Risk prediction method and device

Abstract

The embodiment of the invention provides a risk prediction method and device, and the method comprises the steps: obtaining a resource relation network associated with a target service, and inputting the resource relation network into a risk prediction model which comprises a coding unit and an edge calibration unit; extracting structured features and unstructured features of the resource relation network by using the coding unit, and generating edge features of a plurality of resource interaction edges in the resource relation network according to the structured features and the unstructured features; performing calibration processing on the edge features of the plurality of resource interaction edges by using the edge calibration unit to obtain posterior connection probabilities corresponding to the plurality of resource interaction edges respectively; and updating the resource relation network based on the posterior connection probability, and executing a risk prediction task of the target service based on the updated resource relation network.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of computer technology, and in particular to risk prediction methods and apparatus. Background Technology

[0002] With the development of computer and internet technologies, risk prediction technology has been widely applied in e-commerce, fintech, video browsing platforms, and other scenarios. This technology can effectively ensure the normal operation of any business service, avoiding losses while providing stable services to users. Among existing technologies, risk prediction methods based on relationship networks are widely used. This method constructs a user-centric relationship network through user interactions, utilizing the association paths and risk propagation patterns between nodes in the network to identify potentially high-risk users. Compared to traditional models, this method can effectively utilize network topology information, significantly improving the accuracy of user risk classification. However, relationship networks are usually pre-built by service providers according to their needs. Pre-built relationship networks may contain issues such as identity information tampering and inaccurate device information, which will seriously affect the accuracy of risk prediction. Therefore, an effective solution is urgently needed to address these problems. Summary of the Invention

[0003] In view of this, embodiments of this specification provide a risk prediction method. One or more embodiments of this specification also relate to a risk prediction device, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, a risk prediction method is provided, comprising: Obtain the resource relationship network associated with the target business, and input the resource relationship network into the risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit; The structured and unstructured features of the resource relationship network are extracted using the encoding unit, and edge features of multiple resource interaction edges in the resource relationship network are generated based on the structured and unstructured features. The edge features of the plurality of resource interaction edges are calibrated using the edge calibration unit to obtain the posterior connection probabilities corresponding to the plurality of resource interaction edges respectively. The resource relationship network is updated based on the posterior connection probability, and the risk prediction task of the target service is performed based on the updated resource relationship network.

[0005] According to a second aspect of the embodiments of this specification, another risk prediction method is provided, including: A financial relationship network is acquired and input into a risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit; The structured and unstructured features of the financial relationship network are extracted using the encoding unit, and edge features of multiple financial interaction edges in the financial relationship network are generated based on the structured and unstructured features. The edge features of the plurality of financial interaction edges are calibrated using the edge calibration unit to obtain the posterior connection probabilities corresponding to the plurality of financial interaction edges respectively. The financial relationship network is updated based on the posterior connection probability, and a default risk prediction task is performed based on the updated financial relationship network.

[0006] According to a third aspect of the embodiments of this specification, a risk prediction device is provided, comprising: The acquisition module is configured to acquire the resource relationship network associated with the target business and input the resource relationship network into the risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit; The extraction module is configured to extract the structured and unstructured features of the resource relationship network using the encoding unit, and generate edge features of multiple resource interaction edges in the resource relationship network based on the structured and unstructured features. The processing module is configured to use the edge calibration unit to calibrate the edge features of the plurality of resource interaction edges to obtain the posterior connection probabilities corresponding to the plurality of resource interaction edges respectively. The update module is configured to update the resource relationship network based on the posterior connection probability, and perform the risk prediction task of the target service based on the updated resource relationship network.

[0007] According to a fourth aspect of the embodiments of this specification, another risk prediction device is provided, comprising: The network acquisition module is configured to acquire a financial relationship network and input the financial relationship network into a risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit. The feature extraction module is configured to extract structured and unstructured features of the financial relationship network using the encoding unit, and generate edge features of multiple financial interaction edges in the financial relationship network based on the structured and unstructured features. The feature processing module is configured to use the edge calibration unit to calibrate the edge features of the plurality of financial interaction edges to obtain the posterior connection probabilities corresponding to the plurality of financial interaction edges respectively. The network update module is configured to update the financial relationship network based on the posterior connection probability and perform a default risk prediction task based on the updated financial relationship network.

[0008] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the risk prediction method described above.

[0009] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the risk prediction method described above.

[0010] According to a seventh aspect of the embodiments of this specification, a computer program product is provided, including a computer program or instructions that, when executed by a processor, implement the steps of the risk prediction method described above.

[0011] The risk prediction method provided in this embodiment aims to effectively identify and calibrate relationship networks under conditions of information tampering and noise interference, thereby improving the accuracy of risk prediction. After obtaining the resource relationship network associated with the target business, the resource relationship network is first input into the risk prediction model. The risk prediction model, including an encoding unit and an edge calibration unit, calibrates the resource relationship network and performs subsequent risk prediction. At this stage, the encoding unit first extracts the structured and unstructured features of the resource relationship network. Then, based on these structured and unstructured features, edge features of multiple resource interaction edges in the resource relationship network are generated. This allows the edge features of each resource interaction edge to integrate graph-level and node-level representations, fully reflecting its characteristic information within the relationship network. After obtaining the edge features, the edge calibration unit can be used to calibrate the edge features of multiple resource interaction edges. This calibration process determines the posterior connection probabilities corresponding to each resource interaction edge. The posterior connection probabilities characterize the credibility of each resource interaction edge in the resource interaction network. Therefore, the resource relationship network can be updated based on the posterior connection probabilities to correct the resource interaction network, enabling it to accurately represent the real interaction relationships in the current scenario. Subsequently, risk prediction tasks for the target business can be performed based on the updated resource relationship network, thereby effectively improving the accuracy of risk prediction in business scenarios, reducing interference caused by information tampering and noise, and effectively improving business stability. Attached Figure Description

[0012] Figure 1This is a flowchart of a risk prediction method provided in one embodiment of this specification; Figure 2 This is a schematic diagram of a risk prediction model in a risk prediction method provided in one embodiment of this specification; Figure 3 This is a flowchart of another risk prediction method provided in one embodiment of this specification; Figure 4 This is a flowchart illustrating the processing procedure of a risk prediction method provided in one embodiment of this specification; Figure 5 This is a schematic diagram of the structure of a risk prediction device provided in one embodiment of this specification; Figure 6 This is a schematic diagram of another risk prediction device provided in one embodiment of this specification; Figure 7 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0013] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0014] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0015] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0016] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0017] The technical solutions provided in this application can employ deep learning models with relatively large parameter scales. However, this large model is merely an example; this application does not limit the number of model parameters supported by the deep learning model used, aiming to meet actual needs. The deep learning models involved in this application can be artificial intelligence-based language models (LM) or multimodal models (MM). First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0018] In variational inference, a variational distribution is used to approximate a simplified probability distribution of a complex posterior probability distribution that is difficult to compute directly. Typically, an easily tractable family of distributions (such as the Gaussian distribution) is chosen, and parameters are optimized to make it as close as possible to the true posterior, thereby achieving efficient inference.

[0019] Graph Neural Networks (GNNs) are a class of deep learning models specifically designed for processing graph-structured data. GNNs learn representations of nodes, edges, or the entire graph by passing and aggregating information between nodes and edges. They are widely used in fields such as social network analysis, recommender systems, and financial risk control.

[0020] A multilayer perceptron (MLP) is a type of feedforward artificial neural network consisting of multiple fully connected layers, each containing several neurons connected by a nonlinear activation function. MLPs can learn complex nonlinear relationships between inputs and outputs and are commonly used for tasks such as classification and regression, as well as as components of other models (such as decoders).

[0021] Graph Convolutional Networks (GCNs) are a type of graph neural network that extends convolutional operations to graph structures, weighting and aggregating neighbor information to update node representations. GCNs perform exceptionally well in tasks such as node classification and link prediction, and are one of the fundamental models for processing graph data.

