Dynamic intervention method, system, medium and device for risk loan applications

Abstract

The application discloses a kind of dynamic intervention methods, systems, media and equipment of risk loan application, its method includes the following steps: obtaining the loan application operation data of target user;Real-time risk determination is carried out according to the loan application operation data, and initial dynamic intervention is carried out according to risk determination result;The loan application operation data is extracted using the ST-GCN network of space-time attention enhancement, and long time sequence dependence is captured by fusing bidirectional GRU network and adaptive gating AGU network, to obtain behavior deviation degree;According to the behavior deviation degree, implement risk classification intervention strategy.Therefore the application fuses ST-GCN and adaptive gating AGU, by extracting space-time features and capturing long time sequence dependence, realize the high-precision loan application behavior deviation degree calculation, solve the problem of insufficient multi-dimensional feature fusion, weak long-range dependence capture capability in prior art.

Description

Technical Field

[0001] This invention relates to the field of loan application automation technology, and in particular to a dynamic intervention method, system, medium and equipment for risky loan applications. Background Technology

[0002] Current loan approval systems face multiple technical bottlenecks: 1. In terms of fraud detection, traditional solutions rely on manual verification of historical transaction records and credit reports, with response delays generally exceeding 8 hours. Industry statistics indicate that such delays result in substantial annual fraud losses in the consumer finance sector. 2. The approval process involves a high proportion of manual intervention. For example, a case study from a leading financial institution shows that the average processing time for a single loan is 6.8 hours, leading to a high user churn rate due to waiting times exceeding the limit.

[0003] Meanwhile, in the field of financial risk control, the detection of anomalies in loan application behavior is crucial for preventing fraud and reducing bad debt rates. Traditional methods are mostly based on static features, such as basic user information, historical credit history, or simple time-series models like RNN and LSTM. These methods struggle to simultaneously capture the spatiotemporal correlations and long-term time-series dependencies of user operations, leading to inaccurate calculations of behavioral deviations and high false positive and false negative rates in anomaly detection. Consequently, it becomes difficult to accurately implement risk grading and intervention strategies. Summary of the Invention

[0004] This invention provides a dynamic intervention method, system, medium, and device for risky loan applications. It integrates the spatiotemporal graph convolutional network ST-GCN and the adaptive gated AGU. By extracting spatiotemporal features and capturing long-term dependencies, it achieves high-precision calculation of loan application behavior deviation, solving the problems of insufficient multi-dimensional feature fusion and weak long-term dependency capture capability in the prior art.

[0005] Firstly, a dynamic intervention method for risky loan applications is provided, including:

[0006] Obtain loan application data from target users;

[0007] Risk assessment is performed in real time based on the loan application operation data, and initial dynamic intervention is carried out based on the risk assessment results;

[0008] The loan application operation data is extracted using a spatiotemporal attention-enhanced ST-GCN network, and a bidirectional GRU network and an adaptive gated AGU network are fused to capture long-term temporal dependencies, thus obtaining the behavior deviation degree.

[0009] Implement risk-based intervention strategies based on the degree of behavioral deviation.

[0010] In some embodiments, the real-time risk assessment based on the loan application operation data, and the initial dynamic intervention based on the risk assessment results, include:

[0011] If the target user's income editing time is greater than or equal to the industry benchmark value, and the facial recognition glabellar muscle contraction intensity is greater than or equal to the preset intensity, or the target user applies for a loan more than a preset number of times within a preset time, or the target user's voiceprint fundamental frequency volatility is greater than or equal to the preset volatility, and the voiceprint hash collision rate is greater than or equal to the preset collision rate, then it is judged as high risk.

[0012] Based on the high-risk assessment, the loan application process is frozen, anti-fraud education videos are pushed to users, and risk markers are generated.

[0013] In some embodiments, the ST-GCN network with spatiotemporal attention enhancement is used to extract spatiotemporal features from the loan application operation data, and a bidirectional GRU network and an adaptive gated AGU network are fused to capture long-term temporal dependencies to obtain behavioral deviation, including:

[0014] A dynamic graph construction method is used to construct a dynamic heterogeneous graph for each operation event in the loan application operation data according to the operation time window.

