Training method of decision support model, transaction decision method and related equipment

Abstract

This specification discloses a training method for a decision support model, a transaction decision-making method, and related equipment, relating to the field of model training technology. In this specification, based on an original graph constructed from sample data, multiple coarse-grained initial granular spheres are obtained. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels. The decision support model is further trained based on these multiple granular spheres. The decision support model uses a multi-column feature extraction network to focus on the sets of granular spheres at different granularity levels for feature extraction. Reversible connections are used to fuse cross-granularity feature information, thereby inferring evaluation results for multiple transaction indicators based on the fused feature information. The evaluation results of these multiple transaction indicators are used to assist in making decisions regarding target transactions.

Description

Technical Field

[0001] This manual relates to the field of model training, and in particular to a training method for a decision support model, a transaction decision-making method, and related equipment. Background Technology

[0002] In complex scenarios, transaction data exhibits significant high-order correlation characteristics. For example, the state characteristics of a target object may be associated with multiple dimensions such as its historical behavior records, interaction data of related objects, contextual scene distribution, and external evaluation parameters, while the interaction feedback of the target object may involve the cross-features of multiple modules such as functional attributes, process nodes, and interaction behavior data.

[0003] Traditional graph models can only capture binary pairwise relationships between nodes, failing to effectively model complex high-order interactions and leading to incomplete representation of multi-dimensional features during data analysis. Furthermore, existing hypergraph modeling methods often rely on predefined rules to generate hyperedges, neglecting the topological diversity of multi-source heterogeneous data and the multi-granularity of hyperedges.

[0004] Furthermore, due to the complexity of multi-source heterogeneous data, existing methods are prone to losing key feature information at different granularities during feature information fusion, resulting in insufficient high-order relation representation capabilities, especially in the recall and recognition of long-tail high-order patterns. Therefore, how to match the complex structure of transaction data and mine the correlation feature information between transaction data to support decision-making in complex scenarios has become a key technical challenge to improve decision-making accuracy. Summary of the Invention

[0005] This specification provides a training method for a decision support model, a transaction decision-making method, and related equipment, which can solve the above-mentioned problems. The technical solution is as follows: Firstly, embodiments of this specification provide a method for training a decision support model, wherein the decision support model is used to output evaluation results of multiple transaction indicators to assist in decision-making for a target transaction, and the method includes: Based on the original graph constructed from sample data, multiple initial granularity spheres are obtained; wherein, the initial granularity spheres are obtained by aggregating multiple nodes connected by hyperedges; Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively binary split to obtain multiple granular spheres with different granularity levels. The decision support model to be trained is trained based on multiple granular spheres until a preset convergence condition is met, thereby obtaining the decision support model; wherein, the decision support model to be trained includes a multi-column feature extraction network, and the training process includes extracting features from granular spheres at different granularity levels through the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results of multiple transaction indicators based on the fused feature information.

[0006] Secondly, embodiments of this specification provide a transaction decision-making method, the method comprising: Acquire the data to be processed for decision-making in the target transaction; Multiple granular spheres are obtained from the original graph constructed based on the data to be processed. The multiple granular spheres corresponding to the data to be processed are input into the decision support model to obtain the evaluation results of multiple transaction indicators. The decision support model is trained by the training method of the decision support model described in the first aspect. The decision for the target transaction is made based on the evaluation results of multiple transaction indicators, and the decision result is obtained.

[0007] Thirdly, embodiments of this specification provide a training apparatus for a decision support model, wherein the decision support model is used to output evaluation results of multiple transaction indicators to assist in decision-making for a target transaction, and the training apparatus for the decision support model includes: The hypergraph processing module is used to obtain multiple initial granularity spheres based on the original graph constructed from sample data; wherein, the initial granularity spheres are obtained by aggregating multiple nodes connected by hyperedges; The recursive splitting module is used to recursively split the initial granularity sphere into multiple granularity spheres based on the distance between multiple nodes in the initial granularity sphere, thereby obtaining multiple granularity spheres with different granularity levels. The model training module is used to train the decision support model to be trained based on multiple granular spheres until a preset convergence condition is met to obtain the decision support model; wherein, the decision support model to be trained includes a multi-column feature extraction network, and the training process includes extracting features from granular spheres at different granularity levels through the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results of multiple transaction indicators based on the fused feature information.

[0008] Fourthly, embodiments of this specification provide a transaction decision-making device, the transaction decision-making device comprising: The data acquisition module is used to acquire data to be processed for decision-making in the target transaction. The evaluation result module is used to obtain multiple granular spheres based on the original graph constructed based on the data to be processed, input the multiple granular spheres corresponding to the data to be processed into the decision support model, and obtain the evaluation results of multiple transaction indicators; wherein, the decision support model is trained by the training method of the decision support model described in the first aspect; The transaction decision module is used to make decisions about the target transaction based on the evaluation results of multiple transaction indicators, and obtain the decision results.

[0009] Fifthly, embodiments of this specification provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the above-described method steps.

[0010] Sixthly, embodiments of this specification provide a computer program product that stores multiple instructions adapted for loading by a processor and executing the above-described method steps.

[0011] In a seventh aspect, embodiments of this specification provide an electronic device that may include: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the above-described method steps.

[0012] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following: This specification's embodiments obtain multiple coarse-grained initial granular spheres from an original graph constructed based on sample data. Hyperedges are generated based on these granular spheres, fully preserving high-order interaction information between multiple nodes. Further, each initial granular sphere is split using a fine-grained adaptive splitting mechanism, resulting in multiple granular spheres with different granularity levels corresponding to the initial granular sphere. This adaptive splitting mechanism does not rely on a fixed distribution of sample data or manually preset rules. Instead, it allocates the sub-granular spheres to which nodes belong in each granular sphere based on the consistency of node features representing granularity and the distance of topological density, performing binary splitting on the initial granular spheres. This adaptive splitting mechanism can adapt to differences in association patterns between data from different domains, and also to association patterns in the same domain that change over time or in different scenarios. Further, a decision support model is trained based on the multiple granular spheres to obtain the decision support model. During training, the decision support model to be trained includes a multi-column feature extraction network, which focuses on feature extraction from granular spheres at different granularity levels. Reversible connections are used to fuse cross-granularity feature information, comprehensively capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This significantly improves the efficiency of the decision support model in capturing long-tailed high-order patterns and enhances the expressive power of the fused feature information. Based on this fused feature information, the decision support model infers evaluation results for multiple transaction metrics. These evaluation results are used to assist in making decisions regarding the target transaction, providing a basis for the processing of the target transaction. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the architecture of a training method for a decision support model provided in the embodiments of this specification; Figure 2 This is a schematic diagram of a recursive binary splitting of a particle size sphere provided in the embodiments of this specification; Figure 3 This is a schematic diagram illustrating an embodiment of obtaining multiple spheres of different particle sizes provided in this specification; Figure 4 This is a flowchart illustrating a training method for a decision support model provided in an embodiment of this specification. Figure 5 This is a schematic diagram of a process for obtaining initial particle size spheres provided in the embodiments of this specification; Figure 6This is a flowchart illustrating a training method for a decision support model provided in an embodiment of this specification. Figure 7 This is a flowchart illustrating a training method for a decision support model provided in an embodiment of this specification. Figure 8 This is a schematic diagram of a transaction decision-making process provided in an embodiment of this specification; Figure 9 This is a schematic diagram of the structure of a training device for a decision support model provided in the embodiments of this specification; Figure 10 This is a schematic diagram of the structure of a transaction decision-making device provided in the embodiments of this specification; Figure 11 This is a schematic diagram of the structure of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0015] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0016] In the description of this specification, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise expressly specified and limited, "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. Those skilled in the art can understand the specific meaning of the above terms in this specification based on the specific circumstances. Furthermore, in the description of this specification, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0017] The present specification will now be described in detail with reference to specific embodiments.

