Order data processing method, device and equipment

Abstract

The invention discloses an order data processing method, device and equipment, and the method is applied to the technical field of logistics, and the method comprises the steps: obtaining a user, a commodity, an order and a shop corresponding to each logistics order in a plurality of business platforms; a quaternary heterogeneous graph network of user-commodity-order-shop is constructed; determining a feature encoder corresponding to each node according to the node type corresponding to each node; based on the feature encoder corresponding to each node, attribute features of each node are extracted from the heterogeneous graph network, and based on a relation perception attention mechanism, a weight corresponding to each node type is set; capturing hierarchical structure features in the heterogeneous graph network by adopting a hierarchical graph convolution algorithm to obtain topological structure features; constructing a federated learning framework; and determining a return processing strategy of the return order based on the federated learning framework. According to the invention, the abnormal return order can be accurately identified, and the processing efficiency of the logistics order is improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an order data processing method, apparatus, and device. Background Technology

[0002] Existing e-commerce after-sales automation systems primarily rely on rule engines and simple decision trees, exhibiting the following technical bottlenecks: 1) Static, fixed rules cannot adapt to dynamic business scenarios; 2) Lack of modeling capabilities for complex relationships between users, products, and orders; 3) Data silos across multiple platforms lead to poor model generalization; 4) Disjointed decision-making processes for blocking and refunds prevent global optimization; 5) High coupling in traditional transaction processing architectures makes it difficult to support large-scale concurrency. Existing technical solutions propose rule-based after-sales automation methods, but fail to address the dynamic optimization of strategies; they employ simple machine learning models for risk assessment, but do not consider complex topological relationships between entities. These limitations result in low automation rates, high false positive rates, and poor cost control in existing systems, making it difficult to meet the high-efficiency operational needs of e-commerce platforms. Summary of the Invention

[0003] This invention provides an order data processing method, apparatus, and equipment that can accurately identify abnormal return orders, improve the processing efficiency of logistics orders, and reduce operating costs.

[0004] On one hand, the present invention provides an order data processing method, the method comprising: The system obtains the user, product, order, and store corresponding to each logistics order from multiple business platforms; and constructs a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the strength of the relationship between the entities corresponding to the nodes. Based on the node type corresponding to each node, determine the feature encoder corresponding to each node; Based on the feature encoder corresponding to each node, the attribute features of each node are extracted from the heterogeneous graph network, and the weights corresponding to each node type are set based on the relationship-aware attention mechanism. A hierarchical graph convolution algorithm is used to capture the hierarchical structural features in the heterogeneous graph network, thereby obtaining the topological structural features; A federated learning framework is constructed based on the attribute features of the nodes, the weights corresponding to each node type, and the topological features. The federated learning framework is shared across the multiple business platforms, and a return processing strategy for return orders is determined based on the federated learning framework.

[0005] In some embodiments, determining the return processing strategy for return orders based on the federated learning framework includes: Obtain the return order corresponding to the return request, and determine the target user and target product corresponding to the return order; In the federated learning framework, find the target user node that matches the target user, and determine the user return frequency corresponding to the target user node; In the federated learning framework, find the target product node that matches the target product, and determine the product return frequency of the target product node in the region corresponding to the target user; If the user's return frequency is less than the first preset frequency and the product's return frequency is greater than or equal to the second preset frequency, the target product is determined to be an abnormal product, and the return order is determined to be a normal return order. Obtain the return processing strategy corresponding to the normal return order.

[0006] In some embodiments, the method further includes: If the user's return frequency is greater than or equal to a first preset frequency, and the product's return frequency is less than a second preset frequency, the target user is determined to be an abnormal user, and the return order is determined to be an abnormal return order. Execute the interception strategy corresponding to the abnormal return order.

[0007] In some embodiments, the method further includes: Obtain user satisfaction parameters, platform revenue, operating costs, and risk coefficients of the business platform; Based on the user satisfaction parameters, platform revenue, operating costs, and risk coefficient, a multidimensional reward function is constructed. Reinforcement learning is performed on the federated learning framework based on the multidimensional reward function to obtain a policy generation model; Obtain the return period corresponding to the return order, and determine the first preset frequency, the second preset frequency, and the processing strategy corresponding to the return order based on the strategy generation model and the period type corresponding to the return period.

[0008] The construction of a multi-dimensional reward function based on the user satisfaction parameter, platform revenue, operating costs, and risk coefficient includes: Based on the user satisfaction parameters and the platform revenue, a first reward function is constructed. Based on the aforementioned operating costs and risk coefficients, a second reward function is constructed; A multidimensional reward function is constructed based on the difference between the first reward function and the second reward function.

[0009] In some embodiments, the method further includes: The new orders of the abnormal users are filtered out from the logistics platform, and the target interception delivery station corresponding to the new orders is determined; Construct a three-level logistics map network of the logistics platform, consisting of city-distribution center-delivery station. The nodes in the logistics map network include real-time congestion index and weather influencing factors. A gated spatiotemporal convolution module is used to capture the spatial dependencies and temporal dynamics in the logistics graph network, and to predict the arrival time of the new order at the target interception delivery station. If the arrival time indicates that the new order cannot reach the target interception delivery station within a preset time period, a coordinated strategy of pre-refund and arrival interception is triggered; Based on the aforementioned collaborative strategy, pre-refund processing is performed on the new orders.

[0010] In some embodiments, the method further includes: If the user's return frequency is greater than or equal to the first preset frequency, obtain the return frequency of each item in the returned item set; If the return frequency of each item in the returned goods set is greater than or equal to the second preset frequency, the target user is determined to be a normal user; If the return frequency of each item in the returned goods set is less than the second preset frequency, the target user is determined to be an abnormal user.