[0015] Spatiotemporal node features in the dynamic heterogeneous graph are extracted using a two-layer convolutional network, and the spatiotemporal correlation of the spatiotemporal node features is calculated using a spatiotemporal attention mechanism to obtain a spatiotemporal feature matrix.

[0016] The spatiotemporal feature matrix is ​​sorted according to the operation timestamp to generate an operation sequence, and the operation sequence is divided into multiple sub-sequences by a sliding window.

[0017] A bidirectional GRU network is used to extract temporal features from each of the subsequences, and an adaptive gated AGU network is used to filter out long-term key information from the extracted temporal features to obtain the long-term feature matrix of all subsequences.

[0018] The spatiotemporal feature matrix and the long-time series feature matrix are aligned and standardized to obtain aligned features. Mahalanobis distance and cosine similarity are calculated for the aligned features respectively.

[0019] The behavioral deviation is calculated based on the Mahalanobis distance and the cosine similarity.

[0020] In some embodiments, the step of extracting spatiotemporal node features from the dynamic heterogeneous graph using a two-layer convolutional network and calculating the spatiotemporal correlation of the spatiotemporal node features using a spatiotemporal attention mechanism to obtain a spatiotemporal feature matrix includes:

[0021] Calculate the Laplacian matrix of the dynamic heterogeneous graph, and perform a convolution operation on the Laplacian matrix using the first convolutional network in a two-layer convolutional network to obtain the first layer feature matrix;

[0022] The multi-scale features in the first-layer feature matrix are extracted in parallel using the second-layer multi-scale dilated convolution kernel in a two-layer convolutional network and then concatenated to obtain the second-layer multi-scale feature matrix.

[0023] The spatiotemporal feature matrix is ​​obtained by using the spatiotemporal attention mechanism based on the first layer feature matrix and the second layer multi-scale feature matrix.

[0024] In some embodiments, the step of extracting temporal features from each subsequence using a bidirectional GRU network and filtering out long-term key information from the extracted temporal features using an adaptive gated AGU network to obtain a long-term feature matrix for all subsequences includes:

[0025] The forward sequence of each subsequence is processed using the forward GRU network in the bidirectional GRU network, and the reverse sequence of each subsequence is processed using the backward GRU network in the bidirectional GRU network. The forward sequence and the reverse sequence are then fused to obtain the temporal features.

[0026] The temporal features corresponding to the current time step are calculated using an adaptive gated AGU network to obtain gating information. The temporal features corresponding to the current time step are weighted and fused according to the gating information. The subsequence features of each subsequence are obtained based on all time steps of each subsequence and the weighted fusion calculation results.

[0027] The subsequence features of all subsequences are sorted in chronological order to obtain a long-term time-series feature matrix.

[0028] In some embodiments, the method for calculating the behavioral deviation based on the Mahalanobis distance and the cosine similarity is shown in the following formula:

[0029]

[0030] In the formula, D is the behavioral deviation; α is the weighting coefficient; d is the Mahalanobis distance; and s is the cosine similarity.

[0031] In some embodiments, implementing a risk-based intervention strategy based on the degree of behavioral deviation includes:

[0032] When the deviation of the behavior is greater than or equal to the first deviation threshold and less than the second deviation threshold, a supplementary verification question for the loan application is generated.

[0033] When the deviation of the behavior is greater than or equal to the second deviation threshold and less than the third deviation threshold, iris scanning and blink detection are initiated.

[0034] When the deviation of the behavior is greater than or equal to the third deviation threshold, the loan application process will be frozen, anti-fraud education videos will be pushed, and a risk marker will be generated.

[0035] Secondly, a dynamic intervention system for risky loan applications is provided, including:

[0036] The data acquisition module is used to acquire loan application operation data of target users;

[0037] The initial intervention module is communicatively connected to the data acquisition module and is used to make risk assessments in real time based on the loan application operation data, and to make initial dynamic interventions based on the risk assessment results.

[0038] The deviation module, communicatively connected to the initial intervention module, is used to extract spatiotemporal features from the loan application operation data using a spatiotemporally attention-enhanced ST-GCN network, and to fuse a bidirectional GRU network and an adaptive gated AGU network to capture long-term temporal dependencies, thereby obtaining the behavioral deviation; and,

[0039] The risk grading intervention module is communicatively connected to the deviation module and is used to implement risk grading intervention strategies based on the behavioral deviation.