[0018] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the features, information, and data involved in this specification were all obtained under full authorization.

[0019] like Figure 1 As shown, Figure 1 This is a flowchart illustrating a training method for a decision support model provided in the embodiments of this specification. Figure 1 The system includes at least a server 101 for implementing a training method for a decision support model, and multiple electronic devices for receiving sample data and training requests. These multiple electronic devices include at least electronic devices 1021, 1022, and 1023. It is understood that... Figure 1 The number of servers and electronic devices shown is for illustrative purposes only, and the embodiments in this specification do not impose any limitations on them.

[0020] The aforementioned server 101 can be a standalone server device, such as a rack-mount, blade, tower, or cabinet-type server device, or a workstation, mainframe, or other hardware device with strong computing power; it can also be a server cluster composed of multiple servers. The servers in the service cluster can be composed in a symmetrical manner, where each server is functionally and hierarchically equivalent in the transaction chain, and each server can provide services to the outside world independently. Providing services independently can be understood as not requiring the assistance of other servers.

[0021] For example, a server can be multiple physical servers, each with independent hardware. Alternatively, a server can be multiple virtual servers deployed within the same hardware resource pool. Virtual server deployment methods include, but are not limited to, VMware, VirtualBox, and Virtual PC.

[0022] It is understood that server 101 also possesses other service capabilities and functions to complete the tasks described in the following embodiments. For example, server 101 also provides portal services, resource management services, and CI / CD services, etc.

[0023] Electronic devices include, but are not limited to: wearable devices, handheld devices, personal computers, tablets, in-vehicle devices, smartphones, computing devices, or other processing devices connected to a wireless modem. Electronic devices may have different names in different networks, such as: user equipment, access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication equipment, user agent or user device, cellular phone, cordless phone, personal digital assistant (PDA), and electronic devices in 5G networks or future evolved networks.

[0024] In the embodiments of this specification, electronic devices such as electronic devices 1021, 1022, and 1023 may also be equipped with display devices. The display devices can be various devices capable of displaying functions, such as cathode ray tube displays (CR), light-emitting diode displays (LED), electronic ink screens, liquid crystal displays (LCD), plasma display panels (PDP), etc.

[0025] For example, a user can use the display device on electronic device 1021 to send a training request for a decision support model to server 101, so that server 101 can train the decision support model to be trained based on the request.

[0026] Multiple electronic devices and multiple servers can communicate through communication links established by communication protocols. For example, the network can be a wireless network or a wired network. Wireless networks include, but are not limited to, cellular networks, wireless LANs, infrared networks, or Bluetooth networks. Wired networks include, but are not limited to, Ethernet, universal serial bus (USB), or controller area networks. In one or more embodiments of the specification, technologies and / or formats including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network (such as target compressed packets). Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), and Internet Protocol Security (IPsec) can be used to encrypt all or some of the links. In other embodiments, customized and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0027] In the embodiments described in this specification, the decision support model is used to output the evaluation results of multiple transaction metrics to assist in the decision-making of the target transaction, such as assisting the target transaction in classification, identification, or prediction operations. Transaction metrics are not isolated numerical statistics, but rather integrate the semantics of the target transaction, and can drive the evaluation dimensions of the target transaction decision.

[0028] Taking product interaction as an example, the decision support model can be applied to decision-making for target transactions such as abnormal node identification, interaction pattern mining, node state classification, and product recommendation. Transaction indicators can include node anomaly probability, correlation strength values ​​with other nodes, and state classification parameters. Input data can include user behavior time-series data, interaction scenario data, credit assessment parameters, and interaction feedback data. The above data have high-dimensional and high-order correlation characteristics; for example, the abnormal state of a single node may be associated with multiple related nodes, multiple product types, and multiple historical behavior nodes.

[0029] Taking a target transaction involving interactive recommendation networks as an example, the decision support model can be applied to decision-making for target transactions such as node preference prediction, interaction link identification, and abnormal behavior detection. Transaction indicators can include interaction probability, conversion probability, node interest features, etc. Input data can include user browsing behavior data, social relationship graphs, etc. The above data naturally exhibit high-dimensional and high-order correlation characteristics; for example, the interaction tendency of a single node may be associated with multiple browsing scenarios and multiple historical interaction nodes.

[0030] It is understood that the domain to which the target transaction belongs is merely an example, and this specification also includes the possibility that the target transaction belongs to other domains.

[0031] In one embodiment, such as Figure 2 The diagram shown illustrates a flowchart of a training method for a decision support model provided in an embodiment of this specification. This method can be implemented using a computer program and can run on a training device for decision support models based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application.

[0032] Specifically, the training methods for this decision support model include: S102. Based on the original graph constructed from the sample data, obtain multiple initial granular spheres.

[0033] The sample data consists of historical transaction data used to train the decision support model, specifically data that supports the target transaction decision. The sample data can be obtained from multiple information sources, including user behavior logs, interaction records, and service usage feedback.

[0034] The original graph is constructed based on the sample data. The nodes in the original graph are the transaction entities in the sample data, and the original graph structure is G=(V,E,X,A), where V is the set of nodes, such as users, interactions, products, risk indicators, etc., E is the set of original binary edges, such as the association between users and interaction institutions, and the association between users and products, X is the node features, such as user credit assessment parameters, interaction amount parameters, product attribute parameters, etc., and A is the adjacency matrix, which describes whether there is a connection between nodes.

[0035] like Figure 3 As shown, Figure 3 This is a schematic diagram illustrating an embodiment of obtaining multiple granular spheres provided in this specification. An original graph 201 is constructed based on sample data. The original graph 201 includes multiple nodes, and nodes with related relationships are connected by edges.

[0036] A granular sphere can be understood as a closed sphere subset defined based on the topological structure and node feature information of the original graph. It is determined by the central node and the radius, which must satisfy local connectivity and semantic cohesion. It is used to partition data into different granular subsets. Multiple initial granular spheres constitute a hypergraph.

[0037] A hypergraph can be understood as a graph structure that can connect multiple nodes at the same time. Hyperedges can be associated with any number of nodes and are used to model higher-order relationships that traditional binary graphs cannot represent, such as the multi-node association of "user behavior data - interaction scenario - credit assessment parameters - risk indicators".

[0038] Based on the original graph constructed from the sample data, multiple initial granularity spheres are obtained. These initial granularity spheres are aggregated by connecting multiple nodes with hyperedges. In other words, local association groups are generated on the original graph according to different radii or rules, with each local association group centered around a specific node. Figure 3 As shown, coarse-grained initial granular spheres 202 and 203 are obtained based on the original figure 201.

[0039] Methods for constructing multiple initial granular spheres can be based on hop-based neighborhood expansion. Alternatively, they can be based on community detection methods, using algorithms such as Louvain and Label Propagation to divide the original graph into several "communities," each of which is a granular sphere. Manual grouping based on a specific transaction logic is also possible; for example, connecting the nodes corresponding to multiple users logged in from the same device via hyperedges can create a single granular sphere.

[0040] S104. Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively binary split to obtain multiple granular spheres with different granularity levels.

[0041] The distance between multiple nodes in the initial granularity sphere can be the topological distance or the hop count (Hop) of the shortest path between two points. If two nodes in the granularity sphere require a very long path to connect, it indicates that they are loosely connected, and it is appropriate to split these two nodes into different sub-granularity spheres.

[0042] The distance between multiple nodes in the initial granularity sphere can also be the feature distance, which is the Euclidean or cosine distance between the node attribute vectors in multidimensional space. If the node attributes differ greatly between nodes in the initial granularity sphere, the feature distance between the nodes is large, and the two nodes mentioned above should be split into different granularities.

