A user portrait and transaction permission adaptive control method for tax-free e-commerce

Abstract

This invention provides a user profiling and adaptive transaction permission control method for duty-free e-commerce. It belongs to the fields of big data processing and artificial intelligence technology. The invention includes: collecting user status and operation data, generating a fusion feature matrix through feature fusion; injecting this matrix into heterogeneous graph nodes to construct a dynamic graph, and mining hidden group topological features through a graph attention mechanism; inputting the topological features into a large model to deeply mine the intent of proxy purchasing, and outputting a comprehensive user risk confidence score; constructing an environment-aware threshold adjustment center to dynamically adjust risk control thresholds; matching the comprehensive risk confidence score with the dynamic thresholds to execute tiered permission control. This invention achieves the synergy of big data and a large model, enabling accurate identification of hidden proxy purchasing, reducing false positives while ensuring risk control, balancing security and revenue, and realizing user risk identification and permission control in duty-free e-commerce.

Description

Technical Field

[0001] In the field of e-commerce risk control and artificial intelligence technology, this invention relates in particular to a user profiling and adaptive transaction permission control method for duty-free e-commerce. Based on a pre-processed big data model, this method can profile users’ abnormal behavior and high-risk proxy purchasing networks in duty-free e-commerce scenarios, dynamically adjust thresholds, and adaptively control transaction permissions. Background Technology

[0002] With the development of internet technology and the digitalization of global trade, duty-free e-commerce has become an important channel connecting global consumers and goods. Duty-free shopping involves complex cross-border payments, customs supervision, tax calculation, and membership rights management, which makes its transaction system far more complex and risky than ordinary e-commerce.

[0003] In traditional duty-free e-commerce systems, user access control (such as purchase limits, discount eligibility, and access to specific products) typically employs a static rule-driven model. The system primarily presets permissions based on users' historical spending levels (such as silver, gold, and diamond cards) or fixed identity attributes (such as inbound travelers and departing travelers). While this role-based static access control model is sustainable in fixed scenarios, its limitations become increasingly apparent in the dynamic and complex environment of duty-free transactions. 1) Limited and outdated user profiling dimensions: Existing systems primarily rely on historical statistics and preset membership levels for user profiling, lacking in-depth insights into users' real-time intentions, current browsing behavior, and short-term needs and preferences. For example, they cannot distinguish whether high-net-worth users are shopping for personal use or stockpiling goods for resellers, leading to misjudgments in risk control strategies or inappropriate permission allocation; 2) Rigid access control: Transaction permissions (such as single transaction limits and category permissions) are usually fixed thresholds set in advance. When encountering major promotional events, sudden traffic surges, or significant changes in user demand, rigid access control cannot be dynamically adjusted according to real-time transaction risks and market supply and demand, which can easily lead to transaction failures or missed business opportunities; 3) Disconnect between Big Data and Large Language Model Applications: In existing technologies, big data platforms are mostly used for offline reporting and user segmentation, while the application of large language models is mostly limited to customer service Q&A or general recommendations. There is currently no effective mechanism to coordinate the deep semantic understanding and reasoning capabilities of large language models with the real-time streaming processing capabilities of big data for real-time user intent recognition and dynamic permission decisions. Summary of the Invention

[0004] This invention addresses the problems of limited user profiling dimensions, rigid access control, and the disconnect between big data and large-scale model applications in existing technologies. It provides an adaptive user profiling and transaction permission control method for duty-free e-commerce. By using a large language model to analyze the risk environment and user intent in real time, the method drives the big data platform to dynamically adjust user profile tags, thereby achieving adaptive control of transaction permissions. This improves transaction success rates and enhances risk control capabilities in complex duty-free business scenarios.