[0040] Thirdly, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, it implements the dynamic intervention method for risky loan applications as described above.

[0041] Fourthly, an electronic device is provided, including a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that the processor, when running the computer program, implements the dynamic intervention method for risky loan applications as described above.

[0042] Compared with the prior art, the advantages of the present invention are as follows: by fusing the spatiotemporal graph convolutional network ST-GCN and the adaptive gated AGU, high-precision calculation of loan application behavior deviation is achieved by extracting spatiotemporal features and capturing long-term dependencies, solving the problems of insufficient multi-dimensional feature fusion and weak long-term dependency capture capability in the prior art. Furthermore, a risk classification intervention strategy is implemented through behavior deviation to dynamically intervene in risky loan applications in a timely manner. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating a dynamic intervention method for risky loan applications according to the present invention.

[0044] Figure 2 This is a schematic diagram of the structure of a dynamic intervention system for risky loan applications according to the present invention. Detailed Implementation

[0045] Referring now to specific embodiments of the invention, examples of which are illustrated in the accompanying drawings. Although the invention will be described in conjunction with specific embodiments, it will be understood that it is not intended to limit the invention to the described embodiments. Rather, it is intended to cover variations, modifications, and equivalents included within the spirit and scope of the invention as defined by the appended claims. It should be noted that the method steps described herein can be implemented by any functional block or functional arrangement, and any functional block or functional arrangement can be implemented as a physical entity or a logical entity, or a combination of both.

[0046] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] Note: The examples described below are merely specific examples and are not intended to limit the embodiments of the present invention to the specific steps, values, conditions, data, order, etc. Those skilled in the art can utilize the concept of the present invention to construct more embodiments not mentioned herein by reading this specification.

[0048] See Figure 1 The diagram shown is a flowchart of a dynamic intervention method for risky loan applications.

[0049] Step S100: Obtain the loan application operation data of the target user;

[0050] When applying for a loan, users typically go through six stages: registration and login, identity verification, document submission, credit assessment, electronic signing, and post-loan management. Process nodes can be tracked at each stage. For example, real-time capture of cursor movement heatmaps, page dwell time distribution, and biometric authentication intervals are all characteristic features of loan application behavior.

[0051] Step S200: Perform risk assessment in real time based on the loan application operation data, and conduct initial dynamic intervention based on the risk assessment results.

[0052] When the target user's income editing time is detected to be greater than or equal to the industry benchmark value μ+3σ (the financial industry sets μ=6s, σ=2s, where the probability of tampering and spoofing is relatively high), and the facial recognition glabellar muscle contraction intensity E... micro If the risk level is greater than or equal to the preset strength of 0.6, or if the target user applies for a loan more than 3 times within a preset time of 8 hours, or if the target user's voiceprint base frequency volatility is greater than or equal to the preset volatility of ≥20% and the voiceprint hash collision rate is greater than or equal to the preset collision rate of 85%, then it is considered high risk.

[0053] Based on the high-risk assessment, the loan application process is frozen, anti-fraud education videos are pushed to users, and risk markers are generated.

[0054] Step S300: The ST-GCN (Spatio-Temporal Graph Convolutional Network) network with enhanced spatio-temporal attention extracts spatio-temporal features from the loan application operation data, and fuses a bidirectional Gated Recurrent Unit Network (GRU) and an adaptive Gated Unit Network (AGU) to capture long-term temporal dependencies, thereby obtaining the behavioral deviation. This step specifically includes the following:

[0055] Step S310: Use a dynamic graph construction method to construct a dynamic graph for each operation event in the loan application operation data according to the operation time window, and generate a dynamic heterogeneous graph;

[0056] Using the target user ID as an index, each operation event in the loan application operation data is abstracted into a graph node. The node attributes include operation type, device embedding vector, grid-encoded geographical location, operation timestamp, etc. An edge is defined as a continuous operation of the same user within a time window. The edge weight consists of two parts: temporal correlation and spatial correlation, ultimately forming a dynamic heterogeneous graph G=(V,E), where V is the set of nodes and E is the set of edges.