[0043] Specifically, for each initial granularity sphere, a distance matrix is ​​constructed for its internal nodes. Then, using a farthest-point-pair strategy or spectral bisection, multiple nodes within the initial granularity sphere are divided into two mutually exclusive subsets. The partitioning rule minimizes the internal distance of each subset while maximizing the inter-subset distance, thus generating two sub-granularity spheres. The above distance calculation and splitting judgment are then recursively performed on each generated sub-granularity sphere until a stopping condition is met, such as the number of nodes in a granularity sphere falling below a minimum threshold, the splitting gain being less than a preset value, or the maximum recursion depth being reached.

[0044] Multiple spheres of different particle sizes generated during the splitting process are assigned to the set of spheres corresponding to the initial sphere. For example, in the set of spheres corresponding to each initial sphere, there are n1 coarse-grained spheres, n2 medium-coarse-grained spheres, n3 fine-grained spheres, and n4 extremely fine-grained spheres.

[0045] like Figure 3 As shown, splitting the initial particle size sphere 202 yields particle size spheres 2021 and 2022, and splitting the initial particle size sphere 203 yields particle size spheres 2031 and 2032. Continuing to split particle size sphere 2021 yields two sub-particle size spheres, and splitting particle size sphere 2022 yields two sub-particle size spheres, splitting particle size sphere 2031 yields two sub-particle size spheres, and splitting particle size sphere 2032 yields two sub-particle size spheres, thus achieving recursive binary splitting.

[0046] In one embodiment, the initial granular sphere is recursively binary-splitted based on the distance between multiple nodes in the initial granular sphere, and the splitting quality gain is calculated at each split. If the splitting quality gain is not positive, the splitting is stopped until multiple granular spheres corresponding to the initial granular sphere are obtained.

[0047] For each coarse-grained initial particle size sphere The two distinct nodes with the highest degree within the initial granular sphere are selected as the centers of the sub-granular spheres. Based on the distance metric d, the original granular sphere is bisected into two fine-grained spheres. and .

[0048] The change in the quality index of the granular sphere before and after splitting is calculated, i.e., the splitting quality gain. The quality index characterizes the compactness of the granular sphere and can be represented by the comment distance, profile coefficient, or tag purity of multiple nodes. In this embodiment, the splitting quality gain is the difference between the sum of the quality indices of the two sub-granular spheres after splitting and the quality index of the granular sphere before splitting. The quality index of the granular sphere characterizes the consistency of node features and the compactness of the topological structure.

[0049] If the split mass gain is positive, it means that the nodes inside the granular sphere are more compact after splitting, and the fine-grained sphere can continue to be split to obtain two granular spheres with even finer grains.

[0050] If the splitting mass gain is not positive, it means that the originally tightly packed fine-grained sphere has been broken up, so the splitting of the fine-grained sphere should be stopped.

[0051] Traditional binary splitting methods often predetermine a fixed number of splitting layers, such as forcing splitting into 3 layers or a fixed number of clusters, which can easily lead to over- or under-splitting of the initial granular spheres. This embodiment, however, calculates the splitting quality gain by combining feature information and dynamically adjusts the recursive binary splitting of the granular spheres. For granular spheres with a compact structure and small internal feature differences, splitting is automatically stopped to maintain their integrity. For granular spheres with a loose structure and large internal feature differences, splitting continues until the splitting quality gain is no longer positive, resulting in multiple purer sub-granular spheres that can adapt to the true distribution of the data.

[0052] S106. Train the decision support model to be trained based on multiple granular spheres until the preset convergence condition is met, and obtain the decision support model.

[0053] The decision support model to be trained includes a multi-column feature extraction network. The training process includes extracting features from granular spheres at different granularity levels through the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results of multiple transaction indicators based on the fused feature information.

[0054] Specifically, the decision support model to be trained consists of multiple parallel feature extraction networks, each receiving a full set of multiple granular spheres as input. These full set of multiple granular spheres refer to the multiple granular spheres corresponding to each initial granular sphere.

[0055] Each feature extraction network focuses on a set of granular spheres at different granularity levels, extracting features from the nodes and hyperedges within that set to obtain the feature information of the nodes and hyperedges. For example, the first column of the feature extraction network corresponds to multiple granular spheres at a coarse-grained level, and the second column corresponds to multiple granular spheres at a medium-coarse-grained level. Each column of the feature extraction network can employ a graph neural network, Transformer, or CNN, and the feature information of the nodes and hyperedges is represented using vector forms.

[0056] like Figure 4 As shown, Figure 4 This is a flowchart illustrating a training method for a decision support model provided in an embodiment of this specification. Multiple granular spheres are input into the decision support model 204 to be trained. The feature extraction network 2041 extracts feature information of the hyperedges corresponding to the multiple granular spheres in the hypergraph, including hyperedge feature information e1, e2, e3, and e4, and updates the aforementioned hyperedge feature information until feature information for cross-column fusion is output.

[0057] Furthermore, cross-column fusion of multi-column feature information yields fused feature information, addressing the issue of one-sidedness caused by single-granularity information. For example, the feature information output by the k1-column feature extraction network is combined with k... n The feature information output by the column feature information network is fused, and the feature information output by the k2 column feature extraction network is combined with k m The feature information output by the feature information network is fused. For example... Figure 4 As shown, feature extraction networks 2041, 2042, and 2043, along with other feature extraction networks (not shown in the figure), are reversibly connected, and the fusion feature information extracted by the above multiple feature extraction networks is fused across columns to obtain fusion feature information 205.

[0058] Multiple output nodes are set according to decision-making needs, each corresponding to a different transaction metric. Evaluation results of multiple transaction metrics are output based on fused feature information. For example, transaction metric A is the node anomaly probability, transaction metric B is the association strength value, and transaction metric C is the state evaluation description. The evaluation results 2061 for transaction metric A and 2062 for transaction metric B are both numerical values, while the evaluation result 2063 for transaction metric C can be textual information.

[0059] Finally, a loss function is constructed to calculate the error between the output of the decision support model being trained and the true label. Cross-entropy loss or mean squared error can be used to construct the loss function. Gradient descent is then used for backpropagation, propagating the error back to each layer of the multi-column feature extraction network and updating the weights of the feature extraction network. Training stops when a preset convergence condition is met, resulting in the decision support model.

[0060] This specification describes an embodiment that uses sample data to construct an original graph, obtaining multiple coarse-grained initial granular spheres. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels. This adaptive splitting mechanism assigns sub-granular spheres to nodes within each granular sphere based on distance, representing the consistency of node features and the density of the topological structure. This adapts to differences in association patterns between data from different domains, as well as to association patterns within the same domain that change over time or in different scenarios. Further, a decision support model is trained using these multiple granular spheres. During training, the decision support model includes a multi-column feature extraction network that focuses on extracting features from granular sphere sets at different granularity levels. Reversible connections are used to fuse cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results for multiple transaction indicators based on the fused feature information. The evaluation results of these multiple transaction indicators are used to assist in making decisions regarding target transactions, providing a basis for target transaction processing.

[0061] In one embodiment, based on Figure 2 - Figure 4 Please refer to the embodiments shown in the figure as well. Figure 5 . Figure 5 This is a flowchart illustrating a training decision support model provided in an embodiment of this specification. S102 includes the following steps: S102-1. Obtain the original graph constructed based on the sample data.

[0062] The original graph includes nodes corresponding to multiple transaction entities.

[0063] S102-2. Determine at least one target node among the multiple nodes included in the original graph, and obtain an initial granularity sphere with the target node as the core and connected to at least one node through a hyperedge.

[0064] This step is the seed selection and cluster generation stage. The target node is determined from multiple nodes to serve as the "seed". Then, cluster generation is performed around each target node. That is, the selected target node is used as the center to radiate outwards, and at least one other node with strong correlation with it is pulled in to form a set. The nodes in the set are connected by hyperedges to obtain the initial granularity sphere.

