A cross-platform user trust state identification and migration method and system

By mapping fields and processing cross-platform user behavior data with dual-view time-series correlation graphs, the accuracy and stability issues of cross-platform user trust state identification and migration were resolved. This enabled accurate identification and stable migration of user trust states, reducing the impact of short-term fluctuations and conflicting evidence.

CN123115952APending Publication Date: 2026-11-17AIXING (HAINAN) SMART TRAVEL TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611379106.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The heterogeneity of cross-platform user behavior data leads to uncertainty in account associations. Existing methods struggle to accurately identify user trust status and ensure stable migration, making them susceptible to short-term fluctuations and conflicting evidence.

Method used

By acquiring heterogeneous behavioral data across platforms, performing field mapping and behavior type normalization, constructing a dual-view time-series correlation graph, controlling evidence propagation using correlation confidence, and combining historical state trends and independent confidence, we can achieve accurate identification and stable migration of user trust status.

Benefits of technology

It reduces the impact of differences in data structure between platforms and low-quality data, reduces erroneous account associations, improves the consistency of user trust status identification, reduces false state migrations caused by short-term behavioral fluctuations, and locates and updates error sources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115952A_ABST
    Figure CN123115952A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for cross-platform user trust state identification and migration. It includes: acquiring heterogeneous cross-platform behavioral data and standardizing it into behavioral event data containing account identifiers; generating a set of candidate user association hypotheses and the association confidence of each hypothesis based on account association characteristics and cross-platform behavioral consistency; constructing a dual-view time-series association graph based on the behavioral event data of the accounts contained in each hypothesis; controlling evidence propagation with association confidence to obtain the trust state distribution of each hypothesis; updating the association confidence based on the consistency differences in the trust state distribution; and recalculating the trust state distribution until joint convergence; determining the current user trust state data and independent confidence based on the convergence result; and determining and executing a state migration strategy based on historical states. This invention can reduce the impact of false cross-platform account associations and erroneous evidence propagation, improving the accuracy, stability, and reliability of trust state identification and migration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of user behavior data processing technology, specifically to a cross-platform user trust status identification and migration method and system. Background Technology

[0002] During the dissemination of information and user outreach for car rental services, users can engage in activities such as browsing, saving, commenting, following, and inquiring about content or entities through various new media platform accounts. Due to differences in data structures, account identification systems, behavior types, and recording granularity across platforms, the collected user behavior data is typically heterogeneous, scattered, and of varying quality, making it difficult to uniformly organize the behavior data generated by the same user across different platforms.

[0003] Existing user analysis methods typically merge data directly based on account identifiers or determine unique association results based on limited account characteristics. When cross-platform account identity information is missing, behavioral characteristics are similar, or related evidence conflicts, accounts are prone to incorrect merging or splitting, leading to misattribution of behavioral evidence. Existing methods also often employ fixed-weighted scoring or single-platform behavioral statistics, with account association and trust status analysis operating independently, making it difficult to correct account association relationships based on status analysis results. Furthermore, the lack of collaborative analysis of behavioral evidence from different sources, times, and qualities makes trust status and its credibility susceptible to the influence of missing, abnormal, or conflicting data.

[0004] Furthermore, existing status update methods typically adjust user status directly based on a single new action or current rating, failing to adequately consider historical status change trends, status credibility, and duration, easily leading to frequent changes in user status. Subsequent private domain access, consultation, or service feedback is usually only used for overall adjustment of analysis parameters, making it difficult to pinpoint the source of error in account association, behavioral evidence processing, or status migration. Therefore, improving the accuracy, consistency, and stability of cross-platform account association, user trust status identification, and status migration has become a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-platform user trust status identification and migration method and system to solve the technical problems in the prior art, such as the difficulty in uniformly utilizing heterogeneous behavioral data across platforms, the uncertainty of cross-platform account association leading to incorrect attribution of behavioral evidence, the independence of account association and trust status analysis, and the susceptibility of user trust status updates to short-term fluctuations and conflicting evidence.

[0006] To achieve the above objectives, the present invention provides a cross-platform user trust state identification and migration method, comprising: Acquire heterogeneous cross-platform behavioral data and standardize it into behavioral event data with account identifiers; A set of candidate user association hypotheses and the association confidence of each hypothesis are generated based on account association characteristics and cross-platform behavior consistency. Each hypothesis in the set of candidate user association hypotheses represents the association relationship between cross-platform accounts belonging to the same user. Based on the behavioral event data of the accounts included in each hypothesis, a dual-view time-series correlation diagram is constructed, which includes a view of direct behavioral evidence and a view of relationship corroboration. The evidence propagation of the dual-view temporal correlation graph is controlled by the correlation confidence to obtain the trust state distribution of each hypothesis. The correlation confidence is updated according to the consistency difference between the trust state distributions, and the trust state distribution is recalculated with the updated correlation confidence until the correlation confidence and the trust state distribution converge together. The current user's trust status data and independent confidence are determined based on the converged trust status distribution and the associated confidence. The state transition strategy is determined based on historical and current user trust status data and independent confidence levels. The strategy is then executed and updated user trust status data is output.

