Federal training method and device of user risk assessment model

Abstract

Embodiments of the present specification disclose a federated training method and device of a user risk assessment model. The method involves a server and multiple clients, and is executed by any of the clients, including the following steps: first, processing a locally private user relationship graph by using a locally deployed user risk assessment model to obtain embedded representations of each labeled user node labeled with a risk category label; then, determining a plurality of local prototypes based on the embedded representations, wherein each local prototype corresponds to a risk category-neighborhood hop combination; next, sending the determined plurality of local prototypes to the server and receiving a plurality of target prototypes corresponding thereto from the server, wherein each target prototype is generated by the server based on a plurality of local prototypes uploaded by the multiple clients corresponding to the same risk category-neighborhood hop combination; and then, training the local user risk assessment model based on the received plurality of target prototypes.

Description

Technical Field

[0001] This specification relates to the field of machine learning technology, and in particular to a federated training method and apparatus for a user risk assessment model, a federated training method and apparatus for a business object prediction model, a computer-readable storage medium, and a computing device. Background Technology

[0002] With the rapid development of fintech, user risk assessment has become a core component for financial institutions in preventing risks such as credit defaults, fraudulent transactions, and illegal fund transfers. However, user data held by a single institution often suffers from limited sample coverage and a one-sided view of risk, making it difficult to comprehensively depict a user's true risk profile. For example, a user's borrowing behavior at institution A may be strongly correlated with their fund transfers at institution B; relying solely on data from one institution can easily lead to misjudgments or omissions. Therefore, achieving cross-institutional collaboration and knowledge sharing to build a comprehensive user risk assessment system has become an inevitable trend in the industry.

[0003] In real-world applications, user data held by financial institutions naturally possesses complex relational characteristics, typically existing in the form of a User Relation Graph. In this graph, user nodes contain rich attribute features (such as asset size, credit history, and transaction records), while edges represent topological connections between users, such as transfers, guarantees, and joint shareholdings. This tightly intertwined data structure of "node features and topological structure" makes graph neural network (GNN)-based modeling methods an effective solution for capturing high-order risk propagation patterns.

[0004] Although Federated Learning (FL) provides a privacy-preserving viable path for cross-institutional data collaboration, existing federated training schemes still face severe limitations when dealing with the aforementioned graph data. For example, they require that the GNNs deployed locally by each institution have the same model structure, resulting in huge communication volumes and insufficient privacy protection, etc. Summary of the Invention

[0005] This specification describes a federated training method and apparatus for a user risk assessment model and a business object prediction model, which can solve the aforementioned related technical problems.

[0006] According to the first aspect, a federated training method for a user risk assessment model is provided, involving a server and multiple clients, wherein any client holds a private user relationship graph containing labeled user nodes with risk category tags. The method is executed by any of the clients and includes:

[0007] The user relationship graph is processed using a locally deployed user risk assessment model to obtain the embedded representation of each labeled user node. Several local prototypes are determined, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by aggregating the embedded representations of the following nodes: the user node with that risk category, and the neighbor nodes of that user node with that risk category under that neighbor hop count. These local prototypes are sent to the server. Several target prototypes are received from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes corresponding to the risk category-neighbor hop count combination uploaded by the multiple clients. The user risk assessment model is trained based on these target prototypes.

[0008] In one embodiment, at least two of the plurality of clients deploy user risk assessment models with different model structures.

[0009] In one embodiment, determining several local prototypes includes: for any first combination among several risk category-neighbor hop count combinations, determining several first-labeled user nodes in the user relationship graph that have the first risk category in that combination; for each first-labeled user node, determining several neighbor nodes that have the first risk category under the first neighbor hop count in that combination; performing a first aggregation process on several embedded representations of the several neighbor nodes to obtain the neighborhood representation of the first-labeled user node; and performing a second aggregation process on the several neighborhood representations corresponding to the several first-labeled user nodes to obtain the first local prototype corresponding to the first combination.

[0010] In one embodiment, the user risk assessment model includes an embedding component and a prediction component. The process of processing the user relationship graph using a locally deployed user risk assessment model to obtain the embedding representation of each labeled user node includes: processing the user relationship graph using the embedding component to obtain the embedding representation of each labeled user node. The training of the user risk assessment model includes: processing the embedding representation of each labeled user node using the prediction component to obtain the corresponding predicted risk result; and training the user risk assessment model based on a comprehensive loss, wherein the comprehensive loss includes a first loss term determined based on several target prototypes and several local prototypes, and a second loss term determined based on the predicted risk result and the corresponding risk category label.

[0011] In one embodiment, at least one of the plurality of clients determines a plurality of local prototypes.

[0012] According to the second aspect, a federated training method for a user risk assessment model is provided, involving a server and multiple clients, wherein each client holds a private user relationship graph containing labeled user nodes with risk category tags. The method is executed by the server and includes the following steps:

[0013] Each client uploads several local prototypes, each corresponding to a risk category-neighbor hop count combination, which are obtained by the client through aggregation based on the embedded representations of the following nodes in its local environment: user nodes with that risk category, and neighbor nodes of that user node with that risk category under that neighbor hop count. The embedded representations are obtained by processing the user relationship graph using a locally deployed user risk assessment model. Multiple target prototypes are determined, each generated based on the received local prototypes corresponding to the same risk category-neighbor hop count combination. Several target prototypes corresponding to the uploaded local prototypes are sent to each client, enabling the client to train the user risk assessment model based on these target prototypes.

[0014] In one embodiment, at least one of the plurality of clients uploads a plurality of local prototypes.

[0015] In one embodiment, the target combination corresponding to each target prototype includes a target risk category and a target neighbor hop count. The determination of each target prototype includes: inputting the trainable vector corresponding to the target combination into a locally deployed prototype generation model to obtain an initial global prototype; determining a contrastive loss, which aims to increase the similarity between the initial global prototype and positive sample prototypes, and to increase the discriminativeness between positive sample prototypes and negative sample prototypes; wherein the positive sample prototype is a local prototype corresponding to the target risk category but with a different neighbor hop count than the target neighbor hop count, and the negative sample prototype is a local prototype not corresponding to the target risk category; updating the trainable vector and the prototype generation model using the contrastive loss; inputting the updated trainable vector into the updated prototype generation model to obtain an optimized global prototype; and determining the corresponding target prototype based on the optimized global prototype.

[0016] Furthermore, in a specific embodiment, a marginal parameter is introduced into the contrast loss as a distinguishing boundary between the positive sample prototype and the negative sample prototype. Determining the marginal parameter includes: calculating the central prototype of the local prototype corresponding to each risk category; for the target category, determining the similarity between its corresponding central prototype and the central prototypes corresponding to other risk categories, and selecting the maximum similarity; determining the smaller value between the maximum similarity and a preset cutoff threshold as the marginal parameter.

[0017] Furthermore, in one example, determining the corresponding target prototype based on the optimized global prototype includes: for each client, if it has uploaded a local prototype corresponding to the target combination, calculating the similarity between the local prototype and the local prototypes uploaded by other clients corresponding to the target combination; filtering out several local prototypes with similarity greater than a predetermined threshold; performing a first weighted fusion on the several local prototypes to obtain a local prototype; and performing a second weighted fusion on the global prototype and the local prototype to obtain the corresponding target prototype, which is then sent to the client.

[0018] For example, performing a first weighted fusion on the plurality of local prototypes to obtain a local prototype includes: obtaining the number of labeled user nodes with the target risk category in each client corresponding to the plurality of local prototypes; calculating the proportion of the number of labeled user nodes of each client to the total number of labeled user nodes of all the plurality of clients as their respective weighting coefficients; and performing a weighted summation on the plurality of local prototypes based on the weighting coefficients to obtain the local prototype.

[0019] According to a third aspect, a federated training method for a business object prediction model is provided, involving a server and multiple clients, wherein any client holds a private object relationship graph, the object relationship graph containing labeled object nodes with object category labels. The method is executed by any of the clients and includes:

[0020] The object relationship graph is processed using a locally deployed business object prediction model to obtain the embedded representation of each labeled object node. Several local prototypes are determined, where each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by aggregating the embedded representations of the following nodes: object nodes with that object category, and neighbor nodes of that object node with that object category under that neighbor hop count. These local prototypes are sent to the server. Several target prototypes are received from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes corresponding to the object category-neighbor hop count combination uploaded by the multiple clients. The business object prediction model is trained based on these target prototypes.

[0021] According to the fourth aspect, a federated training method for a business object prediction model is provided, involving a server and multiple clients, wherein each client holds a private object relationship graph, the object relationship graph containing labeled object nodes with object category labels. The method is executed by the server and includes:

[0022] Each client uploads several local prototypes, each corresponding to an object category-neighbor hop count combination, obtained by the client through aggregation of embedded representations of the following nodes in its local environment: object nodes with that object category, and neighbor nodes of that object category with that object category under that neighbor hop count. The embedded representations are obtained by processing the object relationship graph using a locally deployed business object prediction model. Multiple target prototypes are determined, each generated based on local prototypes from the received local prototypes that correspond to the same object category-neighbor hop count combination. Several target prototypes corresponding to the local prototypes uploaded by each client are sent to each client, enabling the client to train the business object prediction model based on these target prototypes.