[0065] The target node is determined based on degree centroid selection, specifically choosing the node with the most connected edges. A higher degree indicates a more active target node in the network, covering more neighbors. For example, the user with the highest interaction frequency can be selected as the target node.

[0066] Specifically, the node degree is initialized and calculated. Node degree represents the number of nodes directly connected to that node, such as the interaction frequency of users (nodes) or the number of users associated with a product (node). Furthermore, the node with the highest degree is selected. N nodes are selected as target nodes, where N is the total number of nodes, ensuring coverage of core related nodes.

[0067] The determination of target nodes can also be based on betweenness centrality, specifically by selecting nodes that are on the shortest path from many other nodes. These nodes are often the "bridges" or "hubs" of the network, controlling the flow of information / risk.

[0068] Target nodes can also be determined by directly using known labels. For example, if some nodes have been marked as "high anomaly probability" or "abnormal sample", then these nodes can be selected as target centers based on these labels.

[0069] Furthermore, a distance metric d between nodes is defined. Based on the characteristics of transaction data in various fields, a hybrid distance function is used to fuse the topological distance of the structural relationship between nodes and the feature distance that characterizes the attribute differences between nodes to obtain the distance between nodes.

[0070] Finally, the remaining nodes are assigned to the nearest central node to the target node using a breadth-first search, forming an initial coarse-grained sphere. Each initial granularity sphere satisfies local connectivity and semantic cohesion, meaning that there are multiple highly correlated nodes within the same initial granularity sphere.

[0071] In this embodiment, the target node is determined by the node degree to ensure coverage of the core transaction entity. Instead of building a sphere for all N nodes, only the target node is selected. The single center significantly reduces the computational cost of subsequent hyperedge construction, improves model training efficiency, and uses breadth-first search to allocate nodes to the initial granular sphere, automatically balancing the sphere size and avoiding extreme distributions.

[0072] In one embodiment, such as Figure 6 The diagram shown illustrates a flowchart of a training method for a decision support model provided in an embodiment of this specification. This method can be implemented using a computer program and can run on a training device for decision support models based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application.

[0073] Specifically, the training methods for this decision support model include: S202. Based on the original graph constructed from the sample data, obtain multiple initial granular spheres.

[0074] See S102 above, which will not be repeated here.

[0075] S202. Among the multiple nodes included in the initial granular sphere, determine two split nodes.

[0076] For each coarse-grained initial particle size sphere The two different splitting nodes with the highest node degree within the initial particle size sphere are selected as the centers of the sub-particle size sphere.

[0077] S202. Calculate the distance between each node in the initial particle size sphere and the two split nodes.

[0078] In this embodiment, the fusion distance between each node in the initial granularity sphere and the target split node is calculated based on multiple dimensions. These multiple dimensions include at least a topological distance dimension characterizing the structural relationships between nodes, and a feature distance dimension characterizing the attribute differences between nodes.

[0079] Define a distance metric d between nodes. Based on the characteristics of transaction data in various domains, a hybrid distance function is used to fuse the topological distance of the structural relationships between nodes and the feature distance representing the attribute differences between nodes to obtain the distance between nodes. For example, the feature distance between nodes can represent the similarity of attribute parameters and interaction behaviors between data objects, while the topological distance can represent the group association relationship and operation sequence similarity between data objects.

[0080] S202. Based on multiple distances, divide the multiple nodes in the initial granularity sphere into two sub-granularity spheres centered on the splitting node.

[0081] The initial granular sphere is split into two fine-grained spheres based on the distance metric d. and .

[0082] S202. Recursively execute the above binary splitting steps on the two sub-granular spheres until the preset stopping condition is met, to obtain multiple granular spheres with different granularity levels corresponding to the initial granular sphere.

[0083] Define particle size distribution index for ball quality assessment ( (where V is the number of edges inside the sphere and V is the number of nodes inside the sphere). The quality index of the granular sphere characterizes the consistency of node features and the density of the topological structure.

[0084] If the masses of the two sub-spheres after splitting are equal to... If the splitting result is not positive, the splitting continues. If the splitting mass gain of the sub-particle size spheres is not positive, the process continues until the mass requirement is met.

[0085] The final multi-granularity sphere set is obtained after iterative splitting. Each granular sphere corresponds to a hyperedge, forming a multi-granularity hypergraph. . For a set of hypergraph nodes, It is a set of superedges.

[0086] S212. Train the decision support model to be trained based on multiple granular spheres until the preset convergence condition is met, and obtain the decision support model.

[0087] See S106 above; it will not be repeated here.

[0088] This specification describes an embodiment that uses sample data to construct an original graph, obtaining multiple coarse-grained initial granular spheres. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels. This adaptive splitting mechanism assigns sub-granular spheres to nodes within each granular sphere based on distance, representing the consistency of node features and the density of the topological structure. This adapts to differences in association patterns between data from different domains, as well as to association patterns within the same domain that change over time or in different scenarios. Further, a decision support model is trained using these multiple granular spheres. During training, the decision support model includes a multi-column feature extraction network that focuses on extracting features from granular sphere sets at different granularity levels. Reversible connections are used to fuse cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results for multiple transaction indicators based on the fused feature information. The evaluation results of these multiple transaction indicators are used to assist in making decisions regarding target transactions, providing a basis for target transaction processing.

[0089] In one embodiment, such as Figure 7 The diagram shown illustrates a flowchart of a training method for a decision support model provided in an embodiment of this specification. This method can be implemented using a computer program and can run on a training device for decision support models based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application.

[0090] Specifically, the training methods for this decision support model include: S302. Based on the original graph constructed from the sample data, obtain multiple initial granular spheres.

[0091] See S102 above, which will not be repeated here.

[0092] S304. Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively binary split to obtain multiple granular spheres with different granularity levels.

[0093] See S304 above; it will not be repeated here.

[0094] S306. Input multiple granular spheres into a multi-column feature extraction network, and extract the feature information of multiple granular spheres at the corresponding granularity level through the feature extraction network to obtain multi-column feature information.

[0095] Different feature extraction networks correspond to different granularity levels. For example, low-column feature extraction networks learn fine-grained hyperedge feature information, while high-column feature extraction networks learn coarse-grained hyperedge feature information.

[0096] In one embodiment, different weights are assigned to multiple granular spheres based on their quality index and the importance of multiple nodes within each granular sphere to the target transaction. The decision support model is then trained using these weighted granular spheres until a preset convergence condition is met, resulting in the final decision support model.

[0097] The quality index of the granular sphere is obtained based on the consistency of node characteristics and the density of the topological structure.

[0098] To determine the importance of multiple nodes in a granular sphere to a target transaction, the metric can be whether the granular sphere contains risk nodes, such as nodes representing abnormal states; a higher proportion of risk nodes indicates greater importance. Another metric is the degree centrality of the granular sphere's nodes; higher degree centrality indicates greater importance. For example, if the target transaction is state assessment, the importance of the granular sphere can be determined based on the contribution or risk correlation of multiple nodes in the granular sphere to the state assessment decision.

[0099] The weight of each granular sphere is calculated based on the two dimensions mentioned above, and the decision support model to be trained is trained based on multiple granular spheres with assigned weights. During model training, when the feature extraction network extracts feature information from each granular sphere, it calculates based on the weight of that granular sphere. The feature information corresponding to hyperedges with high weights is amplified, while the feature information corresponding to hyperedges with low weights is suppressed.

[0100] In this embodiment, weights are assigned to each hyperedge to enhance the feature representation of key high-order relations. This makes it easier for the decision support model to capture hidden features that are small in sample size but have great harm, thus solving the problem of insufficient sensitivity of traditional models to long tails.