[0011] In some embodiments, the method further includes: Obtain the meta-features of the newly added business platform, and construct a general representation space for the newly added business platform based on the meta-features; The policy generation model is trained using the MAML algorithm and the general representation space to obtain an updated model; A knowledge distillation pipeline is constructed to compress the knowledge of the updated model into a lightweight inference engine, resulting in a cross-platform adapted architecture.

[0012] On the other hand, an order data processing apparatus is provided, the apparatus comprising: The order information acquisition module is used to acquire the user, product, order, and store corresponding to each logistics order in multiple business platforms; and to construct a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the strength of the relationship between the entities corresponding to the nodes; The encoder determination module is used to determine the feature encoder corresponding to each node based on the node type corresponding to each node. The weight setting module is used to extract the attribute features of each node from the heterogeneous graph network based on the feature encoder corresponding to each node, and set the weights corresponding to each node type based on the relationship-aware attention mechanism. The topology feature determination module is used to capture the hierarchical structural features in the heterogeneous graph network using a hierarchical graph convolution algorithm to obtain the topology features; The framework construction module is used to construct a federated learning framework based on the attribute features of the nodes, the weights corresponding to each node type, and the topological structure features. The strategy determination module is used to share the federated learning framework across the multiple business platforms and determine the return processing strategy for return orders based on the federated learning framework.

[0013] On the other hand, an electronic device is provided, the device including a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the order data processing method as described above.

[0014] On the other hand, a computer storage medium is provided that stores at least one instruction or at least one program, which is loaded and executed by a processor to implement the order data processing method described above.

[0015] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the order data processing method as described above.

[0016] The order data processing method, apparatus, and equipment provided by this invention have the following technical effects: This invention acquires the user, product, order, and store corresponding to each logistics order in multiple business platforms; and constructs a user-product-order-store quaternary heterogeneous graph network using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the relationship strength between the entities corresponding to the nodes; a feature encoder corresponding to each node is determined according to the node type corresponding to each node; based on the feature encoder corresponding to each node, the attribute features of each node are extracted from the heterogeneous graph network, and the weights corresponding to each node type are set based on a relationship-aware attention mechanism; a hierarchical graph convolution algorithm is used to capture the hierarchical structural features in the heterogeneous graph network to obtain topological structure features; a federated learning framework is constructed based on the attribute features of the nodes, the weights corresponding to each node type, and the topological structure features; the federated learning framework is shared in the multiple business platforms, and the return processing strategy for return orders is determined based on the federated learning framework; thereby, abnormal return orders can be accurately identified, improving the processing efficiency of logistics orders and reducing operating costs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions and advantages 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of an order data processing system provided in the embodiments of this specification; Figure 2 This is a flowchart illustrating an order data processing method provided in an embodiment of this specification; Figure 3 This is a flowchart illustrating a method for determining a return processing strategy for a return order based on the federated learning framework provided in the embodiments of this specification; Figure 4 This is a flowchart illustrating a method for performing pre-refund processing on newly added orders based on a collaborative strategy, as provided in an embodiment of this specification. Figure 5 This is a flowchart illustrating an automated processing method for after-sales order data based on asynchronous segmented transactions, as provided in the embodiments of this specification. Figure 6 This is a schematic diagram of the structure of an order data processing device provided in the embodiments of this specification; Figure 7 This is a schematic diagram of the structure of a server provided in the embodiments of this specification. Detailed Implementation

[0019] 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 the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It is understood that in the specific embodiments of the present invention, data such as user information are involved. When the above embodiments of the present invention are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0022] Please see Figure 1 , Figure 1 This is a schematic diagram of an order data processing system provided in the embodiments of this specification, such as... Figure 1 As shown, the order data processing system may include at least server 01 and client 02.

[0023] Specifically, in the embodiments of this specification, server 01 may include a standalone server, a distributed server, or a server cluster composed of multiple servers. It may also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Server 01 may include a network communication unit, a processor, and a memory, etc. Specifically, server 01 can be used to construct a four-element heterogeneous graph network of user-product-order-store, extract the attribute features of each node from the heterogeneous graph network, and set the weights corresponding to each node type based on a relation-aware attention mechanism; use a hierarchical graph convolution algorithm to capture the hierarchical structural features in the heterogeneous graph network to obtain topological features; construct a federated learning framework based on the attribute features of the nodes, the weights corresponding to each node type, and the topological features; share the federated learning framework among the multiple business platforms, and determine the return processing strategy for returned orders based on the federated learning framework.

[0024] Specifically, in the embodiments of this specification, the client 02 may include physical devices such as smartphones, desktop computers, tablets, laptops, digital assistants, smart wearable devices, smart speakers, in-vehicle terminals, and smart TVs. It may also include software running on the physical device, such as web pages provided to users by service providers, or applications provided to users by those service providers. Specifically, the client 02 can be used to query quaternary heterogeneous graph networks and return processing strategies for returned orders.

[0025] The following describes an order data processing method according to the present invention. Figure 2 This is a flowchart illustrating an order data processing method provided in an embodiment of this specification. This specification provides the operational steps of the method described in the embodiment or flowchart, but based on conventional or non-inventive labor, more or fewer operational steps may be included. The order of steps listed in the embodiment is merely one possible execution order among many and does not represent the only possible execution order. In actual system or server product execution, the method can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment) as shown in the embodiment or drawings. Specifically, as follows... Figure 2 As shown, the method may include: S201: Obtain the user, product, order, and store corresponding to each logistics order in multiple business platforms; and construct a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the strength of the relationship between the entities corresponding to the nodes. In the embodiments of this specification, multiple business platforms may include platforms in different applications; for example, multiple business platforms may be multiple e-commerce platforms. After obtaining the user, product, order, and store corresponding to the logistics order, a four-element heterogeneous graph network of user-product-order-store is constructed, using the entities such as user, product, order, and store as nodes. Each node (user, product, etc.) has unique attribute characteristics, and edges (purchase, return, etc.) represent the strength of the relationship between entities. The return records of different users purchasing the product can be stored in the node corresponding to the product. The return records include the return application time, application area, return description reason (which may include text, images, etc.), and return result, etc. The node corresponding to the user can store the user's location, return application data, etc. The node corresponding to the order can include information such as the region corresponding to the order, etc. The node corresponding to the store can store the cumulative return records of all products in the store. The return records include the return application time, application area, return description reason (which may include text, images, etc.), and return result, etc., which can be used to evaluate the return rate of different stores for store ranking, recommendation, etc.

