Financial business compliance risk identification method based on knowledge graph reasoning

Abstract

The invention discloses a financial business compliance risk identification method based on knowledge graph reasoning, and the method comprises the steps: constructing a heterogeneous knowledge graph, enabling the heterogeneous knowledge graph to be associated with a corresponding collateral node, carrying out the formalized modeling of a supervision rule through a time sequence description logic, carrying out the forward chain reasoning of the knowledge graph through a rule engine, and identifying a dominant compliance risk. An unstructured contract text is converted into vector representation, semantic alignment is performed through a graph attention mechanism and contract element nodes in a knowledge graph, long-term dependence characteristics of equipment state changes are captured, a temporary inference sub-graph is generated through a rapid adaptation module, dominant risk indexes are used as node characteristics of a graph neural network, and the graph neural network is constructed. And optimizing the overall reasoning effect through a joint loss function. Therefore, a collaborative reasoning mode can be formed through hierarchical fusion of explicit and implicit risks, multi-source heterogeneous data processing and rule dynamic adaptation are effectively solved, and the accuracy and efficiency of movable property financing business compliance risk identification of small and medium-sized enterprises are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and specifically to a method for identifying compliance risks in financial business based on knowledge graph reasoning. Background Art

[0002] In the field of financial business compliance risk identification, the movable asset financing scenarios of small and medium-sized enterprises face significant technical challenges due to the characteristics of dispersed collateral, dynamic changes in regulatory rules and complex data types.

[0003] Traditional compliance risk identification methods mainly rely on structured data processing and static rule matching, which makes it difficult to effectively integrate the temporal status data of collateral collected by IoT devices with unstructured contract texts, resulting in insufficient ability to identify hidden risks such as "gradual compliance failure due to equipment aging" and "liability disputes caused by semantic ambiguity in contract terms."

[0004] In addition, when regulatory policies are updated frequently (such as local financial bureaus temporarily adding regional regulatory details), traditional systems require manual reconstruction of the rule engine, which makes it difficult to meet real-time dynamic adaptation needs, and has problems such as low reasoning efficiency and high risk omission rate.

[0005] The limitations of existing technologies in processing multi-source heterogeneous data correlation analysis, time series feature modeling and rapid adaptation of regulatory rules make it difficult for the accuracy and practicality of compliance risk identification in movable asset financing business of small and medium-sized enterprises to meet actual needs. Summary of the Invention

[0006] The present invention aims to solve the technical problems in the above-mentioned technologies at least to some extent.

[0007] To this end, the present invention discloses a method for identifying compliance risks in financial services based on knowledge graph reasoning, comprising the following steps:

[0008] S1. Construct a heterogeneous knowledge graph that includes corporate entities, movable collateral, contract elements, and regulatory rules, and associate the collateral status data collected in real time by IoT devices with the corresponding collateral nodes through a mapping function;

[0009] S2. Use temporal description logic to formally model regulatory rules, and use the rule engine to perform forward chain reasoning on the knowledge graph to identify explicit compliance risks;

[0010] S3. Use a contract semantic encoder based on contrastive learning to convert unstructured contract text into a vector representation, and use a graph attention mechanism to semantically align it with the contract element nodes in the knowledge graph;

[0011] S4. Use the time series graph convolution layer to model the time series data of IoT devices and capture the long-term dependency characteristics of device state changes.

[0012] S5. Introducing a meta-learning mechanism to generate a temporary reasoning subgraph through a fast adaptation module when regulatory rules are dynamically updated;

[0013] S6. Use the explicit risk indicators output from step S2 as node features of the graph neural network, and perform hierarchical fusion with the implicit risk features extracted from steps S3 to S5, and optimize the overall reasoning effect through the joint loss function.

[0014] According to the financial business compliance risk identification method based on knowledge graph reasoning disclosed in the present invention, it is possible to form a collaborative reasoning model through the fusion of explicit and implicit risk stratification, effectively solve the multi-source heterogeneous data processing and dynamic adaptation of rules, and significantly improve the accuracy and efficiency of compliance risk identification in the movable asset financing business of small and medium-sized enterprises.

[0015] In addition, the financial business compliance risk identification method based on knowledge graph reasoning disclosed in the present invention may also have the following additional technical features:

[0016] In one embodiment of the present invention, in step S1:

[0017] The enterprise entity node contains industrial and commercial information and credit records, the movable property collateral node contains the device model and IoT sensor ID, the contract element node contains guarantee terms and repayment period, and the regulatory rules node contains specific terms for movable property financing of small and medium-sized enterprises.

