Cross-border e-commerce business intelligent tracking management method and system based on informatization

Abstract

The invention belongs to the technical field of e-commerce informatization, and particularly relates to an informatization-based cross-border e-commerce business intelligent tracking management method and system, and the method comprises the steps: responding to business abnormality detection, and scheduling participants with independent data sovereignty; building a cross-participant federal causal atlas through a federal causal discovery module in a collaborative manner; dynamically configuring the business digital twinning model based on the causal relationship in the atlas, and generating a causal constraint enhanced digital twinning instance; encoding map structure features in a deep reinforcement learning agent, and taking an enhanced digital twinborn instance as an environment; and the operation agent outputs an intervention action sequence. According to the method, the problems of cross-border e-commerce data island, decision lag and lack of intelligent support are solved, privacy protection collaborative analysis of cross-domain data and active prediction and interpretable intelligent decision based on a causal mechanism are realized, and the overall toughness and decision efficiency of a supply chain are improved.

Description

Technical Field

[0001] This invention relates to the field of e-commerce information technology, and in particular to an intelligent tracking and management method and system for cross-border e-commerce business based on information technology. Background Technology

[0002] With the rapid development of global e-commerce, cross-border e-commerce has a long business chain, many links, and complex participating entities, involving multiple nodes such as domestic consolidation, international transportation, import and export customs clearance, and overseas warehousing and distribution. Currently, common cross-border e-commerce business tracking and management systems mainly suffer from the following problems: Data silos and information barriers: Information such as orders, logistics, warehousing, payment, customs clearance, and customer service is scattered across different platforms and systems, with inconsistent data formats, making integration difficult and hindering the formation of an end-to-end transparent view.

[0003] Passive tracking and delayed early warning: Most systems only provide node status queries based on logistics tracking numbers, and early warnings rely on simple and fixed rules (such as timeout thresholds), making it impossible to conduct proactive and predictive risk identification based on multi-source data fusion (such as predicting that a batch of goods is likely to experience customs clearance delays at a specific port).

[0004] The decision support lacks intelligence: When an anomaly occurs, the system can usually only indicate the anomaly, but cannot provide root cause diagnosis and optimal handling suggestions based on historical data and intelligent analysis (such as when to activate backup logistics channels and when to proactively contact customers and provide compensation solutions). Decision-making relies heavily on human experience.

[0005] Compliance and privacy challenges: Cross-border business involves data compliance requirements in different jurisdictions, making it difficult to maximize the use of data value to enhance the overall supply chain resilience while ensuring legal compliance.

[0006] While existing patented technologies (such as CN118569759B, CN117495229A, and CN119398479A) have made improvements in certain aspects of logistics tracking, data correlation, or anomaly warning, they have failed to systematically solve the above problems and lack a closed-loop management framework that deeply integrates multi-source heterogeneous data and has the ability to make proactive predictions and intelligent decisions. Summary of the Invention

[0007] This invention proposes an information-based intelligent tracking and management method and system for cross-border e-commerce businesses, aiming to solve the technical problem of how to achieve integrated causal cognition and intelligent decision-making among cross-participants under the constraints of decentralized data sovereignty and privacy compliance.

[0008] In a first aspect, embodiments of the present invention provide an information-based intelligent tracking and management method for cross-border e-commerce business, comprising: In response to the detection of anomalies in the target business, at least two participants with independent data sovereignty are scheduled; By deploying federated causal discovery modules among the participating parties, a federated causal graph representing causal dependencies between business entities across participating parties is collaboratively constructed. Based on the causal relationships in the federated causal graph, the pre-built business digital twin model is dynamically parameterized to generate a digital twin instance with enhanced causal constraints. In the deep reinforcement learning agent, the structural features in the federated causal graph are encoded as part of the state vector, and the digital twin instance with the causal constraint enhancement is used as the training and inference environment. By running the deep reinforcement learning agent, a sequence of intervention actions is output for the target business anomaly.