[0057] Step S320: Use a two-layer convolutional network to extract the spatiotemporal node features in the dynamic heterogeneous graph, and use a spatiotemporal attention mechanism to calculate the spatiotemporal correlation of the spatiotemporal node features to obtain the spatiotemporal feature matrix.

[0058] First: Calculate the Laplacian matrix of the dynamic heterogeneous graph, and perform a convolution operation on the Laplacian matrix using the first convolutional layer of a two-layer convolutional network to obtain the first-layer feature matrix X. (1) ;

[0059] Second: The first layer feature matrix X is extracted in parallel using the second multi-scale dilated convolution kernel in the two-layer convolutional network. (1) The multi-scale features in the first layer are concatenated to obtain the second-layer multi-scale feature matrix X. (2) ;

[0060] For each node i, features are extracted using dilated convolution kernels with kernel sizes k∈{3,5,7}:

[0061]

[0062] In the formula, N(i) is the set of neighboring nodes of node i under kernel size k;

[0063] Wk (2) ∈R F1×F1 This is the learnable weight matrix corresponding to each core.

[0064] By concatenating features of three kernel sizes, multi-scale features are obtained:

[0065] X (2) =concat(X3 (2) X5 (2) X7 (2) )∈R N×3F1

[0066] Third: Using the spatiotemporal attention mechanism, the spatiotemporal feature matrix is ​​obtained based on the first layer feature matrix and the second layer multi-scale feature matrix.

[0067] The spatiotemporal correlation between node i and its neighboring node j is calculated as follows:

[0068]

[0069] In the formula, h i (1) The first-layer feature matrix of node i; time difference spatial distance Wa is the learnable weight matrix; This is the normalization factor.

[0070] Then, based on the spatiotemporal correlation α ij The neighboring nodes of each node's features are weighted and aggregated to update each node's features, as shown in the following formula:

[0071]

[0072] In the formula, W b b b It is a learnable matrix;

[0073] Finally, the second-layer multi-scale feature X is... (2) With attention-enhanced node features h i ST Perform convolutional concatenation to output a high-dimensional spatiotemporal feature matrix H. ST .

[0074] The final output spatiotemporal feature matrix H ST each line h i ST It represents the spatiotemporal joint features of the i-th operation event, which integrates local spatiotemporal patterns, multi-scale long-range dependencies, and key correlation information, providing rich input features for subsequent long-term time series modeling.

[0075] Step S330: Sort the spatiotemporal feature matrix according to the operation timestamp to generate an operation sequence S. u ={e1 u e2 u ,...,e T u}, where T is the total number of operations, u is the target user, and the operation sequence is divided into multiple subsequences {s1} by a sliding window. u ,s2 u ,...,s M u}

[0076] Step S340: Use a bidirectional GRU network to extract time-series features from each subsequence, and use an adaptive gated AGU network to filter out long-term key information from the extracted time-series features to obtain the long-term feature matrix of all subsequences.

[0077] The forward temporal sequence of each subsequence is processed using the forward GRU network in the bidirectional GRU network, and the reverse temporal sequence of each subsequence is processed using the backward GRU network in the bidirectional GRU network. The temporal sequence and the reverse temporal sequence are then fused to obtain the temporal feature h. t .

[0078] To filter long-term key information, an AGU adaptive gating AGU network is added after the bidirectional GRU network.

[0079] The temporal feature g corresponding to the current time step is calculated using an adaptive gated AGU network. t According to the gate information g t For the time series feature h t Perform weighted fusion calculations to obtain the fused temporal feature h corresponding to the current time step of each subsequence. t The specific calculation formula is as follows:

[0080]

[0081] In the formula, h t The temporal characteristics of the current time step. This represents the temporal characteristics of the previous time step. This operation dynamically adjusts the weights of historical and current information, preserving crucial contextual information that is critical to long-term temporal dependencies.

[0082] For each subsequence s k u For all time steps t=1,2,...,L, the fused temporal features obtained by using an adaptive gated AGU network are as follows:

[0083] .

[0084] Then, a bidirectional pooling strategy is used to calculate the fused temporal features h′ of all time steps to obtain the subsequence features of M subsequences.

[0085] Finally, the subsequence features of all subsequences are sorted in chronological order to obtain the long-term time-series feature matrix: .