[0005] This was achieved through the following technical solutions: An adaptive control method for user profiling and transaction permissions in duty-free e-commerce includes the following steps: Step S1: Obtain macro-environmental factors and collect user status data and user operation data; Step S2: Perform feature fusion on user status data and user operation data to generate a fused feature matrix; Step S3: Inject the fused feature matrix into the heterogeneous graph nodes of the GNN neural network to construct a dynamic heterogeneous graph and mine the topological features of the group through the graph attention mechanism; Step S4: Input the topological features into the pre-trained large language model to mine the user's intention to purchase on behalf of others; Step S5: Construct an environmentally-aware threshold adjustment center to dynamically adjust risk control thresholds; Step S6: Execute adaptive control actions for transaction permissions based on the dynamic threshold matching results.

[0006] Optionally, the feature fusion process of user status data and user operation data in step S2 includes: User status data is numerically encoded to generate user status feature vectors; user operation data is processed using a sequence processing model to extract temporal features and generate behavioral trajectory feature vectors.

[0007] Optionally, key changes in user status data are detected as anchor points for business events. Based on these anchor points, a flexible time window is determined, and user operation data within the time window is associated with the event. Key changes include at least modification information and abnormal payment information. The user status feature vector is used as a query, and a cross-attention mechanism is used to weight and fuse the behavioral trajectory feature vectors to generate a fused feature matrix. This process establishes a spatiotemporal association between user status changes and dynamic behavioral trajectories using business event anchor points, directly providing initial node features with strong temporal correlation for constructing the dynamic heterogeneous graph in step S3, thus enhancing the sensitivity of the feature matrix to abnormal events.

[0008] Optionally, in step S3, when constructing the dynamic heterogeneous graph, multiple types of nodes are instantiated in the graph database of the GNN neural network. Node types include at least device feature clusters, user IDs, flight numbers, payment accounts, receiving mobile phone numbers, and physical addresses. The fused feature matrix is ​​dynamically injected into the corresponding nodes as a time-varying feature vector, and the feature representation of the nodes is updated in real time as user operation data changes. When the corresponding user operation data changes, the feature representation of each corresponding node is updated accordingly, resulting in the dynamic heterogeneous graph. The corresponding user operation data includes underlying clickstream data or front-end device fingerprints. After the feature representations are updated, connections between nodes are established, and each connection edge is assigned a dynamic weight based on a time decay factor. This edge weight design with a time decay factor allows the graph attention mechanism in step S3 to automatically reduce the influence of historically outdated or invalid connections when aggregating information, thereby ensuring that the mined topological features truly reflect the current gang activity and preventing outdated data from interfering with the judgment.

[0009] Optionally, the process of mining the topological features of hidden groups through graph attention mechanism in step S3 includes: using graph attention mechanism to perform multi-hop information aggregation on heterogeneous graphs, learning node topological relationships and assigning weights, and extracting the abnormally clustered subgraph topology from the normal behavior background. The normal behavior background refers to the regular network topology environment constituted by the vast majority of ordinary users and their daily behaviors in the dynamic heterogeneous graph, used to distinguish it from the abnormally clustered subgraph topology. These identified abnormal subgraph topological features are the core reasoning basis of the large language model in step S4, achieving a seamless connection from data filtering to deep semantic reasoning.

[0010] Optionally, the methods for further exploring the user's potential intention to purchase goods through third-party agents in step S4 include: The behavioral trajectory vector is attached to the user ID node corresponding to the dynamic heterogeneous graph; the node features of the dynamic heterogeneous graph are extracted and converted into group aggregation degree prompt words; the temporal behavior sequence attached to the node is extracted and converted into operation anomaly degree prompt words. The two types of prompt words are fused using an alignment network to generate prompt word vectors. These vectors are then input into a pre-trained large language model, which outputs a comprehensive risk confidence score and interpretable user tags. This provides an interpretable quantitative output for calculating the comprehensive risk confidence score in step S5.