[0009] The technical effect of the information-based intelligent tracking and management method for cross-border e-commerce business disclosed in this invention is as follows: By constructing a federated causal graph across participating parties and thereby enhancing digital twins and guiding reinforcement learning, this method achieves root cause localization, proactive prediction, and generation of interpretable intelligent decision sequences for complex cross-border e-commerce business anomalies under the premise of data privacy compliance.

[0010] Furthermore, the federal causal discovery module performs the following operations: Each participant generates a weighted local causal subgraph based on local data and a differentiable causal structure learning model. The structural gradient of the local causal subgraph is securely exchanged among the participating parties via a homomorphic encrypted channel. On the coordination server, the structural gradients are aggregated, and a global federated causal graph satisfying the directed acyclic graph property is obtained through constrained optimization.

[0011] Furthermore, the differentiable causal structure learning model is the NOTEARS algorithm or a variant thereof based on a neural network; the constrained optimization solution uses the augmented Lagrangian method, and the constraint condition is a directed acyclic graph condition.

[0012] Furthermore, the dynamic parameter configuration of the business digital twin model based on the federated causal graph specifically includes: Identify causal paths in the federal causal graph whose confidence exceeds a preset threshold; Each of the aforementioned causal paths is mapped to a combined logical rule or probability transition function; When initializing the business digital twin model, the logical rules or probability transition functions are injected as part of the core state machine.

[0013] Furthermore, the mapping process employs a combination of rule engines and neural networks: for well-structured deterministic causal relationships, a rule engine is used to generate deterministic logical rules; for complex causal relationships with uncertainty, a trained graph neural network is used to generate new probability transition functions.

[0014] Furthermore, the encoding of structural features in the federated causal graph into state vectors specifically involves: Perform a reverse breadth-first search on the federated causal graph, starting from the target anomaly node, to extract all causal adjacent nodes and edges within K hops; The graph attention network is used to aggregate features of the causal adjacent nodes and edges to generate a fixed-dimensional graph embedding vector. The graph embedding vector is concatenated with the current macroscopic state vector of the digital twin instance to form the complete state observation of the deep reinforcement learning agent.

[0015] Furthermore, the training of the deep reinforcement learning agent employs a causal-aware action masking mechanism: At each decision-making moment, a subset of actions that are allowed to be executed is calculated based on the current state of the federated causal graph and the digital twin instance; By using negative rewards to mask actions outside the aforementioned subset of actions, the deep reinforcement learning agent is guided to use actions only within a pre-defined causal and reasonable action space.

[0016] Furthermore, the method for calculating the subset of actions that are allowed to be executed is as follows: A temporary intervention graph is constructed to simulate the impact of executing candidate actions on key state variables in the digital twin instance; A lightweight causal effect prediction model is used to determine whether the impact violates the strong causal constraints or business security boundaries in the federated causal graph. Candidate actions that do not violate constraints and boundaries are included in the subset of actions that are allowed to be executed.

[0017] Furthermore, the method also includes a strategy evaluation step based on a counterfactual interpretation model: While outputting the sequence of intervention actions, a counterfactual explanation model coupled with the digital twin instance enhanced by the causal constraints is initiated; The counterfactual explanation model generates a simplified comparative intervention scheme that differs from the recommended sequence by minimizing the intervention cost; By simulating the execution of the comparative intervention plan and the recommended sequence, the expected differences between the two in key business indicators are quantified, and these differences are included in the output results as a reference for decision credibility.

[0018] Secondly, embodiments of the present invention provide an information-based intelligent tracking and management system for cross-border e-commerce business to implement the method, the system comprising: A federated causal discovery module, configured with privacy computing middleware, is used to collaboratively construct the federated causal graph; A causal rule injection engine, connected to the federated causal discovery module, is used for dynamic configuration of the business digital twin model; A deep reinforcement learning subsystem integrating a state encoder and an action mask module is connected to the outputs of the federated causal discovery module and the causal rule injection engine, respectively, to generate the intervention action sequence; The strategy interpretation and evaluation module, connected to the deep reinforcement learning subsystem, is used to execute the strategy steps corresponding to the intervention action sequence.