[0086] Therefore, in the above steps, loan application operation data can be abstracted into a dynamic heterogeneous graph containing multi-dimensional attributes such as time, space, and equipment, solving the problem of fragmented spatiotemporal correlation modeling in traditional methods; then, a spatiotemporal attention mechanism is introduced to adaptively capture long-distance spatiotemporal dependencies, improving the representation ability of key operation modes; then, spatiotemporal patterns of different time granularities are extracted in parallel through multi-scale feature convolution kernels to enhance the discriminativeness of features; at the same time, an adaptive gating unit is added on the basis of bidirectional GRU to filter key long-term time-series information and alleviate the gradient vanishing problem.

[0087] Step S350: Align and standardize the spatiotemporal feature matrix with the long-term feature matrix to obtain aligned features, and calculate Mahalanobis distance and cosine similarity for the aligned features respectively.

[0088] Step S360, the method for calculating the behavioral deviation degree based on the Mahalanobis distance and the cosine similarity is shown in the following formula:

[0089]

[0090] In the formula, D is the behavioral deviation; α is the weighting coefficient; d is the Mahalanobis distance; and s is the cosine similarity.

[0091] Step S400: Implement a risk-based intervention strategy based on the degree of behavioral deviation.

[0092] When the deviation of the behavior is greater than or equal to the first deviation threshold of 0.5 and less than the second deviation threshold of 0.7, a supplementary verification question for the loan application is generated.

[0093] When the deviation of the behavior is greater than or equal to the second deviation threshold of 0.7 and less than the third deviation threshold of 0.9, iris scanning and blink detection are initiated.

[0094] When the deviation of the behavior is greater than or equal to the third deviation threshold of 0.9, the loan application process will be frozen, anti-fraud education videos will be pushed, and a risk marker will be generated.

[0095] Therefore, this invention integrates the spatiotemporal graph convolutional network ST-GCN with the adaptive gated AGU, and achieves high-precision calculation of loan application behavior deviation by extracting spatiotemporal features and capturing long-term dependencies. This solves the problems of insufficient multi-dimensional feature fusion and weak long-term dependency capture in the prior art. Furthermore, it implements a risk classification intervention strategy through behavior deviation to dynamically intervene in risky loan applications in a timely manner.

[0096] See also Figure 2 As shown, this embodiment of the invention also provides a dynamic intervention system for risky loan applications, including:

[0097] The data acquisition module is used to acquire loan application operation data of target users;

[0098] The initial intervention module is communicatively connected to the data acquisition module and is used to make risk assessments in real time based on the loan application operation data, and to make initial dynamic interventions based on the risk assessment results.

[0099] The deviation module, communicatively connected to the initial intervention module, is used to extract spatiotemporal features from the loan application operation data using a spatiotemporally attention-enhanced ST-GCN network, and to fuse a bidirectional GRU network and an adaptive gated AGU network to capture long-term temporal dependencies, thereby obtaining the behavioral deviation; and,

[0100] The risk grading intervention module is communicatively connected to the deviation module and is used to implement risk grading intervention strategies based on the behavioral deviation.

[0101] This invention integrates the spatiotemporal graph convolutional network ST-GCN with the adaptive gated AGU to achieve high-precision calculation of loan application behavior deviation by extracting spatiotemporal features and capturing long-term dependencies. This solves the problems of insufficient multi-dimensional feature fusion and weak long-term dependency capture in existing technologies. Furthermore, it implements a risk classification intervention strategy based on behavior deviation to dynamically intervene in risky loan applications in a timely manner.

[0102] Specifically, this embodiment corresponds one-to-one with the above method embodiments. The functions of each module have been described in detail in the corresponding method embodiments, so they will not be repeated here.

[0103] Based on the same inventive concept, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements all or part of the method steps of the above method.

[0104] The present invention can implement all or part of the processes in the above methods, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0105] Based on the same inventive concept, embodiments of this application also provide an electronic device, including a memory and a processor. The memory stores a computer program that runs on the processor. When the processor executes the computer program, it implements all or part of the method steps described above.

[0106] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the computer device, connecting all parts of the computer device through various interfaces and lines.