[0007] As a further improvement to the above technical solution, the standardization includes: mapping fields and normalizing behavior types for the cross-platform heterogeneous behavioral data; determining data quality information based on at least two of the following: field completeness, duplication, anomaly, and source platform reliability; obtaining behavioral event data containing platform identifier, account identifier, behavior object identifier, behavior type, behavior time, and the data quality information, wherein the behavior object identifier is a content identifier or an operating entity identifier; and adjusting the propagation weight of the evidence formed by the behavioral event data in the dual-view time-series correlation graph based on the data quality information and under the association confidence control.

[0008] As a further improvement to the above technical solution, the generation of the candidate user association hypothesis set and the association confidence of each hypothesis includes: determining the deterministic association relationship between cross-platform accounts based on account authorization relationship or account binding relationship; determining the probabilistic association relationship between cross-platform accounts based on at least two of behavior time continuity, content semantic similarity, common operating entity, and historical interaction sequence similarity; combining cross-platform accounts based on the deterministic and probabilistic association relationships to form multiple candidate user association hypotheses, and determining the corresponding initial association confidence based on the deterministic and probabilistic association relationships contained in each hypothesis; maintaining the deterministic association relationship unchanged during the joint convergence process, and updating the association confidence of each hypothesis based on the consistency difference between the trust state distributions.

[0009] As a further improvement to the above technical solution, the dual-view temporal correlation graph includes candidate user nodes, platform nodes, platform account nodes, content nodes, and operating entity nodes corresponding to each hypothesis. The platform account node is connected to its respective platform node. The direct behavioral evidence view consists of behavioral event edges between platform account nodes and content nodes or operating entity nodes. The behavioral event edges carry behavioral type, behavioral time, and data quality information, forming a temporal relationship according to the behavioral time. The data quality information represents the data quality of the corresponding behavioral event data, and the propagation weight of the behavioral event edges is determined based on the data quality information. The relationship corroboration view consists of correlation edges between candidate user nodes and platform account nodes, and behavioral corroboration edges between different platform account nodes. The correlation edges are established based on the deterministic or probabilistic correlation relationship on which the corresponding hypothesis is based, and are weighted by the correlation confidence of the corresponding hypothesis. The behavioral corroboration edges represent the continuity of behavioral time, semantic similarity of content, or relationship of common interaction objects.

[0010] As a further improvement to the above technical solution, the evidence propagation is performed using a dual-branch graph processing network, including: encoding the behavior type, behavior time, and data quality information of the behavior event edge to obtain behavior event feature data; inputting the behavior event feature data into the direct behavior processing branch, performing differential transformation according to the behavior type, generating a time gating value based on the time interval between adjacent behavior events, and adjusting the propagation weight of the behavior event edge with the time gating value to obtain direct behavior features; inputting the association edge and the behavior corroboration edge into the relationship corroboration processing branch, determining the relationship strength based on the historical interaction sequence of cross-platform accounts, aggregating relationship information based on the relationship strength and the association confidence of the corresponding hypothesis to obtain relationship corroboration features; and fusing the direct behavior features and the relationship corroboration features through a state fusion unit to obtain the trust state distribution of each candidate user association hypothesis.

[0011] As a further improvement to the above technical solution, the generation of consistency differences between the trust state distributions includes: obtaining the first state distribution and the second state distribution of each hypothesis based on the direct behavioral evidence view and the relational corroboration view respectively; aggregating the behavioral event edges connected to each platform account node to obtain the account local state distribution; and generating consistency differences between the trust state distributions of each hypothesis based on the differences between the first state distribution and the second state distribution and the differences between each account local state distribution and the corresponding hypothesis's trust state distribution. The process of updating the association confidence and determining joint convergence includes: obtaining the association confidence of the current iteration based on the consistency difference between the association confidence, account association features, data quality information, and trust state distribution of the previous iteration, and recalculating the trust state distribution of each hypothesis based on the association confidence of the current iteration; when the changes in association confidence and trust state distribution of adjacent iterations satisfy the corresponding preset convergence conditions, it is determined that the association confidence and trust state distribution have reached joint convergence.

[0012] As a further improvement to the above technical solution, the determination of the current user trust state data and independent confidence scores includes: weighted aggregation of the corresponding trust state distributions based on the association confidence scores of each candidate user association hypothesis after joint convergence to obtain a user trust state distribution; determining the independent confidence scores based on the concentration of each association confidence score, the consistency difference between the trust state distributions, the coverage of effective behavioral events on the platform and time window, data quality information, and the degree of conflict between positive and negative behavioral evidence; determining the current user trust state data based on the user trust state distribution, wherein the independent confidence scores are used to characterize the credibility of the current user trust state data and are output corresponding to the current user trust state data.