[0022] A Financial Relational Network (FRN) is a graph-structured network used to model complex relationships such as fund transfers, guarantees, and equity among financial institutions, businesses, or individuals. Nodes in an FRN represent financial entities, and edges represent the interactions between them. It can be used in financial applications such as risk propagation analysis, fraud detection, and systemic risk assessment.

[0023] Bayes' formula, also known as Bayes' theorem or Bayes' rule, was originally used to describe the relationship between the conditional probabilities of two events.

[0024] To address the aforementioned technical problems, this specification provides a risk prediction method. This specification also relates to a risk prediction device, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0025] In practical applications, graph neural networks are widely used for node classification tasks in relational networks due to their advantages in modeling complex relationships and aggregating neighbor information. They achieve representation learning of the central node by aggregating feature information from neighboring nodes along higher-order connection paths, where the strength of the connections directly affects the model's classification performance. In real-world relational networks, connections are typically categorized as strong or weak. Strong connections are built upon reliable user identification (such as ID numbers or device serial numbers), reflecting genuine and close interactions, and are highly valuable for risk prediction. Weak connections, on the other hand, rely on less identifiable information (such as IP addresses or GPS locations). While they can serve as supplementary information to maintain network connectivity, they are susceptible to noise interference.

[0026] However, in real-world resource allocation scenarios, the reliability of both strong and weak connections can be challenged. On one hand, some users may tamper with their identity information through irregular means to access resources, thus undermining the trust foundation of strong connections. On the other hand, weak connections, due to their low identification accuracy, often generate incorrect associations due to location errors or shared network environments, introducing a large amount of noise. This, in turn, severely impacts the accuracy of risk prediction in business scenarios. Therefore, an effective solution is urgently needed to address these issues.

[0027] The risk prediction method provided in this embodiment aims to effectively identify and calibrate relationship networks under conditions of information tampering and noise interference, thereby improving the accuracy of risk prediction. After obtaining the resource relationship network associated with the target business, the resource relationship network is first input into the risk prediction model. The risk prediction model, including an encoding unit and an edge calibration unit, calibrates the resource relationship network and performs subsequent risk prediction. At this stage, the encoding unit first extracts the structured and unstructured features of the resource relationship network. Then, based on these structured and unstructured features, edge features of multiple resource interaction edges in the resource relationship network are generated. This allows the edge features of each resource interaction edge to integrate graph-level and node-level representations, fully reflecting its characteristic information within the relationship network. After obtaining the edge features, the edge calibration unit can be used to calibrate the edge features of multiple resource interaction edges. This calibration process determines the posterior connection probabilities corresponding to each resource interaction edge. The posterior connection probabilities characterize the credibility of each resource interaction edge in the resource interaction network. Therefore, the resource relationship network can be updated based on the posterior connection probabilities to correct the resource interaction network, enabling it to accurately represent the real interaction relationships in the current scenario. Subsequently, risk prediction tasks for the target business can be performed based on the updated resource relationship network, thereby effectively improving the accuracy of risk prediction in business scenarios, reducing interference caused by information tampering and noise, and effectively improving business stability.

[0028] See Figure 1 , Figure 1 A flowchart of a risk prediction method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0029] Step S102: Obtain the resource relationship network associated with the target business, and input the resource relationship network into the risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit.

[0030] The risk prediction method provided in this embodiment can be applied to any business service project where risk prediction is based on a relationship network. For example, in a resource allocation scenario, the relationship network built in this scenario can be used to predict the risk of resource demanders (predicting the risk of resource demanders occupying computing resources and the risk of not returning resources). In another example, in a goods rental scenario, the relationship network built in this scenario can be used to predict the risk of lessees (predicting the credit risk of lessees). In yet another example, in an insurance scenario, the relationship network built in this scenario can be used to predict the risk of insured persons (predicting the risk of insured persons committing insurance fraud).

[0031] Specifically, the target business refers to the service items that the user is currently participating in, including but not limited to computing resource usage, item rental, and insurance services. This embodiment does not impose any limitations on these services. Correspondingly, the resource relationship network is a graph network constructed based on user nodes and the resource interaction relationships between them within the target business. User nodes are nodes constructed on a user or business participant basis, and resource interaction relationships are the interaction relationships between users or business participants, such as social relationships (e.g., friend relationships on social platform A), transactional relationships (e.g., buying and selling item B), transfer relationships (e.g., a cooperative relationship between Party A and Party B), and device association relationships (e.g., devices with the same serial number). It should be noted that the resource relationship network used in different business scenarios can be customized according to business needs, and this embodiment does not impose any limitations on this. Accordingly, the risk prediction model specifically refers to a model that calibrates the resource relationship network and is used for risk prediction processing of target business. It includes at least an encoding unit and a side calibration unit. The encoding unit is used to extract the structured and unstructured features of the resource relationship network and to calibrate the resource relationship network. The side calibration unit is specifically a unit that calibrates the resource relationship network to reduce the impact of noise and information tampering in the resource relationship network, so as to ensure that the calibrated network is used to perform risk prediction tasks in the future.

[0032] Therefore, in order to effectively identify and calibrate relationship networks under conditions of information tampering and noise interference, and thereby improve the accuracy of risk prediction, after obtaining the resource relationship network associated with the target business, the resource relationship network can be first input into the risk prediction model. The risk prediction model, which includes encoding units and edge calibration units, can then calibrate the resource relationship network and perform subsequent risk prediction. At this point, the encoding unit can first extract the structured and unstructured features of the resource relationship network. Then, based on these structured and unstructured features, edge features of multiple resource interaction edges in the resource relationship network can be generated. This allows the edge features of each resource interaction edge to integrate graph-level and node-level representations, fully reflecting its characteristic information within the relationship network. After obtaining the edge features, the edge calibration unit can be used to calibrate the edge features of multiple resource interaction edges. This calibration process determines the posterior connection probabilities corresponding to each resource interaction edge. The posterior connection probabilities characterize the credibility of each resource interaction edge in the resource interaction network. Therefore, the resource relationship network can be updated based on the posterior connection probabilities to correct the resource interaction network, enabling it to accurately represent the real interaction relationships in the current scenario. Subsequently, risk prediction tasks for the target business can be performed based on the updated resource relationship network, thereby effectively improving the accuracy of risk prediction in business scenarios, reducing interference caused by information tampering and noise, and effectively improving business stability.

[0033] See Figure 2 The schematic diagram shown illustrates that in the risk prediction method provided in this embodiment, the resource relationship network can be defined as an undirected graph G={V, E}, where V= Let E represent a set containing N user nodes, and E⊆V × V be the corresponding set of edges, representing the resource interaction relationships between users. Let X∈ Let Y be the D-dimensional feature matrix of the nodes, and Y = {0, 1} be the node category label, where... =1 indicates user There is a risk of resource interaction. =0 indicates normal. To model the reliability of connections, a mask matrix M can be introduced to represent the prior strength of the connections between nodes. and The initial connection strength between them is determined by the prior probability p(M) ij =1) determines when p(M) ij When p(M) = 1 and p(M) = α (0 ≤ α ≤ 1), it indicates that the two are associated through a strong connection, while p(M) = 1 and p(M) = α (0 ≤ α ≤ 1). ij =0) =1-α means that the two are only connected by a weak connection. Therefore, α can be regarded as the prior probability of the existence of a strong connection, reflecting the platform's initial trust in the connection built by different identity media.

[0034] Based on this, by inputting the resource relationship network into the risk prediction model, structured and unstructured features can be extracted during the encoding stage using graph neural networks and multilayer perceptrons respectively. Graph neural networks can capture the topological dependencies between nodes in the graph, while multilayer perceptrons focus more on processing independent representations based on the features of the nodes themselves, which can avoid the influence of structural bias. By fusing the two, a joint representation of each edge in the structure and feature spaces can be obtained.

[0035] Furthermore, the model can estimate and calibrate the actual connection probabilities of strong and weak connections in the resource relationship network based on the joint representation of edges, in order to reconstruct a more reliable resource relationship network. Finally, the risk prediction task of business association can be performed based on the reconstructed resource relationship network, thereby completing node-level risk prediction processing and effectively improving the accuracy of risk prediction in business scenarios.