[0026] S203: Determine the feature encoder corresponding to each node based on the node type corresponding to each node.

[0027] In the embodiments of this specification, the node type can be determined based on the entity represented by the node. Since there are four types of entities: user, product, order, and store, this embodiment corresponds to four node types. A feature encoder can be set for each node type to extract the node attribute features of the corresponding node type.

[0028] S205: Based on the feature encoder corresponding to each node, extract the attribute features of each node from the heterogeneous graph network, and set the weights corresponding to each node type based on the relationship-aware attention mechanism.

[0029] In the embodiments of this specification, attribute features of each node can be extracted from the heterogeneous graph network using a feature encoder corresponding to each node, and weights corresponding to each node type can be set according to a relation-aware attention mechanism. For example, a heterogeneous graph transformation layer is designed, employing specific encoders to extract features for different node types; a relation-aware attention mechanism is introduced to assign dynamic weights to edges of different types.

[0030] S207: The hierarchical graph convolution algorithm is used to capture the hierarchical structural features in the heterogeneous graph network to obtain the topological structural features.

[0031] In the embodiments of this specification, hierarchical graph convolution algorithms are a class of neural network methods for learning graph representations by progressively coarsening the graph structure (constructing hierarchical representations). Their core idea is that they not only consider the local neighborhood information of nodes but also capture substructure patterns at different scales in the graph through hierarchical pooling mechanisms, thereby learning more expressive graph-level or node-level representations. Hierarchical graph convolution can be used to capture both local and global topological structures. By capturing hierarchical structures, these algorithms can effectively identify patterns at different scales.

[0032] Enhanced expressiveness and interpretability: Graph-level representations are no longer simply the sum of all node features, but contain rich information from local to global perspectives. The subgraphs represented by the "supernodes" formed after each pooling layer can be observed.

[0033] The main function of hierarchical graph convolution algorithms is to parse heterogeneous graphs, rather than to construct them from scratch. Instead, they leverage a pre-constructed heterogeneous graph and, through their unique hierarchical information aggregation and abstraction mechanisms, delve into the complex structural and semantic information within the graph.

[0034] The main function of hierarchical graph convolution algorithms is to parse heterogeneous graphs, rather than to construct them from scratch. Instead, they leverage a pre-constructed heterogeneous graph and, through their unique hierarchical information aggregation and abstraction mechanisms, delve into the complex structural and semantic information within the graph.

[0035] The powerful ability of hierarchical graph convolution algorithms to parse heterogeneous graphs is mainly reflected in the following aspects: Capturing multi-layered semantic structures: Heterogeneous graphs contain various types of nodes and edges, and different connection paths (i.e., meta-paths, such as "user-product-order-store") represent different semantic relationships. Simple graph convolution may only focus on direct neighbors, while hierarchical graph convolution can learn semantic information based on local neighbors, meta-paths, and even more complex ring structures through hierarchical aggregation, thereby forming a comprehensive, multi-layered representation of nodes and graphs.

[0036] Hierarchical pooling: This is a core operation of hierarchical graph convolution. It constructs a hierarchical graph structure by progressively clustering (or "pooling") fine-grained nodes into more representative supernodes. This process can reveal the inherent functional substructures in the graph (such as identifying functional groups in molecular graphs), which is crucial for graph-level tasks.

[0037] Integrating information from higher-order neighbors: To gain a more comprehensive understanding of a node, it is insufficient to only consider its first-order neighbors. Hierarchical graph convolution algorithms can enrich the representation of the target node by aggregating information from higher-order neighbor nodes. This is particularly crucial in heterogeneous graphs, as different types of higher-order paths may contain rich implicit semantics.

[0038] S209: Construct a federated learning framework based on the attribute features of the nodes, the weights corresponding to each node type, and the topological features.

[0039] In the embodiments of this specification, a federated learning framework can be constructed based on the attribute characteristics of nodes, the weights corresponding to each node type, and the topological structure characteristics. Data from each platform does not leave the domain and only shares encrypted gradients and graph structure patterns.

[0040] In heterogeneous graphs, "node type weight" is not an inherent attribute present in the graph data that can be directly read. Instead, it is a concept or numerical value designed and assigned based on specific tasks and objectives to measure the importance of different node types. In heterogeneous graphs, different types of nodes typically have distinctly different characteristics, meanings, and importance, thus requiring the setting of weights for different node types. The node type weight determines the "intensity" or "proportion" of information aggregation. Weights can be set directly based on domain knowledge. For example, this embodiment can set the weights corresponding to four types of nodes—user, product, order, and store—according to actual circumstances.

[0041] S2011: Share the federated learning framework across the multiple business platforms and determine the return processing strategy for return orders based on the federated learning framework.

[0042] In the embodiments described in this specification, the federated learning framework can be shared across multiple business platforms. Upon receiving a return order, a return processing strategy is determined based on the federated learning framework. This allows for the identification of abnormal users who frequently return items and abnormal products that are frequently returned, and different return processing strategies can be adopted for each. For example, return orders from abnormal users can be pushed to manual processing, while return orders from normal users can be processed normally.