[0101] In one embodiment, multiple granular spheres are input into a multi-column feature extraction network; each feature extraction network includes multiple attention convolutional blocks connected in sequence. Through the multiple attention convolutional blocks in the feature extraction network, iterative feature extraction is performed on the multiple granular spheres corresponding to the granularity level of the feature extraction network, updating the feature data of each node multiple times until feature information including the target feature data corresponding to each of the multiple nodes is obtained.

[0102] Specifically, the feature extraction network can include four sequentially connected attention convolutional blocks, and calculate the attention score between nodes and hyperedges based on two principles: 1. Nodes connected by the same hyperedge have high feature similarity and stronger information propagation; 2. The information of hyperedges with high weights is more reliable.

[0103] Based on the above two principles, the attention score between the node and the hyperedge is calculated: (σ is the non-linear activation function, a is the attention parameter vector, and W is the weight matrix,) For nodes In the feature vector of the current network layer, For super-edge (Feature vectors in the current network layer).

[0104] After calculating the attention scores of a node relative to multiple hyperedges, the attention scores of the node are normalized. The normalization formula includes: . For including nodes The set of superedges.

[0105] Based on the normalized attention score, the feature data of the node is calculated by concatenating multiple feature vectors, and the feature data of the hyperedge is updated based on the feature information corresponding to the multiple nodes connected by the hyperedge.

[0106] In the next attention convolutional block, based on the updated hyperedge data, the feature data of the nodes are updated to obtain the updated feature vector. The formulas for feature updates include: .

[0107] After iteratively extracting features from the multi-granularity spheres through multiple attention convolutional blocks, the feature data of each node is updated multiple times until feature information including target feature data corresponding to multiple nodes is obtained.

[0108] S308. Through the reversible connection layer of the decision support model to be trained, cross-column reversible connection and feature fusion are performed on the multi-column feature information to obtain fused feature information.

[0109] The reversible connection layer of the decision support model to be trained concatenates the feature information from multiple parallel columns from the previous step using a reversible function, and obtains the fused feature information by cross-column fusion.

[0110] In one embodiment, the reversible connection layer of the decision support model to be trained performs cross-column reversible connection and feature fusion on multiple columns of feature information, and retains the feature information corresponding to the cross-column starting layer to participate in the calculation during the forward propagation of information until the fused feature information is obtained.

[0111] Specifically, the forward propagation formula used during training is as follows: (Ft is the nonlinear residual function, γ is the invertible scaling parameter, x) t For the feature information of the current layer, x t-1 For the feature information of the previous layer, x t-m+1 (for the feature information of the intermediate layer across columns), thus preserving the feature information x corresponding to the starting layer across columns during the forward propagation of information. t-m It participates in the calculation, retaining all information while achieving multi-granularity feature fusion.

[0112] And the backpropagation formula is used: This ensures gradient stability.

[0113] S310. Through the inference layer of the decision support model to be trained, the evaluation results of multiple transaction indicators are output based on the fused feature information.

[0114] In the inference layer, multiple output nodes are set according to decision-making requirements, with each output node corresponding to a different transaction metric. The evaluation results of multiple transaction metrics are output based on the fused feature information.

[0115] S312. Based on the difference information between multiple evaluation results and multiple sample evaluation results, iteratively train the decision support model to be trained until the preset convergence condition is met, and obtain the decision support model.

[0116] In this embodiment, the Adam optimizer is used to minimize the cross-entropy loss function, and the regularization parameter is set to... The learning rate was 0.0001, the hidden dimension was 16, and the number of attention heads was 8. Based on the above parameter settings, the decision support model to be trained was iteratively trained according to the differences between multiple evaluation results and multiple sample evaluation results to ensure the generalization ability of the decision support model.

[0117] This specification describes an embodiment that uses sample data to construct an original graph, obtaining multiple coarse-grained initial granular spheres. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels. This adaptive splitting mechanism assigns sub-granular spheres to nodes within each granular sphere based on distance, representing the consistency of node features and the density of the topological structure. This adapts to differences in association patterns between data from different domains, as well as to association patterns within the same domain that change over time or in different scenarios. Further, a decision support model is trained using these multiple granular spheres. During training, the decision support model includes a multi-column feature extraction network that focuses on extracting features from granular sphere sets at different granularity levels. Reversible connections are used to fuse cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results for multiple transaction indicators based on the fused feature information. The evaluation results of these multiple transaction indicators are used to assist in making decisions regarding target transactions, providing a basis for target transaction processing.

[0118] In one embodiment, such as Figure 8 The diagram shown is a flowchart illustrating a transaction decision-making method provided in an embodiment of this specification. This method can be implemented using a computer program and can run on a transaction decision-making device based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone utility application.

[0119] Specifically, the decision-making method for this transaction includes: S402. Obtain the data to be processed for the target transaction decision.

[0120] The data to be processed and the target transaction decision reside in the same domain. Taking an interactive network as an example, the data to be processed may include user behavior time-series data, interaction scenario data, credit assessment parameters, interaction feedback data, etc.

[0121] S404. Obtain multiple granular spheres from the original graph constructed based on the data to be processed, input the multiple granular spheres corresponding to the data to be processed into the decision support model, and obtain the evaluation results of multiple transaction indicators.

[0122] An original graph is constructed based on the data to be processed, yielding multiple initial granularity spheres. Each initial granularity sphere is obtained by aggregating multiple nodes connected by hyperedges. Furthermore, based on the distances between multiple nodes in the initial granularity sphere, a recursive binary split is performed on the initial granularity sphere to obtain multiple granularity spheres with different granularity levels.

[0123] Multiple granular spheres corresponding to the data to be processed are input into the decision support model to obtain evaluation results for multiple transaction indicators. The decision support model is trained using any of the aforementioned decision support model training methods.

[0124] S406. Make a decision on the target transaction based on the evaluation results of multiple transaction indicators, and obtain the decision result.

[0125] Transaction metrics can include node anomaly probability, association strength value, state classification parameters, etc. The evaluation results corresponding to these multiple transaction metrics output by the decision support model are obtained, and decision instructions are generated based on these evaluation results. For example, if the target transaction is anomaly pattern recognition, based on the evaluation result corresponding to the transaction metric of node anomaly probability, the decision instruction is to mark the target node as an anomaly state and trigger the target device to execute a preset response operation.

[0126] In this specification, multiple coarse-grained initial granular spheres are obtained from the original graph constructed based on the data to be processed. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels, fully preserving high-order interaction information between multiple nodes. The decision support model utilizes a multi-column feature extraction network to focus on the granular sphere sets at different granularity levels for feature extraction. Reversible connections are used to achieve the fusion of cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results of multiple transaction indicators based on the fused feature information. Based on these evaluation results, decisions are made for the target transaction, providing a basis for target transaction processing and improving decision accuracy.

[0127] The following are embodiments of the apparatus described in this specification, which can be used to execute the embodiments of the methods described in this specification. For details not disclosed in the apparatus embodiments of this specification, please refer to the embodiments of the methods described in this specification.

[0128] Please see Figure 9 This diagram illustrates a structural schematic of a training apparatus for a decision support model provided in an exemplary embodiment of this specification. The training apparatus for the decision support model can be implemented as all or part of the apparatus through software, hardware, or a combination of both. The training apparatus for the decision support model includes a hypergraph processing module 501, a recursive splitting module 502, and a model training module 503.