[0019] The technical effect of the intelligent tracking and management system for cross-border e-commerce business disclosed in this invention is that the system, through the synergy of four modules—federated causal discovery, causal rule injection, causal guided decision-making, and strategy interpretation—engineers the method and realizes automated closed-loop operation from cross-domain data perception and causal cognitive modeling to adaptive intelligent decision-making, significantly improving the efficiency, accuracy, and reliability of business anomaly handling. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the information-based intelligent tracking and management method for cross-border e-commerce business provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the federal causal graph construction process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the causal enhancement digital twin construction and reinforcement learning decision-making process provided in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0022] The core technical problem of this invention is how to break down information barriers and achieve collaborative analysis and fusion of cross-domain data under the constraints of independent data sovereignty and privacy compliance among participating parties, and how to use the resulting deep understanding (causal mechanism) to drive an interpretable and adaptive intelligent decision-making closed loop to proactively and accurately respond to complex cross-border e-commerce business anomalies. Therefore, embodiments of this invention provide an information-based intelligent tracking and management method for cross-border e-commerce business, referencing... Figures 1 to 3 As shown, the specific steps include: S1, in response to the detection of anomalies in the target business, schedule at least two data-sovereign independent participants; this step is the triggering and initialization phase of the method, and its purpose is to organize the necessary computing and data resources for subsequent cross-domain collaborative analysis under the premise of privacy compliance.

[0023] (1) Anomaly Detection: The system continuously monitors key performance indicators (KPIs) across the entire cross-border e-commerce chain, such as "customs clearance time", "warehouse turnover rate", and "order fulfillment rate". When the real-time or predicted value of any KPI deviates from its normal threshold range calculated based on historical data (such as time series models), a "target business anomaly" event is triggered. This event includes the anomaly type (such as "customs clearance delay"), the associated entity ID (such as order number, logistics tracking number), and the initial context.

[0024] (2) Participant Scheduling: The system maintains a "Participant Registry" that records the data entity types (such as "orders", "logistics tracks", "payment status", and "inventory") and privacy calculation interface addresses that each participant (such as e-commerce platforms, logistics providers A, payment institutions B, and overseas warehouses C) can provide. When an abnormal event is triggered, the system dynamically queries the registry based on the abnormality type and associated entities to determine the minimum and most relevant set of participants required to resolve the abnormality. For example, for "customs clearance delays", the system automatically schedules the e-commerce platform that has the order information, the logistics provider responsible for the transportation segment, and the payment institution that may be involved in tax information.

[0025] (3) Scheduling instructions: Send instructions to the scheduled participants, including session ID, summary of abnormal events, target variables required for this collaborative analysis (such as "whether customs clearance is delayed") and the range of features that can be exchanged. Each participant node prepares local data and computing resources accordingly.

[0026] Through a dynamic, on-demand scheduling mechanism, subsequent privacy-preserving computations are ensured to be performed only among the necessary participants and for specific analytical objectives, minimizing the risk of unnecessary data exposure and complying with the "data minimization" principle of regulations such as GDPR.

[0027] S2, through the federated causal discovery module deployed among the participating parties, collaboratively constructs a federated causal graph representing the causal dependencies between business entities across participating parties; aiming to discover the true causal structure between business entities by combining data from multiple parties without aggregating the original data.

[0028] Suppose there are M participants. Let the joint set of observed variables (i.e., the global set of business variables) for all participants be X = {X1, X2, ..., X...}. p}, corresponding to p entities or indicators, whose underlying causal structure is represented by an adjacency matrix A∈{0,1} p×p Indicates (A) ij =1 indicates X i →X j ), and corresponds to a joint probability distribution P(X).

[0029] Local causal learning: The m-th participant owns a subset of variables X (m) X. This approach employs a differentiable causal structure learning model (such as a variant of NOTEARS). This model defines a differentiable adjacency matrix Θ. (m) (Initial values ​​do not have to be 0 / 1), and the causal structure of the local data is learned by minimizing a loss function: ; The first term represents the reconstruction error, the second term is L1 regularization which promotes sparsity, and the third and fourth terms (from the augmented Lagrangian method) enforce the directed acyclic graph (DAG) constraint. The participants calculate the local loss with respect to Θ using gradient descent. (m) gradient .