[0018] In one embodiment of the present invention, in step S2:

[0019] Temporal description logic transforms regulatory rules into logical expressions containing time variables, and identifies explicit compliance risks by traversing the association path of enterprise-collateral-contract-regulatory rules.

[0020] In one embodiment of the present invention, in step S3:

[0021] The graph attention mechanism generates association weights by calculating the semantic similarity between the contract clause vector and the contract element node features, thereby eliminating the semantic ambiguity of the contract clauses.

[0022] In one embodiment of the present invention, in step S4:

[0023] During the calculation of the temporal graph convolution layer, the adjacency matrix and degree matrix with self-loops are calculated for the subgraph of each time slice, and the temporal dependency features are extracted through Laplace normalization operation and weight matrix multiplication.

[0024] In one embodiment of the present invention, in step S5:

[0025] The meta-learning mechanism trains the fast adaptation module by updating data through historical regulatory rules. When new regulatory rules are added, a temporary reasoning subgraph corresponding to the new rules is generated based on the meta-learning parameters.

[0026] In one embodiment of the present invention, in step S6:

[0027] Explicit risk indicators include abnormal geographical location of the collateral, repeated mortgage of the collateral, and expired supervision period. Implicit risk characteristics include the conflict characteristics between equipment performance degradation trend and contract terms.

[0028] In one embodiment of the present invention, in step S6:

[0029] In the hierarchical fusion process, the Boolean value encoding of the explicit risk indicator is spliced with the vector representation of the implicit risk feature, and a linear transformation is performed through the fusion weight matrix and bias vector to generate the final risk feature representation.

[0030] In one embodiment of the present invention, the collateral status data collected in real time by the IoT device includes a position coordinate sequence and a performance parameter fluctuation curve, and the status attributes of the collateral node in the knowledge graph are updated in real time through a mapping function.

[0031] Additional contents and advantages of the present invention will be given in the following description or can be understood through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The technical solutions and beneficial effects of the present invention will become apparent and easily understood from the following contents in conjunction with the accompanying drawings, in which:

[0033] Figure 1 This is a workflow diagram of the financial business compliance risk identification method based on knowledge graph reasoning of the present invention. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0035] The following will describe the financial business compliance risk identification method based on knowledge graph reasoning disclosed in the present invention with reference to the accompanying drawings.

[0036] like Figure 1 As shown, a financial business compliance risk identification method based on knowledge graph reasoning includes the following steps:

[0037] S1. Construct a heterogeneous knowledge graph that includes corporate entities, movable collateral, contract elements, and regulatory rules, and associate the collateral status data collected in real time by IoT devices with the corresponding collateral nodes through a mapping function;

[0038] S2. Use temporal description logic to formally model regulatory rules, and use the rule engine to perform forward chain reasoning on the knowledge graph to identify explicit compliance risks;

[0039] S3. Use a contract semantic encoder based on contrastive learning to convert unstructured contract text into a vector representation, and use a graph attention mechanism to semantically align it with the contract element nodes in the knowledge graph;

[0040] S4. Use the time series graph convolution layer to model the time series data of IoT devices and capture the long-term dependency characteristics of device state changes.

[0041] S5. Introducing a meta-learning mechanism to generate a temporary reasoning subgraph through a fast adaptation module when regulatory rules are dynamically updated;

[0042] S6. Use the explicit risk indicators output from step S2 as node features of the graph neural network, and perform hierarchical fusion with the implicit risk features extracted from steps S3 to S5, and optimize the overall reasoning effect through the joint loss function.

[0043] It should be noted that, in step S1:

[0044] The enterprise entity node contains industrial and commercial information and credit records, the movable property collateral node contains the device model and IoT sensor ID, the contract element node contains guarantee terms and repayment period, and the regulatory rules node contains specific terms for movable property financing of small and medium-sized enterprises.

[0045] It should be noted that, in step S2:

[0046] Temporal description logic transforms regulatory rules into logical expressions containing time variables, and identifies explicit compliance risks by traversing the association path of enterprise-collateral-contract-regulatory rules.

[0047] It should be noted that, in step S3:

[0048] The graph attention mechanism generates association weights by calculating the semantic similarity between the contract clause vector and the contract element node features, thereby eliminating the semantic ambiguity of the contract clauses.

[0049] It should be noted that, in step S4:

[0050] During the calculation of the temporal graph convolution layer, the adjacency matrix and degree matrix with self-loops are calculated for the subgraph of each time slice, and the temporal dependency features are extracted through Laplace normalization operation and weight matrix multiplication.