[0043] In some embodiments, such as Figure 3 As shown, the method for determining the return processing strategy for returned orders based on the federated learning framework includes: S301: Obtain the return order corresponding to the return request, and determine the target user and target product corresponding to the return order; S303: Find the target user node that matches the target user in the federated learning framework, and determine the user return frequency corresponding to the target user node; S305: Find the target product node that matches the target product in the federated learning framework, and determine the product return frequency of the target product node in the region corresponding to the target user; S307: If the user return frequency is less than the first preset frequency and the product return frequency is greater than or equal to the second preset frequency, the target product is determined to be an abnormal product, and the return order is determined to be a normal return order; S309: Obtain the return processing strategy corresponding to the normal return order.

[0044] In the embodiments of this specification, after receiving a return order, it can be analyzed according to a federated learning framework. By using historical data of the target user and the target product, abnormal products with quality problems and abnormal users who maliciously return goods can be identified. A first preset frequency for intercepting returns from abnormal users and a second preset frequency for identifying abnormal products with quality problems can be set. If the user return frequency is less than the first preset frequency and the product return frequency is greater than or equal to the second preset frequency, the target product can be determined as an abnormal product, and the return order can be determined as a normal return order. The return processing strategy corresponding to the normal return process can be obtained to execute the return request.

[0045] For example, when user A frequently returns product X, and product X is returned by multiple users in the same area, the system can identify potential quality issues rather than simply determining that the user is maliciously returning the product. Experiments show that this model improves accuracy by 23.8% compared to traditional methods and effectively protects data privacy.

[0046] In some embodiments, the method further includes: If the user's return frequency is greater than or equal to a first preset frequency, and the product's return frequency is less than a second preset frequency, the target user is determined to be an abnormal user, and the return order is determined to be an abnormal return order. Execute the interception strategy corresponding to the abnormal return order.

[0047] In the embodiments of this specification, if the user's return frequency is greater than or equal to a first preset frequency and the product's return frequency is less than a second preset frequency, the target user can be identified as an abnormal user with high frequency of returns, and the return order can be identified as an abnormal return order; the interception strategy corresponding to the abnormal return order can be executed; and the user's subsequent orders can be intercepted in the logistics process to reduce losses.

[0048] In some embodiments, the method further includes: If the user's return frequency is greater than or equal to the first preset frequency, obtain the return frequency of each item in the returned item set; If the return frequency of each item in the returned goods set is greater than or equal to the second preset frequency, the target user is determined to be a normal user; If the return frequency of each item in the returned goods set is less than the second preset frequency, the target user is determined to be an abnormal user.

[0049] In the embodiments of this specification, if the user's return frequency is greater than or equal to a first preset frequency, the return frequency of each item in the returned goods set is obtained, and the target user is further judged as an abnormal user based on the return frequency, thereby improving the accuracy of abnormal user identification and reducing the false identification rate.

[0050] For example, if the system identifies that a user's frequent returns are due to the actual quality of the goods not matching the description, rather than malicious behavior by the user, the handling strategy will be adjusted to "priority refund + supplier quality warning" instead of simply refusing the return. This mechanism reduces the false positive rate by 67.3% while improving the accuracy of identifying high-risk cases.

[0051] In some embodiments, the method further includes: Obtain user satisfaction parameters, platform revenue, operating costs, and risk coefficients of the business platform; Based on the user satisfaction parameters, platform revenue, operating costs, and risk coefficient, a multidimensional reward function is constructed. Reinforcement learning is performed on the federated learning framework based on the multidimensional reward function to obtain a policy generation model; Obtain the return period corresponding to the return order, and determine the first preset frequency, the second preset frequency, and the processing strategy corresponding to the return order based on the strategy generation model and the period type corresponding to the return period.

[0052] In the embodiments of this specification, the strategy generation model can be trained using the SAC (Soft Actor-Critic) algorithm, a deep reinforcement learning decision optimizer, and a post-sales decision-making agent constructed based on the SAC algorithm. Long-term returns are continuously optimized through a multi-dimensional reward function and a course learning mechanism. The deep reinforcement learning dynamic decision framework overcomes the limitations of static rule matching, constructing a decision-making agent based on the SAC (Soft Actor-Critic) algorithm, possessing exploration-exploitation balancing capabilities. The weights corresponding to the user satisfaction parameter, platform revenue, operating costs, and risk coefficient can be set separately, and then weighted to construct a multi-dimensional reward function.

[0053] A learning mechanism can be introduced, gradually transitioning from simple scenarios to complex decision-making; a digital twin simulation environment can be built to generate synthetic data to accelerate training; a hierarchical reinforcement learning architecture can be adopted, with high-level policies determining "whether to automate processing" and low-level policies determining "the specific processing method." This framework automatically discovers the optimal strategy combination through continuous interaction with the environment, such as automatically adjusting the interception threshold during peak promotional periods, maximizing platform revenue while ensuring user experience. Real-world testing shows that the system improved strategy returns by 42.6% within 3 months and can adapt to business changes.

[0054] For example, constructing a multidimensional reward function based on the user satisfaction parameter, the platform revenue, the operating cost, and the risk coefficient may include: constructing a first reward function based on the user satisfaction parameter and the platform revenue: α•user satisfaction + β•platform revenue; constructing a second reward function based on the operating cost and the risk coefficient: γ•operating cost + δ•risk coefficient; and then constructing a multidimensional reward function based on the difference between the first reward function and the second reward function.

[0055] For example, a multidimensional reward function can be designed: R = α•user satisfaction + β•platform revenue - γ•operating costs - δ•risk coefficient, where α, β, γ, and δ are all weighting coefficients.