[0036] Step S104: Use the encoding unit to extract the structured and unstructured features of the resource relationship network, and generate edge features of multiple resource interaction edges in the resource relationship network based on the structured and unstructured features.

[0037] Specifically, after inputting the resource relationship network into the risk prediction model, in order to ensure the accuracy of the model's calibration of the resource relationship network, the encoding unit can be used to extract the structured features expressed by the resource relationship network at the graph structure level, as well as the unstructured features expressed by the resource relationship network at the node level. Then, by fusing the structured and unstructured features, it is possible to both preserve the topological relationships in the relationship network and integrate the independent perception capabilities of the nodes' own attributes. This allows us to obtain the edge features corresponding to multiple resource interaction edges in the resource relationship network. This enables us to combine the edge features that integrate information from different dimensions to calibrate the resource relationship network, ensuring that the calibrated resource relationship network truly reflects the actual user relationships, thereby improving the accuracy of risk prediction.

[0038] Structured features specifically refer to the vector representations generated by extracting graph structure information from the resource relationship network, thus characterizing the structural dependencies between nodes. Unstructured features specifically refer to the vector representations constructed based on the attribute information of the nodes themselves in the resource relationship network, avoiding biases introduced by graph structure noise or spurious connections. Correspondingly, resource interaction edges are the connecting edges between any two nodes in the resource relationship network, with different resource interaction relationships corresponding to different edge attributes. Correspondingly, edge features are the vector representations corresponding to each resource interaction edge.

[0039] Furthermore, to ensure accurate feature representation during the extraction of structured and unstructured features, different coding sub-units can be used. In this embodiment, the extraction of structured and unstructured features of the resource relationship network using the coding unit includes: The resource relationship network is input into the graph encoding subunit and node encoding subunit of the encoding unit; the graph encoding subunit processes the resource relationship network to obtain graph features corresponding to the graph network dimension, and the graph features are used as structured features; the node encoding subunit processes the resource relationship network to obtain node features corresponding to the node dimension, and the node features are used as unstructured features.

[0040] Specifically, the graph coding subunit refers to an encoder that can extract graph structure information to construct structured features, which can be implemented by a graph neural network (GNN); the node coding subunit refers to an encoder that can extract node information to construct unstructured features, which can be implemented by a multilayer perceptron (MLP); the graph features of the graph network dimension are vector representations constructed from the graph level, and the node features are vector representations constructed from the node level.

[0041] Based on this, when constructing structured and unstructured features, the resource relationship network can be first input into the graph coding sub-unit and node coding sub-unit in the coding unit. At this time, the graph coding sub-unit can be used to process the resource relationship network to obtain the graph features corresponding to the graph network dimension, and the graph features are used as structured features. At the same time, the node coding sub-unit can be used to process the resource relationship network to obtain the node features corresponding to the node dimension, and the node features are used as unstructured features.

[0042] In practical applications, for any resource relationship network containing strong and weak connections, an adjacency matrix A∈{0,1} can be defined. N*N , where M ij When A > 0, ij =1 indicates a node and A connection exists, otherwise A ij =0. Based on this, assume that in A ij Given that g = 1, the edges of the resource relationship network form the topological structure of the graph. Therefore, let g = 1. E As a graph encoding subunit (GNN Encoder), it can extract graph structure information through message passing mechanism, thereby obtaining a graph representation that can characterize the structural dependencies between nodes.

[0043] On the other hand, in A ij When f = 0, we can assume that the nodes are independent and identically distributed, meaning that feature modeling does not rely on graph structure. In this case, a multilayer perceptron (MLP) can be used as the node encoding subunit f. E This enables unstructured representation that relies solely on the features of the nodes themselves, thereby avoiding biases introduced by graph structure noise or spurious connections.

[0044] In practice, for each node Their graph structure representations can be obtained respectively. ∈ (Graph features corresponding to each node), and feature-independent representations. ∈ (The node features corresponding to each node), the corresponding calculation can be achieved through the following formula (1): =g E (A, x) i ; ), =f E (x) i ; (1) in, and These are the learnable parameters for the graph coding subunit and the node coding subunit, respectively. This dual-path coding structure enables the model to capture topological dependencies in relational networks while retaining independent awareness of node attributes, providing complementary information for subsequent reliable connection evaluation.

[0045] In summary, to comprehensively understand user relationships, the model employs a dual-path encoding mechanism. On one hand, it utilizes graph encoding subunits to analyze a user's location within the network and their neighbor relationships, capturing social risk propagation patterns formed by interactions and resource flows. This structured learning approach effectively leverages high-confidence connections to transmit risk signals. On the other hand, considering that some connections may be distorted due to abnormal means or noise, the model simultaneously uses node encoding subunits to independently model each user's unique characteristics, independent of any network structure. This approach avoids biases caused by spurious connections and retains the ability to objectively assess individual creditworthiness. Through this structure, the model not only grasps the contextual information of users within the relationship network but also retains sensitivity to their independent attributes, providing complementary evidence for subsequent connection credibility assessments.

[0046] Furthermore, the fusion of structured and unstructured features can be based on variational distribution. In this embodiment, generating edge features of multiple resource interaction edges in the resource relationship network based on the structured and unstructured features includes: The structured features are divided into a structured mean vector and a structured standard deviation vector, and the unstructured features are divided into an unstructured mean vector and an unstructured standard deviation vector. A structured variational distribution is constructed based on the structured mean vector and the structured standard deviation vector, and an unstructured variational distribution is constructed based on the unstructured mean vector and the unstructured standard deviation vector. The structured variational distribution and the unstructured variational distribution are fused, and the edge features of multiple resource interaction edges in the resource relationship network are generated based on the fusion result.

[0047] Specifically, the structured mean vector and structured standard deviation vector refer to the vector representations of the mean and standard deviation obtained after decomposing structured features, and are used to construct structured variational distributions. Correspondingly, the unstructured mean vector and unstructured standard deviation vector refer to the vector representations of the mean and standard deviation obtained after decomposing unstructured features, and are used to construct unstructured variational distributions.

[0048] Based on this, after obtaining structured and unstructured features, the risk prediction model, in order to fully integrate the feature representations of both dimensions for each resource interaction edge, can first divide the structured features into a structured mean vector and a structured standard deviation vector, and divide the unstructured features into an unstructured mean vector and an unstructured standard deviation vector. Then, a structured variational distribution can be constructed based on the structured mean vector and the structured standard deviation vector, and an unstructured variational distribution can be constructed based on the unstructured mean vector and the unstructured standard deviation vector. The variational distribution facilitates the subsequent calculation of posterior connection probabilities. At this point, the structured and unstructured variational distributions can be fused, and the fusion result generates the edge features of multiple resource interaction edges in the resource relationship network for subsequent use.

[0049] This embodiment uses the process of risk prediction for a user requesting computing resources as an example. Descriptions of other scenarios can refer to the same or corresponding descriptions in this embodiment. Specifically, when user A requests computing resources from a cloud computing service provider, the provider can perform risk prediction for user A. First, a resource relationship network including user A can be determined. This network is constructed with users as nodes and resource interaction relationships between users (such as device serial numbers, ID numbers, IP addresses, GPS locations, etc.) as edges. Based on this, the resource relationship network can be input into a risk prediction model, where a graph neural network (GNN) and a multilayer perceptron (MLP) encode the network. During this process, the GNN can extract graph structure information through a message passing mechanism to obtain a graph representation that depicts the structural dependencies between nodes. The MLP can perform unstructured construction based on the node's own features to obtain a feature-independent representation. Based on this, the graph representation and the feature-independent representation can be fused to obtain the edge features corresponding to each resource interaction edge in the resource relationship network. These edge features can then be used to calibrate the resource relationship network to accurately predict the risk for user A.

[0050] In summary, by fusing graph representation and node representation, edge features can fully reflect their attribute meaning in the resource relationship network, which can effectively improve the accuracy of subsequent resource relationship network calibration.

[0051] Step S106: Use the edge calibration unit to calibrate the edge features of the multiple resource interaction edges to obtain the posterior connection probabilities corresponding to the multiple resource interaction edges respectively.