[0011] Optionally, the process of constructing the environmentally aware automatic adjustment center for risk control thresholds in step S5 includes: Based on external macro-environmental data, an environmental context vector is constructed through an environmental encoder. This vector includes the scope of impact, time-dependent decay curves, and tolerance guidance factors. Global topological perturbation features of the dynamic heterogeneous graph are extracted to quantify the graph network oscillation index. The proportion and growth rate of users with high-risk confidence scores output by the large language model are statistically analyzed and quantified as the overall black market intent concentration. The graph network oscillation index and the overall black market intent concentration are then integrated to form a global risk level. This characterizes the intensity of the proxy buying attack currently being suffered by the platform, thereby enabling dynamic adjustment of user thresholds.

[0012] Optionally, the process of constructing an environmentally aware automatic adjustment center for risk control thresholds also includes: The environmental context vector and global risk level are fed into the threshold calibration network to output the dynamically adjusted risk control threshold.

[0013] Optionally, the matching and control actions in step S6 include: obtaining the user's comprehensive risk confidence score and dynamically adjusted risk control threshold; matching the comprehensive risk confidence score with the dynamically adjusted risk control threshold in real time; executing tiered access control actions based on the matching results, including identity verification, transaction restriction, or transaction interception; the new data generated by the execution results of this process will also flow back to step S1, participating in the next round of training and iteration as new user status data, forming the system's self-evolution capability.

[0014] The beneficial effects of this invention compared to the prior art are: 1) Multidimensional heterogeneous data alignment and graph denoising improve the accuracy and robustness of large language models.

[0015] Existing technologies often directly allow large language models to process raw, dirty data, or simply splice together multi-source features, which is like directly feeding crude oil mixed with oil stains into a precision engine. This results in high noise and low accuracy in the model input, making it prone to misjudgment in complex duty-free business scenarios.

[0016] This invention preprocesses the data, performing temporal alignment, cleaning, denoising, and semantic fusion of multi-source data such as flights, payments, and logistics before inputting it into a large language model, and constructs a dynamic heterogeneous graph containing entity associations. This deep fusion provides a high signal-to-noise ratio feature base for the large language model, significantly reducing the model error rate.

[0017] 2) Large language model deep intent reasoning, effectively penetrates the disguise of black market and uncovers hidden proxy shopping networks.

[0018] Existing technologies rely on manually preset thresholds and rule engines, primarily identifying explicit, one-dimensional abnormal behaviors (such as high-frequency purchases). They struggle to uncover the interconnected networks behind covert fraudulent activities like dispersed purchases, group buying, and cross-regional shipping. This invention utilizes an entity association network constructed from a dynamic heterogeneous graph. It integrates graph topological features with long-order user behavior sequences into structured prompts, which are then input into a large language model fine-tuned for tax-free business purposes for deep semantic understanding. This large model can penetrate the disguises of black market activities, accurately extracting abnormally clustered subgraph topologies from normal behavioral backgrounds, enabling early detection and intervention of hidden proxy buying networks. 3) Adaptive threshold adjustment based on environmental awareness reduces false positive rates during periods of business fluctuation.

[0019] Existing technologies use static thresholds for risk control and interception. When faced with sudden business fluctuations such as large-scale flight delays or duty-free sales, they act like an unchanging gate to the ebb and flow of tides. This leads to either large-scale false positives and customer complaints due to thresholds that are too low, or missed opportunities due to thresholds that are too high. This invention uses an environmentally aware threshold adjustment center to access macro-environmental factors (such as flight status and promotional activities) in real time, while simultaneously monitoring the graph network oscillation index and the overall concentration of malicious intent, constructing a global risk level. When legitimate environmental fluctuations occur, the system automatically calibrates the thresholds. When abnormal attacks are detected, the system dynamically tightens the thresholds for precise interception.

[0020] 4) Achieve synergistic optimization of security protection and business transformation through tiered flexible control.