[0051] It should be noted that, in step S5:

[0052] The meta-learning mechanism trains the fast adaptation module by updating data through historical regulatory rules. When new regulatory rules are added, a temporary reasoning subgraph corresponding to the new rules is generated based on the meta-learning parameters.

[0053] It should be noted that, in step S6:

[0054] Explicit risk indicators include abnormal geographical location of the collateral, repeated mortgage of the collateral, and expired supervision period. Implicit risk characteristics include the conflict characteristics between equipment performance degradation trend and contract terms.

[0055] It should be noted that, in step S6:

[0056] In the hierarchical fusion process, the Boolean value encoding of the explicit risk indicator is spliced with the vector representation of the implicit risk feature, and a linear transformation is performed through the fusion weight matrix and bias vector to generate the final risk feature representation.

[0057] It should be noted that the collateral status data collected by IoT devices in real time includes location coordinate sequences and performance parameter fluctuation curves, and the status attributes of the collateral nodes in the knowledge graph are updated in real time through mapping functions.

[0058] In an embodiment of the present invention, the overall execution logic of the method is based on a sequential process, and completes knowledge graph construction, symbolic reasoning execution, unstructured text processing, time series data modeling, dynamic rule adaptation and fusion reasoning in sequence; wherein, the IoT data mapping is cyclically connected to the real-time data stream according to the timestamp, the forward chain reasoning is cyclically traversed through the rule base until the fact set is not updated (the selection logic of the internal nested rule antecedent matching), and the time series graph convolution extracts time series dependency features through multi-layer loops; the selection logic is also reflected in the meta-learning process in which the corresponding reasoning subgraph is adapted to different regulatory rule update scenarios, forming a nested execution architecture of "sequence-dominated, cyclic processing of repetitive tasks, and selection of implementation conditional judgments".

[0059] For the basic model, the implementation steps of the knowledge graph symbolic reasoning engine enhanced by temporal logic are as follows:

[0060] Construct a knowledge graph, specifically,

[0061] The first step is to define four types of core entities and initialize nodes, namely: input enterprise entity data (business information, credit records), generate enterprise node set V corp , input the basic information of movable collateral (device model, sensor ID), and generate collateral node set V collateral , parse the key elements of the financing contract (guarantee terms, repayment period), and generate the contract node set V contract And extract the regulatory rules text (such as specific movable property financing clauses) to generate the regulatory rule node set V reg ;

[0062] The second step is to establish the relationship edges between entities, connecting the enterprise and the collateral ("ownership" relationship), the collateral and the contract ("association" relationship), and the contract and the regulatory rules ("applicability" relationship) according to the business logic, forming the initial edge set E0;

[0063] The third step is to dynamically associate IoT device data (loop mapping). For each timestamp t (obtained sequentially from the real-time data stream), the collateral status data (such as location coordinates and performance parameters) collected by the device are extracted and associated with the corresponding collateral node v through the mapping function φ. collateral ∈V collateral , update the node real-time status attributes.

[0064] The symbolic reasoning algorithm is executed, specifically,

[0065] The first step is to formalize the regulatory rules and convert each regulatory rule r k ∈R is converted into a temporal logic expression (such as the formula corresponding to “N consecutive days of location anomaly triggering an early warning”) and stored in the rule base.

[0066] Step 2: Forward chain reasoning loop (until no new facts are generated)

[0067] Initialize the fact set F to all observable facts in the current knowledge graph (such as node attributes and activated relationship edges);

[0068] Loop through each rule r in the rule base k , if the rule antecedent antecedent(r k ) is completely included in the current fact set F, and the rule consequent(r k ) (e.g., "Generate risk warning event") is added to the fact set F to mark the rule as triggered to avoid repeated execution;

[0069] If the fact set F is not updated in this cycle, the reasoning is terminated and the explicit risk facts (such as "repeated mortgage of collateral" and "exceeded supervision period") are output.

[0070] Unstructured contract text processing, specifically,

[0071] The first step is to encode the semantics of the contract text. The financing contract text is input and the clause-level semantic vector sequence {h1,h2,…,h u}, where each h i Corresponding to a clause paragraph.

[0072] The second step is to align the semantics of the graph nodes. For each contract node v collateral ∈V collateral, extract the clause text associated with the node (such as the paragraph corresponding to the “guarantee clause”) and determine its position set I∈{1,2,…,u} in the semantic vector sequence;

[0073] The association weight is calculated by the graph attention mechanism. For each association term i∈I, the node feature e is calculated. contract With the term vector h i The attention weight α i (based on node neighbor relationships and semantic similarity);

[0074] Aggregate the clause vector by weight and update the semantic feature e′ of the contract node contract =Σα i h i , eliminating ambiguity in terms (such as clarifying the context of the "floating mortgage" clause).