[0052] Specifically, after obtaining the edge features of each resource interaction edge in the resource relationship network, considering that the resource relationship network may have expressions that do not match the actual situation due to abnormal information tampering by users or noise interference, the edge calibration unit can be used to calibrate the table features of each resource interaction edge before risk prediction. This calibration is used to calculate the posterior connection probability corresponding to each resource interaction edge, so that the resource relationship network can be corrected according to the posterior connection probability, making the resource relationship network match the actual situation, thereby effectively improving the accuracy of risk prediction.

[0053] The calibration process specifically refers to the process of verifying the authenticity of each resource interaction edge. It can calculate the posterior connection probability corresponding to each resource interaction edge. This posterior connection probability can reflect the confidence level of each resource interaction edge, which can facilitate subsequent decisions on whether each resource interaction edge should be retained based on the posterior connection probability, so as to remove resource interaction edges that have been tampered with or have high noise.

[0054] Furthermore, when calculating the posterior connectivity probability, it can be achieved by sampling candidate features and then calculating the confidence level. In this embodiment, the step of using the edge calibration unit to calibrate the edge features of the multiple resource interaction edges to obtain the posterior connectivity probabilities corresponding to the multiple resource interaction edges includes: The edge calibration unit samples a set number of candidate features in a first variational distribution and a second variational distribution according to the edge features of the multiple resource interaction edges; the set number of candidate features are fused, and the confidence level corresponding to each of the multiple resource interaction edges is determined according to the fusion result; the confidence level is used as the posterior connection probability corresponding to each of the multiple resource interaction edges.

[0055] Specifically, the first variational distribution refers to the variational distribution of structured features constructed after the risk prediction model is trained, and the second variational distribution refers to the variational distribution of unstructured features constructed after the risk prediction model is trained. The set number of candidate features refers to the sample features obtained by independently sampling from the two variational distributions. The set number can be selected according to actual needs, and this embodiment does not impose any limitations. Correspondingly, the confidence score is the confidence score corresponding to each resource interaction edge, used to characterize its importance relative to the resource relationship network.

[0056] Based on this, when calculating the posterior connection probability, the edge calibration unit can first sample a set number of candidate features in the first variational distribution and the second variational distribution according to the edge features of multiple resource interaction edges. Since the selected candidate features are all associated with the edge features of resource interaction edges, the set number of candidate features can be fused to determine the confidence level corresponding to multiple resource interaction edges based on the fusion result. Thus, the confidence level can be used as the posterior connection probability corresponding to multiple resource interaction edges for subsequent update processing of the resource relationship network.

[0057] In practical applications, the most suitable set of parameters θ will be obtained after the risk prediction model has been trained. * According to Bayes' theorem, the posterior connection probability p(M=1|Y,X,A) of each resource interaction edge can be approximately calculated using Monte Carlo estimation, and its calculation formula is as follows: Formula (2): p(M=1|Y,X,A)≈ (M=1|Y,X,A, , (2) in, and From the trained variational distribution (z|Y,X,A) and In (z|Y,X), there are s independently sampled samples, where S is the number of samplings. By averaging multiple samples, a more stable and robust connection confidence assessment result can be obtained, thereby completing the reconstruction of each resource interaction edge in the resource relationship network for subsequent risk prediction.

[0058] In practical implementation, when calculating the posterior connection probability for each resource interaction edge, it can be assumed that the nodes are independent and identically distributed. The aim is to learn a probabilistic classification function (edge ​​calibration unit), which is modeled as a time-varying process F with a specific distribution p(f). The marginal distribution of the note Y can be expressed as p(Y|X) = The random function f can be generated by a neural network l φ Approximation, we introduce a set of implicit edge variables z belonging to the density family Z, i.e., f(X) = l φ (X; Z), where the prior distribution p(z) of the latent variable Z is set as a multivariate standard normal distribution. Combining the above representation, we can obtain the following formula (3): P(Y|X) = (3) in, Let be the posterior distribution of the latent variable z, but its analytical form is difficult to calculate directly. Therefore, we can introduce a distribution derived from the parameter φ. l Variational distribution of control (z|Y,X) is used to approximate the posterior. In the variational inference framework, maximizing the log-likelihood logp(Y|X) is equivalent to optimizing its Evidence Lower Bound (ELBO), which is expressed by the following formula (4): logp(Y|X) = ELBO(θ) l , φ l )+KL( )≥ELBO l (θ) l , φ l (4) Among them, the KL divergence term makes the variational distribution approximate the true posterior, and the specific form of the ELBO objective function is as follows: Formula (5): ELBO l (θ) l , φ l )= [logp(Y|X)]-KL( (5) When it is assumed that p(Y|X) follows a binomial distribution, the first term can be regarded as a binary cross-entropy loss.

[0059] Furthermore, considering the influence of the resource relationship network (graph structure A) and the connection mask M, we can obtain the conditional probability: P(Y|X,A,M) = Accordingly, its ELBO can be expressed as the following formula (6): ELBO g (θ) g , φ g )= [logp(Y|X,A,M)]-KL( (6) Wherein, posterior pθ g (z| This can be further expanded to: pθ g (z| )∝p(M|z, p(z|Y,X,A); Based on Bayes' theorem, the posterior connection probability p(M=1|z, This can be expressed as the following formula (7): p(M=1|z, )= (7) in, This can be viewed as an implicit distribution (i.e., effective connections) of "high-confidence resource interaction edges that should be retained," while This corresponds to "low-confidence resource interaction edges should be removed" (i.e., noisy connections), which can be approximated as: Since the true posterior is difficult to calculate, the variational approximation of the Gaussian distribution can be used, as expressed by the following formula (8): ~N(z; μ) g , ); ~N(z; μ) l , (8) Where, μ and Let represent the mean and covariance matrices, respectively.

[0060] According to the above formula, the posterior connection probability of an edge connection can be dynamically updated by fusing node attributes, label information and prior connection probability p (M=1). After obtaining the posterior connection probability of each resource interaction edge, subsequent resource relationship network reconstruction operations can be performed.

[0061] Continuing with the previous example, after inputting the resource relationship network into the risk prediction model, and processing it through a multilayer perceptron and graph neural network, the edge features corresponding to each edge can be obtained. Furthermore, by inputting these edge features into the edge calibration unit within the risk prediction model, the edge calibration unit can calculate the posterior connectivity probability of each resource interaction edge in the resource relationship network according to the aforementioned processing logic. For example, if there are n resource interaction edges in the resource relationship network, and the posterior connectivity probability of each resource interaction edge is p... n Subsequently, the resource relationship network can be reconstructed according to the posterior connection probability of each resource interaction edge.

[0062] In summary, to identify the reliability of each resource interaction edge in the resource relationship network, the model introduces a calibration mechanism based on Bayesian inference. This mechanism comprehensively considers user characteristics, label information, and initial connection types to dynamically calculate the credibility of each connection. For connections initially considered strong but exhibiting abnormal behavior, their weights are reduced; while for weak connections with consistent behavior, their influence is appropriately increased. This allows for the accurate determination of the posterior connection probability for each resource interaction edge, facilitating subsequent reconstruction of the resource relationship network.

[0063] Step S108: Update the resource relationship network based on the posterior connection probability, and perform the risk prediction task of the target service based on the updated resource relationship network.

[0064] Specifically, after obtaining the posterior connection probability corresponding to each resource interaction edge, considering that the posterior connection probability of each resource interaction edge can characterize its reliability in the resource relationship network, it is possible to analyze whether each resource interaction edge in the resource relationship network needs to be retained based on the posterior connection probability. Based on this mechanism, the resource relationship network can be updated. The updated resource relationship network can more accurately and realistically express the resource interaction relationship between users in the target business. Subsequently, the risk prediction task of the target business can be performed based on the updated resource relationship network to improve the accuracy of risk prediction.