[0023] According to the fifth aspect, a federated training device for a user risk assessment model is provided, wherein federated training involves a server and multiple clients, wherein any client holds a private user relationship graph, the user relationship graph containing labeled user nodes with risk category tags. The device is integrated into any of the clients and includes:

[0024] An embedding processing module is configured to process the user relationship graph using a locally deployed user risk assessment model to obtain the embedding representation of each labeled user node. A local prototype determination module is configured to determine several local prototypes, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by aggregating the embedding representations of the following nodes: user nodes with that risk category, and neighbor nodes of that user node with that risk category under that neighbor hop count. A local prototype sending module is configured to send the several local prototypes to the server. A target prototype receiving module is configured to receive several target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes uploaded by the multiple clients corresponding to the risk category-neighbor hop count combination. A local model training module is configured to train the user risk assessment model based on the several target prototypes.

[0025] According to the sixth aspect, a federated training device for a user risk assessment model is provided, wherein federated training involves a server and multiple clients, each client holding a private user relationship graph containing labeled user nodes with risk category tags. The device is integrated into the server and includes:

[0026] The local prototype receiving module is configured to receive several local prototypes uploaded by each client, wherein each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by the client by aggregating the embedded representations corresponding to the following nodes in the local machine: user nodes with the risk category, and neighbor nodes of the user node with the risk category under the neighbor hop count; the embedded representations are obtained by processing the user relationship graph using a locally deployed user risk assessment model. The target prototype determining module is configured to determine multiple target prototypes, wherein each target prototype is generated based on the local prototypes received that correspond to the same risk category-neighbor hop count combination. The target prototype sending module is configured to send several target prototypes corresponding to the several local prototypes uploaded by each client, so that the client trains the user risk assessment model based on the several target prototypes.

[0027] According to the seventh aspect, a federated training apparatus for a business object prediction model is provided, wherein federated training involves a server and multiple clients, wherein any client holds a private object relationship graph, the object relationship graph containing labeled object nodes with object category labels. The apparatus is integrated into any of the clients and includes:

[0028] An embedding processing module is configured to process the object relationship graph using a locally deployed business object prediction model to obtain the embedding representation of each labeled object node. A local prototype determination module is configured to determine several local prototypes, where each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by aggregating the embedding representations of the following nodes: object nodes with that object category, and neighbor nodes of that object node with that object category under that neighbor hop count. A local prototype sending module is configured to send the several local prototypes to the server. A target prototype receiving module is configured to receive several target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes corresponding to the object category-neighbor hop count combination uploaded by the multiple clients. A local model training module is configured to train the business object prediction model based on the several target prototypes.

[0029] According to the eighth aspect, a federated training apparatus for a business object prediction model is provided, wherein federated training involves a server and multiple clients, each client holding a private object relationship graph containing labeled object nodes with object category labels. The apparatus is integrated into the server and includes:

[0030] The local prototype receiving module is configured to receive several local prototypes uploaded by each client, wherein each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by the client by aggregating the embedding representations corresponding to the following nodes in the local machine: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count; the embedding representations are obtained by processing the object relationship graph using a locally deployed business object prediction model. The target prototype determining module is configured to determine multiple target prototypes, wherein each target prototype is generated based on local prototypes from the received local prototypes that correspond to the same object category-neighbor hop count combination. The target prototype sending module is configured to send several target prototypes corresponding to the several local prototypes uploaded by each client, so that the client trains the business object prediction model based on the several target prototypes.

[0031] According to a ninth aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method provided in any one of the first to fourth aspects.

[0032] According to a tenth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method provided by any one of the first to fourth aspects.

[0033] In summary, by adopting the federated training method and apparatus disclosed in the embodiments of this specification, the generalization ability and business prediction accuracy of the federated learning model in heterogeneous graph data scenarios are significantly improved by introducing a two-dimensional granular prototype construction mechanism of "risk category-neighbor hop count", combined with the adaptive generation strategy of the server-side target prototype and the targeted design of the client-side prototype alignment loss. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This diagram illustrates the communication interaction steps involved in any round of iterative training of the user risk assessment model disclosed in the embodiments of this specification under federated training.

[0036] Figure 2 The embodiments disclosed in this specification are for Figure 1 A schematic diagram of the sub-step sequence of step S120;

[0037] Figure 3 This is a schematic diagram of the process steps for generating a global model based on contrastive learning on the server side, as disclosed in the embodiments of this specification.

[0038] Figure 4 This is a functional structure diagram of the federated training device for the user risk assessment model disclosed in the embodiments of this specification. The device is integrated into any of the multiple clients.

[0039] Figure 5 This is a functional structure diagram of the federated training device for the user risk assessment model disclosed in the embodiments of this specification. The device is integrated on the server side.

[0040] Figure 6 This is a functional structure diagram of the federated training device for the business object prediction model disclosed in the embodiments of this specification. The device is integrated into any one of the multiple clients.

[0041] Figure 7 This is a functional structure diagram of the federated training device for the business object prediction model disclosed in the embodiments of this specification. The device is integrated on the server side. Detailed Implementation

[0042] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0043] As mentioned earlier, while federated learning (FL) offers a privacy-preserving path for cross-institutional data collaboration, existing federated training schemes still face significant limitations when processing user relationship graph data from financial institutions. Specifically:

[0044] 1. Immense pressure regarding privacy protection and compliance. Financial data contains highly sensitive customer information (such as account statements and credit records). Directly sharing raw data or transmitting model gradients containing sensitive information greatly increases the risk of privacy breaches and makes it difficult to meet increasingly stringent regulatory requirements. Traditional parameter aggregation methods still carry the risk of inferring the original data under certain attacks.

[0045] 2. Model heterogeneity hinders collaborative implementation. Different financial institutions often deploy evaluation models with significantly different architectures and parameter counts based on their own technology stacks, data scales, and business needs. Traditional distributed training methods such as FedAvg strictly require all participants to have a unified model structure in order to directly aggregate weights. This makes knowledge fusion between heterogeneous models extremely difficult and limits the applicability of cross-institutional collaboration.

[0046] 3. Data heterogeneity leads to the failure to capture global patterns. Financial data exhibits significant non-independent and identically distributed characteristics across different institutions: not only do customer characteristic distributions (such as asset level and credit score) vary greatly, but their transaction network topologies (such as fund flow patterns, cooperation density, and guarantee chains) are also drastically different. Existing solutions struggle to effectively decouple and integrate these differentiated local features and topological information when integrating knowledge, resulting in global models that fail to accurately capture cross-domain global risk propagation patterns.

[0047] 4. High communication costs limit real-time performance. Cross-institutional communication is constrained by bandwidth resources and privacy compliance reviews. Traditional methods require frequent exchange of multi-layer model weights or high-dimensional node embeddings, causing communication volume to increase exponentially with model depth and the number of participating institutions. In scenarios with extremely high real-time requirements, such as high-frequency trading risk monitoring, the huge communication latency makes it difficult to implement existing solutions.

[0048] Based on the above, the applicant proposes an improved scheme for federated training of user risk assessment models (hereinafter referred to as this scheme). This scheme aims to: effectively adapt to heterogeneous model architectures, overcome differences in data distribution and topology, and significantly reduce communication overhead, while strictly protecting the data privacy of each institution, thereby achieving high-precision, high-efficiency, and compliant cross-institutional joint user risk assessment.

[0049] To more clearly illustrate the technical principles and innovations of this scheme, we will first briefly introduce the basic prototype federated learning scheme (hereinafter referred to as the basic scheme) to which this scheme is based. In prototype-based federated learning, the participants usually no longer exchange model parameters, but instead exchange "prototype" vectors representing the feature centers of each category, which serve as the medium for knowledge transfer.

[0050] In the basic scheme, it is assumed that there are a total of Each participant (usually corresponding to) (Multiple different financial institutions), each participant maintains a client. The construction of the client's local prototype typically follows this logic:

[0051] For a certain risk category Client of any participating institution By simply aggregating the feature representations (or feature vectors, embeddings) of all labeled samples under this category, we obtain the local prototype representation of this category. Its mathematical expression can be illustrated as follows:

[0052] (1)

[0053] in, Indicates client Previous category The local prototype; Indicates client The middle label is the category The user sample set; Indicates the first A user sample, which includes sample features and sample labels ; This refers to the embedded component (Encoder) in the graph processing model deployed locally on the client, responsible for processing the input samples. Mapped to feature vectors ;symbol This indicates the number of elements in the set.

[0054] Furthermore, after receiving the local prototypes from each client, the traditional global prototype generation method often employs a weighted average strategy to integrate knowledge from multiple sources:

[0055] (2)

[0056] in, Indicates the server-side The categories generated by the round Global prototype; Indicates the first Round participated in the category A collection of clients for prototype uploading; Indicates client Previous category The number of samples is used to calculate the weighting coefficients; Indicates client Uploaded categories The local prototype.

[0057] However, after in-depth research, the applicant discovered that the above-mentioned basic solution still has the following shortcomings when facing the complex challenges of financial scenarios:

[0058] 1. Formula (1) corresponds to the method of directly averaging the embedding vectors of samples of the same class. It assumes that each sample is independent and only focuses on its position in the feature space. This is a simple and intuitive method, but it ignores the graph structure.

[0059] 2. Formula (2) corresponds to the method of weighting all prototypes of the same category by the amount of client data. It assumes that the prototypes of all clients are linearly additive in semantic space, and the weights are determined only by the number of samples, without considering topological context or semantic differences. This is an intuitive but naive strategy that ignores graph structure and heterogeneity.