[0056] In some embodiments, such as Figure 4 As shown, the method further includes: S401: Filter out the new orders of the abnormal users from the logistics platform, and determine the target interception delivery station corresponding to the new orders; S403: Construct a three-level logistics map network of the logistics platform, consisting of city-distribution center-delivery station, wherein the nodes in the logistics map network include real-time congestion index and weather influencing factors; S405: A gated spatiotemporal convolution module is used to capture the spatial dependencies and temporal dynamics in the logistics graph network, and to predict the arrival time of the new order at the target interception delivery station; S407: If the arrival time indicates that the new order cannot reach the target interception delivery station within a preset time period, trigger the coordinated strategy of pre-refund and arrival interception; S409: Perform a pre-refund process on the new order based on the aforementioned collaborative strategy.

[0057] In the embodiments of this specification, spatiotemporal graph convolutional logistics prediction and collaborative optimization addresses the interception failure problem caused by inaccurate logistics status prediction. The technical solution proposes a fusion architecture of ST-GCN and multi-objective optimization: A three-level logistics graph network is constructed, consisting of a city-distribution center-delivery station network, with nodes including real-time congestion indices and weather influencing factors. A gated spatiotemporal convolutional module is employed to simultaneously capture spatial dependencies and temporal dynamics. Spatial dependency capture: Graph Convolutional Networks (GCN) or Graph Attention Networks (GAT) are typically used to model the spatial relationships between nodes. Temporal dynamics capture: One-dimensional convolutional or recurrent neural networks (such as LSTM and GRU) are used to model time series. Gating mechanisms (such as GLU and GRU) can help control the flow of information, prevent gradient vanishing, and capture long-term dependencies.

[0058] The Gated Spatiotemporal Convolution module is a powerful deep learning component specifically designed to simultaneously model spatial dependencies and temporal dynamics in data. It effectively processes complex spatiotemporal data by cleverly combining graph convolution (capturing spatial relationships) and gated temporal convolution (capturing temporal patterns) and introducing a gating mechanism to control the flow of information. The goal of temporal dynamic capture (gated temporal convolution) is to capture the patterns of change of each node over time (e.g., the trend of traffic flow at an intersection over time). The goal of the gating mechanism is to control the flow of information in the spatiotemporal dimension, learn long-term dependencies, and stabilize the training process.

[0059] The gated spatiotemporal convolution module organically combines the spatial modeling capabilities of graph convolution, the local feature extraction capabilities of temporal convolution, and the long-range dependency learning capabilities of the gating mechanism, providing an efficient and powerful spatiotemporal data modeling solution.

[0060] Integrating the NSGA-III evolutionary algorithm, the system generates a Pareto optimal solution set. For example, if the system predicts that a high-value package will arrive at an uninterceptable delivery station within 2 hours, it immediately triggers a collaborative strategy of "pre-refund + arrival interception" instead of waiting for the traditional process. This technology increases the interception success rate to 96.7%, a 31.5 percentage point improvement over traditional methods.

[0061] In the embodiments of this specification, a risk control mechanism enhanced by causal inference is employed. This technical solution overcomes the limitations of traditional correlation analysis by innovatively integrating causal inference and deep learning: It constructs a post-sales causal graph to distinguish between confounding factors and true causal chains; it uses a dual machine learning framework (Orthogonal Random Forest) to estimate treatment effects; it introduces counterfactual reasoning to evaluate the results "if different strategies were adopted"; and it designs an interpretability module to generate a risk factor contribution report. For example, if the system identifies that a user's frequent returns are due to the actual quality of the goods not matching the description, rather than malicious user behavior, it adjusts the handling strategy to "priority refund + supplier quality warning" instead of simply refusing the product. This mechanism reduces the false positive rate by 67.3% while improving the accuracy of high-risk case identification.

[0062] In some embodiments, the method further includes: Obtain the meta-features of the newly added business platform, and construct a general representation space for the newly added business platform based on the meta-features; The policy generation model is trained using the MAML algorithm and the general representation space to obtain an updated model; A knowledge distillation pipeline is constructed to compress the knowledge of the updated model into a lightweight inference engine, resulting in a cross-platform adapted architecture.

[0063] In the embodiments of this specification, the newly added business platform refers to a platform other than multiple business platforms. The meta-features of the newly added business platform can be platform-independent features. "Platform-independent meta-features" refer to quantitative indicators extracted from the data or model itself to describe its intrinsic attributes. These indicators are independent of any specific software framework, hardware environment, or computing platform. MAML (Model-Agnostic Meta-Learning) is a meta-learning algorithm designed to enable a model to quickly adapt to new tasks using very little data by training it on a large number of tasks, making it particularly suitable for few-shot learning scenarios. The core idea of ​​MAML is to learn a set of general initialization parameters (i.e., "meta-knowledge"), allowing the model to quickly fit a new task with only a small number of gradient updates. This is achieved through a two-layer optimization process: the inner loop performs rapid adaptation (fine-tuning) for each task, while the outer loop optimizes the initial parameters to minimize the average loss of all tasks after adaptation.

[0064] A meta-learning-driven cross-platform adaptation architecture utilizes the MAML algorithm to enable models to quickly adapt to new platform rules. The meta-learning-driven cross-platform adaptation architecture addresses the issue of rule differences across multiple platforms by proposing a MAML (Model-Agnostic Meta-Learning) enhanced cross-platform adaptation architecture: extracting platform-independent meta-features to construct a universal representation space; and using the MAML algorithm to train the model to quickly adapt to new platforms. Specific types of extracted meta-features: 1. Statistical characteristics: Features: mean, variance, skewness, kurtosis; category distribution entropy (for classification tasks); statistics on correlation coefficients between features. 2. Geometric / topological features: estimation of the inherent dimension of the data distribution, measure of the separability of categories / clusters, and manifold complexity index.