[0013] As a further improvement to the above technical solution, the determination and execution of the state transition strategy includes: determining the state change trend based on historical user trust state data from multiple consecutive update cycles; determining candidate transition directions based on the change in current user trust state data relative to the historical user trust state data; when, within a preset duration window, the candidate transition direction continuously aligns with the state change trend and the independent confidence level continuously reaches the corresponding transition threshold, performing state upgrade or downgrade according to the candidate transition direction; if the state upgrade or downgrade is not performed, and if the positive behavioral evidence conflicts with the negative behavioral evidence, the state transition strategy is set to pending confirmation; otherwise, the trust state represented by the historical user trust state data is maintained; the state upgrade and state downgrade employ different transition thresholds or duration windows to form a state transition hysteresis, and updated user trust state data is generated based on the execution result of the state transition strategy.

[0014] As a further improvement to the above technical solution, the method further includes: generating a trust state package based on the updated user trust state data, wherein the trust state package includes a user association identifier, the association hypotheses of each candidate user after joint convergence and the corresponding association confidence, the updated user trust state data, the independent confidence, the valid evidence time window, the evidence path identifier, the state version, and the validity period; associating subsequent private domain access feedback, consultation feedback, or service feedback obtained within the validity period with the trust state package according to the user association identifier and the state version, and performing state mapping on each associated feedback to obtain verified trust state data; determining the state error based on the state difference between the verified trust state data and the updated user trust state data and the independent confidence, and locating the error source of the state error according to the association hypotheses of each candidate user in the trust state package and the corresponding association confidence, the valid evidence time window, and the evidence path identifier; and updating the corresponding account association parameters, direct behavior evidence processing parameters, relationship corroboration processing parameters, or state transition parameters according to the error source.

[0015] On the other hand, the present invention also provides a cross-platform user trust state identification and migration system, comprising: The event handling module is used to acquire cross-platform heterogeneous behavioral data and standardize it into behavioral event data with account identifiers; The candidate association module is used to generate a set of candidate user association hypotheses and the association confidence of each hypothesis based on account association characteristics and cross-platform behavior consistency. Each hypothesis in the candidate user association hypothesis set represents the association relationship between cross-platform accounts belonging to the same user. The dual-view construction module is used to construct a dual-view time-series correlation diagram, which includes a view of direct behavioral evidence and a view of relationship corroboration, based on the behavioral event data of the accounts contained in each hypothesis. The joint convergence module is used to control the evidence propagation of the dual-view temporal correlation graph with the correlation confidence, obtain the trust state distribution of each hypothesis, update the correlation confidence based on the consistency difference between the trust state distributions, and recalculate the trust state distribution with the updated correlation confidence until the correlation confidence and the trust state distribution jointly converge. The state generation and transition module is used to determine the current user trust state data and independent confidence based on the converged trust state distribution and associated confidence, determine the state transition strategy based on the historical and current user trust state data and independent confidence, execute the state transition strategy and output the updated user trust state data.

[0016] The beneficial effects of the cross-platform user trust state identification and migration method and system provided by this invention are as follows: 1. By mapping fields, normalizing behavior types, and processing data quality across heterogeneous cross-platform behavioral data, behavioral event data with account identifiers is generated. By adjusting the evidence propagation weight using data quality information, the impact of data structure differences between platforms and low-quality behavioral data on subsequent analysis can be reduced.

[0017] 2. By retaining multiple candidate user association hypotheses, the evidence propagation in the dual-view time-series association graph is controlled by the association confidence. The association confidence is updated in reverse based on the consistency difference between the trust state distributions, so that the account association and trust state analysis form an iterative closed loop, which helps to reduce the false propagation of behavioral evidence caused by incorrect account associations.

[0018] 3. By processing account behavior relationships and cross-platform associations through direct behavioral evidence views and relationship corroboration views respectively, and by using time gating values, relationship strength and association confidence to regulate information dissemination, the system can comprehensively utilize the time characteristics, relationship characteristics and quality characteristics of behavioral evidence to improve the consistency of user trust status identification.

[0019] 4. By separating user trust state data from independent confidence levels and combining historical state change trends, migration thresholds, and persistence windows to implement state upgrades, downgrades, maintenance, or pending confirmation, it is possible to reduce false state migrations and frequent state changes caused by short-term behavioral fluctuations and conflicting evidence.