[0021] Existing technologies employ a black-and-white blocking strategy, treating all risks with a one-size-fits-all approach. While this can intercept high-risk transactions, it can easily inadvertently harm genuine high-net-worth users, leading to significant business losses and customer complaints. This invention transforms the black-and-white blocking strategy into a tiered access control system, allowing for fine-grained access demotion (such as restricting logistics methods, mandating liveness detection, and blocking sensitive SKUs). While ensuring compliance with national duty-free policies and cutting off the daigou (personal shopping agent) chain, it maximizes the protection of the shopping experience for genuine high-net-worth travelers and recovers normal business revenue. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of a user profiling and transaction permission adaptive control method for duty-free e-commerce provided by an embodiment of the present invention; Figure 2 This invention provides a schematic diagram of the heterogeneous data deep fusion and intent reasoning process for a user profiling and transaction permission adaptive control method for duty-free e-commerce. Figure 3 This invention provides an internal refinement process for risk matching and judgment in a user profiling and transaction permission adaptive control method for duty-free e-commerce. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Figure 1 This is an overall flowchart of a user profiling and transaction permission adaptive control method for duty-free e-commerce provided in an embodiment of the present invention.

[0025] Figure 2 This invention provides a detailed flowchart of the heterogeneous data deep fusion and intent reasoning process for a user profiling and transaction permission adaptive control method for duty-free e-commerce, illustrating the entire process from acquiring user data to outputting risk confidence.

[0026] Combination Figure 1 and Figure 2 As shown, the adaptive control method includes the following steps: Step S1: Obtain macro-environmental factors and collect user status data and user operation data. Macro-environmental factors include data such as e-commerce promotion activity indicators that represent business fluctuations and flight delay warnings that represent changes in the external environment. User status data is structured data, including real-name identity information, customs quota reduction records, flight ticket status, and historical delivery addresses. User operation data is unstructured or semi-structured data, including device fingerprint data and page click flow. Step S2: Perform feature fusion on user status data and user operation data to generate a fused feature matrix; Based on the multi-source data collected in step S1, the data is first aligned in time sequence and then cleaned and denoised. Subsequently, feature extraction and fusion training are performed on the user status data and user operation data to generate a fused feature matrix. Specific sub-steps include: Step S201: Classify the acquired user status data and user operation data. User status data is structured data, and user operation data is unstructured or semi-structured data. Step S202: Numerically encode the user status data to generate a user status feature vector; Step S203: Use a sequence processing model to extract temporal features from user operation data and generate behavioral trajectory feature vectors; Step S204: Detect key changes in user status data as anchor points for business events. Key changes include changes in ticket status, initiation of high-frequency payment requests (change information, abnormal payment information), etc. Step S205: Determine a flexible time window based on the business event anchor point, and associate user operation data within the time window with the event. A flexible time window refers to a time interval whose length is dynamically adjusted according to the event type, centered on the business event anchor point. For example, for premeditated events such as ticket changes, the time window extends forward (e.g., 24 hours before the change) to capture planned behavior; for sudden events such as high-frequency payments, the time window focuses on a short time slice before and after the event (e.g., 10 minutes before and after). The design of the flexible time window can cover both long-term premeditated behavior and short-term explosive behavior, ensuring that the associated operation data and the triggering event are highly correlated in causal logic, and eliminating irrelevant noise.

[0027] Step S206: Using the user state feature vector as the query, the behavior trajectory feature vector is weighted and fused through a cross-attention mechanism to generate a fused feature matrix. In this step, the user state feature vector representing static identity is used as the query vector, and the behavior trajectory feature vector representing dynamic behavior is used as the key and value. The working principle is: calculate the similarity score between the query and the key; the higher the score, the more relevant the behavior at that moment is to the current user state, thus assigning it a higher weight. Its function is to allow the static identity to actively focus on the dynamic behavior segments most relevant to it (for example, a "high-frequency payment" state anchor will automatically focus on the rapid click behaviors before and after it with high weight), thereby achieving deep semantic alignment between structured events and unstructured behaviors. The generated fused feature matrix can simultaneously represent "who did what in what state," significantly improving the feature's ability to identify abnormal patterns.

[0028] This fused feature matrix not only preserves the temporal dependencies of the data, but also provides a standardized input for subsequent map construction.