[0065] 3. Information Theory Features: Mutual information between features and the target variable; eigenvalues ​​of the mutual information matrix between features. 4. Model-oriented Landmarking: Performance metrics for quickly training simple models (such as linear models, 1-nearest neighbors, decision stumps) on subsets of data; these performance metrics serve as proxy features for the "easy" or "difficult" level of the dataset. Key takeaway: All of these features are platform-independent—they do not depend on a specific software framework, hardware, or data processing flow; they only describe the intrinsic properties of the data itself.

[0066] Constructing a universal representation space: Objective: To map high-dimensional meta-feature vectors to a structured low-dimensional space where similar tasks are close to each other. In this space, the similarity between tasks can be measured (e.g., Euclidean distance, cosine similarity). The MAML algorithm is used to find a set of initial model parameters that allow for good performance with only a few gradient update steps when facing new tasks.

[0067] The design domain features an adaptive gating mechanism that dynamically adjusts specific platform parameters; a knowledge distillation pipeline is built to compress large model knowledge into a lightweight inference engine. This architecture reduces the new platform access time from 2 weeks to 8 hours, achieves a policy migration accuracy of 92.4%, and significantly reduces operation and maintenance costs.

[0068] In some embodiments, such as Figure 5 As shown, Figure 5This is an automated processing method for after-sales order data based on asynchronous segmented transactions. The corresponding processing system includes a platform after-sales order terminal, a callback interface / query platform, a consumer Kalfka, a strategy Kalfka, and a relational database management system (MySQL). First, the platform after-sales order terminal sends raw after-sales order data to the callback interface / query platform. Then, the callback interface / query platform processes the raw after-sales order data to obtain tokens, which are then sent to the consumer Kalfka. Next, the consumer Kalfka queries the MySQL database for tokens. Specifically, it queries the business platform for order information and after-sales order information, checks if the after-sales order already exists, assembles the after-sales order data if it exists, queries the after-sales order logistics information, and updates the data. It also organizes the push of data to the strategy Kalfka and sends the data to the strategy Kalfka. Finally, the strategy Kalfka determines the strategy to be executed, queries the store token, and updates the MySQL database.

[0069] This embodiment establishes a federated graph neural network strategy engine to construct a heterogeneous graph network of users, products, and orders. It extracts topological features through a relationship-aware attention mechanism and achieves cross-platform knowledge sharing while protecting data privacy. A deep reinforcement learning decision optimizer, based on the SAC algorithm, constructs an after-sales decision-making agent, continuously optimizing long-term returns through a multi-dimensional reward function and a course learning mechanism. A spatiotemporal graph convolutional logistics prediction module uses the ST-GCN model to predict package status in real time and combines a multi-objective evolutionary algorithm to optimize interception decisions. A causal inference anomaly detection mechanism integrates Do-Calculus and variational autoencoders to accurately identify abnormal applications and analyze their root causes. This achieves optimal global resource allocation and reduces operating costs.

[0070] Traditional data processing flows are all confined to a single transaction, resulting in a massive amount of code and difficult maintenance. Asynchronous segmented processing breaks down a large, cumbersome transaction into a set of smaller, autonomous transactions, each focusing on a specific business function to complete the entire transaction. This simplifies project management and makes data processing more flexible and scalable.

[0071] As can be seen from the technical solutions provided in the embodiments of this specification above, the embodiments of this specification obtain the user, product, order, and store corresponding to each logistics order in multiple business platforms; and construct a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the relationship strength between the entities corresponding to the nodes; a feature encoder corresponding to each node is determined according to the node type corresponding to each node; based on the feature encoder corresponding to each node, the attribute features of each node are extracted from the heterogeneous graph network, and the weights corresponding to each node type are set based on the relationship-aware attention mechanism; a hierarchical graph convolution algorithm is used to capture the hierarchical structure features in the heterogeneous graph network to obtain the topological structure features; a federated learning framework is constructed based on the attribute features of the nodes, the weights corresponding to each node type, and the topological structure features; the federated learning framework is shared in the multiple business platforms, and the return processing strategy for return orders is determined based on the federated learning framework; thereby, abnormal return orders can be accurately identified, improving the processing efficiency of logistics orders and reducing operating costs.

[0072] This specification also provides an order data processing device, such as... Figure 6 As shown, the device includes: The order information acquisition module 610 is used to acquire the user, product, order and store corresponding to each logistics order in multiple business platforms; and to construct a four-element heterogeneous graph network of user-product-order-store using the user, product, order and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the strength of the relationship between the entities corresponding to the nodes. The encoder determination module 620 is used to determine the feature encoder corresponding to each node based on the node type corresponding to each node. The weight setting module 630 is used to extract the attribute features of each node from the heterogeneous graph network based on the feature encoder corresponding to each node, and set the weights corresponding to each node type based on the relationship-aware attention mechanism. The topology feature determination module 640 is used to capture the hierarchical structural features in the heterogeneous graph network using a hierarchical graph convolution algorithm to obtain the topology features; The framework construction module 650 is used to construct a federated learning framework based on the attribute features of the nodes, the weights corresponding to each node type, and the topological structure features. The strategy determination module 660 is used to share the federated learning framework among the multiple business platforms and determine the return processing strategy for return orders based on the federated learning framework.

[0073] In some embodiments, the strategy determination module includes: The order acquisition unit is used to acquire the return order corresponding to the return application and determine the target user and target product corresponding to the return order; The first determining unit is used to find the target user node that matches the target user in the federated learning framework, and to determine the user return frequency corresponding to the target user node; The second determining unit is used to find the target product node that matches the target product in the federated learning framework, and to determine the product return frequency of the target product node in the region corresponding to the target user. An abnormal product identification unit is used to determine that the target product is an abnormal product and to identify the return order as a normal return order if the user's return frequency is less than a first preset frequency and the product's return frequency is greater than or equal to a second preset frequency. The strategy acquisition unit is used to acquire the return processing strategy corresponding to the normal return order.