[0020] 5. By storing candidate user association hypotheses, association confidence levels, effective evidence time windows, and evidence path identifiers through the trust state package, and mapping subsequent feedback to verify trust state data, it is possible to locate the data processing links corresponding to state errors and make targeted updates to the corresponding parameters. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the operation of the cross-platform user trust state identification and migration method provided in this embodiment of the invention. Figure 2 This is a structural diagram of the cross-platform user trust state identification and migration system module provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the application deployment architecture of the cross-platform user trust state identification and migration system provided in this embodiment of the invention; Figure 4 This is a schematic diagram illustrating the structure of behavioral event data and candidate user association hypotheses provided in an embodiment of the present invention, wherein... Figure 4 (a) is a schematic diagram of the behavioral event data structure. Figure 4 (b) is a schematic diagram of the relationship structure of multiple candidate user association hypotheses and corresponding association confidence levels; Figure 5This is a schematic diagram of the structure of the dual-view timing correlation diagram provided in an embodiment of the present invention, wherein, Figure 5 (a) is a view of direct behavioral evidence. Figure 5 (b) is a view supporting the relationship; Figure 6 This is a schematic diagram of the network structure for evidence propagation in a bi-branch graph processing network provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the joint convergence of the association confidence and trust state distribution provided in an embodiment of the present invention. Figure 8 This is a schematic diagram illustrating the distribution of trust states, association confidence, and independent confidence of each candidate user association hypothesis provided in this embodiment of the invention. Figure 9 This is a timing diagram illustrating the user trust state transition hysteresis provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the path-level update architecture for the trust state packet and subsequent verification feedback provided in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific implementation process. The implementation process is used to illustrate the data flow relationship between various data objects, processing mechanisms and modules, and is not intended to limit the platform type, server model, model layer number or parameter values; without changing the data processing relationship, the access protocol, storage medium and computing resources of each platform can be replaced according to the actual business scale.

[0023] This implementation method is designed for new media outreach scenarios in car rental services. The data source consists of a short video platform A, a text and image community platform B, and a video platform C. On each platform, personalized operators continuously publish vehicle information, driving guides, user experiences, and service Q&A content using fixed operational identifiers. Users interact with the content or operators through their platform accounts, engaging in activities such as browsing, watching, liking, saving, forwarding, commenting, following, sending private messages, and inquiring. The personalized operators function as operational nodes in data processing, rather than virtual roles performing all computations; account association, trust status identification, and migration are handled by backend computing services.

[0024] Figure 1The overall operational sequence of the method is illustrated. The system first receives heterogeneous behavioral data from across platforms and forms unified behavioral events, then generates several candidate user association hypotheses. Subsequently, a dual-view time-series association graph is constructed using the accounts and events covered by the candidate hypotheses. Evidence is propagated under association confidence control, and the association confidence is corrected in reverse. After joint convergence, the current user trust status data and independent confidence are formed. Finally, the migration result—maintain, pending confirmation, upgrade, or downgrade—is determined by combining historical states. The processing boxes in the diagram are executed in a top-down order. When streaming events arrive, they can be incrementally triggered from the event formation stage, without having to replay all historical data.

[0025] Figure 2 The diagram illustrates the connections between system modules. The event handling module and the candidate association module provide behavioral event data and candidate user association hypotheses to the dual-view construction module, respectively. The dual-view construction module outputs graph data with shared node indices but different edge sets. The joint convergence module reads the graph data and iteratively outputs the converged hypothesis-level state distribution and association confidence. The state generation and migration module merges this data with historical state data to generate updated user trust state data. Modules can reside within the same server process or be split into microservices connected via message queues and remote procedure calls.

[0026] Figure 3 The application deployment architecture is illustrated. Data from the three platforms enters the access gateway via authorization interfaces, Webhook callbacks, or log exchange files. After authentication, rate limiting, and receiving time stamps, the gateway writes the data to the message layer. Stream processing nodes perform idempotent deduplication based on event identifiers and create ordered partitions by account, sending the events to the trust state analysis service. The analysis service internally configures event processing and candidate association services, dual-view construction and graph processing network, and joint convergence and state transition services, accessing the event repository, graph data storage, and state version repository respectively. The updated trust state package is provided to the private domain access end, consultation end, or car rental service application via interfaces. Subsequent results generated by these applications enter the verification feedback end, which then returns the corresponding parameter update object along the evidence path.

[0027] Table 1. Examples of System Deployment and Data Processing Environment Configuration

[0028] Before entering the analysis domain, the platform account identifier uses a hashed alias with a business key, and the original account is only stored in the controlled mapping service; the event text and content text are read according to the minimum necessary range. Internally, the system calculates the chronological order using UTC millisecond timestamps, while retaining the platform time zone and original time fields to detect platform clock offsets. For minute-level online recognition scenarios, the stream processing node forms a micro-batch every 5 minutes; for daily operation scenarios, incremental events can be summarized by natural day. Both modes use the same data objects and model interfaces.

[0029] In one overall implementation, cross-platform heterogeneous behavioral data is acquired and standardized into behavioral event data containing account identifiers; a set of candidate user association hypotheses and the association confidence of each hypothesis are generated based on account association characteristics and cross-platform behavioral consistency; a dual-view time-series association graph containing direct behavioral evidence and relational corroboration views is constructed based on the behavioral event data of the accounts contained in each hypothesis; evidence propagation is controlled by association confidence, and the association confidence is updated and the trust state distribution is recalculated based on the consistency difference between trust state distributions until the two converge; then the current user trust state data and independent confidence are determined, and a state transition strategy is determined and executed in combination with historical states.