[0029] Step S3: Inject the fused feature matrix into the heterogeneous graph nodes of the GNN neural network to construct a dynamic heterogeneous graph, and mine the topological features of the groups through the graph attention mechanism; the specific sub-processes for constructing the dynamic heterogeneous graph are as follows: Step S301: Instantiate various types of nodes in the graph database of the GNN neural network to construct a basic heterogeneous graph; the node types include at least device feature clusters, user IDs, flight numbers, payment accounts, receiving mobile phone numbers, and physical addresses; Step S302: The fusion feature matrix generated in step S2 is used as a time-varying feature vector and dynamically injected into each node in the heterogeneous graph that corresponds to the user ID and device feature cluster, so as to update the feature representation of each node. When the user's underlying click stream data or user operation data such as the front-end device fingerprint changes, the feature representation of the corresponding node will also be updated accordingly, forming a dynamic heterogeneous graph.

[0030] Step S303: After the feature representation in step S302 is updated, association edges between each node are established. Each association edge includes explicit strong association edges and implicit strong association edges. Explicit strong association edges refer to connections established based on explicit business rules, such as "the same user ID uses the same payment account" or "the same order corresponds to the same receiving mobile phone number." Explicit strong association edges are used to confirm established attribution relationships and form the skeleton of the graph. Implicit strong association edges refer to connections established based on statistical patterns or potential commonalities, such as "the same device fingerprint logs in to multiple different user IDs in a short period of time," "multiple user IDs access the same IP segment frequently," or "different users share similar click behavior sequences." Implicit strong association edges are used to uncover potential clues of organized crime and discover professional resellers controlling multiple accounts.

[0031] Step S304: Assign a weight W(t) to each associated edge based on a time decay factor; wherein the weight W(t) of the associated edge increases exponentially with time t. Decreasing (of which) (This is the decay coefficient). During the weight allocation process, recently occurring related behaviors (such as recent shared device logins) are assigned high weights, considered to have a significant risk contribution; for older related behaviors, the weights are automatically reduced, considered to have less reference value. The principle is that cybercrime groups often change their devices and accounts frequently, and recent associations better reflect the current criminal activity. The time decay mechanism can effectively filter historical noise and prevent outdated data from interfering with the current identification of the group.

[0032] Step S305: Employ Graph Attention (GAT) to perform multi-hop information transfer and aggregation on the heterogeneous graph, automatically assigning weights to different neighbor nodes, and accurately extracting and separating the abnormally clustered subgraph topology from the normal behavior background. The normal behavior background refers to the regular network topology environment composed of the vast majority of ordinary users and their daily behaviors in the dynamic heterogeneous graph, used to distinguish it from the abnormally clustered subgraph topology.

[0033] Step S3 effectively solves the problem that single-point analysis cannot detect large-scale black market groups, and the extracted topological features will serve as the core basis for the next step of deep reasoning.

[0034] Step S4: Input the topological features into the pre-trained Large Language Model (LLM) to mine the user's intent to use a proxy shopping service; utilize the strong semantic understanding capability of the Large Language Model (LLM) for inference to deeply mine the user's potential proxy shopping intent. Specific methods include: Step S401: Attach the behavior trajectory vector to the user ID node corresponding to the dynamic heterogeneous graph; extract the feature representation of each node in the dynamic heterogeneous graph. The feature representation refers to the high-dimensional feature vector stored on the corresponding node after multi-hop aggregation and updating via the graph attention mechanism in step S3. This feature vector not only contains the node's own initial attributes but also integrates the structural information and interaction weights of its neighboring nodes, which can characterize the user's topological position and potential gang association attributes in the tax-free transaction network.

[0035] Step S402: Convert the feature representation of each node in the dynamic heterogeneous graph into structured cue words reflecting the degree of gang aggregation; extract the temporal behavior sequence attached to each node and convert it into structured cue words reflecting the degree of operational anomaly; Step S403: The two types of prompt words obtained in Step S402 are concatenated and aligned using a specific alignment network to generate multimodal / structured prompt word vectors. These vectors are then input into a pre-trained large language model. The pre-trained large language model refers to a specialized model suitable for duty-free e-commerce risk control scenarios, obtained by adapting a general-purpose base model to the domain and fine-tuning its instructions using historical risk control cases, customs audit documents, and business rule corpora from the duty-free e-commerce sector. This model has internalized duty-free industry terminology and risk assessment logic, specifically designed to parse structured prompt words and output risk assessment results.