[0129] The hypergraph processing module 501 is used to obtain multiple initial granularity spheres based on the original graph constructed based on sample data; wherein the initial granularity spheres are obtained by aggregating multiple nodes connected by hyperedges; The recursive splitting module 502 is used to recursively split the initial granular sphere into multiple granular spheres with different granularity levels based on the distance between multiple nodes in the initial granular sphere. The model training module 503 is used to train the decision support model to be trained based on multiple granular spheres until a preset convergence condition is met, thereby obtaining the decision support model. The decision support model to be trained includes a multi-column feature extraction network. The training process includes extracting features from granular spheres at different granularity levels using the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results for multiple transaction indicators based on the fused feature information.

[0130] In one embodiment, the recursive splitting module 502 includes: The first recursive splitting unit is used to determine two splitting nodes among the multiple nodes included in the initial granular sphere; The second recursive splitting unit is used to calculate the distance between each node in the initial granular sphere and the two splitting nodes; The third recursive splitting unit is used to divide multiple nodes in the initial granularity sphere into two sub-granularity spheres centered on the splitting node, according to the multiple distances. The fourth recursive splitting unit is used to recursively execute the above binary splitting steps on the two sub-granular spheres until a preset stopping condition is met, thereby obtaining multiple granular spheres with different granularity levels corresponding to the initial granular sphere.

[0131] In one embodiment, the second recursive splitting unit is specifically used to calculate the fusion distance between each node in the initial granularity sphere and the target splitting node based on multiple dimensions; wherein, the multiple dimensions include at least a topological distance dimension characterizing the structural relationship between nodes, and a feature distance dimension characterizing the attribute differences between nodes.

[0132] In one embodiment, the recursive splitting module 502 includes: The fifth recursive splitting unit is used to recursively split the initial granular sphere into binary spheres based on the distance between multiple nodes in the initial granular sphere, and to calculate the splitting quality gain at each split. If the splitting quality gain is not positive, the splitting is stopped until multiple granular spheres corresponding to the initial granular sphere are obtained.

[0133] In one embodiment, the split mass gain is the difference between the sum of the mass indices of the two sub-spheres after splitting and the mass index of the sphere before splitting, whereby the mass index of the sphere characterizes the consistency of node features and the compactness of the topology of the sphere.

[0134] In one embodiment, the hypergraph processing module 501 includes: The first hypergraph processing unit is used to obtain the original graph constructed based on sample data; wherein the original graph includes nodes corresponding to multiple transaction entities; The second hypergraph processing unit is used to determine at least one target node among the multiple nodes included in the original graph, and to obtain an initial granularity sphere with the target node as the core and connected to at least one of the nodes through hyperedges.

[0135] In one embodiment, the model training module 503 includes: The first model training unit is used to assign different weights to multiple granular balls based on the quality index of the granular ball and the importance of multiple nodes in the granular ball to the target transaction. The second model training unit is used to train the decision support model to be trained based on multiple granular spheres with assigned weights until a preset convergence condition is met, thereby obtaining the decision support model.

[0136] In one embodiment, the model training module 503 includes: The third model training unit is used to input multiple granular spheres into multiple feature extraction networks, and extract feature information of multiple granular spheres corresponding to the granularity level through the feature extraction networks to obtain multiple columns of feature information; wherein, different feature extraction networks correspond to different granularity levels; The fourth model training unit is used to perform cross-column reversible connections and feature fusion on multiple columns of feature information through the reversible connection layer of the decision support model to be trained, so as to obtain fused feature information. The fifth model training unit is used to output the evaluation results of multiple transaction indicators based on the fused feature information through the inference layer of the decision support model to be trained. The sixth model training unit is used to iteratively train the decision support model to be trained based on the difference information between the multiple evaluation results and the multiple sample evaluation results until the preset convergence condition is met, thereby obtaining the decision support model.

[0137] In one embodiment, the third model training unit is specifically used to: input multiple granular spheres into multiple columns of feature extraction networks; wherein each feature extraction network includes multiple attention convolutional blocks connected in sequence; through the multiple attention convolutional blocks in the feature extraction network, iterative feature extraction is performed on multiple granular spheres at the corresponding granularity level of the feature extraction network to update the feature data of each node multiple times until feature information including target feature data corresponding to multiple nodes is obtained.

[0138] In one embodiment, the fourth model training unit is specifically used to: perform cross-column reversible connection and feature fusion on multiple columns of feature information through the reversible connection layer of the decision support model to be trained, and retain the feature information corresponding to the cross-column starting layer to participate in the calculation during the forward propagation of information until the fused feature information is obtained.

[0139] This specification describes an embodiment that uses sample data to construct an original graph, obtaining multiple coarse-grained initial granular spheres. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels. This adaptive splitting mechanism assigns sub-granular spheres to nodes within each granular sphere based on distance, representing the consistency of node features and the density of the topological structure. This adapts to differences in association patterns between data from different domains, as well as to association patterns within the same domain that change over time or in different scenarios. Further, a decision support model is trained using these multiple granular spheres. During training, the decision support model includes a multi-column feature extraction network that focuses on extracting features from granular sphere sets at different granularity levels. Reversible connections are used to fuse cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results for multiple transaction indicators based on the fused feature information. The evaluation results of these multiple transaction indicators are used to assist in making decisions regarding target transactions, providing a basis for target transaction processing.

[0140] It should be noted that the decision support model training device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the decision support model training method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the decision support model training device and the decision support model training method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0141] Please see Figure 10 This diagram illustrates the structure of a transaction decision-making device provided in an exemplary embodiment of this specification. The transaction decision-making device can be implemented as all or part of a device through software, hardware, or a combination of both. The transaction decision-making device includes a data acquisition module 601, an evaluation result module 602, and a transaction decision-making module 603.

[0142] The data acquisition module 601 is used to acquire data to be processed for the target transaction decision. The evaluation result module 602 is used to obtain multiple granular spheres based on the original graph constructed based on the data to be processed, input the multiple granular spheres corresponding to the data to be processed into the decision support model, and obtain the evaluation results of multiple transaction indicators; wherein, the decision support model is trained by the training method of any one of the decision support models of claims 1-10; The transaction decision module 603 is used to make a decision on the target transaction based on the evaluation results of multiple transaction indicators, and obtain a decision result.

[0143] In this specification, multiple coarse-grained initial granular spheres are obtained from the original graph constructed based on the data to be processed. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels, fully preserving high-order interaction information between multiple nodes. The decision support model utilizes a multi-column feature extraction network to focus on the granular sphere sets at different granularity levels for feature extraction. Reversible connections are used to achieve the fusion of cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results of multiple transaction indicators based on the fused feature information. Based on these evaluation results, decisions are made for the target transaction, providing a basis for target transaction processing and improving decision accuracy.

[0144] It should be noted that the transaction decision-making device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the transaction decision-making method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the transaction decision-making device and the transaction decision-making method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0145] The example numbers in this specification are for descriptive purposes only and do not represent the superiority or inferiority of the examples.

[0146] This specification also provides a computer storage medium that can store multiple instructions adapted to be loaded and executed by a processor as described above. Figure 1 - Figure 7 The training method of the decision support model in the illustrated embodiment or Figure 8 The transaction decision-making method of the illustrated embodiment can be found in the following document for its specific execution process. Figure 1 - Figure 8 The specific details of the illustrated embodiments will not be elaborated here.

[0147] This specification also provides a computer program product that stores at least one instruction, which is loaded and executed by a processor as described above. Figure 1 - Figure 7 The training method of the decision support model in the illustrated embodiment or Figure 8 The transaction decision-making method of the illustrated embodiment can be found in the following document for its specific execution process. Figure 1 - Figure 8 The specific details of the illustrated embodiments will not be elaborated here.

[0148] Please see Figure 11 This document provides a schematic diagram of the structure of an electronic device as an embodiment of the present specification. Figure 11 As shown, the electronic device 700 may include: at least one processor 701, at least one network interface 704, a user interface 703, a memory 705, and at least one communication bus 702.

[0149] The communication bus 702 is used to enable communication between these components.