[0030] Privacy-preserving gradient exchange and aggregation: Participants do not exchange original gradients. Or local data X (m) Instead, homomorphic encryption is used. Each party uses the public key pk distributed by the coordinating server to encrypt its local gradient information, obtaining... And send it to the coordination server.

[0031] Global graph solution: The coordinating server utilizes the additive homomorphism of homomorphic encryption to calculate the aggregate gradient of the encrypted state. The server sends this encrypted aggregate gradient back to each participant. Each participant decrypts it using their private key to obtain the global gradient. This is used to update a global, consensus-based causal structure matrix Θ. Through multiple iterations, a convergent Θ satisfying the DAG constraint is finally obtained. After binarizing it (e.g., by passing a threshold), the federated causal graph G=(v,e) is obtained, where node v corresponds to variable X, and edge e is determined by A (the binarized version of Θ).

[0032] Algorithms such as NOTEARS provide differentiable DAG constraints, enabling gradients to be securely propagated among parties and achieving joint optimization; while homomorphic encryption ensures information security during gradient aggregation. This resolves the contradiction between "data not leaving the domain" and "causal discovery requiring global data."

[0033] S3. Based on the causal relationships in the federated causal graph, the pre-built business digital twin model is dynamically parameterized to generate a digital twin instance with enhanced causal constraints. This step injects the "causal cognition" obtained in the previous step into the business simulation environment, enabling the simulation model to reflect the real business mechanism.

[0034] Pre-built business digital twin model: This is a benchmark model of cross-border e-commerce business processes built based on technologies such as discrete event simulation or intelligent agent modeling. Its formal representation is a tuple M. base =(S,A,T,R), where S is the set of states, A is the set of actions, T:S×A→Δ(S) is the state transition probability function (Δ represents the probability distribution), and R:S×A→R is the reward function.

[0035] Causal Relationship Identification and Mapping: From the federated causal graph G, extract the set P of all causal paths that terminate at the target anomalous variable (e.g., "customs clearance time_abnormal") and whose edge weights (causal strength) exceed the threshold t. For each path P: X i1 →X i2 →...→X 异常 ∈P, which is mapped to a rule or function in the digital twin model.

[0036] Dynamic parameter configuration: (1) Deterministic rule injection: For the parts of the path with clear causal relationships and high certainty (such as "product category = 'electronic products' → 3C certificate required"), use a rule engine (such as Drools) to express them as deterministic logic rules of IF<condition>THEN<action> and directly embed them into the state transition logic of the model.

[0037] (2) Probabilistic Function Learning: For complex causal relationships with high uncertainty (such as "port congestion level → customs clearance processing time"), a graph neural network (GNN) is constructed using the variables involved in the path as input. This GNN is trained using historical data to predict the conditional distribution parameters of the outcome variables. For example, learning a function... , where (μ,σ) are the normal distribution parameters of the assumed customs clearance time. During simulation, this GNN is used to dynamically generate the probability distribution parameters of the corresponding steps in the T function.

[0038] The above-mentioned injection rules and the trained GNN function are then compared with the baseline model M. base By combining these elements, a digital twin instance M is generated that is causally constrained and deeply integrated with causal knowledge, specifically tailored to the current exceptional scenario. causal .

[0039] Unlike traditional simulation models that use fixed or statistical parameters, this invention employs two differentiated approaches—rule injection and neural function approximation—based on the strength of the determinism of causal relationships. This approach ensures the accuracy of key business logic while retaining the flexibility to handle complex and uncertain relationships. This hybrid modeling method of "white-box rules + black-box learning" is key to achieving high-fidelity, interpretable business simulation.

[0040] S4, In the deep reinforcement learning agent, the structural features in the federated causal graph are encoded as part of the state vector, and the digital twin instance with causal constraint enhancement is used as the training and inference environment; this step uses causal knowledge to shape the "perception" and "learning environment" of the decision agent, guiding it to make efficient and reasonable decision learning.

[0041] State encoding: State observations s of a deep reinforcement learning agent at time t t It is composed of two parts: .

