, k ₁ Is a sum of k ₃ Is a hyper-parameter.

After several events related to the target precondition P or the target guess event H are acquired, they can be recorded as

And

it is also denoted as an anchor event. The processor then searches for cause and result events within the theta hops in the target event map, referred to as evidence events, using a breadth first algorithm based on these anchor events. Finally, the evidence events, the anchor point events, the target precondition events and the target guess events are connected into a sub-graph of the target event graph according to the causal relationship, namely a causal evidence graph, which is marked as G _sub ＝(V _sub ,R _sub ) Wherein V is _sub ＝{V ^P ,V ^H P, H }. And obtaining the related event of the target risk event by obtaining the related subgraph of the target risk event.

In this embodiment of the present application, re-inputting the relevant sub-graph into the pre-training model to obtain the initial event representation of the relevant sub-graph may include:

encoding event nodes in the relevant subgraphs to obtain event sequences;

Specifically, after searching out a relevant subgraph of the target risk event, that is, after obtaining an evidence graph, the processor needs to encode event nodes of the relevant subgraph and encode the event in a text form into a dense vector. The embodiment of the application may use a pre-training model, such as a BERT model, as an encoder, and re-input the relevant sub-graph into the pre-training model to obtain a context-dependent vector of each event in the relevant sub-graph, that is, a relevant event corresponding to the target risk event.

First, the relevant subgraph can be expanded to form an event sequence, and the processed event sequence can be expressed as: [ CLS]C[SEP]I ₁ …[SEP]I _n [SEP]E[SEP]. Wherein, [ CLS]For the representation character of the entire sequence, [ SEP]The event separation function can also be taken as a character for representing each event, C is used as a reason, C is used as a guess event if Ask-for is used as a reason, and A is used as a guess eventsk-for is the result, C is the precondition, I = { I = { (I) ₁ ,…,I _n E is the result, E is the precondition if Ask-for is the cause, and E is the guess event if Ask-for is the result.

After the graph is expanded into the above sequence, the processor may input it into a pre-trained model. After the output of the pre-training model is obtained, [ SEP ] after each event]The representation vector of the character serves as the representation vector of this event, the representation vector of all events is denoted as x = { x = { x = } ₁ ,…,x _i ,…,x _n+2 And (c) the step of (c) in which,

is a vector representation of the first event in the correlation graph, and x ₁ For vector representation of causal events, x _n+2 Is a vector representation of the resulting event.

In this embodiment of the present application, the transferring and aggregating the evidence information of the initial event representation through the directed hyperbolic curve attention layer to obtain the target event representation may include:

after the evidence information represented by any two initial events is propagated and aggregated, any two updated event representations are transformed back into a hyperbolic space;

Specifically, for a constructed evidence graph, namely a related subgraph, the processor may map an initial event representation in the related subgraph to a hyperbolic space, and then use the causal reasoner to transfer and update evidence information. Since the relevant subgraphs are directed and the causal relationships are also directed, the transfer and updating of evidence information is considered from two aspects. The DHGAT layer may be used by the processor for each round of information update in the entire cause and effect reasoner.

The vector of the hyperbolic attention layer input to the DHGAT layer is H', which is an event representation matrix in the hyperbolic space. As with the attention mechanism in european space, the DHGAT layer also passes and aggregates the evidence information in two ways. First, the processor may compute an attention score for the vector representation of the central node and the neighboring nodes; second, the processor may use the attention weights to perform a weighted summation of the neighboring nodes and update the representation of the central node. However, hyperbolic space is not a vector space, so both partial operations are computationally intensive in hyperbolic space. To solve the computationally intensive problem, i.e. to reduce the computational burden, an approximation method may be employed for the two part calculations.

Specifically, for example, the information of the cause event is updated with the result event information, and the hyperbolic vector h 'of the ith event in the hyperbolic space is given' _i And a hyperbolic vector h 'for the jth event' _j And it computes the hyperbolic vector h 'in the hyperbolic space' _i And hyperbolic vector h' _j By selecting a hyperbolic vector h' _i And hyperbolic vector h' _j Projected Return to Euclidean tangent space

And calculating the attention value between them, and regarding the attention value calculated in the Euclidean space as an approximate calculation method of the attention value calculated in the hyperbolic space. The attention weight in hyperbolic space satisfies equation (9),

wherein, w _ij Attention weight in hyperbolic space, eff (i) is the resulting event of event i, h' _i Hyperbolic vector, h 'of ith event in hyperbolic space' _j Is a hyperbolic vector of the jth event in the hyperbolic space, and c is the Euclidean space.