[0065] Updating the resource relationship network can be understood as removing resource interaction edges from the network. This process eliminates resource interaction edges that could negatively impact the accuracy of risk prediction, as these edges may be subject to noise or tampering. Correspondingly, the risk prediction task is the task of predicting risk for any user within the target business, used to assess a user's default risk, etc. Risk prediction tasks differ across scenarios, and this embodiment does not impose any limitations.

[0066] Furthermore, the update of the resource relationship network can be achieved by comparing the posterior connection probability with a threshold. In this embodiment, updating the resource relationship network based on the posterior connection probability and performing the risk prediction task of the target service based on the updated resource relationship network includes: The posterior connection probabilities corresponding to the multiple resource interaction edges are compared with a preset probability threshold; based on the comparison results, resource interaction edges with probabilities less than the preset probability threshold are deleted from the resource relationship network, and a target resource relationship network is generated based on the deletion results; the risk prediction task of the target service is performed based on the target resource relationship network.

[0067] Specifically, the preset probability threshold refers to the threshold compared with each posterior connection probability, which can filter out resource interaction edges that need to be retained or removed. Correspondingly, the target resource relationship network is the resource relationship network obtained after reconstructing the resource relationship network, which more closely matches the actual resource interaction relationships.

[0068] Based on this, after obtaining the posterior connection probability corresponding to each resource interaction edge, the posterior connection probabilities corresponding to multiple resource interaction edges can be compared with a preset probability threshold. Then, resource interaction edges with probabilities less than the preset probability threshold can be deleted from the resource relationship network according to the comparison results, thereby generating a target resource relationship network based on the deletion results. After that, risk prediction tasks for the target business can be performed based on the target resource relationship network.

[0069] In practical applications, a threshold δ∈[0,1] can be given. When p(M=1|z,Y,X,A)>δ, the resource interaction edge can be determined as a valid edge connection and retained. Otherwise, the resource interaction edge can be regarded as a low-confidence connection and removed from the resource relationship network. This achieves dynamic calibration and purification of the resource relationship network.

[0070] In summary, by comparing the posterior connection probability with a set threshold, unreasonable resource interaction edges can be accurately eliminated, thereby reconstructing a target resource relationship network that better matches the real resource interaction relationship. Based on this, the accuracy of risk prediction can be effectively improved.

[0071] In practice, different types of risk prediction tasks can be completed using different decoding units in the model. In this embodiment, the risk prediction task for the target service based on the updated resource relationship network includes: Based on the risk prediction task of the target business, a target node is determined in the updated resource relationship network; if the target node is an independent node, the node characteristics of the target node are processed by the node decoding unit in the risk prediction model to obtain the risk prediction information corresponding to the target node; if the target node is not an independent node, the node characteristics of the target node are processed by the graph decoding unit in the risk prediction model to obtain the risk prediction information corresponding to the target node.

[0072] Specifically, a target node refers to the node corresponding to the user or business participant whose risk needs to be predicted at the current moment. An independent node refers to a target node that exists independently in the resource relationship network and has no connection edges with other user nodes. Correspondingly, the node decoding unit refers to a decoder that decodes and processes independent target nodes to predict risks, which can be implemented using a multilayer perceptron decoder. Conversely, a non-independent node refers to a target node that has connections with other user nodes in the resource relationship network. Correspondingly, the graph decoding unit refers to a decoder that decodes and processes non-independent target nodes to predict risks, which can be implemented using a graph neural network decoder. The corresponding risk prediction information is the risk prediction result corresponding to the target node.

[0073] Based on this, when performing risk prediction tasks corresponding to the target business, the target node can first be determined in the updated resource relationship network according to the risk prediction task of the target business. If the target node is determined to be an independent node, it means that it has no connection relationship with other nodes. Therefore, the node decoding unit in the risk prediction model can be used to process the node characteristics of the target node to obtain the risk prediction information corresponding to the target node. If the target node is not an independent node, it means that it has a connection relationship with other stages. Therefore, the graph decoding unit in the risk prediction model can be used to process the node characteristics of the target node to obtain the risk prediction information corresponding to the target node.

[0074] In summary, by selecting different decoding units for processing different types of target nodes, the risk prediction results can be accurately represented in the resource relationship network, thereby effectively improving the accuracy of risk prediction.

[0075] Furthermore, different processing methods can be applied to different risk prediction results based on business scenarios. In this embodiment, after the step of performing the risk prediction task for the target business based on the updated resource relationship network is executed, the method further includes: If the target node passes the risk detection based on the risk prediction information, a target resource is determined from the resource set corresponding to the target service, and the target resource is allocated to the target node; if the target node fails the risk detection based on the risk prediction information, a risk detection reminder is sent to the target node.

[0076] Specifically, the target resource refers to the resources that need to be allocated to the target node, such as computing resources, items, transaction limits, etc., which are not limited in this embodiment. Correspondingly, the risk detection alert information refers to the alert information informing the target node that it cannot receive the allocated resources.

[0077] Based on this, if the target node passes the risk detection based on the risk prediction information, it indicates that the target node has a low risk of default. Therefore, the target resource can be determined from the resource set corresponding to the target business and allocated to the target node. If the target node fails the risk detection based on the risk prediction information, it indicates that the target node has a high risk of default. In order to avoid losses to the business party, resources can not be allocated to the target node, and risk detection reminder information can be fed back to the target node.

[0078] Continuing with the previous example, let's determine the posterior connectivity probability p of each resource interaction edge in a resource relationship network. nThen, the posterior connection probability can be compared with a set probability threshold δ. Based on the comparison result, it is determined that *a* edges out of the *n* resource interaction edges are less than the threshold. Therefore, *a* resource interaction edges can be removed from the *n* resource interaction edges. Based on the removal result, the target resource relationship network can be obtained. Subsequently, the risk prediction model can be used to predict the risk of user A according to the target resource relationship network. If the node corresponding to user A is an independent node in the target resource relationship network, the multilayer perceptron (MLP) decoder in the risk prediction model can be used to classify user A as a risky or non-risky user. If the node corresponding to user A is a non-independent node in the target resource relationship network, the graph neural network (GNN) decoder in the risk prediction model can be used to predict user A as a risky or non-risky user. If user A is a non-risky user, the requested computing resources can be allocated to user A according to their request, such as allocating 1TB of storage space.

[0079] In summary, adjusting the handling strategy according to different risk prediction results can prevent losses from the target business.

[0080] In practical applications, to ensure a strong predictive capability, a risk prediction model needs to be thoroughly trained. In this embodiment, the training of the risk prediction model includes: The sample resource relationship network is input into the initial risk prediction model for processing to obtain a reset resource relationship network. A first loss function is used to calculate a first loss value for the non-independent nodes in the reset resource relationship network, and a second loss function is used to calculate a second loss value for the independent nodes in the reset resource relationship network. The first loss value and the second loss value are fused to obtain a target loss value. The initial risk prediction model is tuned based on the target loss value until the risk prediction model that meets the training stopping condition is obtained.

[0081] Specifically, the sample resource relationship network refers to the resource relationship network used in the model training phase, and the reset resource relationship network refers to the resource relationship network output by the risk prediction model, i.e., the resource relationship network generated after reconstructing the sample resource relationship network. Correspondingly, the first loss function refers to the loss function used to calculate the loss value for non-independent nodes, and the second loss function refers to the loss function used to calculate the loss value for independent nodes. The training stopping condition refers to the conditions for stopping the training of the risk prediction model, including but not limited to loss value comparison conditions, iteration count conditions, or validation set verification conditions; this embodiment does not impose any limitations on these conditions.

[0082] Based on this, when training the risk prediction model, the sample resource relationship network can be input into the initial risk prediction model for processing to obtain a reset resource relationship network. At this time, the first loss function can be used to calculate the first loss value for the non-independent nodes in the reset resource relationship network, and the second loss function can be used to calculate the second loss value for the independent nodes in the reset resource relationship network. Then, the first loss value and the second loss value can be fused to obtain the target loss value. The initial risk prediction model can then be tuned based on the target loss value until a risk prediction model that meets the training stopping condition is obtained.