[0075] IoT time series data feature modeling, specifically,

[0076] The first step is to build a time series graph structure, dividing the time series data collected by IoT devices into T time slices according to timestamps. Each time slice t corresponds to a subgraph G. t =(V,E t ), where E t Contains the state-related edges of the collateral node at that moment (such as the "location-time" edge).

[0077] The second step is to extract the convolutional layer features of the time series graph. For each time slice t, the initial node features (including real-time status data), looping through L layers of graph convolution operations to calculate the adjacency matrix with self-loops Sum degree matrix ;

[0078] By formula Extract time-dependent features (such as the trend of changes in the status of collateral) until the final layer features are output

[0079] Dynamic regulatory rule adaptation, specifically,

[0080] The first step is to define the regulatory rule update task. When new or modified regulatory rules (such as temporary rules of local financial bureaus) are added, a meta-learning task T is generated. i , including the reasoning subgraph structure corresponding to the new rules (such as the newly added “energy consumption data verification” node and associated edges).

[0081] The second step is to quickly adapt the module training, for each task T i ,sample the support set (old rule scenario) and query set (new rule scenario) from the historical rule update data;

[0082] By meta-learning the objective function Optimize the model parameters θ so that the model can quickly generate the inference subgraph parameters corresponding to the new rules (without rebuilding the entire rule engine), and repeat until convergence.

[0083] Feature fusion preparation, specifically,

[0084] The first step is to obtain the explicit risk characteristics x output by the basic model s (e.g. Boolean encoding of “abnormal geographical location of collateral”);

[0085] The second step is to obtain the implicit risk features x output by the enhanced model d (Such as the vector representation of “equipment performance degradation and conflict with contract terms”).

[0086] Layered fusion calculation, specifically,

[0087] The first step is to concatenate the explicit and implicit feature vectors to form the joint feature x f =Concat(x s ,x d );

[0088] The second step is to fuse the weight matrix W f and bias b f Perform linear transformation on the joint features to generate the final risk feature representation;

[0089] The third step is to calculate the joint loss function (The weight ratio of explicit and implicit losses is selected through hyperparameters λ1 and λ2), and backpropagation optimizes the parameters of the entire system.

[0090] Specifically, let the knowledge graph be G = (V, E), where the node set V = {v1, v2, ..., v n}, including entities such as corporate entities, movable collateral, contract elements, and regulatory rules, and the edge set E = {(v i ,v j )} represents the relationship between entities.

[0091] IoT device data is associated with the knowledge graph. Let the time series data sequence collected by IoT devices be D = {d 1 ,d 2 ,…,d T},in, is the m-dimensional data vector at time t. Define the mapping function φ:D→V to map the device data to the collateral node in the knowledge graph, that is, v t =φ(d t ), where v k The corresponding collateral node.

[0092] Regulatory rules are formalized, and the regulatory rule set R = {r1, r2, ..., r s}, expressed using Temporal Description Logic (TDL). Taking the rule "a risk warning is triggered when the collateral exceeds the agreed storage area for N consecutive days" as an example, let the location function Location(v,t) represent the location of node v at time t, the agreed area is AgreedArea, and the duration function Duration(t1,t2) represents the time interval. The mathematical expression of this rule is:

[0093] Among them, V collateral A collection of collateral nodes.

[0094] Risk transmission path identification, let the risk transmission path set be P = {p1,p2,...,p l}, path p i ={v i1 ,v i2 ,...,v in} is the node sequence in the knowledge graph. Define the risk transmission probability function Prob(p i ) represents the path p i The probability of triggering risk, the problem to be solved is to find a path that meets a certain risk threshold θ, that is

[0095] Knowledge graph construction, entity embedding representation, using TransE and other models to embed nodes in the knowledge graph i Mapping to a low-dimensional vector space Relationship ij Mapped to Satisfy e i +r ij ≈e j .

[0096] Symbolic reasoning algorithm, let the rule engine input be the fact set F={f1,f2,…,f q}, fact f i Corresponding to nodes or edges in the knowledge graph. Rule r k The antecedent is an antecedent(r k ), consequent(r k ), the reasoning process is expressed as: By iteratively applying the above rules until the fact set F no longer changes, explicit risk identification is completed.