Then, based on the calculated attention weight w _ij For event i, using an aggregation operation on event i willThe information of its evidence event neighbors is aggregated onto event i. During the course of a polymerization operation, we will refer to the hyperbolic vector h' _i In the tangential space of

As an approximation of a hyperbolic space. It is not difficult to find that the hyperbolic space is about a hyperbolic vector h' _i Is a hyperbolic vector h' _i The best approximation of the surrounding hyperbolic space. Then, the information aggregation operation is also in hyperbolic space with respect to hyperbolic vector h' _i In the tangential space of

For the ith event u _i The represented aggregation operation can be represented as formula (10):

wherein u is _i For the ith event, eff (i) is the resulting event of event i, w _ij Is attention weight, h 'in hyperbolic space' _i Hyperbolic vector, h 'of ith event in hyperbolic space' _j Is a hyperbolic vector of the jth event in the hyperbolic space, and c is the Euclidean space.

It can be seen that in all of the above formulas, the neighbor node of event i is defined as its result node. Therefore, the event in the direction of the cause of the event i is not confused with the event in the direction of the result of the event i. In this way, directionality is introduced into the attention mechanism, and evidential information can be passed from outcome to cause in a retrospective form.

The embodiment of the application can also introduce a multi-head attention mechanism comprising n heads, wherein the multi-head attention mechanism is mainly used for capturing more various information, and the multi-head attention calculation method meets the formula (11):

wherein

For the ith event u _i With respect to the output of the mth head in the multi-head attention mechanism,

[·]a vector splicing operation is represented as a vector splicing operation,

is a trainable matrix.

In Euclidean space, after propagation and aggregation of evidence information, each event u _i The updated representation is transformed back into the hyperbolic space by equation (12):

wherein u' _i Is an updated event, h' _i Hyperbolic vector, u, for the ith event in hyperbolic space _i Is the ith event, and c is Euclidean space.

Finally, the updated event represents U '= [ U' ₁ ,…,u′ _i+2 ]Is taken as the output of the DHGAT layer. Through the DHGAT layer, evidence information may be passed from result to cause to prevent confusion between cause and result. However, the information of the evidence event is difficult to be fully utilized in the process of reasoning. Therefore, in order to fully utilize the information of the evidence event, the embodiment of the present application may introduce a parallel DHGAT layer, in which the information of the reason event is updated to the information of the result event in a deductive manner. And to facilitate propagation of evidence event information, two parallel DHGAT layers may be stacked multiple times, with the input to the DHGAT layer of each layer being the output of the DHGAT layer of the previous layer. According to the embodiment of the application, the causal logical relationship among risk events is considered through event embedding, the interpretability is stronger, and the requirements of financial application scenes are better met.

In the embodiment of the present application, joint embedding may satisfy formula (1):

Η＝αu _k +(1-α)u _e ； (1)

wherein H is combined insertion, u _k For knowledge embedding, u _e For event embedding, α is a hyperparameter.

In particular, joint embedding refers to merging knowledge representations of different spaces in the same task. The processor may derive joint embedding from knowledge embedding and event embedding projected into the hyperbolic space. In the embodiment of the application, the entity representation of the node obtained according to the initial knowledge graph and the target event representation obtained according to the initial event graph can be weighted, and finally the vector representation containing the joint embedding of knowledge and events is obtained. It should be noted that the processor may also update the information of the entity representation and the target event representation of the node by means of a gate mechanism.

In an embodiment of the present application, the step 110 of transforming the joint embedding into the euclidean space and inputting the predictor to predict the enterprise risk may include:

transforming the joint embedding into a Euclidean space to obtain joint embedding vector representation;

inputting the joint embedding vector representation into a predictor to determine a prediction probability that the target risk event and the corresponding related event form a causal relationship;

and determining the score of the target risk event according to the prediction probability.

Specifically, considering that the final joint embedded vector representation H is a vector representation in a hyperbolic space, in order to facilitate the final prediction task, the joint embedded vector representation in the hyperbolic space may be first transformed back to the euclidean space, and the joint embedded vector representation in the hyperbolic space may be transformed back to the euclidean space, which satisfies formula (13):

wherein u is _H Joint insert vector representation for transformation back into Euclidean space, H being doubleAnd c is Euclidean space.