[0060] Based on the above observations and analysis, the improved scheme retains the advantage of low communication volume in prototype aggregation, and introduces "fine-grained multi-hop prototype construction" to overcome the problems of formula (1). Optionally, it also introduces "global prototype generator based on contrastive learning" to overcome the problems of formula (2).

[0061] The specific implementation steps of this solution will be described in detail below with reference to the accompanying drawings. First, it should be noted that this solution implements federated training of the user risk assessment model, which involves multiple rounds of iterative training. The following explanation will take any one round of iterative training as an example.

[0062] The parties involved in any round of iterative training include a server and multiple clients. It should be understood that these multiple clients can be the entire set or a subset of clients deployed by all institutions participating in the federated training; the server can be deployed and maintained by a neutral third party, or it can be deployed and maintained by an institution. Furthermore, for the sake of brevity, the terms "institution" and "client" can be used interchangeably below.

[0063] To aid understanding, we will first introduce the private graph data held by each of the multiple clients and the local graph processing models deployed, and then introduce the communication and interaction steps involved in any round of iterative training.

[0064] Each client maintains its own private user relationship graph. In the user relationship graph, nodes represent specific user entities (such as individual customers or corporate accounts), and connecting edges represent business relationships between users (such as fund transfers, joint guarantees, equity holdings, etc.).

[0065] The user relationship graphs maintained by different clients may exhibit heterogeneity in data distribution: on the one hand, there may be overlap between the graphs of different institutions, that is, they may contain the same user nodes or connecting edges (for example, a user has accounts at both Bank A and Bank B), but the attribute characteristics (such as asset size and transaction frequency) of the same user in different institutions are often different; on the other hand, even for the same relationship, the attributes or weights of the connecting edges recorded by different institutions (such as transfer amount and guarantee strength) may also differ.

[0066] In the user relationship graph, some user nodes are labeled user nodes. This means that these nodes have risk category labels that have been manually reviewed or historically confirmed (e.g., normal, fraud, overdue, suspected illegal fund transfer, etc.), which can be used as supervision signals for model training.

[0067] It should be noted that although the data distribution and graph structure differ among institutions, all clients participating in federated training follow a unified risk labeling system. This means that the risk category names used by different clients and their underlying semantic definitions are consistent (for example, the criteria for determining "fraud" or "overdue" are aligned in business logic across all institutions). This consistency in label semantics ensures that the local prototypes generated by each client can be effectively aggregated and compared within the same feature space, thereby achieving cross-institutional knowledge sharing.

[0068] The graph processing models deployed on multiple clients can be heterogeneous. Specifically, this graph processing model, also known as a user risk assessment model, typically includes an embedding component (Encoder) and a prediction component (Decoder / Classifier). The embedding component is responsible for mapping each user node to an embedded representation (i.e., a feature vector) based on the topology and node features of the user relationship graph. The prediction component receives this embedded representation, maps it to a specific risk category space, and outputs the probability distribution of the user belonging to each risk category.

[0069] In terms of model size and parameter magnitude, the models deployed by different clients can vary significantly: some institutions with sufficient computing power or large data scales can deploy deep, wide-layer complex graph neural networks (or even large-scale graph pre-trained models) to capture high-order nonlinear risk patterns; while resource-constrained institutions can deploy shallow, lightweight graph convolutional networks. In the federated training process of this scheme, because only "prototype" vectors are interacted with rather than model weights, it can accommodate this heterogeneity in model structure and parameter quantity, without forcing all participants to unify the model architecture.

[0070] The above describes the private graph data held by multiple clients and the user risk assessment models deployed by each. Next, we will combine... Figure 1 Describe the communication flow for any round of iterative training. It should be noted that, given the consistent interaction logic executed by each client, Figure 1 With any client only Taking an example, the specific interaction steps include:

[0071] Step S110, Client By utilizing the locally deployed user risk assessment model, the user relationship graph it holds is processed to obtain the embedded representation of each labeled user node carrying a risk category label.

[0072] Specifically, the user risk assessment model mainly consists of an embedding component and a prediction component in its architecture. In this step, the embedding component is primarily invoked to perform the feature extraction task.

[0073] Client The local user relationship graph is input into this embedded component. This embedded component is configured to process graph-structured data. By perceiving the topological connections in the user relationship graph, it aggregates and transforms node information: for each labeled user node, the embedded component not only extracts the node's own attribute features (such as user profile, account status, etc.), but also aggregates the feature information of its neighboring nodes and even multi-hop neighboring nodes based on the edge connections in the graph.

[0074] In practice, the specific implementation algorithms of the embedded component can be diverse. For example, it can use graph neural networks (GNNs) for message passing and aggregation, or it can use sequence modeling methods based on random walks, or any other graph learning algorithm that can generate node vector representations using graph topology. Through this neighborhood information fusion based on topology, the embedded component can capture the structural contextual dependencies of users in their local transaction network, mapping each user node to a corresponding embedded representation.

[0075] The embedded representation output by the embedded component retains the user's static attributes while incorporating its dynamic topological features in the relationship graph. This allows for a more comprehensive characterization of the user's potential risk patterns, providing high-quality feature input for subsequent prototype construction and risk prediction.

[0076] Step S120, Client Several local prototypes are identified, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by aggregation based on the embedded representations of the following nodes: user nodes with that risk category, and neighbor nodes of that user node with that risk category under that neighbor hop count.

[0077] This step breaks through the limitation of the traditional approach of "determining a single prototype for each risk category only". In the traditional approach, the construction of the prototype often assumes that the samples are independent and directly performs an arithmetic average of the embedding vectors of all nodes in the same category, completely ignoring the topological structure of the graph data. For this, please refer to the relevant description of the above formula (1).

[0078] By employing this step, user nodes within a specific risk category are no longer treated as isolated points. Instead, they are used as central nodes, and information about neighboring nodes with a specific number of hops and the same risk label is explicitly aggregated to construct a "neighborhood representation" for that central node. Then, the "neighborhood representations" of all central nodes within that category are aggregated again to form the final local graph prototype. This two-layer mechanism of "node-level neighborhood aggregation + category-level global aggregation" deeply integrates the topological structure information of the user relationship graph into the prototype generation process. This ensures that the generated local prototype not only contains the static characteristics of the nodes themselves but also embodies their higher-order structural context within the risk propagation network, resulting in a more informative and discriminative prototype representation.

[0079] It should be noted that the "several" in the aforementioned "several local prototypes," as well as the "several" elsewhere in the text, all refer to one or more in quantity. In practical application scenarios, in order to fully cover the correlation between different risk types and different orders, at least one of the multiple clients determines that the number of the several local prototypes is multiple (i.e., corresponding to multiple different combinations of "risk category - neighbor hop count").

[0080] For example, the implementation of this step may include Figure 2 The following sub-steps are shown in the figure:

[0081] S201, for any first combination among several risk category-neighbor hop count combinations, determine several first labeled user nodes in the user relationship graph that have the first risk category in the first combination.

[0082] Specifically, in the aforementioned combinations of risk categories and neighbor hop counts, the risk categories can be derived from all known risk categories (e.g., fraud, delinquency, etc.) involved by the user nodes marked in the local user relationship graph; the neighbor hop count depends on the topological depth that the user risk assessment model can perceive during the feature extraction stage (e.g., the number of aggregation layers or the receptive field of the model, such as 1 hop, 2 hops, or 3 hops). By performing a Cartesian product cross-combination of all the available risk categories and all feasible neighbor hop counts, several independent combinations of risk categories and neighbor hop counts can be obtained.

[0083] In this step, the client A combination-based processing strategy is adopted, meaning that only one combination (denoted as the first combination, containing a specific first risk category) is processed at a time. Jump number with the first neighbor ) Perform subsequent operations. It should be noted that the "first" in "first combination" and "first risk category", as well as the "second", "third" and similar terms elsewhere in the text, are all for distinguishing similar things and do not have any other limiting function such as ranking.

[0084] For example, the client You can first traverse its local user relationship graph and filter out all tags that fall under the first risk category. Several first-label user nodes form the set of first-label user nodes in the current combination, which can be denoted as .

[0085] S203, for each first labeled user node, determine a number of neighbor nodes that have the first risk category under the first neighbor hop count in the first combination.

[0086] For example, a collection can be Each of the first labeled user nodes in the data serves as the central node. The neighborhood search operation is performed, which aims to construct a topology-aware local context environment, specifically including the following two dimensions of filtering:

[0087] 1. Topological distance filtering: based on the central node Starting from the central node, expand outwards along the edges of the user relationship graph, filtering out nodes that are related to the central node. The shortest path length is equal to the number of hops to the first neighbor. The node. Here Corresponding to the jump value defined in the first combination (e.g.) ), representing the depth of risk propagation that the model needs to perceive.

[0088] 2. Category consistency screening: In Among all the neighboring nodes of the jump, further filter out those that are also marked as the first risk category. This ensures that subsequent aggregation processes only focus on neighbors with the same risk attributes as the central node, thereby extracting pure, homogeneous risk propagation patterns and avoiding interference from outliers.

[0089] The set of nodes obtained after the above double screening constitutes the central node. Homogeneous neighborhoods of the same type can be denoted as a set. .

[0090] S205, perform a first aggregation process on the embedded representations of the neighboring nodes to obtain the first labeled user node (i.e., the central node). The neighborhood representation of ).