[0074] In some embodiments, the apparatus further includes: An abnormal user identification module is used to identify the target user as an abnormal user and identify the return order as an abnormal return order if the user's return frequency is greater than or equal to a first preset frequency and the product return frequency is less than a second preset frequency. The interception strategy execution module is used to execute the interception strategy corresponding to the abnormal return order.

[0075] In some embodiments, the apparatus further includes: The parameter acquisition module is used to acquire user satisfaction parameters, platform revenue, operating costs, and risk coefficients of the business platform. The function construction module is used to construct a multi-dimensional reward function based on the user satisfaction parameter, the platform revenue, the operating cost, and the risk coefficient. The model determination module is used to perform reinforcement learning on the federated learning framework based on the multidimensional reward function to obtain a policy generation model. The processing strategy determination module is used to obtain the return period corresponding to the return order, and determine the first preset frequency, the second preset frequency, and the processing strategy corresponding to the return order based on the strategy generation model and the period type corresponding to the return period.

[0076] In some embodiments, the apparatus further includes: The interception delivery station determination module is used to filter out the new orders of the abnormal users from the logistics platform and determine the target interception delivery station corresponding to the new orders; The graph construction module is used to construct a three-level logistics graph network of the logistics platform, consisting of city-distribution center-delivery station. The nodes in the logistics graph network include real-time congestion index and weather influencing factors. The prediction module is used to capture the spatial dependencies and temporal dynamics in the logistics graph network using a gated spatiotemporal convolution module, and to predict the arrival time of the new order at the target interception delivery station; The strategy triggering module is used to trigger a coordinated strategy of pre-refund and arrival interception if the arrival time indicates that the new order cannot reach the target interception delivery station within a preset time period. The execution module is used to perform pre-refund processing on the new order based on the collaborative strategy.

[0077] In some embodiments, the apparatus further includes: The return frequency acquisition module is used to acquire the return frequency of each product in the returned product set if the user's return frequency is greater than or equal to a first preset frequency. The first determining module is used to determine that the target user is a normal user if the return frequency of each item in the returned goods set is greater than or equal to the second preset frequency. The second determining module is used to determine the target user as an abnormal user if the return frequency of each item in the returned goods set is less than the second preset frequency.

[0078] In some embodiments, the apparatus further includes: A space construction module is used to obtain the meta-features of the new business platform and construct a general representation space of the new business platform based on the meta-features. The model training module is used to train the strategy generation model using the MAML algorithm and the general representation space to obtain an updated model; The architecture generation module is used to build a knowledge distillation pipeline, which compresses the knowledge of the updated model into a lightweight inference engine to obtain a cross-platform adapted architecture.

[0079] The apparatuses in the described apparatus embodiments are based on the same inventive concept as the apparatus embodiments.

[0080] This specification provides an electronic device including a processor and a memory. The memory stores at least one instruction or at least one program, which is loaded and executed by the processor to implement the order data processing method provided in the above method embodiments.

[0081] Embodiments of the present invention also provide a computer storage medium, which can be disposed in a terminal to store at least one instruction or at least one program related to implementing an order data processing method in the method embodiments. The at least one instruction or at least one program is loaded and executed by the processor to implement the order data processing method provided in the above method embodiments.

[0082] Embodiments of the present invention also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the order data processing method provided in the above-described method embodiments.

[0083] Optionally, in the embodiments of this specification, the storage medium may be located at at least one of the multiple network servers in a computer network. Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0084] The memory described in the embodiments of this specification can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for the functions, etc.; the data storage area may store data created according to the use of the device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory may also include a memory controller to provide the processor with access to the memory.

[0085] The order data processing method embodiments provided in this specification can be executed on mobile terminals, computer terminals, servers, or similar computing devices. Taking running on a server as an example, Figure 7 This is a hardware structure block diagram of a server for an order data processing method provided in an embodiment of this specification. Figure 7As shown, the server 700 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 710 (CPUs 710 may include, but are not limited to, microprocessors (MCUs) or programmable logic devices (FPGAs), a memory 730 for storing data, and one or more storage media 720 (e.g., one or more mass storage devices) for storing application programs 723 or data 722. The memory 730 and storage media 720 may be temporary or persistent storage. The program stored in the storage media 720 may include one or more modules, each module may include a series of instruction operations on the server. Furthermore, the CPU 710 may be configured to communicate with the storage media 720 and execute the series of instruction operations stored in the storage media 720 on the server 700. Server 700 may also include one or more power supplies 760, one or more wired or wireless network interfaces 750, one or more input / output interfaces 740, and / or one or more operating systems 721, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0086] The input / output interface 740 can be used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of server 700. In one example, the input / output interface 740 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the input / output interface 740 may be a radio frequency (RF) module used for wireless communication with the Internet.

[0087] Those skilled in the art will understand that Figure 7 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, server 700 may also include... Figure 7 The more or fewer components shown, or having the same Figure 7 The different configurations shown.

[0088] As can be seen from the embodiments of the order data processing method, apparatus, electronic device, or storage medium provided by the present invention, the present invention acquires the user, product, order, and store corresponding to each logistics order in multiple business platforms; and constructs a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the relationship strength between the entities corresponding to the nodes; a feature encoder corresponding to each node is determined according to the node type corresponding to each node; based on the feature encoder corresponding to each node, the attribute features of each node are extracted from the heterogeneous graph network, and the weights corresponding to each node type are set based on the relationship-aware attention mechanism; a hierarchical graph convolution algorithm is used to capture the hierarchical structural features in the heterogeneous graph network to obtain topological structural features; a federated learning framework is constructed based on the attribute features of the nodes, the weights corresponding to each node type, and the topological structural features; the federated learning framework is shared in the multiple business platforms, and the return processing strategy for return orders is determined based on the federated learning framework; thereby, abnormal return orders can be accurately identified, improving the processing efficiency of logistics orders and reducing operating costs.