[0150] The user interface 703 may include a display screen and a camera. Optionally, the user interface 703 may also include a standard wired interface and a wireless interface.

[0151] The network interface 704 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0152] The processor 701 may include one or more processing cores. The processor 701 connects to various parts within the electronic device 700 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 705, and by calling data stored in the memory 705. Optionally, the processor 701 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 701 may integrate one or a combination of several of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 701 and may be implemented as a separate chip.

[0153] The memory 705 may include random access memory (RAM) or read-only memory. Optionally, the memory 705 may include a non-transitory computer-readable storage medium. The memory 705 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 705 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 705 may also be at least one storage device located remotely from the aforementioned processor 701. Figure 11 As shown, the memory 705, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a decision support model training or transaction decision application.

[0154] exist Figure 11In the illustrated electronic device 700, the user interface 703 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 701 can be used to call the training application of the decision support model stored in the memory 705 and specifically perform the following operations: Based on the original graph constructed from sample data, multiple initial granularity spheres are obtained; wherein, the initial granularity spheres are obtained by aggregating multiple nodes connected by hyperedges; Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively binary split to obtain multiple granular spheres with different granularity levels. The decision support model to be trained is trained based on multiple granular spheres until a preset convergence condition is met, thereby obtaining the decision support model; wherein, the decision support model to be trained includes a multi-column feature extraction network, and the training process includes extracting features from granular spheres at different granularity levels through the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results of multiple transaction indicators based on the fused feature information.

[0155] In one embodiment, processor 701 performs a recursive binary split on the initial granularity sphere based on the distances between multiple nodes in the initial granularity sphere to obtain multiple granularity spheres with different granularity levels. Specifically, the following steps are executed: Among the multiple nodes included in the initial granular sphere, two split nodes are identified; Calculate the distance between each node in the initial granular sphere and the two split nodes; Based on the multiple distances, the multiple nodes in the initial granular sphere are respectively divided into two sub-granular spheres centered on the splitting node; The binary splitting steps described above are recursively performed on the two sub-granular spheres until a preset stopping condition is met, resulting in multiple granular spheres with different granularity levels corresponding to the initial granular sphere.

[0156] In one embodiment, the processor 701 performs the calculation of the distances between each node in the initial granularity sphere and the two split nodes, including: The fusion distance between each node in the initial granularity sphere and the target split node is calculated based on multiple dimensions; wherein, the multiple dimensions include at least the topological distance dimension characterizing the structural relationship between nodes, and the feature distance dimension characterizing the attribute differences between nodes.

[0157] In one embodiment, processor 701 performs a recursive binary split on the initial granularity sphere based on the distances between multiple nodes in the initial granularity sphere to obtain multiple granularity spheres with different granularity levels. Specifically, the following steps are executed: Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively split into two binary spheres, and the splitting quality gain is calculated at each split. If the splitting quality gain is not positive, the splitting stops until multiple granular spheres corresponding to the initial granular sphere are obtained.

[0158] In one embodiment, the split mass gain is the difference between the sum of the mass indices of the two sub-spheres after splitting and the mass index of the sphere before splitting, whereby the mass index of the sphere characterizes the consistency of node features and the compactness of the topology of the sphere.

[0159] In one embodiment, processor 701 executes the process of obtaining multiple initial granular spheres based on the original graph constructed from sample data, specifically by: Obtain the original graph constructed based on the sample data; wherein the original graph includes nodes corresponding to multiple transaction entities; Among the multiple nodes included in the original graph, at least one target node is determined, and an initial granularity sphere is obtained with the target node as the core and connected to at least one of the nodes through a hyperedge.

[0160] In one embodiment, processor 701 executes the process of training the decision support model to be trained based on a plurality of granular spheres until a preset convergence condition is met, thereby obtaining the decision support model. Specifically, the following steps are performed: Based on the quality index of the granular sphere and the importance of multiple nodes in the granular sphere to the target transaction, different weights are assigned to multiple granular spheres. The decision support model to be trained is trained using multiple granular spheres with assigned weights until a preset convergence condition is met, thus obtaining the decision support model.

[0161] In one embodiment, processor 701 executes the process of training the decision support model to be trained based on a plurality of granular spheres until a preset convergence condition is met, thereby obtaining the decision support model. Specifically, the following steps are performed: Multiple granular spheres are respectively input into multiple feature extraction networks, and feature information of multiple granular spheres at corresponding granularity levels is extracted by the feature extraction networks to obtain multiple columns of feature information; wherein, different feature extraction networks correspond to different granularity levels; Through the reversible connection layer of the decision support model to be trained, the feature information of multiple columns is reversibly connected and fused across columns to obtain fused feature information; The inference layer of the decision support model to be trained outputs evaluation results of multiple transaction indicators based on the fused feature information. Based on the differences between the multiple evaluation results and the multiple sample evaluation results, the decision support model to be trained is iteratively trained until the preset convergence condition is met, and the decision support model is obtained.

[0162] In one embodiment, processor 701 executes the step of inputting multiple granular spheres into a multi-column feature extraction network, and extracting feature information of multiple granular spheres at corresponding granularity levels through the feature extraction network to obtain multi-column feature information. Specifically, the execution is as follows: Multiple granular spheres are respectively input into multiple columns of feature extraction networks; wherein each feature extraction network includes multiple attention convolutional blocks connected in sequence; The feature extraction network processes multiple attention convolutional blocks to iteratively extract features from multiple granular spheres at the corresponding granularity level, thereby updating the feature data of each node multiple times until feature information including target feature data corresponding to multiple nodes is obtained.

[0163] In one embodiment, processor 701 executes the reversible connection layer through the decision support model to be trained, performing cross-column reversible connections and feature fusion on multiple columns of feature information to obtain fused feature information, specifically: Through the reversible connection layer of the decision support model to be trained, the feature information of multiple columns is reversibly connected and fused across columns. During the forward propagation of information, the feature information corresponding to the cross-column starting layer is retained to participate in the calculation until the fused feature information is obtained.

[0164] In one embodiment, processor 701 can be used to invoke an application storing physical decisions in memory 705, and specifically perform the following operations: Acquire the data to be processed for decision-making in the target transaction; Multiple granular spheres are obtained from the original graph constructed based on the data to be processed. The multiple granular spheres corresponding to the data to be processed are input into the decision support model to obtain the evaluation results of multiple transaction indicators. The decision support model is trained using the above-mentioned decision support model training method. The decision for the target transaction is made based on the evaluation results of multiple transaction indicators, and the decision result is obtained.

[0165] This specification describes an embodiment that uses sample data to construct an original graph, obtaining multiple coarse-grained initial granular spheres. Each initial granular sphere is then split using a fine-grained adaptive splitting mechanism to obtain multiple granular spheres with different granularity levels. This adaptive splitting mechanism assigns sub-granular spheres to nodes within each granular sphere based on distance, representing the consistency of node features and the density of the topological structure. This adapts to differences in association patterns between data from different domains, as well as to association patterns within the same domain that change over time or in different scenarios. Further, a decision support model is trained using these multiple granular spheres. During training, the decision support model includes a multi-column feature extraction network that focuses on extracting features from granular sphere sets at different granularity levels. Reversible connections are used to fuse cross-granularity feature information, fully capturing hierarchical feature information from fine-grained local associations to coarse-grained global associations. This improves the expressive power of the fused feature information, enabling the decision support model to infer evaluation results for multiple transaction indicators based on the fused feature information. The evaluation results of these multiple transaction indicators are used to assist in making decisions regarding target transactions, providing a basis for target transaction processing.

[0166] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented. Each of the above methods can be executed by a computer program instructing related hardware. The program corresponding to the above methods can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium of the electronic device 700 can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0167] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0168] The above-disclosed embodiments are merely preferred embodiments of this specification and should not be construed as limiting the scope of this specification. Therefore, any equivalent variations made in accordance with the claims of this specification shall still fall within the scope of this specification.