[0083] In practical applications, during the model training phase, after calibrating the edge connection confidence in the resource relationship network, it is considered that some nodes may become isolated nodes because all adjacent edges are judged as low-confidence. To address this situation, differentiated optimization objectives can be designed for non-isolated nodes and isolated nodes respectively, so as to take into account the advantages of graph structure learning and unstructured feature modeling.

[0084] For non-isolated nodes (i.e., nodes with at least one high-confidence connection), assume that the lower bound of evidence for their corresponding logp(Y|X,A,M) is ELBO. g (θ) g , φ g Based on this, the loss term can be defined as follows (9): J g =-ELBO g (θ) g , φ g )·I(p(M=1|z, )>δ)(9) Here, I(·) is an indicator function, indicating that an edge is included in the graph-based optimization objective only if its posterior connectivity probability exceeds a threshold δ. This design ensures that GNN branches are trained only on high-confidence graph structures, effectively avoiding the interference of noisy connections on model performance.

[0085] Conversely, for isolated nodes formed due to the complete removal of all connections, i.e., satisfying p(M=1|z, For nodes whose values ​​are less than or equal to δ, prediction can be made based on their own features. In this case, the ELBO objective corresponding to the MLP encoder can be used, and its loss term can be defined as follows (10): J l =-ELBO l (θ) l , φ l )·I(p(M=0|z, )≤δ)(10) This loss term encourages the model to revert to robust classification capabilities based on node-independent features when the graph structure is unreliable.

[0086] Finally, the two losses can be weighted and fused to obtain the overall loss function as shown in formula (11): J=βJ g +(1-β)J l (11) Here, β∈(0,1) is a balancing coefficient used to adjust the model's dependence on the graph structure hypothesis (GNN) and the unstructure hypothesis (MLP). When β is close to 1, the model relies more on high-confidence graph structures for prediction; when β is small, it enhances the ability to model features of isolated or weakly connected nodes. This mechanism realizes an adaptive learning strategy of "using graphs when the structure is reliable and features when the structure is unreliable," significantly improving the model's robustness and generalization ability in complex and noisy environments.

[0087] The risk prediction method provided in this embodiment, during the training phase of the risk prediction model, (z|Y,X,A) and (z|Y,X) can be implemented using two independent edge encoding modules, with parameters W and W respectively. g and W l This module is used to model the latent variable representation of each edge (i, j) ∈ E, specifically in the form of formula (12): =W g [ ],e g = (i,j)∈E =W l [ ],e l = (i,j)∈E(12) in, and They represent the graph-based structure representation h respectively. g and independent feature representation h l The generated edge-level implicit vector, then, e g It is equally divided into two parts, the mean vector μ g and standard deviation vector σ g Furthermore, the variational distribution is analytically constructed. (z|Y,X,A). Assume its covariance has a diagonal structure, i.e. I, similarly, e l It is also divided into μ l and σ l Used for parameterization (z|Y,X).

[0088] Furthermore, to achieve differentiability of the latent variable generation process, the reparameterization trick from variational autoencoders can be employed. Specifically, from... and Mid-sampling yields latent variable z g and z l Differentiable sampling can be performed using z=μ+σ⊙ϵ, where ϵ~N(0,1), thus enabling the entire model to be trained end-to-end.

[0089] Based on these differentiable latent vectors, the posterior connectivity probability of each edge can be calculated using the aforementioned Bayesian formula, and it can be further determined whether to retain the connection (compared with a threshold δ), ultimately constructing the calibrated adjacency matrix A′. On this basis, the overall loss function can be effectively optimized.

[0090] Furthermore, it can be assumed that the label distributions p(Y|X) and p(Y|X,A,M) follow a binomial distribution, and their parameters are determined by the GNN decoder g. D ( (A′) and MLP decoder f D ( The output is ψ. g and ψ l For each node, after edge connection calibration, if it is a non-isolated node, the connection representation output by the GNN encoder is used. Make a prediction; if it is an isolated node, use the independent representation obtained from the MLP decoder. Classify them.

[0091] The risk prediction method provided in this embodiment aims to effectively identify and calibrate relationship networks under conditions of information tampering and noise interference, thereby improving the accuracy of risk prediction. After obtaining the resource relationship network associated with the target business, the resource relationship network is first input into the risk prediction model. The risk prediction model, including an encoding unit and an edge calibration unit, calibrates the resource relationship network and performs subsequent risk prediction. At this stage, the encoding unit first extracts the structured and unstructured features of the resource relationship network. Then, based on these structured and unstructured features, edge features of multiple resource interaction edges in the resource relationship network are generated. This allows the edge features of each resource interaction edge to integrate graph-level and node-level representations, fully reflecting its characteristic information within the relationship network. After obtaining the edge features, the edge calibration unit can be used to calibrate the edge features of multiple resource interaction edges. This calibration process determines the posterior connection probabilities corresponding to each resource interaction edge. The posterior connection probabilities characterize the credibility of each resource interaction edge in the resource interaction network. Therefore, the resource relationship network can be updated based on the posterior connection probabilities to correct the resource interaction network, enabling it to accurately represent the real interaction relationships in the current scenario. Subsequently, risk prediction tasks for the target business can be performed based on the updated resource relationship network, thereby effectively improving the accuracy of risk prediction in business scenarios, reducing interference caused by information tampering and noise, and effectively improving business stability.

[0092] See Figure 3 , Figure 3 A flowchart of a risk prediction method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0093] Step S302: Obtain the financial relationship network and input the financial relationship network into the risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit.

[0094] Step S304: Use the encoding unit to extract the structured and unstructured features of the financial relationship network, and generate edge features of multiple financial interaction edges in the financial relationship network based on the structured and unstructured features.

[0095] Step S306: Use the edge calibration unit to calibrate the edge features of the multiple financial interaction edges to obtain the posterior connection probabilities corresponding to the multiple financial interaction edges respectively.

[0096] Step S308: Update the financial relationship network based on the posterior connection probability, and perform a default risk prediction task based on the updated financial relationship network.

[0097] This embodiment provides another risk prediction method that can be applied to any financial risk prediction scenario provided by a service provider, such as financial resource usage scenarios provided by various platforms like e-commerce, fintech, and video browsing platforms. It is used to predict default risk for users and then decide whether to provide them with corresponding financial resource usage services. The description of this other risk prediction method can be found in the descriptions of the risk prediction methods in the above embodiments, and will not be elaborated upon further in this embodiment.

[0098] The following is in conjunction with the appendix Figure 4 Taking the application of the risk prediction method provided in this specification in a risk prediction scenario within an item rental context as an example, the risk prediction method will be further explained. Figure 4 The present specification illustrates a flowchart of a risk prediction method according to an embodiment, which includes the following steps.

[0099] Step S402: Obtain the resource relationship network associated with the target business and input the resource relationship network into the risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit.

[0100] Step S404: Input the resource relationship network into the graph coding sub-unit and node coding sub-unit in the coding unit.

[0101] Step S406: The resource relationship network is processed using graph coding sub-units to obtain graph features corresponding to the graph network dimension, and the graph features are used as structured features.

[0102] Step S408: The resource relationship network is processed using the node coding sub-unit to obtain the node features of the corresponding node dimension, and the node features are used as unstructured features.

[0103] Step S410: Divide the structured features into a structured mean vector and a structured standard deviation vector, and divide the unstructured features into an unstructured mean vector and an unstructured standard deviation vector.

[0104] Step S412: Construct a structured variational distribution based on the structured mean vector and the structured standard deviation vector, and construct an unstructured variational distribution based on the unstructured mean vector and the unstructured standard deviation vector.

[0105] Step S414: The structured variational distribution and the unstructured variational distribution are fused together, and the edge features of multiple resource interaction edges in the resource relationship network are generated based on the fusion result.

[0106] Step S416: Using the edge calibration unit, a set number of candidate features are sampled in the first variational distribution and the second variational distribution according to the edge features of multiple resource interaction edges.

[0107] Step S418: Fusion is performed on a set number of candidate features, and the confidence levels corresponding to multiple resource interaction edges are determined based on the fusion results.