After transforming the joint embedded vector representation back into euclidean space, the processor may begin the task of predicting the target risk event. The processor may pass the vector through a linear layer and through a normalization operation to determine a prediction probability that the target risk event and the corresponding related event form a causal relationship, thereby determining a score for the target risk event based on the prediction probability. Wherein the prediction of the predictor satisfies equation (14):

Y＝softmax(W _y u _H +b _y )； (14)

wherein Y is the prediction probability that the target risk event and the corresponding related event form a causal relationship, u _H For joint embedding vector representation, W, transformed back into Euclidean space _y Is a matrix of coefficients, b _y Is a bias matrix. Knowledge and event information are comprehensively considered, and information related to risks is fully utilized.

FIG. 2 schematically illustrates a flow chart for predicting risk of an enterprise according to a particular embodiment of the present application. As shown in FIG. 2, in one embodiment, the overall process for predicting risk of a business includes three steps, data preparation A, model training B, and assessment C. The data preparation A comprises entity, event and event element extraction A-1, relation extraction A-2, event and knowledge graph construction A-3 and event and knowledge graph updating A-3'; model training B comprises event retrieval B-2-1, event embedding B-2-2, joint embedding B-3 and prediction B-4. Wherein, knowledge embedding B-1 is to embed knowledge in the initial knowledge map and perform hyperbolic transformation. And event retrieval B-2-1 and event embedding B-2-2 are to carry out event embedding in the initial event map and carry out hyperbolic transformation. The joint embedding B-3 is to obtain joint embedding containing events and knowledge according to the event embedding and the knowledge embedding. And finally, the combined embedding is returned to Euclidean space for risk prediction so as to determine the prediction probability of the causal relationship formed by the target risk event and the corresponding related event. The evaluation module C determines a score for the target risk event based on the predicted probability.

The embodiments of the present application have many advantages over the prior art. First, compared with the need of establishing a risk prediction model in the beginning of the prior art, the embodiment of the present application starts from causal logic between risk events, and has stronger interpretability and better meets the actual requirements of financial application scenarios. Secondly, the prior art only utilizes the context knowledge of the event, but rarely comprehensively considers the logic relationship between the background knowledge and the event, and the embodiment of the application comprehensively considers the knowledge and the event information and fully utilizes the information related to the risk. Moreover, at present, more knowledge embedding is projected to an Euclidean space, but the Euclidean space cannot reflect the hierarchy of the knowledge, the knowledge and the event are projected to the hyperbolic space to be jointly embedded, and the knowledge embedding is better processed by utilizing the characteristic of the hyperbolic space. Finally, in the knowledge embedding process, the embodiment of the application focuses on the problem of heterogeneous graphs of edges of various types of nodes existing in event background knowledge, so that the heterogeneous background knowledge is processed by adopting a layered HGAT, and the topological structure of the knowledge graph is better reserved.

It should be understood that although the steps in the flowcharts of fig. 1 and 2 are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least some of the steps in fig. 1 and 2 may include multiple sub-steps or multiple stages that are not necessarily performed at the same time, but may be performed at different times, and the order of performing the sub-steps or stages is not necessarily sequential, but may be performed alternately or alternately with other steps or at least some of the sub-steps or stages of other steps.

Fig. 3 is a block diagram schematically illustrating an apparatus for predicting risk of an enterprise according to an embodiment of the present application. As shown in fig. 3, an embodiment of the present application provides an apparatus for predicting risk of an enterprise, which may include:

a memory 310 configured to store instructions; and

a processor 320 configured to invoke the instructions from the memory 310 and upon execution of the instructions is capable of implementing the methods for predicting risk of the enterprise described above.

Specifically, in the embodiment of the present application, the processor 320 may be configured to:

Further, the processor 320 may be further configured to:

the method for acquiring the initial affair map and the initial knowledge map of the risk prediction scene comprises the following steps:

Further, the processor 320 may be further configured to:

embedding knowledge in the initial knowledge graph and projecting the knowledge into a hyperbolic space comprises the following steps:

Further, the processor 320 may be further configured to:

embedding events in the initial event graph and projecting the events to a hyperbolic space comprises the following steps:

Further, the processor 320 may be further configured to:

searching the target risk event in the event graph to obtain a relevant subgraph of the target risk event comprises the following steps:

Further, the processor 320 may be further configured to:

re-inputting the relevant sub-graph into the pre-trained model to obtain an initial event representation of the relevant sub-graph comprises:

encoding event nodes in the relevant subgraphs to obtain event sequences;

Further, the processor 320 may be further configured to:

the method for transferring and aggregating the evidence information of the initial event representation through the directed hyperbolic graph attention layer to obtain the target event representation comprises the following steps:

Η＝αu _k +(1-α)u _e ； (1)

Further, the processor 320 may be further configured to:

transforming the joint embedding into Euclidean space and inputting into a predictor to predict enterprise risk includes:

According to the technical scheme, the initial matter graph and the initial knowledge graph of the risk prediction scene are obtained, then knowledge embedding is carried out in the initial knowledge graph and is projected to the hyperbolic space, event embedding is carried out in the initial matter graph and is projected to the hyperbolic space, then joint embedding is obtained in the hyperbolic space, and finally the joint embedding is transformed to the Euclidean space and is input into the predictor so as to predict the enterprise risk. The method and the system integrate the background knowledge in the knowledge graph and the logical relationship between the events in the affair graph, so that the influence of specific events on enterprises is reflected more comprehensively; the joint embedding of knowledge and events is projected to the hyperbolic space, and the knowledge embedding can be better processed by utilizing the good properties of knowledge layering and large parameter capacity of the hyperbolic space. In conclusion, the enterprise risk prediction model has the advantages of being good in effect and high in generalization capability.

Embodiments of the present application also provide a machine-readable storage medium having stored thereon instructions that, when executed by a processor, cause the processor to be configured to perform the above-described method for predicting risk of an enterprise.

An embodiment of the application further provides a computer program product comprising a computer program adapted to perform, when executed on a processor, a procedure for initializing the following method steps: acquiring an initial matter graph and an initial knowledge graph of a risk prediction scene; embedding knowledge in the initial knowledge graph and projecting the knowledge to a hyperbolic space; embedding an event in the initial event map and projecting the event to a hyperbolic space; in a hyperbolic space, joint embedding is obtained according to knowledge embedding and event embedding; the joint embedding is transformed into Euclidean space and input into a predictor to predict enterprise risk.

In one embodiment, entities, risk events and risk event elements are extracted from the original financial corpus; extracting causal relationships of the risk events and relationships between the entities to determine relationships between the risk events and relationships between the entities; constructing an initial affair map according to the risk events, the risk event elements and the relationship among the risk events; and constructing an initial knowledge graph according to the entities and the relationship among the entities.

In one embodiment, in the initial knowledge-graph, an initial embedding is derived by a TransE algorithm; projecting the initial embedding to a hyperbolic space through a hyperbolic transform; an initial embedding is input to the graph attention model of the two-tier attention mechanism to derive an entity representation of each node of the knowledge-graph.

In one embodiment, searching the target risk event in the initial event graph to obtain a related subgraph of the target risk event; re-inputting the relevant sub-graph into the pre-training model to obtain an initial event representation of the relevant sub-graph; mapping an initial event representation of the correlated subgraph into a hyperbolic space; and (4) transferring and aggregating the evidence information of the initial event representation through a directed hyperbolic graph attention layer to obtain a target event representation.

In one embodiment, the similarity between any event in the initial event map and the target risk event is determined through a BM25 algorithm to obtain an anchor point event; determining an evidence event of an anchor event in an initial affair map through a breadth-first algorithm; connecting the target risk event, the anchor event and the evidence event according to a causal relationship to obtain a relevant subgraph of the target risk event; wherein the target risk event comprises a target precondition event and a target guess event.

In one embodiment, event nodes in the relevant subgraph are encoded to obtain an event sequence; the sequence of events is input to a pre-trained model to obtain an initial event representation of the associated sub-graph.

In one embodiment, hyperbolic vectors represented by any two initial events are projected back into Euclidean space and an approximation of the attention value is calculated; after the evidence information represented by any two initial events is propagated and aggregated, transforming any two updated event representations back into a hyperbolic space; and determining any two updated event representations as target event representations.

In one embodiment, the joint embedding is transformed into Euclidean space to obtain a joint embedding vector representation; the joint embedding vector representation is input into a predictor to determine the prediction probability of the causal relationship formed by the target risk event and the corresponding related event; determining a score for the target risk event based on the predicted probability.

Fig. 4 schematically shows an internal structure diagram of a computer device according to an embodiment of the present application. In one embodiment, as shown in fig. 4, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. Y. The computer apparatus includes a processor a01, a network interface a02, a display screen a04, an input device a05, and a memory (not shown in the figure) connected through a system bus. Wherein the processor a01 of the computer device is adapted to provide computing and control capabilities. The memory of the computer apparatus includes an internal memory a03 and a nonvolatile storage medium a06. The nonvolatile storage medium a06 stores an operating system B01 and a computer program B02. The internal memory a03 provides an environment for running the operating system B01 and the computer program B02 in the nonvolatile storage medium a06. The network interface a02 of the computer apparatus is used for communicating with an external terminal through a network connection. The computer program is executed by processor a01 to implement a method for predicting risk of an enterprise. The display screen a04 of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device a05 of the computer device may be a touch layer covered on the display screen, a key, a trackball or a touch pad arranged on a casing of the computer device, or an external keyboard, a touch pad or a mouse.