[0030] In step S101, the access gateway receives a JSON event stream from platform A, a Webhook callback from platform B, and an hourly generated CSV or Parquet file from platform C. Each raw record can at least parse out the platform-side account, the behavior object, the time of occurrence of the behavior, and the behavior action; the behavior object can be operational content or a personalized operational entity. If the platform only provides aggregated counts, the system marks the aggregated records as statistical events and retains the statistical window, without forging a single count value into multiple successive events. The received results are written to the raw area located by raw_ref, and the successfully parsed records are sent to the event handling module.

[0031] In step S102, the event handling module calls the platform adapter to convert different field names and enumeration values ​​into a unified event object. The unified event object maintains its original attribution using platform and account identifiers, maintains its time sequence using event time, connects to the content or operating entity using object identifiers, and carries event attributes for subsequent dissemination. The output of this step is a set of behavioral event data that can be retrieved by account, time window, and behavioral object; records that fail to be standardized are placed in an isolation queue, recording the failed fields and original references, and do not participate in this round of trust state calculation.

[0032] In step S103, the candidate association module searches for other platform accounts that may belong to the same user for the account to be processed. Instead of immediately fixing the highest-scoring account combination as a unique identity, it outputs candidate hypotheses H1 to HM. Each hypothesis stores the set of accounts it covers, the association evidence that generated the combination, and the initial association confidence level. The association confidence level represents the current degree of credibility that the candidate account combination belongs to the same user; it applies to the candidate hypothesis. It is not equivalent to the user's trust status, nor does it directly indicate whether the user will complete the transaction.

[0033] In step S104, the dual-view construction module reads behavioral events within the corresponding time window according to candidate hypotheses. Direct behavioral evidence view organizes the actual interactions between the account and the content or operating entity, while relationship corroboration view organizes candidate hypotheses, platform accounts, and cross-account consistency relationships. Both views use the same batch of account node identifiers, thus aligning the representation of the same account during the fusion phase; each view maintains its own adjacency relationship, avoiding the direct propagation of account association evidence as real behavioral events. The output graph object includes node indices, two sets of edge indices, edge attributes, candidate hypothesis pointers, and valid time windows.

[0034] In step S105, the joint convergence module reads the association confidence of each hypothesis in the r-th round, uses this value to limit the impact of the corresponding account evidence on the candidate user nodes, and calculates the hypothesis-level trust state distribution; then it compares the consistency between the local results of different views and different accounts and the hypothesis-level results. If an account in a candidate combination provides contradictory state evidence for a long period of time, its association confidence decreases; if the evidence from each account supports each other in time and state, its association confidence increases. The updated association confidence enters the (r+1)-th round of graph propagation until the change in association confidence and the change in state distribution are both below their respective thresholds, or the safe iteration limit is reached.

[0035] In step S106, the state generation and transition module aggregates the converged results of multiple candidate hypotheses, outputting a four-dimensional user trust state distribution, the current discrete trust state, and independent confidence scores. The current discrete trust state can be categorized as L1 unfamiliar public domain user, L2 potential user, L3 trusted quasi-transaction user, and L4 highly loyal private domain user. Independent confidence scores separately characterize the credibility of the state conclusion under the combined influence of account ambiguity, evidence coverage, evidence quality, and conflicting opinions. Therefore, even if two users are both identified as L3, they can have independent confidence scores of 0.86 and 0.52 respectively, allowing subsequent transition strategies to be applied differently.

[0036] In step S107, the migration module reads the historical state, current state distribution, and independent confidence level of the continuous update cycle. It first generates candidate migration directions, then determines whether the direction aligns with the historical trend and meets the continuity condition. If the condition is met, it writes an upgrade or downgrade version; if the direction is unstable, it maintains the original state; if the evidence conflict meets the condition, it writes a pending confirmation flag. Each output simultaneously records the calculation window, the set of candidate hypotheses used, the model version, and the migration reason code for the next cycle to read.

[0037] For example, account A1 from platform A, account B1 from platform B, and account C1 from platform C form the main candidate hypothesis H1 in this round, with another hypothesis H2 including account B2 from platform B. The system computes H1 and H2 in parallel, without deleting H2 from the input. After multiple rounds of propagation, the state distribution of H1 is concentrated at L3 and the association confidence increases, while the state distribution of H2 differs significantly from the local states of its constituent accounts and the association confidence decreases. After aggregation, the current state is still L3, but it only upgrades to L4 when the independent confidence remains sufficiently high over multiple periods. This data stream illustrates that the uncertainty of candidate account combinations, the uncertainty of state conclusions, and the stability of state transitions are carried by different data objects.