[0108] Step S420: Use the confidence level as the posterior connection probability corresponding to each of the multiple resource interaction edges, and compare the posterior connection probabilities corresponding to each of the multiple resource interaction edges with a preset probability threshold.

[0109] Step S422: Based on the comparison results, delete resource interaction edges in the resource relationship network that are less than a preset probability threshold, and generate the target resource relationship network based on the deletion results.

[0110] Step S424: Perform risk prediction task for the target business based on the target resource relationship network.

[0111] The risk prediction method provided in this embodiment aims to effectively identify and calibrate relationship networks under conditions of information tampering and noise interference, thereby improving the accuracy of risk prediction. After obtaining the resource relationship network associated with the target business, the resource relationship network is first input into the risk prediction model. The risk prediction model, including an encoding unit and an edge calibration unit, calibrates the resource relationship network and performs subsequent risk prediction. At this stage, the encoding unit first extracts the structured and unstructured features of the resource relationship network. Then, based on these structured and unstructured features, edge features of multiple resource interaction edges in the resource relationship network are generated. This allows the edge features of each resource interaction edge to integrate graph-level and node-level representations, fully reflecting its characteristic information within the relationship network. After obtaining the edge features, the edge calibration unit can be used to calibrate the edge features of multiple resource interaction edges. This calibration process determines the posterior connection probabilities corresponding to each resource interaction edge. The posterior connection probabilities characterize the credibility of each resource interaction edge in the resource interaction network. Therefore, the resource relationship network can be updated based on the posterior connection probabilities to correct the resource interaction network, enabling it to accurately represent the real interaction relationships in the current scenario. Subsequently, risk prediction tasks for the target business can be performed based on the updated resource relationship network, thereby effectively improving the accuracy of risk prediction in business scenarios, reducing interference caused by information tampering and noise, and effectively improving business stability.

[0112] Corresponding to the above method embodiments, this specification also provides embodiments of a risk prediction device. Figure 5 A schematic diagram of a risk prediction device according to one embodiment of this specification is shown. Figure 5 As shown, the device includes: The acquisition module 502 is configured to acquire the resource relationship network associated with the target business and input the resource relationship network into the risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit; The extraction module 504 is configured to extract the structured and unstructured features of the resource relationship network using the encoding unit, and generate edge features of multiple resource interaction edges in the resource relationship network based on the structured and unstructured features. Processing module 506 is configured to use the edge calibration unit to calibrate the edge features of the plurality of resource interaction edges to obtain the posterior connection probabilities corresponding to the plurality of resource interaction edges respectively. The update module 508 is configured to update the resource relationship network based on the posterior connection probability, and perform the risk prediction task of the target service based on the updated resource relationship network.

[0113] In an optional embodiment, the step of extracting the structured and unstructured features of the resource relationship network using the encoding unit includes: The resource relationship network is input into the graph encoding subunit and node encoding subunit of the encoding unit; the graph encoding subunit processes the resource relationship network to obtain graph features corresponding to the graph network dimension, and the graph features are used as structured features; the node encoding subunit processes the resource relationship network to obtain node features corresponding to the node dimension, and the node features are used as unstructured features.

[0114] In an optional embodiment, generating edge features of multiple resource interaction edges in the resource relationship network based on the structured features and the unstructured features includes: The structured features are divided into a structured mean vector and a structured standard deviation vector, and the unstructured features are divided into an unstructured mean vector and an unstructured standard deviation vector. A structured variational distribution is constructed based on the structured mean vector and the structured standard deviation vector, and an unstructured variational distribution is constructed based on the unstructured mean vector and the unstructured standard deviation vector. The structured variational distribution and the unstructured variational distribution are fused, and the edge features of multiple resource interaction edges in the resource relationship network are generated based on the fusion result.

[0115] In an optional embodiment, the step of calibrating the edge features of the plurality of resource interaction edges using the edge calibration unit to obtain the posterior connectivity probabilities corresponding to the plurality of resource interaction edges includes: The edge calibration unit samples a set number of candidate features in a first variational distribution and a second variational distribution according to the edge features of the multiple resource interaction edges; the set number of candidate features are fused, and the confidence level corresponding to each of the multiple resource interaction edges is determined according to the fusion result; the confidence level is used as the posterior connection probability corresponding to each of the multiple resource interaction edges.

[0116] In an optional embodiment, updating the resource relationship network based on the posterior connection probability and performing the risk prediction task for the target service based on the updated resource relationship network includes: The posterior connection probabilities corresponding to the multiple resource interaction edges are compared with a preset probability threshold; based on the comparison results, resource interaction edges with probabilities less than the preset probability threshold are deleted from the resource relationship network, and a target resource relationship network is generated based on the deletion results; the risk prediction task of the target service is performed based on the target resource relationship network.

[0117] In an optional embodiment, the risk prediction task for the target service based on the updated resource relationship network includes: Based on the risk prediction task of the target business, a target node is determined in the updated resource relationship network; if the target node is an independent node, the node characteristics of the target node are processed by the node decoding unit in the risk prediction model to obtain the risk prediction information corresponding to the target node; if the target node is not an independent node, the node characteristics of the target node are processed by the graph decoding unit in the risk prediction model to obtain the risk prediction information corresponding to the target node.

[0118] In an optional embodiment, training the risk prediction model includes: The sample resource relationship network is input into the initial risk prediction model for processing to obtain a reset resource relationship network. A first loss function is used to calculate a first loss value for the non-independent nodes in the reset resource relationship network, and a second loss function is used to calculate a second loss value for the independent nodes in the reset resource relationship network. The first loss value and the second loss value are fused to obtain a target loss value. The initial risk prediction model is tuned based on the target loss value until the risk prediction model that meets the training stopping condition is obtained.

[0119] In an optional embodiment, after the step of performing the risk prediction task of the target service based on the updated resource relationship network is executed, the method further includes: If the target node passes the risk detection based on the risk prediction information, a target resource is determined from the resource set corresponding to the target service, and the target resource is allocated to the target node; if the target node fails the risk detection based on the risk prediction information, a risk detection reminder is sent to the target node.

[0120] The risk prediction device provided in this embodiment, in order to effectively identify and calibrate relationship networks under conditions of information tampering and noise interference, and thereby improve the accuracy of risk prediction, can first input the resource relationship network associated with the target business into the risk prediction model after obtaining the resource relationship network. The risk prediction network, including an encoding unit and an edge calibration unit, calibrates the resource relationship network and performs subsequent risk prediction. At this time, the encoding unit can first extract the structured and unstructured features of the resource relationship network. Then, based on the structured and unstructured features, edge features of multiple resource interaction edges in the resource relationship network can be generated. This allows the edge features of each resource interaction edge to integrate graph-level and node-level expressions, fully reflecting its characteristic information in the relationship network. After obtaining the edge features, the edge calibration unit can be used to calibrate the edge features of multiple resource interaction edges. This calibration process determines the posterior connection probabilities corresponding to each resource interaction edge. The posterior connection probabilities characterize the credibility of each resource interaction edge in the resource interaction network. Therefore, the resource relationship network can be updated based on the posterior connection probabilities to correct the resource interaction network, enabling it to accurately represent the real interaction relationships in the current scenario. Subsequently, risk prediction tasks for the target business can be performed based on the updated resource relationship network, thereby effectively improving the accuracy of risk prediction in business scenarios, reducing interference caused by information tampering and noise, and effectively improving business stability.

[0121] The above is an illustrative scheme of a risk prediction device according to this embodiment. It should be noted that the technical solution of this risk prediction device and the technical solution of the risk prediction method described above belong to the same concept. For details not described in detail in the technical solution of the risk prediction device, please refer to the description of the technical solution of the risk prediction method described above.