[0107] Memory can be used to store computer programs and / or modules. The processor performs various functions of the computer device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can primarily include a program storage area and a data storage area. The program storage area can store the operating system and at least one application program required for a function (e.g., sound playback, image playback, etc.); the data storage area can store data created based on the use of the mobile phone (e.g., audio data, video data, etc.). Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMedia Cards (SMC), Secure Digital (SD) cards, Flash Cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0108] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, servers, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0109] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), servers, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can 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, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0111] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0112] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A dynamic intervention method for risky loan applications, characterized in that, include: Obtain loan application data from target users; Risk assessment is performed in real time based on the loan application operation data, and initial dynamic intervention is carried out based on the risk assessment results; The loan application operation data is extracted using a spatiotemporal attention-enhanced ST-GCN network, and a bidirectional GRU network and an adaptive gated AGU network are fused to capture long-term temporal dependencies, thus obtaining the behavior deviation degree. Implement risk-based intervention strategies based on the degree of behavioral deviation; The ST-GCN network, which utilizes spatiotemporal attention enhancement, extracts spatiotemporal features from the loan application operation data and integrates a bidirectional GRU network with an adaptive gated AGU network to capture long-term temporal dependencies, thereby obtaining behavioral deviation, including: A dynamic graph construction method is used to construct a dynamic heterogeneous graph for each operation event in the loan application operation data according to the operation time window. Spatiotemporal node features in the dynamic heterogeneous graph are extracted using a two-layer convolutional network, and the spatiotemporal correlation of the spatiotemporal node features is calculated using a spatiotemporal attention mechanism to obtain a spatiotemporal feature matrix. The spatiotemporal feature matrix is ​​sorted according to the operation timestamp to generate an operation sequence, and the operation sequence is divided into multiple sub-sequences by a sliding window. A bidirectional GRU network is used to extract temporal features from each of the subsequences, and an adaptive gated AGU network is used to filter out long-term key information from the extracted temporal features to obtain the long-term feature matrix of all subsequences. The spatiotemporal feature matrix and the long-time series feature matrix are aligned and standardized to obtain aligned features. Mahalanobis distance and cosine similarity are calculated for the aligned features respectively. The behavioral deviation is calculated based on the Mahalanobis distance and the cosine similarity. The process involves using a bidirectional GRU network to extract temporal features from each subsequence, and then using an adaptive gated AGU network to filter out long-term key information from the extracted temporal features, resulting in a long-term feature matrix for all subsequences, including: The forward sequence of each subsequence is processed using the forward GRU network in the bidirectional GRU network, and the reverse sequence of each subsequence is processed using the backward GRU network in the bidirectional GRU network. The forward sequence and the reverse sequence are then fused to obtain the temporal features. The temporal features corresponding to the current time step are calculated using an adaptive gated AGU network to obtain gating information. The temporal features corresponding to the current time step are weighted and fused according to the gating information. The subsequence features of each subsequence are obtained based on all time steps of each subsequence and the weighted fusion calculation results. The subsequence features of all subsequences are sorted in chronological order to obtain a long-term time-series feature matrix.

2. The dynamic intervention method for risky loan applications as described in claim 1, characterized in that, The real-time risk assessment based on the loan application operation data, and the initial dynamic intervention based on the risk assessment results, include: If the target user's income editing time is greater than or equal to the industry benchmark value, and the facial recognition glabellar muscle contraction intensity is greater than or equal to the preset intensity, or the target user applies for a loan more than a preset number of times within a preset time, or the target user's voiceprint fundamental frequency volatility is greater than or equal to the preset volatility, and the voiceprint hash collision rate is greater than or equal to the preset collision rate, then it is judged as high risk. Based on the high-risk assessment, the loan application process is frozen, anti-fraud education videos are pushed to users, and risk markers are generated.

3. The dynamic intervention method for risky loan applications as described in claim 1, characterized in that, The process involves extracting spatiotemporal node features from the dynamic heterogeneous graph using a two-layer convolutional network and calculating the spatiotemporal correlation of these features using a spatiotemporal attention mechanism to obtain a spatiotemporal feature matrix, including: Calculate the Laplacian matrix of the dynamic heterogeneous graph, and perform a convolution operation on the Laplacian matrix using the first convolutional network in a two-layer convolutional network to obtain the first layer feature matrix; The multi-scale features in the first-layer feature matrix are extracted in parallel using the second-layer multi-scale dilated convolution kernel in a two-layer convolutional network and then concatenated to obtain the second-layer multi-scale feature matrix. The spatiotemporal feature matrix is ​​obtained by using the spatiotemporal attention mechanism based on the first layer feature matrix and the second layer multi-scale feature matrix.