[0097] Unstructured data processing, let the contract text C = {c1,c2,…,c u}, encoded into a vector sequence by the RoBERTa financial field fine-tuning model

[0098] Calculate node v through the graph attention mechanism (GAT) j The association weight with the contract text vector Among them, a is the attention parameter vector, W is the weight matrix, and N(j) is the node v j The neighbor node set of , || represents vector concatenation. The updated node features are

[0099] Time series characteristics and dynamic rules, for IoT device time series data graph G t =(V,E t ), define the temporal graph convolution layer (Temporal GCN): Among them, X (l) is the feature matrix of the l-th layer node, is an adjacency matrix with self-loops, for The diagonal node degree matrix, W (l) is the weight matrix and σ is the activation function.

[0100] Introducing the meta-learning mechanism, let the meta-learning task set be Each task T i Corresponding to different regulatory rule update scenarios, the objective function is optimized through meta-learning: Among them, f θ is the prediction function with model parameter θ, y i is the true label, is the loss function.

[0101] Fusion reasoning mechanism, let the explicit risk feature vector output by the symbolic reasoning engine be x s , the implicit risk feature vector extracted by DS-GCNN is x d , the fused feature vector x f =Concat(x s ,x d )·W f +b f , where Concat represents vector concatenation, W f is the fusion weight matrix, b f is the bias vector. Through the joint loss function Optimization model, where Classify losses for explicit risks, is the implicit risk embedding loss, λ1,λ2 are weight coefficients.

[0102] To sum up, the financial business compliance risk identification method based on knowledge graph reasoning disclosed in the present invention can form a collaborative reasoning model through the fusion of explicit and implicit risk stratification, effectively solve the multi-source heterogeneous data processing and dynamic adaptation of rules, and significantly improve the accuracy and efficiency of compliance risk identification in movable asset financing business of small and medium-sized enterprises.

[0103] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A financial business compliance risk identification method based on knowledge graph reasoning, characterized by: The following steps are involved: S1. Construct a heterogeneous knowledge graph that includes corporate entities, movable collateral, contract elements, and regulatory rules, and associate the collateral status data collected in real time by IoT devices with the corresponding collateral nodes through a mapping function; S2. Use temporal description logic to formally model regulatory rules, and use the rule engine to perform forward chain reasoning on the knowledge graph to identify explicit compliance risks; S3. Use a contract semantic encoder based on contrastive learning to convert unstructured contract text into a vector representation, and use a graph attention mechanism to semantically align it with the contract element nodes in the knowledge graph; S4. Use the time series graph convolution layer to model the time series data of IoT devices and capture the long-term dependency characteristics of device state changes. S5. Introducing a meta-learning mechanism to generate a temporary reasoning subgraph through a fast adaptation module when regulatory rules are dynamically updated; S6. Use the explicit risk indicators output from step S2 as node features of the graph neural network, and perform hierarchical fusion with the implicit risk features extracted from steps S3 to S5, and optimize the overall reasoning effect through the joint loss function.

2. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S1: The enterprise entity node contains industrial and commercial information and credit records, the movable property collateral node contains the device model and IoT sensor ID, the contract element node contains guarantee terms and repayment period, and the regulatory rules node contains specific terms for movable property financing of small and medium-sized enterprises.

3. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S2: Temporal description logic transforms regulatory rules into logical expressions containing time variables, and identifies explicit compliance risks by traversing the association path of enterprise-collateral-contract-regulatory rules.

4. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S3: The graph attention mechanism generates association weights by calculating the semantic similarity between the contract clause vector and the contract element node features, thereby eliminating the semantic ambiguity of the contract clauses.

5. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S4: During the calculation of the temporal graph convolution layer, the adjacency matrix and degree matrix with self-loops are calculated for the subgraph of each time slice, and the temporal dependency features are extracted through Laplace normalization operation and weight matrix multiplication.

6. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S5: The meta-learning mechanism trains the fast adaptation module by updating data through historical regulatory rules. When new regulatory rules are added, a temporary reasoning subgraph corresponding to the new rules is generated based on the meta-learning parameters.

7. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S6: Explicit risk indicators include abnormal geographical location of the collateral, repeated mortgage of the collateral, and expired supervision period. Implicit risk characteristics include the conflict characteristics between equipment performance degradation trend and contract terms.

8. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: In the step S6: In the hierarchical fusion process, the Boolean value encoding of the explicit risk indicator is spliced with the vector representation of the implicit risk feature, and a linear transformation is performed through the fusion weight matrix and bias vector to generate the final risk feature representation.

9. The financial business compliance risk identification method based on knowledge graph reasoning according to claim 1 is characterized in that: The collateral status data collected in real time by the IoT device includes a position coordinate sequence and a performance parameter fluctuation curve, and the status attributes of the collateral node in the knowledge graph are updated in real time through a mapping function.

Images