[0042] From digital twin instance M causal The macroscopic state (such as queue length at each stage, resource utilization rate, etc.). It is the structural feature encoding of the federated causal graph. Specifically, with the target anomaly node as the root, a reverse breadth-first search (BFS) is performed on G to obtain its K-hop subgraph G. t (K) Then, a Graph Attention Network (GAT) is used to encode the subgraph. After L layers of propagation, graph pooling is used to obtain the global representation vector of the graph: .

[0043] Environment settings: M causalThis serves as the training and inference environment for the intelligent agent. The interaction between the agent and the environment follows the standard reinforcement learning paradigm: in s t Choose action a t ∈A, environment transition st+1 T causal (s t ,a t And return the reward r t =R causal (s t ,a t ).

[0044] Causal-aware action masking: At each step, not all actions a∈A are available. The system dynamically computes a subset A of allowed actions. t allowed A. The specific method is: for candidate action a, in M... causal The short-term impact is quickly simulated to form a "temporary intervention map". A pre-trained, lightweight causal effect prediction model (e.g., a linear or shallow network) is used to determine whether the impact violates strong causal constraints in G (e.g., "cannot cause the state of upstream critical nodes to deteriorate beyond the safety boundary"). Only actions that do not violate the constraints are added to A. t allowed During training, for cases where 'a' does not belong to A... t allowed The action can be given a very large negative reward (mask), or its probability can be set to zero directly at the output layer of the policy network.

[0045] The lightweight causal effect prediction model is a regression model based on a multilayer perceptron. Its training data is generated through the following method: in the causal constraint-enhanced digital twin instance M... causal In this model, a large number of states are randomly sampled, and candidate actions are randomly applied to each state. The changes in key state variables before and after the action are recorded, forming (state, action, effect) sample pairs. The model uses the encodings of states and actions as input to predict the effect vector. During runtime, for a candidate action 'a', the model quickly predicts its effect and determines whether the effect vector causes the precondition of any strong causal edge in G (such as 'inventory depletion → order cancellation') to be satisfied in an undesirable way. If so, the action is excluded.

[0046] 1) Causal Augmentation of the State Space: Through GAT encoding graphs, the agent can "perceive" the causal network structure behind the current state, which is not available in traditional RL state representations. 2) Causal Constraints in the Decision Space: The action masking mechanism transforms causal and security constraints from "soft rewards" to "hard constraints," fundamentally preventing the agent from learning strategies that violate business mechanisms or security, greatly improving learning efficiency, security, and interpretability.

[0047] S5, by running the deep reinforcement learning agent, output a sequence of intervention actions for the target business anomaly.

[0048] Policy execution and sequence generation: After M causal A fully trained agent, whose policy network It can provide the optimal action probability distribution in any state s. For the current initial abnormal state s0, in M... causal In the inference: the intelligent agent based on The agent selects action a0, the environment state transitions to s1, and the agent selects a1 again, repeating this cycle until a termination state is reached (e.g., an anomaly is resolved or the maximum number of steps is exceeded). The final output is an action-state sequence τ={(s0,a0),(s1,a1),...,(sT,aT)}, where a0,a1,...,aT are the intervention action sequences.

[0049] Policy evaluation based on counterfactual explanation model: To enhance decision credibility, in the output recommendation sequence τ rec Simultaneously, a counterfactual explanation model is initiated. This model uses M... causal Based on G, a value related to τ is found through search or optimization methods. rec Different but with the lowest intervention costs (e.g., cost, operational complexity) comparative sequences τ counter For example, by solving the following optimization problem: ; Where Cost is the cost of intervention, and Perf is the performance indicator (such as the amount of delay reduction). Subsequently, in M... causal Parallel simulation execution Quantitatively calculate the differences between the two in core business metrics. The output not only contains τ rec Also included is τ counter The differences between them form the basis for decision-making.