[0091] This step aims to utilize the already identified set of similar, homogeneous neighbors. , as the central node This generates a local representation rich in contextual information. In practice, the aggregation method can be flexibly chosen based on different requirements for model accuracy and computational complexity. Two aggregation methods are given below as examples.

[0092] Method 1: Weighted aggregation based on self-attention mechanism

[0093] To differentiate the contributions of different neighboring nodes to the risk assessment of the central node, a self-attention mechanism can be introduced. This mechanism adaptively allocates weights based on the feature correlation between neighboring nodes and the central node.

[0094] The following is an example of the calculation process:

[0095] 1. Interactive Feature Encoding: Constructing the Central Node With each of its neighboring nodes Interaction representation between As shown in formula (3), the original feature vectors of the two can be concatenated and then input into a multi-layer perceptron (MLP) for transformation processing:

[0096] (3)

[0097] in, This indicates a vector concatenation operation.

[0098] 2. Attention weight calculation: using a learnable scoring function. (Typically linear layers or small networks) for interactive representation Scoring is performed, and the results are normalized using the Softmax function to calculate the score for each neighbor node. attention weights :

[0099] (4)

[0100] This means that neighboring nodes with risk patterns more similar to or stronger correlation with the central node will receive higher weight. .

[0101] 3. Weighted aggregation: using the calculated weights For this central node Embedding vectors of all similar neighbors We perform a weighted summation to obtain the final neighborhood representation:

[0102] (5)

[0103] Method 2: Parametric Aggregation Based on Simple Averaging

[0104] In scenarios with extremely high computational efficiency requirements or small data volumes, the first aggregation process can also employ a simpler mean aggregation strategy, assuming that all eligible neighbor nodes have equal importance. In this case, the arithmetic mean of the embedding vectors of all neighbor nodes can be directly calculated, as follows:

[0105] (6)

[0106] Regardless of the aggregation method used, the output neighborhood representation not only includes the static attributes of the central node itself, but also integrates the structured information of its neighbors of the same type under a specific number of hops, laying the foundation for the subsequent generation of highly discriminative local prototypes.

[0107] S207, perform a second aggregation process on the neighborhood representations corresponding to the plurality of first labeled user nodes to obtain the first local prototype corresponding to the first combination.

[0108] The goal of this step is to further aggregate several neighborhood representations to form a representation that can represent the entire client. The vector of the overall risk pattern under the current "risk category - neighbor hop count" (first combination), i.e., the first local prototype.

[0109] For example, the client It is possible to traverse the first set of labeled user nodes. For all elements, their respective neighborhood representations are collected. Then, a second aggregation process is performed on these neighborhood representations, such as using a mean aggregation strategy, which involves summing all neighborhood representation vectors and dividing by the total number of center nodes. Normalization is performed. Its mathematical expression can be written as follows:

[0110] (7)

[0111] Based on the above, the first local prototype is generated through this two-layer mechanism of "node-level neighborhood aggregation (step S205) + category-level global aggregation (step S207)". It not only preserves the static attributes of individual users but also deeply integrates them into the higher-order topological context of the risk propagation network. This makes the prototype more discriminative and information-rich than traditional solutions, enabling it to more accurately depict the local risk distribution patterns of an institution under specific risk categories and specific association depths. This provides high-quality foundational data for determining the target prototype for the server and training the local model on the client side.

[0112] The above combination Figure 2 Introducing the client The determination of several local prototypes. Continuing back... Figure 1 .

[0113] Step S130, Client Send the aforementioned local prototypes to the server.

[0114] It should be noted that, in order for the server to correctly parse and aggregate these prototypes, the client... The risk category-neighbor hop count combination (or combination identifier) ​​corresponding to each local prototype needs to be sent to the server.

[0115] In this step, the client No original user relationship graph data, node features, or edge information are transmitted, nor are the weight parameters or gradients of the local model transmitted. Since the prototype's dimensionality is typically much lower than the number of model parameters and much smaller than the original graph data, this transmission method significantly reduces bandwidth consumption and latency in cross-institutional communication. Furthermore, because the prototype represents statistical features aggregated over multiple layers, it is difficult to reverse engineer sensitive information about individual users, thus further enhancing privacy protection at the communication level and meeting compliance requirements in financial scenarios.

[0116] In step S140, the server determines multiple target prototypes, each of which is generated based on the local prototypes received that correspond to the same risk category-neighbor hop count combination.

[0117] The server aggregates several local prototypes uploaded by multiple clients, along with a "risk category - neighbor hop count" combination identifier for each local prototype. The core task of this step is to perform cross-organizational knowledge fusion, that is, for each combination (e.g., "fraud - 2 hops"), aggregating all local prototypes uploaded by all clients belonging to that combination to generate a target prototype that can represent the convergence of knowledge from multiple parties.

[0118] It should be noted that the strategy for generating the target prototype is flexible:

[0119] In some implementations, the target prototype is shared globally, meaning that for the same "risk category-neighbor hop count" combination, all clients receive the same target prototype.

[0120] In other implementations, the target prototype incorporates personalization in addition to referencing globally shared knowledge. The server generates a separate target prototype for each client that better matches its data distribution (e.g., by merging prototypes from highly similar clients).

[0121] The following describes a more natural implementation: a weighted average aggregation strategy based on sample size. Under this strategy, the server aims to generate a globally shared target prototype for each "risk category - neighbor hop count" combination. For example, the specific operation is as follows:

[0122] The server first groups all local prototypes uploaded by clients according to their corresponding combinations. For any specific first combination (including risk categories)... Jumping numbers with neighbors The server extracts multiple (denoted as) The corresponding local prototype uploaded by the client (one) .

[0123] Subsequently, the server uses a weighted average strategy to generate a global prototype for this combination. The weights are usually determined by the number of samples from each client in that risk category, to reflect that the client with more data contributes more to the overall knowledge. Its mathematical expression is shown in formula (8):

[0124] (8)

[0125] in, Indicates risk category Next Neighbor Jump Count The corresponding target prototype; Indicates client Previous category The number of samples is used to calculate the weighting coefficients; Indicates client Uploaded categories Next Neighbor Jump Count The corresponding local prototype, This represents the total number of clients participating in this round of iterative training.

[0126] In this way, the server can quickly obtain a baseline global model that integrates knowledge from multiple sources, and use it as the target model for the corresponding combination.

[0127] In addition to the simple weighted averaging strategy described above, this solution also proposes more advanced generation mechanisms. For example, it uses a "global prototype generator based on contrastive learning" to generate a global model, or, based on the global prototype, it further combines a personalized fusion mechanism to customize a unique target prototype for each client. The implementation steps of these advanced mechanisms will be described in detail later and will be omitted here.

[0128] Regardless of the strategy employed, this step ultimately outputs multiple target prototypes, each corresponding to a "risk category - neighbor hop count" combination (and, in personalized scenarios, possibly also a specific client). These target prototypes will serve as knowledge anchors for this round of federated training, and will be distributed back to each client to guide the optimization of their local models.

[0129] In step S150, the server sends several target prototypes corresponding to the several local prototypes uploaded by each client.

[0130] The core of this step is to establish a "upload-download" correspondence, ensuring that the target prototype received by each client is strictly aligned semantically with its own uploaded local prototype.

[0131] For example, the implementation process of this step may include:

[0132] 1. Matching and Mapping: For any client The server retrieves all local prototype collections uploaded in this round. For each local prototype in the collection The server extracts the target prototype with the same "risk category-neighbor hop count" combination identifier from the multiple generated target prototypes. (Or, in a personalized scenario, a target prototype specific to that client) ).

[0133] 2. Construct and distribute data packets: The server will send data packets to the client. All matched target prototypes are packaged into a unique knowledge update package. This package contains the target prototype vector and its corresponding "risk category-neighbor hop count" combination identifier, but does not contain any original user data or sensitive information from other clients.

[0134] 3. Secure transmission: The server sends the data packet to the client through an encrypted communication channel. .

[0135] Through this step, the client Several target prototypes can be obtained that integrate cross-organizational global risk knowledge.

[0136] Step S160, Client The user risk assessment model is trained based on the aforementioned target prototypes.

[0137] The core of this step is to use the received target prototype, which incorporates cross-institutional knowledge, to guide the parameter updates of the local model, so that the updated model can both fit the local real-label data and remain consistent with the risk distribution across institutions.

[0138] The training process can be optimized through backpropagation based on a comprehensive loss function. For example, the comprehensive loss function can be expressed as follows:

[0139] (9)

[0140] Classification loss (or the second loss term) and prototype alignment loss (Or the first loss item) constitutes; These are hyperparameters used to control... The intensity of the risk classification task is used to balance the weights of the risk classification task and the knowledge transfer task, and the value range is generally (0,1).

[0141] Classification loss To ensure the user risk model's ability to fit locally labeled data, its computational logic includes: using the prediction component in the locally deployed user risk assessment model to process the embedded representation of each labeled user node and obtain the predicted probability distribution. Then, calculations were performed. With the node's true risk category label The difference between (which can be in the form of one-hot encoded vectors). The classification loss can be achieved using the cross-entropy loss function, calculated as follows:

[0142] (10)

[0143] in, Indicates client The set of labeled user nodes; Represents user node Category The true label (if it belongs, the value is 1; if it does not belong, the value is 0). Indicates the prediction node of the user risk assessment model Category The probability of.