4. The dynamic intervention method for risky loan applications as described in claim 1, characterized in that, The method for calculating the behavioral deviation based on the Mahalanobis distance and the cosine similarity is shown in the following formula: In the formula, D is the behavioral deviation; α is the weighting coefficient; d is the Mahalanobis distance; and s is the cosine similarity.

5. The dynamic intervention method for risky loan applications as described in claim 1, characterized in that, The risk-based intervention strategy based on the behavioral deviation includes: When the deviation of the behavior is greater than or equal to the first deviation threshold and less than the second deviation threshold, a supplementary verification question for the loan application is generated. When the deviation of the behavior is greater than or equal to the second deviation threshold and less than the third deviation threshold, iris scanning and blink detection are initiated. When the deviation of the behavior is greater than or equal to the third deviation threshold, the loan application process will be frozen, anti-fraud education videos will be pushed, and a risk marker will be generated.

6. A dynamic intervention system for risky loan applications, characterized in that, include: The data acquisition module is used to acquire loan application operation data of target users; The initial intervention module is communicatively connected to the data acquisition module and is used to make risk assessments in real time based on the loan application operation data, and to make initial dynamic interventions based on the risk assessment results. The deviation module, which is communicatively connected to the initial intervention module, is used to extract spatiotemporal features from the loan application operation data using a spatiotemporally attention-enhanced ST-GCN network, and to fuse a bidirectional GRU network and an adaptive gated AGU network to capture long-term temporal dependencies, thereby obtaining the behavioral deviation. as well as, The risk grading intervention module is communicatively connected to the deviation module and is used to implement risk grading intervention strategies based on the behavioral deviation. The ST-GCN network, which utilizes spatiotemporal attention enhancement, extracts spatiotemporal features from the loan application operation data and integrates a bidirectional GRU network with an adaptive gated AGU network to capture long-term temporal dependencies, thereby obtaining behavioral deviation, including: A dynamic graph construction method is used to construct a dynamic heterogeneous graph for each operation event in the loan application operation data according to the operation time window. Spatiotemporal node features in the dynamic heterogeneous graph are extracted using a two-layer convolutional network, and the spatiotemporal correlation of the spatiotemporal node features is calculated using a spatiotemporal attention mechanism to obtain a spatiotemporal feature matrix. The spatiotemporal feature matrix is ​​sorted according to the operation timestamp to generate an operation sequence, and the operation sequence is divided into multiple sub-sequences by a sliding window. A bidirectional GRU network is used to extract temporal features from each of the subsequences, and an adaptive gated AGU network is used to filter out long-term key information from the extracted temporal features to obtain the long-term feature matrix of all subsequences. The spatiotemporal feature matrix and the long-time series feature matrix are aligned and standardized to obtain aligned features. Mahalanobis distance and cosine similarity are calculated for the aligned features respectively. The behavioral deviation is calculated based on the Mahalanobis distance and the cosine similarity. The process involves using a bidirectional GRU network to extract temporal features from each subsequence, and then using an adaptive gated AGU network to filter out long-term key information from the extracted temporal features, resulting in a long-term feature matrix for all subsequences, including: The forward sequence of each subsequence is processed using the forward GRU network in the bidirectional GRU network, and the reverse sequence of each subsequence is processed using the backward GRU network in the bidirectional GRU network. The forward sequence and the reverse sequence are then fused to obtain the temporal features. The temporal features corresponding to the current time step are calculated using an adaptive gated AGU network to obtain gating information. The temporal features corresponding to the current time step are weighted and fused according to the gating information. The subsequence features of each subsequence are obtained based on all time steps of each subsequence and the weighted fusion calculation results. The subsequence features of all subsequences are sorted in chronological order to obtain a long-term time-series feature matrix.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the dynamic intervention method for risky loan applications as described in any one of claims 1 to 5.

8. An electronic device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that, When the processor runs the computer program, it implements the dynamic intervention method for risky loan applications as described in any one of claims 1 to 5.

Images