[0038] For the event standardization process, field mapping and behavior type normalization are performed on cross-platform heterogeneous behavioral data. Data quality information is determined based on field completeness, duplication, anomaly, and source platform reliability, forming behavioral event data that includes platform identifier, account identifier, behavior object identifier, behavior type, behavior time, and data quality information. The data quality information adjusts the propagation weight of behavioral evidence under the control of association confidence.

[0039] The platform adapter maintains a versioned field dictionary. For example, a platform A raw record might be {uid, item_id, action, ts, duration_ms}, a platform B callback might be {member, note, interaction_code, occurred_at, text}, and a platform C exchange file might be {open_account, target, event_name, event_time, source_batch}. The adapter maps uid, member, and open_account to account_id, item_id, note, and target to object_id, and converts seconds, milliseconds, or dates with time zones to event_time. Fields that cannot be directly read are left as null values, and their names are written to the missing_fields array.

[0040] The behavior type dictionary maps synonymous actions across platforms to a finite enumeration. For example, play, exposure, and read are unified as VIEW; finish and complete as COMPLETE; collect and favorite as FAVORITE; and direct_message and private_chat as MESSAGE. Inquiries about vehicle models, prices, self-driving needs, or service processes are further tagged as CONSULT, and the consultation category is stored as consult_topic. Unfollowing, explicit rejection, complaints, or a large number of invalid clicks in a short period of time can enter the reverse behavior category, but the original action code is still preserved, making it easier for the model to distinguish between general exit and substantial negative feedback according to business semantics.

[0041] The event identifier `event_id` preferentially uses the platform event serial number; if a serial number is missing, a hash is calculated by concatenating `platform_id`, `account_id`, `object_id`, `event_type`, `event_time`, and the original batch number. Stream processing nodes query the idempotent cache using `event_id` within a set retention period; if a match is found, `duplicate_flag` is set to 1 and not counted as duplicate evidence. Anomaly detection combines regular rules with an isolation forest: rules are used to identify future times, negative durations, enumeration out-of-bounds, and a large number of repeated actions within the same millisecond; the isolation forest generates anomaly scores based on the number of events per unit time, action intervals, and object dispersion.

[0042] Data quality is composed of four normalized components: field integrity (the proportion of non-empty required fields); non-duplicate quality (decreasing with repeated events); non-abnormal quality (conversely transformed by the anomalous score); and source reliability (determined by platform interface type, callback signature verification results, and historical packet loss rate). The quality value of a single behavioral event can be calculated using the following formula: (1) in, Represents behavioral events Data quality values; Indicates the first The normalized values ​​of each quality component correspond to field integrity, non-repetition, non-abnormality, and source reliability, respectively. This represents the non-negative weight of each mass component. In the example, values ​​of 0.30, 0.20, 0.25, and 0.25 can be used. Adjustments can be made according to the platform after offline calibration.

[0043] Standard events are validated using Avro or JSON Schema and then written to the event subject. Commonly used field structures are: event_id:string, platform_id:string, account_id:string, object_id:string, object_kind:enum, event_type:enum, event_time:int64, numeric_value:float, text_ref:string, quality_info:struct, raw_ref:string. object_kind can be either CONTENT or OPERATOR; quality_info stores four components, anomaly codes, and field dictionary versions in addition to the quality value. Figure 4 (a) Represent the organizational relationship of these fields from the event identifier to the original trace reference in the form of a structure box.

[0044] When constructing the graph, the basic propagation weight of an event is first multiplied by the association confidence of the corresponding candidate hypothesis and the data quality value. Edges formed by high-quality consultation events from high-confidence candidate hypotheses receive higher basic weights; the weights of browsing events from low-confidence account combinations or those with time anomalies are suppressed.

[0045] If the data quality is below the isolation threshold of 0.20, the event only enters the audit storage and no propagation edge is established; events between 0.20 and 0.50 still participate in the calculation, but carry the LOW_QUALITY flag so that valid evidence can be distinguished when calculating the coverage of independent confidence scores later.

[0046] Taking a collection record from platform B as an example, the original time is 2026-08-18T21:16:04+08:00, the content identifier is note_307, and the account alias is b_82f1. After standardization, event_type is FAVORITE, event_time is UTC milliseconds, object_kind is CONTENT, field integrity is 1.00, non-repetition is 1.00, non-abnormality is 0.92, and source reliability is 0.90. The quality value obtained by weighting by example is 0.956. This result, along with raw_ref, enters the candidate association and dual-view construction stage. Any subsequent state errors can be traced back to the access batch and original fields through raw_ref.

[0047] During the candidate association formation process, deterministic association relationships between cross-platform accounts are determined based on account authorization relationships or account binding relationships, and probabilistic association relationships are determined based on at least two of the following: behavioral temporal continuity, content semantic similarity, common operating entity, and historical interaction sequence similarity. The two types of relationships are combined to form multiple candidate user association hypotheses and the initial association confidence is calculated. In the subsequent joint convergence process, the deterministic relationships remain unchanged, and only the association confidence corresponding to the probabilistic relationships is corrected.