[0089] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0090] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0091] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer storage medium, such as a read-only memory, a disk, or an optical disk.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing order data, characterized in that, The method includes: The system obtains the user, product, order, and store corresponding to each logistics order from multiple business platforms; and constructs a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the strength of the relationship between the entities corresponding to the nodes. Based on the node type corresponding to each node, determine the feature encoder corresponding to each node; Based on the feature encoder corresponding to each node, the attribute features of each node are extracted from the heterogeneous graph network, and the weights corresponding to each node type are set based on the relationship-aware attention mechanism. A hierarchical graph convolution algorithm is used to capture the hierarchical structural features in the heterogeneous graph network, thereby obtaining the topological structural features; A federated learning framework is constructed based on the attribute features of the nodes, the weights corresponding to each node type, and the topological features. The federated learning framework is shared across the multiple business platforms, and a return processing strategy for return orders is determined based on the federated learning framework.

2. The method according to claim 1, characterized in that, The method for determining the return processing strategy for returned orders based on the federated learning framework includes: Obtain the return order corresponding to the return request, and determine the target user and target product corresponding to the return order; In the federated learning framework, find the target user node that matches the target user, and determine the user return frequency corresponding to the target user node; In the federated learning framework, find the target product node that matches the target product, and determine the product return frequency of the target product node in the region corresponding to the target user; If the user's return frequency is less than the first preset frequency and the product's return frequency is greater than or equal to the second preset frequency, the target product is determined to be an abnormal product, and the return order is determined to be a normal return order. Obtain the return processing strategy corresponding to the normal return order.

3. The method according to claim 2, characterized in that, The method further includes: If the user's return frequency is greater than or equal to a first preset frequency, and the product's return frequency is less than a second preset frequency, the target user is determined to be an abnormal user, and the return order is determined to be an abnormal return order. Execute the interception strategy corresponding to the abnormal return order.

4. The method according to claim 3, characterized in that, The method further includes: Obtain user satisfaction parameters, platform revenue, operating costs, and risk coefficients of the business platform; Based on the user satisfaction parameters, platform revenue, operating costs, and risk coefficient, a multidimensional reward function is constructed. Reinforcement learning is performed on the federated learning framework based on the multidimensional reward function to obtain a policy generation model; Obtain the return period corresponding to the return order, and determine the first preset frequency, the second preset frequency, and the processing strategy corresponding to the return order based on the strategy generation model and the period type corresponding to the return period.

5. The method according to claim 3, characterized in that, The method further includes: The new orders of the abnormal users are filtered out from the logistics platform, and the target interception delivery station corresponding to the new orders is determined; Construct a three-level logistics map network of the logistics platform, consisting of city-distribution center-delivery station. The nodes in the logistics map network include real-time congestion index and weather influencing factors. A gated spatiotemporal convolution module is used to capture the spatial dependencies and temporal dynamics in the logistics graph network, and to predict the arrival time of the new order at the target interception delivery station. If the arrival time indicates that the new order cannot reach the target interception delivery station within a preset time period, a coordinated strategy of pre-refund and arrival interception is triggered; Based on the aforementioned collaborative strategy, pre-refund processing is performed on the new orders.

6. The method according to claim 3, characterized in that, The method further includes: If the user's return frequency is greater than or equal to the first preset frequency, obtain the return frequency of each item in the returned item set; If the return frequency of each item in the returned goods set is greater than or equal to the second preset frequency, the target user is determined to be a normal user; If the return frequency of each item in the returned goods set is less than the second preset frequency, the target user is determined to be an abnormal user.

7. The method according to claim 4, characterized in that, The method further includes: Obtain the meta-features of the newly added business platform, and construct a general representation space for the newly added business platform based on the meta-features; The policy generation model is trained using the MAML algorithm and the general representation space to obtain an updated model; A knowledge distillation pipeline is constructed to compress the knowledge of the updated model into a lightweight inference engine, resulting in a cross-platform adapted architecture.

8. An order data processing device, characterized in that, The device includes: The order information acquisition module is used to acquire the user, product, order, and store corresponding to each logistics order in multiple business platforms; and to construct a four-element heterogeneous graph network of user-product-order-store using the user, product, order, and store as nodes; each node in the heterogeneous graph network stores the attribute information of the entity represented by the node, and the edges in the heterogeneous graph network represent the strength of the relationship between the entities corresponding to the nodes; The encoder determination module is used to determine the feature encoder corresponding to each node based on the node type corresponding to each node. The weight setting module is used to extract the attribute features of each node from the heterogeneous graph network based on the feature encoder corresponding to each node, and set the weights corresponding to each node type based on the relationship-aware attention mechanism. The topology feature determination module is used to capture the hierarchical structural features in the heterogeneous graph network using a hierarchical graph convolution algorithm to obtain the topology features; The framework construction module is used to construct a federated learning framework based on the attribute features of the nodes, the weights corresponding to each node type, and the topological structure features. The strategy determination module is used to share the federated learning framework across the multiple business platforms and determine the return processing strategy for return orders based on the federated learning framework.

9. An electronic device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or at least one program being loaded and executed by the processor to implement the order data processing method as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores at least one instruction or at least one program, which is loaded and executed by a processor to implement the order data processing method as described in any one of claims 1-7.

Images