[0122] Corresponding to the above method embodiments, this specification also provides another embodiment of a risk prediction device. Figure 6 A schematic diagram of another risk prediction device provided in one embodiment of this specification is shown. Figure 6 As shown, the device includes: The network acquisition module 602 is configured to acquire a financial relationship network and input the financial relationship network into a risk prediction model, wherein the risk prediction model includes an encoding unit and an edge calibration unit; The feature extraction module 604 is configured to extract the structured and unstructured features of the financial relationship network using the encoding unit, and generate edge features of multiple financial interaction edges in the financial relationship network based on the structured and unstructured features. The feature processing module 606 is configured to use the edge calibration unit to calibrate the edge features of the plurality of financial interaction edges to obtain the posterior connection probabilities corresponding to the plurality of financial interaction edges respectively. The network module 608 is configured to update the financial relationship network based on the posterior connection probability and perform a default risk prediction task based on the updated financial relationship network.

[0123] The above is an illustrative scheme of another risk prediction device in this embodiment. It should be noted that the technical solution of this risk prediction device and the technical solution of the risk prediction method described above belong to the same concept. For details not described in detail in the technical solution of the risk prediction device, please refer to the description of the technical solution of the risk prediction method described above.

[0124] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0125] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0126] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0127] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.

[0128] The processor 720 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the risk prediction method described above.

[0129] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the risk prediction method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the risk prediction method described above.

[0130] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the risk prediction method described above.

[0131] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the risk prediction method described above belong to the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the risk prediction method described above.

[0132] An embodiment of this specification also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the risk prediction method described above.

[0133] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the risk prediction method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the risk prediction method described above.

[0134] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0135] The computer program / instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, 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 included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0136] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0137] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0138] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A risk prediction method, comprising: obtaining a resource relationship network associated with a target business, and inputting the resource relationship network into a risk prediction model, wherein the risk prediction model comprises an encoding unit and an edge calibration unit; extracting structured features and unstructured features of the resource relationship network by using the encoding unit, and generating edge features of multiple resource interaction edges in the resource relationship network according to the structured features and the unstructured features; performing calibration processing on the edge features of the multiple resource interaction edges by using the edge calibration unit, to obtain posterior connection probabilities respectively corresponding to the multiple resource interaction edges; updating the resource relationship network based on the posterior connection probabilities, and performing a risk prediction task of the target business based on the updated resource relationship network.

2. The risk prediction method of claim 1, wherein the extracting the structured features and the unstructured features of the resource relationship network by using the encoding unit comprises: inputting the resource relationship network into a graph encoding subunit and a node encoding subunit in the encoding unit; processing the resource relationship network by using the graph encoding subunit to obtain graph features corresponding to graph network dimensions, and taking the graph features as structured features; processing the resource relationship network by using the node encoding subunit to obtain node features corresponding to node dimensions, and taking the node features as unstructured features.

3. The risk prediction method of claim 1, wherein the generating the edge features of the multiple resource interaction edges in the resource relationship network according to the structured features and the unstructured features comprises: dividing the structured features into a structured mean vector and a structured standard deviation vector, and dividing the unstructured features into an unstructured mean vector and an unstructured standard deviation vector; constructing a structured variational distribution according to the structured mean vector and the structured standard deviation vector, and constructing an unstructured variational distribution according to the unstructured mean vector and the unstructured standard deviation vector; fusing the structured variational distribution and the unstructured variational distribution, and generating the edge features of the multiple resource interaction edges in the resource relationship network according to a fusion result.

4. The risk prediction method of claim 1, wherein the performing the calibration processing on the edge features of the multiple resource interaction edges by using the edge calibration unit, to obtain the posterior connection probabilities respectively corresponding to the multiple resource interaction edges comprises: sampling a set number of candidate features in first and second variational distributions according to the edge features of the multiple resource interaction edges by using the edge calibration unit; fusing the set number of candidate features, and determining confidence degrees respectively corresponding to the multiple resource interaction edges according to a fusion result; taking the confidence degrees as the posterior connection probabilities respectively corresponding to the multiple resource interaction edges.

5. The risk prediction method of claim 1, wherein the updating the resource relationship network based on the posterior connection probabilities, and performing the risk prediction task of the target business based on the updated resource relationship network comprises: comparing the posterior connection probability corresponding to each of the plurality of resource interaction edges with a preset probability threshold; deleting, according to a comparison result, a resource interaction edge smaller than the preset probability threshold in the resource relationship network, and generating a target resource relationship network according to a deletion result; performing a risk prediction task of the target business based on the target resource relationship network.

6. The risk prediction method of claim 1, wherein performing the risk prediction task of the target business based on the updated resource relationship network comprises: determining a target node in the updated resource relationship network according to the risk prediction task of the target business; in a case where the target node is an independent node, processing a node feature of the target node by using a node decoding unit in the risk prediction model to obtain risk prediction information corresponding to the target node; in a case where the target node is a non-independent node, processing the node feature of the target node by using a graph decoding unit in the risk prediction model to obtain the risk prediction information corresponding to the target node.

7. The risk prediction method of any one of claims 1 to 6, wherein training the risk prediction model comprises: inputting a sample resource relationship network into an initial risk prediction model for processing to obtain a reset resource relationship network; calculating a first loss value for a non-independent node in the reset resource relationship network by using a first loss function, and calculating a second loss value for an independent node in the reset resource relationship network by using a second loss function; fusing the first loss value and the second loss value to obtain a target loss value; adjusting the initial risk prediction model based on the target loss value until the risk prediction model satisfying a training stop condition is obtained.

8. The risk prediction method of claim 6, wherein after performing the step of performing the risk prediction task of the target business based on the updated resource relationship network, the method further comprises: in a case where the target node passes risk detection according to the risk prediction information, determining a target resource in a resource set corresponding to the target business and allocating the target resource to the target node; and in a case where the target node fails to pass risk detection according to the risk prediction information, feeding back risk detection reminding information to the target node.

9. A risk prediction method, comprising: obtaining a financial relationship network and inputting the financial relationship network into a risk prediction model, wherein the risk prediction model comprises an encoding unit and an edge calibration unit; extracting a structured feature and an unstructured feature of the financial relationship network by using the encoding unit, and generating an edge feature of a plurality of financial interaction edges in the financial relationship network according to the structured feature and the unstructured feature; performing calibration processing on the edge feature of the plurality of financial interaction edges by using the edge calibration unit to obtain a posterior connection probability corresponding to each of the plurality of financial interaction edges; updating the financial relationship network based on the posterior connection probability, and performing a default risk prediction task based on an updated financial relationship network.

10. A risk prediction device, comprising: An acquisition module is configured to acquire a resource relationship network associated with a target business and input the resource relationship network into a risk prediction model, wherein the risk prediction model comprises an encoding unit and an edge calibration unit; An extraction module is configured to extract structured features and unstructured features of the resource relationship network by using the encoding unit, and generate edge features of multiple resource interaction edges in the resource relationship network according to the structured features and the unstructured features; A processing module is configured to perform calibration processing on the edge features of the multiple resource interaction edges by using the edge calibration unit, and obtain posterior connection probabilities respectively corresponding to the multiple resource interaction edges; An updating module is configured to update the resource relationship network based on the posterior connection probabilities, and perform a risk prediction task of the target business based on the updated resource relationship network.

11. A risk prediction apparatus, comprising: An acquisition network module is configured to acquire a financial relationship network and input the financial relationship network into a risk prediction model, wherein the risk prediction model comprises an encoding unit and an edge calibration unit; An extraction feature module is configured to extract structured features and unstructured features of the financial relationship network by using the encoding unit, and generate edge features of multiple financial interaction edges in the financial relationship network according to the structured features and the unstructured features; A processing feature module is configured to perform calibration processing on the edge features of the multiple financial interaction edges by using the edge calibration unit, and obtain posterior connection probabilities respectively corresponding to the multiple financial interaction edges; An updating network module is configured to update the financial relationship network based on the posterior connection probabilities, and perform a default risk prediction task based on the updated financial relationship network.

12. A computing device, comprising: a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, and the computer executable instructions, when executed by the processor, implement steps of the method in any one of claims 1 to 9.

13. A computer readable storage medium storing computer executable instructions, and the computer executable instructions, when executed by a processor, implement steps of the method in any one of claims 1 to 9.

14. A computer program product comprising a computer program or instructions, and the computer program or instructions, when executed by a processor, implement steps of the method in any one of claims 1 to 9.

Images