[0048] Deterministic associations are generated by verifiable authorization chains. For example, if a user enters a car rental application from platform A using a one-time redirect token and completes authorization within the token's validity period, the access service binds the subject alias of platform A to the user_link_id within the application. When the same user_link_id is then authorized to log in from platform B, a deterministic association is formed between the two platform accounts. Existing user-confirmed account bindings within the enterprise can also serve as deterministic relationships. This relationship edge is set to hard_link=1 and stores the authorization serial number, signature verification result, and effective time, and is not deleted due to state differences during joint convergence.

[0049] Probabilistic associations are formed by feature services. Behavioral temporal continuity is calculated by statistically analyzing the event intervals of accounts within adjacent time periods related to the same travel theme, with excessively short, physically impossible overlaps treated as negative features. Content semantic similarity is achieved by encoding titles, comments, and consultation summaries into 768-dimensional vectors using a general Chinese pre-trained language model, then projecting them to 64 dimensions via a linear layer and calculating cosine similarity. Shared operating entities are determined by the weighted intersection-union ratio (IUU) of the sets of operating entities interacted by the two accounts. Historical interaction sequence similarity employs dynamic time warping with time tolerance for behavioral type sequences, or calculates cosine similarity for fixed-length sequence embeddings. All features are scaled to 0 to 1 within the quantile range of the training set.

[0050] For a candidate account combination h, the feature service concatenates the mean, minimum, missing rate, and conflict markers of the features of each pair of accounts within the combination into an associated feature vector, and calculates the initial association confidence using logistic regression or a two-layer perceptron. When using logistic regression, it can be expressed as: (2) in, This represents the initial association confidence level of the candidate user association hypothesis; This represents the account association feature vector that supports this hypothesis. The bias term was obtained by training with labeled cross-platform accounts; This represents the Sigmoid mapping. When supervised samples are lacking, positive samples can be constructed first based on confirmed binding relationships, and negative samples can be constructed based on relationships that cannot coexist in the same time period. Then, the threshold can be calibrated by manual sampling.

[0051] Candidate generation employs a combination of blocking and beam search. First, candidate buckets are created based on shared user_link_id, identical authorization tokens, or high-probability account pairs. Then, at most the top K accounts from each platform are retained. The beam search score is based on cross-platform coverage and associated feature scores, and the top M hypotheses are retained from the combinations. An example configuration uses K=5 and M=3, requiring each hypothesis to cover at least two platforms. Hypothesis objects store the fields: hypothesis_id, account_set, hard_edges, soft_edges, feature_vector, initial_confidence, and status. If all probabilistic candidates have a probability below 0.15, only the single-account hypothesis for the current platform account is output to prevent forced merging.

[0052] Figure 4 (b) illustrates the relationships between three candidate hypotheses: account A1 on platform A, accounts B1 and B2 on platform B, and account C1 on platform C. H1, comprising A1, B1, and C1, has an initial association confidence of 0.78 due to a single authorization binding and strong behavioral consistency. H2, replacing B1 with B2, is primarily based on temporal and semantic similarity, with an initial value of 0.16. H3, consisting of B1 and B2, has weaker evidence and an initial value of 0.06. Solid lines in the figure represent deterministic relationships, while dashed lines represent probabilistic relationships. The value 0.78 will not be directly interpreted as the probability of a trust state, but will only be used to control the extent to which the account evidence contained in H1 propagates to the candidate user nodes in H1.

[0053] When a new event arrives, the system does not need to re-enumerate all historical accounts. The candidate association module reads the soft relationship neighbors and valid deterministic relationships within the last 30 days, centered on the affected account, and only updates the assumptions that include that account; unaffected assumptions retain the previous state version. Deterministic relationship revocation can only be triggered by authorized unbinding events or controlled mapping services. After revocation, a new candidate set and state version are generated, and the old version retains audit references but is no longer used in the current migration.

[0054] In the implementation of the dual-view temporal association graph, a candidate user node is established for each candidate hypothesis, along with a platform node, platform account node, content node, and operating entity node; the platform account node is connected to its respective platform node. The direct behavioral evidence view consists of behavioral event edges from the account to the content or operating entity, while the relational corroboration view consists of association edges from the candidate user to the platform account and behavioral corroboration edges between cross-platform accounts. Both types of views carry edge attributes adapted to their semantics.

[0055] The node table uses a global index combining `node_id` and `node_type`. Platform nodes store platform category and interface version; platform account nodes store alias accounts, active windows, and platform affiliation; content nodes store content topic vectors, publication time, and corresponding operating entity; operating entity nodes store stable operation identifiers and content sets; candidate user nodes store `hypothesis_id`, pointers to included accounts, and current association confidence. Initial node features are formed by concatenating category embeddings, statistical features, and textual semantic projections. Missing dimensions are masked instead of being padded with zeros before directly participating in attention calculations.