[0144] The above categories of loss items This is used to ensure that the local model does not forget local specific risk patterns in the pursuit of global consistency, thereby guaranteeing basic classification accuracy.

[0145] Prototype alignment loss This is key to achieving federated knowledge transfer in this scheme, aiming to bridge the gap between the local prototype and the target prototype, and forcing the local feature space to align with global risk patterns. The calculation logic for this loss term includes: for each local risk category... (from 1 to ) and the number of hops for each neighbor (From 0 to ), calculate the Frobenius norm (i.e., the L2 distance of the vectors) between the local prototype and the target prototype. ), and sum over all combinations. The calculation formula can be:

[0146] (11)

[0147] in , These represent the local prototype and the corresponding target prototype, respectively. The double summation symbol in the formula... This demonstrates the fine-grained alignment mechanism of this scheme. It not only requires alignment of the final risk category prototypes, but also different topology depths (hop counts) The intermediate layer prototypes are also aligned. This ensures that the model captures consistent global structural information at different order levels of neighborhood aggregation.

[0148] The above allows for the completion of any round of federated training in this scheme. Through multiple rounds of iterative training, each client can obtain a locally trained user risk assessment model.

[0149] In summary, this solution significantly improves the generalization ability and risk assessment accuracy of federated learning models in heterogeneous graph data scenarios by introducing a two-dimensional granular prototype construction mechanism based on "risk category - neighbor hop count," coupled with an adaptive generation strategy for the server-side target prototype and a targeted design for client-side prototype alignment loss. The main benefits of this solution include:

[0150] 1. Overcomes the shortcomings of traditional prototype construction that ignores graph topology, and enhances feature discrimination power.

[0151] Traditional federated prototyping methods (see formula (1) above) are essentially an arithmetic average of the embedding vectors of nodes of the same category. This approach assumes that the samples are independent and identically distributed, completely severing the edge connection information and neighborhood topology in the user relationship graph. In contrast, this scheme innovatively proposes a two-layer aggregation mechanism of "node-level neighborhood aggregation + category-level global aggregation" in step S120. This mechanism not only utilizes the features of the labeled user nodes themselves, but also explicitly integrates their features. Structured information about similar neighbors within the jump range. This generates a local prototype. It is no longer a simple statistical center, but rather contains a high-order structural context of users in the risk propagation network. This prototype representation, rich in topological semantics, significantly enhances the model's ability to capture complex risk propagation patterns, thereby obtaining more informative and discriminative feature representations.

[0152] 2. Supports fine-grained, multi-scale knowledge alignment, adapting to real-world financial risk control scenarios.

[0153] By defining and decoupling the "risk category - neighbor hop count" combination (e.g., "fraud risk - 1-hop nearest neighbor", "overdue risk - 2-hop distant neighbor"), this scheme achieves multi-granularity, multi-scale prototype alignment. In steps S140 and S160, the server and client respectively target different hop counts. The prototype undergoes independent aggregation calculations and loss constraints to ensure that the model can capture consistent global risk patterns across different topological depths. This design precisely aligns with the business characteristics of "strong influence from nearest neighbors and high noise from distant neighbors" in financial risk control scenarios: it retains strong signals from close-range correlations while avoiding interference from distant noise on the global model. Compared to the single-dimensional global averaging in traditional solutions, this approach effectively avoids semantic ambiguity caused by topological scale confusion, thus improving the precision of risk identification.

[0154] 3. Compatibility with highly heterogeneous models and data significantly improves deployment flexibility.

[0155] Because this solution only transmits anonymized prototype vectors during federated interactions, rather than the original graph data or high-dimensional model weights / gradients, it naturally possesses strong heterogeneous compatibility: 1) Data heterogeneity compatibility: Differences in data such as low overlap of user relationship graphs and non-independent and identically distributed label distributions held by participating institutions do not affect the prototype-based knowledge alignment process; 2) Model heterogeneity compatibility: Different clients can flexibly deploy graph processing models of different architectures (such as graph neural networks GNN, random walk, etc.) and different scales (lightweight to large-scale pre-trained models) according to their own computing power and business needs. As long as the output dimension of the embedded representation is consistent, it can seamlessly participate in federated training; 3) Flexible and secure deployment: The server can be a neutral third party or the leading institution. The protocol itself does not depend on a specific model structure, ensuring the security and feasibility of cross-institutional collaboration.

[0156] 4. Dual-loss driven collaborative training balances local fitting accuracy and global consistency.

[0157] This scheme constructs a comprehensive loss function consisting of a "classification loss" and a "prototype alignment loss," achieving synergistic optimization driven by dual objectives. The classification loss term (the second loss term), based on the predicted results and the true labels, ensures the model's ability to fit local, unique data distributions, preventing catastrophic forgetting. Meanwhile, the prototype alignment loss term, based on the distance between the local prototype and the target prototype, forces the local feature space to converge towards global risk patterns. This dual constraint mechanism effectively balances the contradiction between "preserving local specificity" and "absorbing global commonality," resulting in a user risk assessment model that possesses both high-precision local classification performance and strong cross-institutional generalization ability, accurately identifying potential risk transmission paths across domains.

[0158] Although the aforementioned two-dimensional granularity mechanism of "risk category-neighbor hop count" and the basic adaptive generation strategy have significantly improved the model's performance in conventional heterogeneous scenarios, there is still room for improvement in the server's method of determining the target prototype by relying solely on the sample size or simple similarity weighted average aggregation when facing the extreme non-independent and identically distributed characteristics of financial data and potential malicious noise interference.

[0159] Specifically, the weighting strategy in formula (8) assumes that the global prototype is a linear superposition of the local prototypes, ignoring the deep topological differences in risk characteristics between different institutions in the semantic space. This may cause the global prototype to be biased by a few abnormal clients, or lose its discriminative power when the category boundaries are blurred.

[0160] To further overcome this performance bottleneck and better explore the potential of federated learning in complex risk control scenarios, this solution builds two additional advanced generation mechanisms on top of the basic architecture:

[0161] First, we introduce a "global prototype generator based on contrastive learning". By constructing a contrastive loss for positive and negative sample pairs, we can explicitly compress intra-class distance and expand inter-class margin, thereby extracting a more robust global risk pattern.

[0162] Secondly, based on Mechanism 1, a "personalized adaptive fusion mechanism" is further designed to dynamically select high-value neighbors based on the local prototype of each client and customize a unique target prototype for each node, so as to achieve a leap from "generalization" to "precise adaptation".

[0163] The two advanced generation mechanisms will be introduced in detail below.

[0164] Mechanism 1: Global Prototype Generator Based on Comparative Learning

[0165] Specifically, for the target combination corresponding to the target model determined by the server in step S140 above, the risk category and neighbor hop count in this combination are respectively referred to as the target risk category and the target neighbor hop count. The determination of each target prototype includes... Figure 3 The following sub-steps are shown:

[0166] Step S301 involves inputting the trainable vectors corresponding to the target combination into the locally deployed prototype generation model to obtain the initial global prototype. This can be denoted as:

[0167] (12)

[0168] For trainable vectors It corresponds to the current target portfolio (i.e., a specific target risk category). Hop count with target neighbor (A combination of data). This vector is not directly from the client's raw data, but rather a parameterized representation existing in the latent space. It serves as the input seed for the prototype generation model, aiming to capture the general semantic features of that specific risk category at a specific topological depth through subsequent learning processes. It can be understood that if this iteration of training is the first round, then the trainable vector... The vector elements in the vector can be obtained through random initialization.

[0169] The Global Prototype Generator (GPG), or simply prototype generation model, uses learnable model parameters denoted as... For example, the GPG model can be a multilayer perceptron (MLP) or a Transformer network. It should be understood that GPG is generally responsible for processing low-dimensional, abstract trainable vectors. It is mapped to a high-dimensional feature space.

[0170] After conversion using the GPG model, the initial global prototype is obtained as the output. This represents the server's understanding of the "target risk category" in the current iteration round. Target neighbor hop count A preliminary assessment of the "downward" risk model. It should be noted that at this point... It has not yet undergone comparative learning optimization correction, and is therefore called the “initial” global prototype, which will serve as the benchmark object for constructing the comparative loss function in the next step (S303).

[0171] Furthermore, the same set of GPG model parameters is reused for the initial global prototype generation process for different target combinations. This parameter sharing mechanism not only ensures the consistency of the feature space, but also promotes knowledge transfer and collaborative learning between different risk categories and topological depths.

[0172] Step S303: Determine the contrast loss, which aims to increase the similarity between the initial global prototype and the positive sample prototype, and to increase the distinguishability between the positive sample prototype and the negative sample prototype; wherein, the positive sample prototype is a local prototype that corresponds to the target risk category but has a different number of neighbor hops than the target neighbor hops, and the negative sample prototype is a local prototype that does not correspond to the target risk category.

[0173] This step aims to construct an objective function that can simultaneously optimize both "intra-class compactness" and "inter-class separation"—comparative loss function. The core idea of ​​this function is to enhance intra-class consistency by maximizing the similarity between the initial global prototype and the positive sample prototype, and to strengthen inter-class discriminative power by maximizing the distinguishability between positive and negative sample prototypes.

[0174] The aforementioned positive sample prototype was designed to belong to the target risk category. But its neighbors jumped several times Unlike the current target neighbor hop count The local prototypes. For example, assuming the current target combination is "fraud category + 2 hops", then all local prototypes of "fraud category + 1 hop" or "fraud category + 3 hops" can be considered as positive samples. The intention of this design is to force the model to learn robust class-specific features that do not depend on a specific number of hops by bringing the prototypes of the same class closer together at different topological depths, thereby achieving "intra-class multi-granularity consistency".