[0050] The counterfactual explanation model uses a Monte Carlo tree search-based method to generate contrast schemes. It takes the current state as the root node and the first action different from the recommended sequence as the branching point, in M... causalA finite-step simulation is performed. The cost function Cost(τ) is defined as the weighted sum of the economic cost and operational complexity of all actions in the sequence. The algorithm searches for a minimum performance threshold Perf. min Under the constraint, the action sequence with the minimum cost is τ. counter .

[0051] The core inventive concept of the complete technical solution of this application lies in "using the cross-domain causal cognition (federated causal graph) obtained by federated learning as the basis for constructing the digital twin model and the decision guidance for the reinforcement learning agent", forming a closed loop of deep coupling of "cognition-simulation-decision".

[0052] On the one hand, by resolving the contradiction between privacy protection and causal structure learning at the algorithmic level (such as the synergy between gradient swapping and DAG constraints), effective causal discovery is achieved within the federated learning framework.

[0053] On the other hand, the synergistic effect arises: causal graphs are not only used for "explanation," but are also actively injected into the simulation environment and decision-making process, upgrading digital twin simulation from "correlation-based simulation" to "causal mechanism-based deduction," and transforming reinforcement learning exploration from "random trial and error" to "causally guided directed search." This deep coupling significantly improves simulation fidelity, decision-making efficiency, and interpretability simultaneously.

[0054] It solves long-standing technical challenges: The solution directly and specifically addresses two interconnected industry challenges in the cross-border e-commerce sector: "achieving cross-enterprise joint root cause analysis under compliance" and "providing explainable and adaptive decision-making for complex supply chain anomalies."

[0055] Based on the same inventive concept, embodiments of the present invention also provide an information-based intelligent tracking and management system for cross-border e-commerce business to implement the method described above, the system comprising: A federated causal discovery module, configured with privacy computing middleware, is used to collaboratively construct the federated causal graph; A causal rule injection engine, connected to the federated causal discovery module, is used for dynamic configuration of the business digital twin model; A deep reinforcement learning subsystem integrating a state encoder and an action mask module is connected to the outputs of the federated causal discovery module and the causal rule injection engine, respectively, to generate the intervention action sequence; The strategy interpretation and evaluation module, connected to the deep reinforcement learning subsystem, is used to execute the strategy steps corresponding to the intervention action sequence.

[0056] To verify the performance of the present invention, all comparative experiments were conducted under uniform and reproducible parameter settings. The key algorithm modules and their hyperparameter settings involved in this experiment are as follows: 1. Parameters of the Federal Causal Discovery Module.

[0057] Causal structure learning model: The NOTEARS-variant model based on a two-layer fully connected neural network is adopted. The input layer dimension of the local model is dynamically adjusted according to the number of features of each participant, and the hidden layer dimension is 64, using the ReLU activation function.

[0058] Optimization parameters: Learning rate is 1e-3; L1 regularization coefficient (λ) is 0.1; the penalty coefficient (ρ) in the augmented Lagrange method is initially 1.0, and multiplied by a growth factor of 1.1 in each iteration; the convergence threshold is to stop when the Frobenius norm change of the adjacency matrix Θ is less than 1e-6 or the number of iterations reaches 500.

[0059] Homomorphic encryption: The Paillier homomorphic encryption algorithm is used, and the key length is 2048 bits.

[0060] Federated learning: The FedAvg aggregation rounds are 50, and the local training rounds are 5.

[0061] 2. Digital twin construction parameters with enhanced causal constraints.

[0062] Causal path selection threshold (τ): The confidence threshold is set to 0.85. A causal path is only used to enhance the digital twin if the average weight of all edges in the causal path exceeds this threshold.

[0063] Rule engine: Drools 7.x, rule matching mode is "Stateful".

[0064] Graph Neural Network (GNN) parameters: The GNN used to learn the probability transition function is a 3-layer GraphSAGE model with 128 hidden dimensions per layer, aggregated in "mean" mode, using the Adam optimizer, with a learning rate of 1e-3, and trained until the loss on the validation set converges.

[0065] 3. Parameters of deep reinforcement learning agents.

[0066] State coding: The number of hops (K) in reverse breadth-first search (BFS) is 3.