Those skilled in the art will appreciate that the architecture shown in fig. 4 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include forms of volatile memory in a computer readable medium, random Access Memory (RAM) and/or non-volatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.

Computer-readable media include permanent and non-permanent, removable and non-removable media and may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a … …" does not exclude the presence of another identical element in a process, method, article, or apparatus that comprises the element.

The above are merely examples of the present application and are not intended to limit the present application. Various modifications and changes may occur to those skilled in the art. Any modification, equivalent replacement, improvement or the like made within the spirit and principle of the present application shall be included in the scope of the claims of the present application.

Claims

1. A method for predicting risk of a business, the method comprising:

event embedding is carried out in the initial affair map and the initial affair map is projected to the hyperbolic space;

in the hyperbolic space, obtaining joint embedding according to the knowledge embedding and the event embedding;

transforming the joint embedding into Euclidean space and inputting a predictor to predict enterprise risks.

2. The method of claim 1, wherein obtaining the initial fact graph and the initial knowledge graph of the risk prediction scenario comprises:

extracting causal relationships of the risk events and relationships between entities to determine relationships between the risk events and relationships between the entities;

constructing the initial affair map according to the risk events, the risk event elements and the relationship among the risk events;

and constructing the initial knowledge graph according to the entities and the relation between the entities.

3. The method of claim 1, wherein the embedding and projecting knowledge into the initial knowledge-graph to hyperbolic space comprises:

in the initial knowledge graph, obtaining initial embedding through a TransE algorithm;

projecting the initial embedding to the hyperbolic space through hyperbolic transformation;

the initial embedding is input to a graph attention model of a two-tier attention mechanism to derive an entity representation of each node of the knowledge-graph.

4. The method of claim 1, wherein the event embedding in the initial event graph and projecting to the hyperbolic space comprises:

searching a target risk event in the initial event graph to obtain a related subgraph of the target risk event;

re-inputting the relevant subgraph into a pre-training model to obtain an initial event representation of the relevant subgraph;

mapping the initial event representation of the correlated subgraph into a hyperbolic space;

and transferring and aggregating the evidence information of the initial event representation through a directed hyperbolic graph attention layer to obtain a target event representation.

5. The method of claim 4, wherein searching the initial event graph for the target risk event to obtain the associated sub-graph of the target risk event comprises:

determining an evidence event of the anchor event in the initial event map through a breadth first algorithm;

6. The method of claim 4, wherein re-inputting the relevant subgraph into a pre-trained model to obtain an initial event representation of the relevant subgraph comprises:

encoding event nodes in the relevant subgraphs to obtain event sequences;

inputting the sequence of events to the pre-trained model to obtain an initial event representation of the correlated sub-graph.

7. The method according to claim 4, wherein the transferring and aggregating evidence information of the initial event representation through a directed hyperbolic attention layer to obtain a target event representation comprises:

transforming the updated any two event representations back into the hyperbolic space after propagation and aggregation of evidence information of the any two initial event representations;

and determining any two updated event representations as the target event representation.

8. The method of claim 1, wherein the joint embedding satisfies formula (1):

Η＝αu _k +(1-α)u _e ； (1)

wherein H is combined insertion, u _k For the knowledge embedding, u _e For the event embedding, α is a hyper-parameter.

9. The method of claim 1, wherein transforming the joint embedding into Euclidean space and inputting a predictor to predict enterprise risk comprises:

inputting the joint embedded vector representation into the predictor to determine a prediction probability that the target risk event and the corresponding related event constitute a causal relationship;

10. An apparatus for predicting risk of an enterprise, comprising:

a memory configured to store instructions; and

a processor configured to invoke the instructions from the memory and when executing the instructions to implement the method for predicting risk of an enterprise according to any of claims 1 to 9.

11. A machine-readable storage medium having stored thereon instructions which, when executed by a processor, cause the processor to be configured to perform a method for predicting risk of a business as claimed in any one of claims 1 to 9.

12. A computer program product comprising a computer program, characterized in that the computer program, when executed by a processor, implements a method for predicting a risk of an enterprise according to any of claims 1-9.