[0175] The aforementioned negative sample prototype was designed to be: not belonging to the target risk category. (i.e., category label is) The local prototype of any neighbor hop count of a class is used. For example, assuming the current target category is "fraud", the prototypes of all categories such as "fraudulent transactions" and "normal transactions", regardless of their hop count, can be used as negative samples. The intention of this design is to strengthen the discriminative boundary of the global prototype in the feature space by increasing the distance between prototypes of different categories, thereby improving the model's ability to distinguish heterogeneous risks and achieving "inter-class separability".

[0176] In one embodiment, the contrast loss function The calculation can be performed using the following formula:

[0177] (13)

[0178] (14)

[0179] In another embodiment, it is proposed to introduce marginal parameters. This serves as the distinguishing boundary between positive and negative prototypes, effectively differentiating highly similar heterogeneous prototypes. Therefore, formula (13) can be replaced with the following:

[0180] (15)

[0181] Furthermore, in one specific embodiment, These can be hyperparameters that are set manually.

[0182] In another specific embodiment, it can be designed To adapt to different values ​​and for clarity, we will rewrite it as... Its calculation process may include:

[0183] 1) Calculate the central prototype of the local prototype corresponding to each risk category.

[0184] For example, for any risk category All its corresponding local prototypes Take the average to obtain the central prototype of this category. :

[0185] (16)

[0186] 2) Target category Calculate its central prototype and all other non-identical categories The similarity between the central prototypes is calculated, and the maximum value is selected:

[0187] (17)

[0188] This value represents the similarity between the two categories that are most likely to be misclassified in this round of training.

[0189] 3) Truncation is performed to obtain stable marginal values.

[0190] To prevent excessive gradient excitation due to excessively high maximum similarity, a preset truncation threshold is introduced. The smaller of the two values ​​is taken as the final marginal parameter:

[0191] (18)

[0192] Adaptive marginal parameters Substituting into formula (15), we can obtain the following contrast loss function:

[0193] (19)

[0194] Based on the above, the comparative loss can be determined.

[0195] Step S305: Update the trainable vectors and prototype generation model using the contrastive loss.

[0196] It should be understood that this step can utilize contrastive loss. Gradient descent is used to train the vectors. Model parameters of the GPG model Perform joint updates to enable the use of and The generated global prototype is closer to the positive sample prototype and farther from the negative sample prototype, thus gradually approaching the better global consensus.

[0197] Step S307: Input the updated trainable vectors into the updated prototype generation model to obtain the optimized global prototype. This can be denoted as:

[0198] (20)

[0199] in, and These are the trainable vectors and GPG model parameters updated in this training round, respectively. For the target portfolio (target risk category) -Target neighbor hop count The corresponding optimized global prototype.

[0200] Step S309: Based on the optimized global prototype, determine the corresponding target prototype.

[0201] This step aims to build upon the optimized global prototype generated in the previous steps. Determine the final target prototype for each client. To balance model generalization ability and local adaptability, two optional implementation paths are proposed below:

[0202] Implementation Path 1 (Global Unification): Directly implement the optimized global prototype. This serves as the target prototype for all client-side target combinations. It can be written as:

[0203] (twenty one)

[0204] Implementation Path Two (Personalized Customization): Considering the differences in data distribution among clients, and to improve adaptability to heterogeneous data, a personalized fusion mechanism is proposed. Based on the optimized global prototype, a target prototype that better matches the data distribution of each client is generated separately. Specifically:

[0205] 1) For any client Calculate its local prototype and other clients The similarity (e.g., cosine similarity) between uploaded local prototypes of the same target combination is calculated, and those with a similarity not lower than a preset threshold are selected. The client set corresponding to several local prototypes:

[0206] (twenty two)

[0207] in, For the client The local prototype; A similarity threshold (e.g., 0.7) is used to control the range of neighbors; Indicates with the client A set of semantically similar neighbors.

[0208] 2) Perform a first weighted fusion on the aforementioned local prototypes to obtain local prototypes.

[0209] For example, the first weighted fusion can be expressed as follows:

[0210] (twenty three)

[0211] in Represents a set Chinese client Previous category The number of samples, Represents the corresponding client in several local prototypes. The local prototype.

[0212] It should be understood that the first weighted fusion can also be achieved by other methods such as simple averaging.

[0213] 3) Perform a second weighted fusion on the optimized global prototype and local prototype to obtain the corresponding target prototype, which is then sent to the client.

[0214] For example, the second weighted fusion can be expressed as follows:

[0215] (twenty four)

[0216] in, This is the balancing coefficient, and its specific value can be set manually, with a range of (0,1).

[0217] From the above, we can obtain a target model that combines global and personalized features for each client.

[0218] The above introduces two advanced generation mechanisms for the target model in this solution. These two mechanisms are designed in a progressive architecture, together forming a two-layer optimization paradigm of "global consensus extraction + local personalized adaptation".

[0219] Mechanism 1: Global prototype generator based on contrastive learning.

[0220] This mechanism focuses on "commonality mining" on the server side. By introducing a topology-guided contrastive learning loss function, it drives the collaborative optimization of trainable vectors and prototype generation models (GPG) to adaptively extract robust and discriminative global risk patterns from heterogeneous local prototypes uploaded by multiple clients. Its core innovation lies in: instead of directly aggregating original prototypes, it implicitly models category semantics through neural generative networks; and it utilizes positive samples with hop counts to construct and adaptive marginal parameters to enhance intra-class consistency and inter-class separation, thereby overcoming the semantic ambiguity and noise sensitivity problems of traditional weighted averaging in non-independent and identically distributed scenarios.

[0221] Mechanism 2: Personalized Adaptive Fusion Mechanism (built upon Mechanism 1)

[0222] Building upon a high-quality global prototype, this mechanism further implements "personalized customization" for the client side. It abandons a uniform global distribution strategy for all clients, instead dynamically selecting high-confidence partners based on the similarity between each client's local prototype and other neighboring prototypes. This constructs local prototypes, and a weighted fusion of global consensus and local features is achieved through an adjustable balancing coefficient, generating a unique target prototype for each client. This approach not only preserves the generalization capability of the global model but also significantly improves the adaptation accuracy to local data distribution, achieving a leap from "general risk control" to "precise prevention and control."

[0223] In summary, the two mechanisms complement each other: Mechanism one ensures "accurate identification"—extracting stable and reliable global risk representations; Mechanism two ensures "skillful application"—allowing each node to obtain guidance signals most suited to its own business scenario. The combination of these two mechanisms makes this solution more robust, adaptable, and practical in complex financial risk control federated learning scenarios.

[0224] It should be noted that the aforementioned embodiments are mainly described in detail using a financial risk control scenario as an example. The applicant has also found that the federated training method proposed in this solution can be transferred to other business domains involving complex relationships. Specifically, in a financial risk control scenario, the object relationship graph can be a user transaction network (nodes are connected by transfers, payments, etc.), the business object is the user's financial account, and the object category is a risk label (such as fraud, money laundering, normal). In other business scenarios:

[0225] For example, in the context of social network security and public opinion analysis, the business object can be a user's social account, the object category can be a user status tag (such as: real user, marketing robot, high-influence user), and the object relationship graph can be constructed as a social relationship network (nodes are connected by relationships such as following, liking, and commenting).

[0226] It should be noted that while the attack patterns of online trolls faced by different social media platforms are often similar, their manifestations vary. Using this solution, each platform can upload local prototypes of marketing bots based on different levels of follower engagement (e.g., first-degree follower circle, second-degree follower circle). The server-side, through global aggregation, extracts the "common behaviors of the bots" (global prototype) across platforms, helping each platform more accurately identify new types of online trolls. Simultaneously, a personalized integration mechanism allows each platform to retain user characteristics unique to its own community culture (local prototype), thereby avoiding the mistaken targeting of active, genuine users.

[0227] For example, in e-commerce intelligent recommendation and product classification scenarios, the business object can be a product item, the object category corresponds to the product category tag (such as: digital products, fresh food, counterfeit and shoddy products), and the object relationship graph can be constructed as a product co-occurrence or related purchase graph (nodes are connected by implicit relationships such as "purchased by the same user" or "browsed by the same session").

[0228] It should be noted that for small and medium-sized e-commerce platforms, samples for certain long-tail categories are scarce (cold start problem). This solution allows multiple e-commerce platforms to collaborate on training. For example, for the category of "counterfeit luxury goods," each platform aggregates and uploads prototypes of its associated products. The global prototype generated on the server can capture cross-platform counterfeit association features (such as specific bundled sales patterns), and after distribution, it can significantly improve the counterfeit identification rate of smaller platforms. Furthermore, the "neighbor hop count" dimension here means: first-degree neighbors represent directly bundled products, and second-degree neighbors represent indirectly associated products; the two-dimensional prototype can more finely characterize the contextual semantics of the products.

[0229] For example, in the scenario of medical and health auxiliary diagnosis, the business object can be a patient's medical record or a medical image case, the object category can be the corresponding disease diagnosis result (such as: influenza, diabetes, health), and the object relationship graph can be constructed as a patient similarity graph or a symptom concurrency graph (nodes are connected by symptom similarity, gene homology or complication relationship).