[0067] Graph Attention Network (GAT): 2 layers, each with 4 attention heads and 64 hidden dimensions. The output layer is followed by a global average pooling layer to generate a 128-dimensional graph embedding vector.

[0068] Both the policy and value networks employ a multilayer perceptron (MLP) with two hidden layers (dimensions of 256 and 128, respectively) and use the ReLU activation function.

[0069] Reinforcement learning algorithm: Proximity Policy Optimization (PPO) algorithm is used.

[0070] The learning rate is 3e-4, the discount factor (γ) is 0.99, the generalized advantage estimation (GAE) parameter (λ) is 0.95, the number of update steps per round is 2048, the minimum batch size is 64, the policy clipping range (ε) is 0.2, the value function loss coefficient is 0.5, and the entropy reward coefficient is 0.01.

[0071] Causal Action Mask: A lightweight causal effect prediction model: a 3-layer MLP (64-32-16) pre-trained in a digital twin environment with 10,000 randomly sampled samples. The mask penalty reward is 100.

[0072] 4. Counterfactual explanation model parameters.

[0073] Search algorithm: Monte Carlo Tree Search (MCTS) is used.

[0074] Number of simulations: 200.

[0075] Search depth: Maximum 5 steps.

[0076] Explore constant (C) puct ): 1.0.

[0077] Cost function weights: economic cost weight 0.7, operational complexity weight 0.3.

[0078] Minimum performance threshold (Perf) min ): Set to 70% of the expected performance of the recommended strategy.

[0079] 5. Experimental environment and baseline comparison settings.

[0080] Hardware environment: All experiments were run on a server configured with an Intel Xeon Gold 6248R CPU, an NVIDIA Tesla V100 GPU, and 256GB of memory.

[0081] Software environment: Python 3.8, PyTorch 1.12, Federated AI Technology Enabler (FATE) 1.10, Ray RLlib 2.2.

[0082] Traditional rule system baseline: 15 fixed rules are set based on the experience of business experts (e.g., if the customs clearance time is >72 hours, an early warning will be triggered).

[0083] Basic machine learning (ML) baseline: XGBoost classification model is used for anomaly warning, random forest is used for root cause classification, and model parameters are tuned to the optimal level through grid search.

[0084] In a specific implementation scenario, the dataset consists of 50,000 abnormal order records from a cross-border e-commerce platform between 2022 and 2023, involving 8 logistics providers and 12 ports. A comparative analysis of this solution, a traditional rule-based system (baseline), and a typical machine learning (ML) early warning system is presented, with the structure shown in the sub-table: As shown in the table above, under the premise of protecting privacy, the causal graph constructed by this invention covers cross-domain relationships that traditional methods cannot obtain, significantly improving the accuracy of root cause analysis. The action masking mechanism reduces invalid exploration in reinforcement learning by 76%, accelerating convergence; causal rewards make the agent focus more on fundamental solutions. The fully automated process from anomaly detection to decision output reduces human intervention from necessary to optional, greatly improving processing efficiency.

[0085] In addition, ablation experiments were conducted to verify the contributions of each module: Complete system: All indicators are as above.

[0086] Without federal causal discovery (using only unilateral data): accuracy dropped to 79.3%, and decision quality decreased to 7.8.

[0087] Uncausal constraint injection (ordinary digital twin): decision quality drops to 8.1, training time increases by 2.4 times.

[0088] Without action masking: Decision quality drops to 8.4, and convergence speed is 2.1 times slower.

[0089] No counterfactual explanation: User satisfaction dropped to 83.2%.

[0090] Experiments show that the modules of this invention work together to significantly improve the overall performance of the system.

[0091] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. An information-based intelligent tracking and management method for cross-border e-commerce business, characterized in that: include: In response to the detection of anomalies in the target business, at least two participants with independent data sovereignty are scheduled; By deploying federated causal discovery modules among the participating parties, a federated causal graph representing causal dependencies between business entities across participating parties is collaboratively constructed. Based on the causal relationships in the federated causal graph, the pre-built business digital twin model is dynamically parameterized to generate a digital twin instance with enhanced causal constraints. In the deep reinforcement learning agent, the structural features in the federated causal graph are encoded as part of the state vector, and the digital twin instance with the causal constraint enhancement is used as the training and inference environment. By running the deep reinforcement learning agent, a sequence of intervention actions is output for the target business anomaly.