[0036] Step S404: The large language model comprehensively evaluates the graph features and temporal trajectory, comprehensively assessing the user's position in the gang network (graph features) and the user's current operational intent (temporal trajectory). The model finally outputs a risk confidence score (R) between 0 and 1, as well as the corresponding interpretable user tags (such as "suspected professional reseller", "gang member", etc.).

[0037] Step S5: Construct an environment-aware threshold adjustment center to dynamically adjust risk control thresholds; this step aims to solve the problem that static thresholds cannot adapt to dynamic environments. The specific steps for constructing an environment-aware automatic risk control threshold adjustment center include: Step S501: Based on external macro-environmental data, construct an environmental context vector through an environmental encoder. The environmental context vector includes the scope of impact, the time-effect decay curve, and the tolerance guidance factor for specific behaviors. The scope of impact includes at least the global scope, specific routes, or specific SKUs. Specific behaviors include at least high-frequency visits associated with marketing activities, remote logins associated with flight changes, or concentrated purchases associated with the launch of specific products. The tolerance guidance factor indicates the degree of tolerance of the risk control system for preset types of user behaviors in a specific macro-environment and is one of the bases for dynamically adjusting the risk control threshold. Step S502: Extract global topological perturbation features from the dynamic graph, calculate the number of multi-hop connected subgraphs that increase rapidly at high frequencies per unit time, and quantify them as the graph network oscillation index; synchronously count all user requests output by the large language model within the time window, calculate the proportion and growth rate of users with high-risk confidence comprehensive scores, and quantify them as the overall black market intent concentration; merge the graph network oscillation index and the overall black market intent into a continuous value, which serves as the global risk level of the current duty-free business, representing the intensity of the proxy purchase attack that the platform is currently suffering.

[0038] Step 503: Construct a real-time sliding analysis window to automatically calculate the current global risk level from the bottom up. Then, the environmental context vector, including the impact range, time-effect decay curve, and tolerance guiding factor, along with the global risk level, is fed into a threshold calibration network for cross-collision calculation (comparison calculation). The output deviation index is used to determine whether the rise in the global risk level is due to objective fluctuations caused by legitimate macroeconomic conditions, thus achieving automated dynamic adjustment of the risk control threshold. The threshold calibration network is a specially trained lightweight neural network (such as a multilayer perceptron MLP). Its working principle is that the network learns the normal coupling relationship between environmental factors and risk levels in historical data. If the current rise in risk level is consistent with the trend of environmental vector changes (such as a natural rise in risk level during a major promotion), the network outputs a normal matching signal, and the threshold remains unchanged or is slightly adjusted. If the risk level spikes but the environmental vector remains stable, or the environment indicates a major promotion but the risk level is abnormally lower than expected, the network outputs an abnormal deviation signal. The threshold calibration network automatically guides the system to raise or lower the risk control threshold, thereby achieving adaptive adjustment of the threshold and avoiding false positives or false negatives caused by environmental fluctuations.

[0039] Step 504: Based on the threshold drift obtained in step S503, the large language model will generate a dynamically adjusted set of risk control thresholds [T1, T2, T3, T4].

[0040] Step S6: Based on the dynamic threshold matching results, execute adaptive control actions for transaction permissions. Specific steps include: Step S601: Obtain the user's comprehensive risk confidence score R and the dynamically adjusted risk control threshold set [T1,T2,T3,T4], and match the comprehensive risk confidence score with the dynamically adjusted risk control threshold in real time; Step S602: Based on the matching result, generate the corresponding access control instruction, and execute the specific tiered transaction access control action according to the access control instruction; Step S603: The execution results (including erroneous samples from manual review or customer complaints) are fed back as negative samples. This data is recorded in the log system for audit traceability, and fed back to the environmental awareness threshold adjustment stage for iterative optimization of the threshold model. It is also fed back to the large language model for model fine-tuning, thereby achieving a self-evolving closed loop of risk control strategy.