[0230] It should be noted that medical data privacy is extremely sensitive, making federated learning particularly suitable. Different hospitals hold private patient maps. Using this approach, hospitals can collaborate to optimize disease prediction models. For example, for a rare disease (object category), by aggregating prototypes of multi-hop neighbors (patients with similar comorbidities), the model can learn more robust representations of pathological features. The introduction of adaptive marginal parameters is particularly important, as it helps the model distinguish between two different diseases with highly similar symptoms (such as the common cold and early influenza), improving diagnostic discriminative power without sharing any patients' original medical records.

[0231] In summary, regardless of whether the business object is a user, a product, or a medical case, and regardless of whether the object relationship is a cash flow, a social chain, or a pathological association, this solution can effectively solve the problem of non-independent and identically distributed data in distributed graph data in federated learning scenarios through a fine-grained prototype alignment mechanism of "object category - neighbor hop count".

[0232] Corresponding to the above-described federated method, the embodiments of this specification also disclose the following federated learning apparatus.

[0233] Figure 4A federated training device 400 illustrating a user risk assessment model is provided, wherein federated training involves a server and multiple clients, each client holding a private user relationship graph containing labeled user nodes with risk category tags. The device is integrated into any of these clients and includes the following functional modules:

[0234] Embedding processing module 410 is configured to process the user relationship graph using a locally deployed user risk assessment model to obtain the embedded representation of each labeled user node. Local prototype determination module 420 is configured to determine several local prototypes, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by aggregating the embedded representations of the following nodes: user nodes with that risk category, and neighbor nodes of that user node with that risk category under that neighbor hop count. Local prototype sending module 430 is configured to send the several local prototypes to the server. Target prototype receiving module 440 is configured to receive several target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes corresponding to the risk category-neighbor hop count combination uploaded by the multiple clients. Local model training module 450 is configured to train the user risk assessment model based on the several target prototypes.

[0235] In one embodiment, at least two of the plurality of clients deploy user risk assessment models with different model structures.

[0236] In one embodiment, the local prototype determination module 420 is specifically configured as follows: for any first combination among several risk category-neighbor hop count combinations, determine several first labeled user nodes in the user relationship graph that have the first risk category in the combination; for each first labeled user node, determine several neighbor nodes that have the first risk category under the first neighbor hop count in the combination; perform a first aggregation process on several embedded representations of the several neighbor nodes to obtain the neighborhood representation of the first labeled user node; and perform a second aggregation process on several neighborhood representations corresponding to the several first labeled user nodes to obtain the first local prototype corresponding to the first combination.

[0237] In one embodiment, the user risk assessment model includes an embedding component and a prediction component. The embedding processing module 410 is specifically configured to: process the user relationship graph using the embedding component to obtain the embedding representation of each labeled user node. The local model training module 450 is specifically configured to: process the embedding representation of each labeled user node using the prediction component to obtain the corresponding predicted risk result; and train the user risk assessment model based on a comprehensive loss, wherein the comprehensive loss includes a first loss term determined based on several target prototypes and several local prototypes, and a second loss term determined based on the predicted risk result and the corresponding risk category label.

[0238] In one embodiment, at least one of the plurality of clients determines a plurality of local prototypes.

[0239] Figure 5 A federated training device 500 for a user risk assessment model is shown, wherein federated training involves a server and multiple clients, each client holding a private user relationship graph containing labeled user nodes with risk category tags. The federated training device 500 is integrated into the server and includes the following functional modules:

[0240] The local prototype receiving module 510 is configured to receive several local prototypes uploaded by each client, wherein each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by the client by aggregating the embedded representations corresponding to the following nodes in the local environment: user nodes with the risk category, and neighbor nodes of the user node with the risk category under the neighbor hop count; the embedded representations are obtained by processing the user relationship graph using a locally deployed user risk assessment model. The target prototype determining module 520 is configured to determine multiple target prototypes, wherein each target prototype is generated based on the local prototypes corresponding to the same risk category-neighbor hop count combination among the received local prototypes. The target prototype sending module 530 is configured to send several target prototypes corresponding to the several local prototypes uploaded by each client, so that the client trains the user risk assessment model based on the several target prototypes.

[0241] In one embodiment, at least one of the plurality of clients uploads a plurality of local prototypes.

[0242] In one embodiment, the target combination corresponding to each target prototype includes a target risk category and a target neighbor hop count. The target prototype determination module 520 is specifically configured to: input the trainable vector corresponding to the target combination into a locally deployed prototype generation model to obtain an initial global prototype; determine a contrastive loss, which aims to increase the similarity between the initial global prototype and positive sample prototypes, and to increase the discriminative power between positive and negative sample prototypes; wherein the positive sample prototype is a local prototype corresponding to the target risk category but with a different neighbor hop count than the target neighbor hop count, and the negative sample prototype is a local prototype not corresponding to the target risk category; update the trainable vector and the prototype generation model using the contrastive loss; input the updated trainable vector into the updated prototype generation model to obtain an optimized global prototype; and determine the corresponding target prototype based on the optimized global prototype.

[0243] Furthermore, in a specific embodiment, a marginal parameter is introduced into the contrast loss as a distinguishing boundary between the positive sample prototype and the negative sample prototype. Determining the marginal parameter includes: calculating the central prototype of the local prototype corresponding to each risk category; for the target category, determining the similarity between its corresponding central prototype and the central prototypes corresponding to other risk categories, and selecting the maximum similarity; determining the smaller value between the maximum similarity and a preset cutoff threshold as the marginal parameter.

[0244] Furthermore, in one example, the target prototype determination module 520 is configured to determine the corresponding target prototype based on the optimized global prototype. Specifically, this includes: for each client, if it has uploaded a local prototype corresponding to the target combination, calculating the similarity between that local prototype and local prototypes uploaded by other clients corresponding to the target combination; filtering out several local prototypes with similarity scores greater than a predetermined threshold; performing a first weighted fusion on the several local prototypes to obtain a local prototype; and performing a second weighted fusion on the global prototype and the local prototypes to obtain the corresponding target prototype, which is then sent to the client.

[0245] Furthermore, in a specific example, the target prototype determination module 520 is configured to perform a first weighted fusion on the plurality of local prototypes to obtain a local prototype, specifically including: obtaining the number of labeled user nodes with the target risk category in each client corresponding to the plurality of local prototypes; calculating the proportion of the number of labeled user nodes of each client to the total number of labeled user nodes of all the plurality of clients, as their respective weighting coefficients; and performing a weighted summation on the plurality of local prototypes based on the weighting coefficients to obtain the local prototype.

[0246] Figure 6A federated training apparatus 600 for a business object prediction model is shown, wherein federated training involves a server and multiple clients, each client holding a private object relationship graph containing labeled object nodes with object category labels. The federated training apparatus 600 is integrated into any of the clients and includes the following functional modules:

[0247] The embedding processing module 610 is configured to process the object relationship graph using a locally deployed business object prediction model to obtain the embedding representation of each labeled object node.

[0248] The local prototype determination module 620 is configured to determine several local prototypes, wherein any local prototype corresponds to an object category-neighbor hop count combination, and is obtained by aggregation based on the embedded representations corresponding to the following nodes: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count.

[0249] The local prototype sending module 630 is configured to send the plurality of local prototypes to the server.

[0250] The target prototype receiving module 640 is configured to receive a plurality of target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes uploaded by the plurality of clients that correspond to the object category-neighbor hop count combination.

[0251] The local model training module 650 trains the business object prediction model based on the aforementioned target prototypes.

[0252] Figure 7 A federated training apparatus 700 for a business object prediction model is shown, wherein federated training involves a server and multiple clients, each client holding a private object relationship graph containing labeled object nodes with object category labels. The federated training apparatus 700 is integrated into the server and includes the following functional modules:

[0253] The local prototype receiving unit 710 is configured to receive several local prototypes uploaded by each client, wherein each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by the client by aggregating the embedding representations corresponding to the following nodes in the local environment: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count; the embedding representations are obtained by processing the object relationship graph using a locally deployed business object prediction model. The target prototype determining unit 720 is configured to determine multiple target prototypes, wherein each target prototype is generated based on local prototypes from the received local prototypes that correspond to the same object category-neighbor hop count combination. The target prototype sending unit 730 is configured to send several target prototypes corresponding to the several local prototypes uploaded by each client, so that the client trains the business object prediction model based on the several target prototypes.

[0254] It should be noted that for a description of the above functional modules, please refer to the relevant description of the process method in the foregoing embodiments.

[0255] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform... Figure 1 or Figure 2 or Figure 3 The method described.

[0256] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement... Figure 1 or Figure 2 or Figure 3 The method described.

[0257] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0258] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A federated training method for a user risk assessment model, involving a server and multiple clients, wherein each client holds a private user relationship graph, the user relationship graph containing labeled user nodes with risk category labels; The method is executed by any of the clients and includes: The user relationship graph is processed using a locally deployed user risk assessment model to obtain the embedded representation of each labeled user node; Several local prototypes are identified, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by aggregation based on the embedded representations of the following nodes: user nodes with the risk category, and neighbor nodes of the user node with the risk category under the neighbor hop count. Send the aforementioned local prototypes to the server; The server receives several target prototypes, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes uploaded by the multiple clients that correspond to the risk category-neighbor hop count combination; The user risk assessment model is trained based on the aforementioned target prototypes.