2. The method according to claim 1, characterized in that, The federal causality discovery module performs the following operations: Each participant generates a weighted local causal subgraph based on local data and a differentiable causal structure learning model. The structural gradient of the local causal subgraph is securely exchanged among the participating parties via a homomorphic encrypted channel. On the coordination server, the structural gradients are aggregated, and a global federated causal graph satisfying the directed acyclic graph property is obtained through constrained optimization.

3. The method according to claim 2, characterized in that, The differentiable causal structure learning model is the NOTEARS algorithm or a variant thereof based on a neural network; the constrained optimization solution uses the augmented Lagrangian method, and the constraint condition is the directed acyclic graph condition.

4. The method according to claim 1, characterized in that, The dynamic parameter configuration of the business digital twin model based on the federated causal graph is as follows: Identify causal paths in the federal causal graph whose confidence exceeds a preset threshold; Each of the aforementioned causal paths is mapped to a combined logical rule or probability transition function; When initializing the business digital twin model, the logical rules or probability transition functions are injected as part of the core state machine.

5. The method according to claim 4, characterized in that, The mapping process employs a combination of rule engines and neural networks: for well-structured deterministic causal relationships, a rule engine is used to generate deterministic logical rules; for complex causal relationships with uncertainty, a trained graph neural network is used to generate new probability transition functions.

6. The method according to claim 1, characterized in that, The process of encoding structural features in the federated causal graph into state vectors specifically involves: Perform a reverse breadth-first search on the federated causal graph, starting from the target anomaly node, to extract all causal adjacent nodes and edges within K hops; The graph attention network is used to aggregate features of the causal adjacent nodes and edges to generate a fixed-dimensional graph embedding vector. The graph embedding vector is concatenated with the current macroscopic state vector of the digital twin instance to form the complete state observation of the deep reinforcement learning agent.

7. The method according to claim 1, characterized in that, The deep reinforcement learning agent is trained using a causal-aware action masking mechanism. At each decision-making moment, a subset of actions that are allowed to be executed is calculated based on the current state of the federated causal graph and the digital twin instance; By using negative rewards to mask actions outside the aforementioned subset of actions, the deep reinforcement learning agent is guided to use actions only within a pre-defined causal and reasonable action space.

8. The method according to claim 7, characterized in that, The method for calculating the subset of actions that can be executed is as follows: A temporary intervention graph is constructed to simulate the impact of executing candidate actions on key state variables in the digital twin instance; A lightweight causal effect prediction model is used to determine whether the impact violates the strong causal constraints or business security boundaries in the federated causal graph. Candidate actions that do not violate constraints and boundaries are included in the subset of actions that are allowed to be executed.

9. The method according to claim 1, characterized in that, The method also includes a strategy evaluation step based on a counterfactual interpretation model: While outputting the sequence of intervention actions, a counterfactual explanation model coupled with the digital twin instance enhanced by the causal constraints is initiated; The counterfactual explanation model generates a simplified comparative intervention scheme that differs from the recommended sequence by minimizing the intervention cost; By simulating the execution of the comparative intervention plan and the recommended sequence, the expected differences between the two in key business indicators are quantified, and these differences are included in the output results as a reference for decision credibility.

10. An information-based intelligent tracking and management system for cross-border e-commerce business, used to implement the method of any one of claims 1 to 9, characterized in that, The system includes: A federated causal discovery module, configured with privacy computing middleware, is used to collaboratively construct the federated causal graph; A causal rule injection engine, connected to the federated causal discovery module, is used for dynamic configuration of the business digital twin model; A deep reinforcement learning subsystem integrating a state encoder and an action mask module is connected to the outputs of the federated causal discovery module and the causal rule injection engine, respectively, to generate the intervention action sequence; The strategy interpretation and evaluation module, connected to the deep reinforcement learning subsystem, is used to execute the strategy steps corresponding to the intervention action sequence.

Images