[0041] Figure 3 This is a detailed flowchart of risk assessment and access control provided for embodiments of the present invention. It specifically illustrates the entire process from obtaining environmental factors to final step-by-step execution, and is a further elaboration of steps S5 (threshold adjustment) and S6 (access control) mentioned above. Figure 3 As shown, the process includes the following steps: The system receives real-time external macro-environmental data streams, including e-commerce promotional event markers and flight delay statuses. A pre-defined environmental encoder constructs an environmental context vector from this data. This vector contains three dimensions: impact range, time-dependent decay curve, and tolerance guidance factor, used for subsequent threshold calibration.

[0042] Acquire real-time cleansing features and historical graphs of users: Obtain the real-time behavioral features of the current user and the historical association information of the user in the dynamic heterogeneous graph, providing a data foundation for large model inference.

[0043] Large-scale model deduces and issues real-time control threshold set: Based on the environmental context vector and global risk level, the dynamically adjusted risk control threshold set [T1, T2, T3, T4] is deduced in real time. In this embodiment, T1, T2, T3, and T4 are low-risk threshold, medium-high-risk threshold, high-risk threshold, and extremely high-risk threshold, respectively.

[0044] The large model inference outputs the user's current comprehensive risk score R: The large language model also outputs the user's current risk confidence comprehensive score R.

[0045] Risk matching and judgment: The user's risk score R is compared with the dynamic threshold set [T1, T2, T3, T4]: if R < T1, then it is judged as low risk; If T1 ≤ R < T2, then it is classified as medium risk; If T2 ≤ R < T3, then it is classified as medium to high risk; If T3 ≤ R < T4, then it is judged as high risk; If R ≥ T4, it is considered to be extremely high risk.

[0046] Tiered access control: When low risk is determined, a no-crime clearance action is executed: allowing purchase and mailing of all categories; When a medium-risk condition is determined, a downgraded verification action is executed: the user is required to undergo identity verification (face liveness detection). When a medium-to-high risk level is determined, a downgrade action is taken: high-value SKUs are blocked. When a risk level is determined to be high, implement a downgraded access control action: or restrict mailing methods; When an extremely high risk is determined, a blocking action is taken: the transaction is intercepted and the transaction is added to a blacklist; End the assessment and log: Record the risk assessment results, the matched thresholds, the executed access control actions, and related context information to the log system for subsequent model iteration and audit traceability.

[0047] In summary, this invention addresses the problems of single-dimensional user profiling, rigid access control, and the disconnect between big data and large-scale model applications in duty-free e-commerce scenarios. It proposes an adaptive control method for user profiling and transaction permissions in duty-free e-commerce. By constructing a dynamic heterogeneous graph and a deep fusion mechanism of cross-attention, it achieves temporal alignment and feature denoising of multi-source heterogeneous data. It leverages graph attention to mine the topological features of hidden groups and inputs them into a pre-trained large-scale language model for deep intent reasoning, outputting a comprehensive user risk confidence score. An environmentally aware threshold adjustment center is introduced, dynamically calibrating risk control thresholds based on global risk levels and macro-environmental factors. Finally, tiered access control is implemented based on the matching results of the score and threshold. This invention achieves synergy between real-time big data processing and deep reasoning using a large-scale language model, significantly improving the ability to identify hidden proxy purchasing behavior. While ensuring risk control effectiveness, it reduces the false positive rate during business fluctuations and balances security protection and business conversion through a flexible control strategy, demonstrating strong practicality and progressiveness.