2. The method according to claim 1, wherein, Among the multiple clients, at least two clients deploy user risk assessment models with different model structures.

3. The method according to claim 1, wherein, Several local prototypes were identified, including: For any first combination among several risk category-neighbor hop count combinations, determine several first labeled user nodes in the user relationship graph that have the first risk category in that combination; For each first-labeled user node, identify several neighbor nodes that have the first risk category under the first neighbor hop count in the combination; A first aggregation process is performed on several embedded representations of the neighboring nodes to obtain the neighborhood representation of the first labeled user node. A second aggregation process is performed on the neighborhood representations corresponding to the plurality of first labeled user nodes to obtain the first local prototype corresponding to the first combination.

4. The method according to claim 1, wherein, The user risk assessment model includes an embedding component and a prediction component; wherein, the user relationship graph is processed using the locally deployed user risk assessment model to obtain the embedding representation of each labeled user node, including: The user relationship graph is processed using the embedded component to obtain the embedded representation of each labeled user node; Training the user risk assessment model includes: The embedding representation of each labeled user node is processed using the prediction component to obtain the corresponding prediction risk result; The user risk assessment model is trained based on a comprehensive loss, which includes a first loss term determined based on several target prototypes and several local prototypes, and a second loss term determined based on the predicted risk result and the corresponding risk category label.

5. The method according to claim 1, wherein, Among the multiple clients, at least one client determines a number of local prototypes.

6. A federated training method for a user risk assessment model, involving a server and multiple clients, wherein each client holds a private user relationship graph containing labeled user nodes with risk category labels; The method is executed by the server and includes: Each client uploads several local prototypes, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by the client by aggregating the embedded representations of the following nodes in the local machine: user nodes with the risk category, and neighbor nodes of the user node with the risk category under the neighbor hop count; the embedded representations are obtained by processing the user relationship graph using a locally deployed user risk assessment model; Multiple target prototypes are identified, each of which is generated based on the local prototypes received that correspond to the same risk category-neighbor hop number combination. Each client is sent with several target prototypes corresponding to several local prototypes uploaded by the client, so that the client can train the user risk assessment model based on the several target prototypes.

7. The method according to claim 6, wherein, Among the multiple clients, at least one client has uploaded a number of local prototypes.

8. The method according to claim 6, wherein, The target combination corresponding to each target prototype includes the target risk category and the target neighbor hop count; wherein, the determination of each target prototype includes: The trainable vectors corresponding to the target combination are input into the locally deployed prototype generation model to obtain the initial global prototype. The contrast loss is determined with the aim of increasing the similarity between the initial global prototype and the positive sample prototype, and increasing the discriminativeness between the positive sample prototype and the negative sample prototype; wherein the positive sample prototype is a local prototype that corresponds to the target risk category but has a different number of neighbor hops than the target neighbor hops, and the negative sample prototype is a local prototype that does not correspond to the target risk category. The trainable vectors and the prototype generation model are updated using the contrastive loss. The updated trainable vectors are input into the updated prototype generation model to obtain the optimized global prototype. Based on the optimized global prototype, the corresponding target prototype is determined.

9. The method according to claim 8, wherein, The contrastive loss incorporates a marginal parameter, which serves as the distinguishing boundary between the positive and negative sample prototypes; the determination of the marginal parameter includes: Calculate the central prototype of the local prototype corresponding to each risk category; For the target category, determine the similarity between its corresponding central prototype and the central prototypes corresponding to other risk categories, and select the maximum similarity among them; The smaller value between the maximum similarity and the preset truncation threshold is determined as the marginal parameter.

10. The method according to claim 9, wherein, Based on the optimized global prototype, the corresponding target prototype is determined, including: For each client, if it uploads a local prototype corresponding to the target combination, calculate the similarity between the local prototype and the local prototypes corresponding to the target combination uploaded by other clients. Select several local prototypes corresponding to similarities greater than a predetermined threshold; The aforementioned local prototypes are subjected to a first weighted fusion to obtain local prototypes; The global prototype and the local prototype are then subjected to a second weighted fusion to obtain the corresponding target prototype, which is then sent to the client.

11. The method according to claim 10, wherein, The aforementioned local prototypes are subjected to a first weighted fusion to obtain local prototypes, including: Obtain the number of labeled user nodes with the target risk category in each client corresponding to the plurality of local prototypes; Calculate the proportion of the number of labeled user nodes of each client to the total number of labeled user nodes of all the aforementioned clients, and use it as the weighting coefficient for each client; Based on the weighting coefficients, the local prototypes are summed using a weighted method to obtain the local prototype.

12. A federated training method for a business object prediction model, involving a server and multiple clients, wherein any client holds a private object relationship graph, the object relationship graph containing labeled object nodes with object category labels; The method is executed by any of the clients and includes: The object relationship graph is processed using a locally deployed business object prediction model to obtain the embedded representation of each labeled object node; Several local prototypes are identified, where each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by aggregation based on the embedding representations of the following nodes: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count. Send the aforementioned local prototypes to the server; Receive several target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes uploaded by the multiple clients that correspond to the object category-neighbor hop count combination; The business object prediction model is trained based on the aforementioned target prototypes.

13. A federated training method for a business object prediction model, involving a server and multiple clients, wherein each client holds a private object relationship graph, the object relationship graph containing labeled object nodes with object category labels; The method is executed by the server and includes: Each client uploads several local prototypes, where each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by the client by aggregating the embedded representations of the following nodes in the local machine: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count; the embedded representations are obtained by processing the object relationship graph using a locally deployed business object prediction model; Multiple target prototypes are identified, each of which is generated based on the local prototypes received that correspond to the same object category-neighbor hop count combination. Each client is sent with several target prototypes corresponding to several local prototypes uploaded thereon, so that the client can train the business object prediction model based on the several target prototypes.

14. A federated training apparatus for a user risk assessment model, wherein federated training involves a server and multiple clients, wherein each client holds a private user relationship graph containing labeled user nodes with risk category labels; The device is integrated into any of the client applications and includes: The embedding processing module is configured to process the user relationship graph using a locally deployed user risk assessment model to obtain the embedding representation of each labeled user node. The local prototype determination module is configured to determine several local prototypes, where each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by aggregation based on the embedded representations corresponding to the following nodes: user nodes with the risk category, and neighbor nodes of the user node with the risk category under the neighbor hop count. A local prototype sending module is configured to send the plurality of local prototypes to the server. The target prototype receiving module is configured to receive a number of target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes uploaded by the multiple clients that correspond to the risk category-neighbor hop count combination; The local model training module is configured to train the user risk assessment model based on the aforementioned target prototypes.

15. A federated training apparatus for a user risk assessment model, wherein federated training involves a server and multiple clients, wherein each client holds a private user relationship graph containing labeled user nodes with risk category labels; The device is integrated into the server and includes: The local prototype receiving module is configured to receive several local prototypes uploaded by each client, wherein each local prototype corresponds to a risk category-neighbor hop count combination, and is obtained by the client by aggregating the embedded representations corresponding to the following nodes in the local environment: user nodes with the risk category, and neighbor nodes of the user node with the risk category under the neighbor hop count; the embedded representations are obtained by processing the user relationship graph using a locally deployed user risk assessment model; The target prototype determination module is configured to determine multiple target prototypes, wherein each target prototype is generated based on the local prototypes that correspond to the same risk category-neighbor hop number combination from the received local prototypes; The target prototype sending module is configured to send several target prototypes corresponding to several local prototypes uploaded by each client, so that the client can train the user risk assessment model based on the several target prototypes.

16. A federated training apparatus for a business object prediction model, wherein federated training involves a server and multiple clients, wherein any client holds a private object relationship graph, the object relationship graph containing labeled object nodes with object category labels; the apparatus is integrated into any of the clients, comprising: The embedding processing module is configured to process the object relationship graph using a locally deployed business object prediction model to obtain the embedding representation of each labeled object node. The local prototype determination module is configured to determine several local prototypes, where each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by aggregation based on the embedding representations corresponding to the following nodes: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count. A local prototype sending module is configured to send the plurality of local prototypes to the server. The target prototype receiving module is configured to receive a number of target prototypes from the server, including a target prototype corresponding to any of the local prototypes, which is generated by the server based on the local prototypes uploaded by the multiple clients that correspond to the object category-neighbor hop count combination; The local model training module is configured to train the business object prediction model based on the aforementioned target prototypes.

17. A federated training apparatus for a business object prediction model, wherein federated training involves a server and multiple clients, wherein each client holds a private object relationship graph containing labeled object nodes with object category labels; The device is integrated into the server and includes: The local prototype receiving module is configured to receive several local prototypes uploaded by each client, wherein each local prototype corresponds to an object category-neighbor hop count combination, and is obtained by the client by aggregating the embedded representations corresponding to the following nodes in the local environment: object nodes with the object category, and neighbor nodes of the object node with the object category under the neighbor hop count; the embedded representations are obtained by processing the object relationship graph using a locally deployed business object prediction model; The target prototype determination module is configured to determine multiple target prototypes, wherein each target prototype is generated based on the local prototypes in the received local prototypes that correspond to the same object category-neighbor hop number combination; The target prototype sending module is configured to send several target prototypes corresponding to several local prototypes uploaded by each client, so that the client can train the business object prediction model based on the several target prototypes.

18. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed in a computer, it causes the computer to perform the method according to any one of claims 1-13.

19. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-13.

Images