Claims

1. A method for training a decision support model, wherein the decision support model outputs evaluation results of multiple transaction indicators to assist in decision-making for a target transaction, the method comprising: Based on the original graph constructed from sample data, multiple initial granularity spheres are obtained; wherein, the initial granularity spheres are obtained by aggregating multiple nodes connected by hyperedges; Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively binary split to obtain multiple granular spheres with different granularity levels. The decision support model to be trained is trained based on multiple granular spheres until a preset convergence condition is met, thereby obtaining the decision support model; wherein, the decision support model to be trained includes a multi-column feature extraction network, and the training process includes extracting features from granular spheres at different granularity levels through the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results of multiple transaction indicators based on the fused feature information.

2. The training method for the decision support model according to claim 1, wherein the recursive binary splitting of the initial granularity sphere based on the distance between multiple nodes in the initial granularity sphere to obtain multiple granularity spheres with different granularity levels includes: Among the multiple nodes included in the initial granular sphere, two split nodes are identified; Calculate the distance between each node in the initial granular sphere and the two split nodes; Based on the multiple distances, the multiple nodes in the initial granular sphere are respectively divided into two sub-granular spheres centered on the splitting node; The binary splitting steps described above are recursively executed on the two sub-granular spheres until a preset stopping condition is met, resulting in multiple granular spheres with different granularity levels corresponding to the initial granular sphere.

3. The training method for the decision support model according to claim 2, wherein calculating the distance between each node in the initial granularity sphere and the two split nodes comprises: The fusion distance between each node in the initial granularity sphere and the target split node is calculated based on multiple dimensions; wherein, the multiple dimensions include at least the topological distance dimension characterizing the structural relationship between nodes, and the feature distance dimension characterizing the attribute differences between nodes.

4. The training method for the decision support model according to claim 1, wherein the recursive binary splitting of the initial granularity sphere based on the distance between multiple nodes in the initial granularity sphere to obtain multiple granularity spheres with different granularity levels includes: Based on the distance between multiple nodes in the initial granular sphere, the initial granular sphere is recursively split into two binary spheres, and the splitting quality gain is calculated at each split. If the splitting quality gain is not positive, the splitting stops until multiple granular spheres corresponding to the initial granular sphere are obtained.

5. The training method for the decision support model according to claim 4, wherein the split quality gain is the difference between the sum of the quality indices of the two sub-granular spheres after splitting and the quality index of the granular sphere before splitting, and the quality index of the granular sphere characterizes the node feature consistency and topological density of the granularity.

6. The training method for the decision support model according to claim 1, wherein obtaining multiple initial granularity spheres based on the original graph constructed based on sample data includes: Obtain the original graph constructed based on the sample data; wherein the original graph includes nodes corresponding to multiple transaction entities; Among the multiple nodes included in the original graph, at least one target node is determined, and an initial granularity sphere is obtained with the target node as the core and connected to at least one of the nodes through a hyperedge.

7. The training method for the decision support model according to claim 1, wherein training the decision support model to be trained based on a plurality of said granular spheres until a preset convergence condition is met to obtain the decision support model, comprises: Based on the quality index of the granular sphere and the importance of multiple nodes in the granular sphere to the target transaction, different weights are assigned to multiple granular spheres. The decision support model to be trained is trained using multiple granular spheres with assigned weights until a preset convergence condition is met, thus obtaining the decision support model.

8. The training method for the decision support model according to claim 1, wherein training the decision support model to be trained based on a plurality of said granular spheres until a preset convergence condition is met to obtain the decision support model, comprises: Multiple granular spheres are respectively input into multiple feature extraction networks, and feature information of multiple granular spheres at corresponding granularity levels is extracted by the feature extraction networks to obtain multiple columns of feature information; wherein, different feature extraction networks correspond to different granularity levels; Through the reversible connection layer of the decision support model to be trained, the feature information of multiple columns is reversibly connected and fused across columns to obtain fused feature information; The inference layer of the decision support model to be trained outputs evaluation results of multiple transaction indicators based on the fused feature information. Based on the differences between the multiple evaluation results and the multiple sample evaluation results, the decision support model to be trained is iteratively trained until the preset convergence condition is met, and the decision support model is obtained.

9. The training method for the decision support model according to claim 8, wherein the step of inputting the plurality of granular spheres into the multi-column feature extraction network, and extracting feature information of the plurality of granular spheres at the corresponding granularity level through the feature extraction network to obtain multi-column feature information, includes: Multiple granular spheres are respectively input into multiple columns of feature extraction networks; wherein each feature extraction network includes multiple attention convolutional blocks connected in sequence; The feature extraction network processes multiple attention convolutional blocks to iteratively extract features from multiple granular spheres at the corresponding granularity level, thereby updating the feature data of each node multiple times until feature information including target feature data corresponding to multiple nodes is obtained.

10. The training method for the decision support model according to claim 8, wherein the step of performing cross-column reversible connections and feature fusion on multiple columns of feature information through the reversible connection layer of the decision support model to be trained to obtain fused feature information includes: Through the reversible connection layer of the decision support model to be trained, the feature information of multiple columns is reversibly connected and fused across columns. During the forward propagation of information, the feature information corresponding to the cross-column starting layer is retained to participate in the calculation until the fused feature information is obtained.

11. A transaction decision-making method, the method comprising: Acquire the data to be processed for decision-making in the target transaction; Multiple granular spheres are obtained from the original graph constructed based on the data to be processed. The multiple granular spheres corresponding to the data to be processed are input into the decision support model to obtain the evaluation results of multiple transaction indicators. The decision support model is trained by any of the decision support models described in claims 1-10. The decision for the target transaction is made based on the evaluation results of multiple transaction indicators, and the decision result is obtained.

12. A training apparatus for a decision support model, wherein the decision support model is used to output evaluation results of multiple transaction indicators to assist in decision-making for a target transaction, the apparatus comprising: The hypergraph processing module is used to obtain multiple initial granularity spheres based on the original graph constructed from sample data; wherein, the initial granularity spheres are obtained by aggregating multiple nodes connected by hyperedges; The recursive splitting module is used to recursively split the initial granularity sphere into multiple granularity spheres based on the distance between multiple nodes in the initial granularity sphere, thereby obtaining multiple granularity spheres with different granularity levels. The model training module is used to train the decision support model to be trained based on multiple granular spheres until a preset convergence condition is met to obtain the decision support model; wherein, the decision support model to be trained includes a multi-column feature extraction network, and the training process includes extracting features from granular spheres at different granularity levels through the multi-column feature extraction network to obtain multi-column feature information, fusing the multi-column feature information across columns to obtain fused feature information, and outputting evaluation results of multiple transaction indicators based on the fused feature information.

13. A transaction decision-making apparatus, the apparatus comprising: The data acquisition module is used to acquire data to be processed for decision-making in the target transaction. The evaluation result module is used to obtain multiple granular spheres based on the original graph constructed based on the data to be processed, input the multiple granular spheres corresponding to the data to be processed into the decision support model, and obtain the evaluation results of multiple transaction indicators; wherein, the decision support model is trained by the training method of any of the decision support models in claims 1-10; The transaction decision module is used to make decisions about the target transaction based on the evaluation results of multiple transaction indicators, and obtain the decision results.

14. A computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method steps of any one of claims 1 to 11.

15. A computer program product storing a plurality of instructions adapted for loading by a processor and executing the method steps of any one of claims 1 to 11.

16. An electronic device comprising: a processor and a memory; wherein the memory stores a computer program adapted to be loaded and executed by the processor and to perform the method steps of any one of claims 1 to 11.

Images