[0048] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A user profiling and transaction permission adaptive control method for duty-free e-commerce, characterized by the following steps: include: Step S1: Obtain macro-environmental factors and collect user status data and user operation data; Step S2: Perform feature fusion on user status data and user operation data to generate a fused feature matrix; Step S3: Inject the fused feature matrix into the heterogeneous graph nodes of the GNN neural network to construct a dynamic heterogeneous graph and mine the topological features of the group through the graph attention mechanism; Step S4: Input the topological features into the pre-trained large language model to mine the user's intention to purchase on behalf of others; Step S5: Construct an environmentally-aware threshold adjustment center to dynamically adjust risk control thresholds; Step S6: Execute adaptive control actions for transaction permissions based on the dynamic threshold matching results.

2. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 1, characterized in that, The process of feature fusion between user status data and user operation data in step S2 includes: Numerical encoding of user status data generates user status feature vectors; A sequence processing model is used to extract temporal features from user operation data to generate behavioral trajectory feature vectors.

3. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 2, characterized in that, The process of generating the fused feature matrix includes: Key changes in user status data are detected as anchor points for business events. Elastic time windows are determined based on these anchor points, and user operation data within the time window is associated with the events. The user state feature vector is used as the query, and the behavior trajectory feature vector is weighted and fused through a cross-attention mechanism to generate a fused feature matrix.

4. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 1, characterized in that, In step S3, when constructing the dynamic heterogeneous graph, multiple types of nodes are instantiated in the graph database of the GNN neural network. The node types include at least device feature clusters, user IDs, flight numbers, payment accounts, receiving mobile phone numbers, and physical addresses. The fused feature matrix is ​​dynamically injected into the corresponding node as a time-varying feature vector, and the feature representation of the node is updated in real time as the user operation data changes. When the corresponding user operation data changes, the feature representation of each node is updated accordingly, resulting in a dynamic heterogeneous graph; the corresponding user operation data includes underlying clickstream data or front-end device fingerprints; In this process, after the feature representation is updated, the associated edges between nodes are established, and each associated edge is assigned a dynamic weight based on a time decay factor.

5. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 1, characterized in that, Step S3, which involves mining the topological features of a group through graph attention, includes: A graph attention mechanism is used to aggregate multi-hop information in heterogeneous graphs, learn the topological relationships of nodes and assign weights, and extract the subgraph topology of abnormal clusters from the background of normal behavior.

6. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 1, characterized in that, The methods for in-depth analysis of users' potential purchasing intentions in step S4 include: Attach the behavior trajectory vector to the user ID node corresponding to the dynamic heterogeneous graph; Extract node features from dynamic heterogeneous graphs and convert them into group aggregation degree prompts; Extract the temporal behavior sequence attached to the node and convert it into operation anomaly warning words; The two types of cue words are fused using an alignment network to generate cue word vectors; Input the prompt word vectors into a pre-trained large language model, and output a comprehensive risk confidence score and interpretable user tags.

7. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 1, characterized in that, The process of constructing the environmentally aware automatic adjustment center for risk control thresholds in step S5 includes: Based on external macro-environmental data, an environmental context vector is constructed through an environmental encoder. The environmental context vector includes the scope of impact, the time-effect decay curve, and the tolerance guiding factor. Extract global topological perturbation features from dynamic heterogeneous graphs and quantify the graph network oscillation index; The proportion and growth rate of users with high-risk confidence scores output by the statistical big language model are quantified as the overall concentration of black market intent. The graph network oscillation index and the concentration of black market intentions are integrated to assess the overall risk level.

8. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 7, characterized in that, The process of building an environmentally aware, automated risk control threshold adjustment center also includes: The environmental context vector and global risk level are fed into the threshold calibration network to output the dynamically adjusted risk control threshold.

9. The adaptive control method for user profiling and transaction permissions for duty-free e-commerce as described in claim 1, characterized in that, The matching and control actions in step S6 include: The risk confidence score is matched in real time with the dynamically adjusted risk control threshold. Based on the matching results, tiered access control actions are executed, including identity verification, transaction restriction, or transaction blocking.

Images