[0056] In the direct behavioral evidence view, each standard behavioral event generates a directed edge from the platform account node to the content node or operating entity node. Edge attributes include behavioral type embedding index, event time, data quality value, numerical strength, text digest reference, and raw_ref. Multiple events from the same account for the same object are not merged into a single static count edge; instead, they are sorted by event_time to form an edge sequence. To control the graph size, identical exposure events without text can be merged within a 24-hour window, and the event_count and start / end time are saved.

[0057] In the relational corroboration view, candidate user nodes are linked to their contained account nodes via edge connections. Deterministic edge connections have an edge_mode of HARD and a propagation upper bound of 1, while probabilistic edge connections have an edge_mode of SOFT and a weight based on the confidence level of the candidate hypothesis. Three types of behavioral corroboration edges can also be established between account nodes from different platforms: TEMP_CONT represents the temporal continuity of behavior, SEM_SIM represents the semantic similarity of interaction content, and COMMON_OP represents a common operating entity or common interaction object. Multiple edge connections can coexist for the same account pair, and edge attributes store similarity and calculation windows.

[0058] Figure 5 (a) Using A1 and B1 as account nodes, the browsing, collection, comment and consultation events are shown to be connected to content C01, content C02 and operating entity O01 respectively, and a timeline is formed at the bottom by t1 to t4. Figure 5 (b) Maintain the correspondence between account nodes such as A1 and B1. The candidate user node H1 is connected to the account nodes through association edges with association confidence. The accounts are then connected through corroborating edges such as time continuity, semantic similarity, and common operating entity. The same node position and number enable the two views to be aligned in the state fusion unit, but no corroborating edge will be disguised as a user's actual consultation or collection in the direct behavior view.

[0059] In engineering, graph databases can be used to store traceable relationships, and sparse tensors can be used to perform model calculations. The graph construction service serializes `direct_edge_index`, `direct_edge_attr`, `relation_edge_index`, `relation_edge_attr`, `node_feature`, and `hypothesis_ptr` into graph batches; `hypothesis_ptr` records the node range of each candidate hypothesis in the batch. For single-user events exceeding the limit, all inquiries, private messages, follows, and reverse behaviors are retained. Regular browsing is stratified by time, and the sampling ratio is written into the edge attributes to prevent high-frequency exposure from overshadowing low-frequency, high-information-content behaviors.

[0060] This example uses the most recent 30 days as the main evidence window and the previous 60 days as the background window. Events in the main window propagate sequentially according to the complete edge sequence, while the background window only generates decaying statistical features. When an operating entity continuously publishes vehicle information and self-driving content, content nodes are connected to the same operating entity via operator_id. This allows the system to distinguish between occasional browsing of a specific trending topic and continuous attention, repeat visits, and inquiries to a fixed, personalized operating entity. Thus, the three-dimensional trust foundation of "recognizing the brand, recognizing the service, and recognizing the person" is implemented as computable nodes and edges such as brand or platform affiliation, service inquiries, and fixed operating entities, rather than abstract labels.

[0061] In the evidence propagation implementation, a dual-branch graph processing network is adopted. First, the behavior type, behavior time, and data quality information of the behavior event edges are encoded; the direct behavior processing branch performs differential transformation according to the behavior type, forms a time gating value based on the time interval between adjacent behavior events, and adjusts the propagation weight of the behavior event edges accordingly; the relationship corroboration processing branch determines the relationship strength based on the historical interaction sequence of cross-platform accounts, and aggregates relationship information under the control of association confidence; the state fusion unit fuses the features of the two branches and outputs the trust state distribution of each candidate hypothesis.

[0062] The event encoder consists of discrete embeddings and continuous feature projections. `event_type` is mapped to a 16-dimensional type embedding, while `object_kind` and `platform_id` are each mapped to an 8-dimensional embedding. Event hour, day of the week, and relative window position are encoded using sine and cosine time encoding. Viewing duration, completion rate, consultation rounds, and data quality values ​​are truncated to quantiles before being input into a two-layer perceptron. Comment or consultation text is used to generate semantic vectors from general Chinese encoders and cached in a content feature library; events without text are masked. The concatenated result is processed through a 64-dimensional linear layer and layer normalization to obtain event features.

[0063] The direct behavior processing branch sets a shared foundation plus a type adaptation matrix for behaviors such as VIEW, COMPLETE, FAVORITE, COMMENT, FOLLOW, MESSAGE, CONSULT, and NEGATIVE. The shared foundation learns common patterns across behaviors, while the type adaptation matrix preserves the differences between consultation, favorites, and normal browsing. After sorting events for the same account by time, adjacent events are subject to exponential time gating. (3) in, Indicates the time gating value of the behavioral event; This indicates the time interval between this event and the previous valid event on the same account; Indicates the type of behavior; A learnable or calibrated decay scale representing the corresponding behavior type; The overall propagation weight of the event among the candidate hypotheses is represented by the association confidence score. Data quality value This is combined with the time-gated value. Consulting applications can have a longer attenuation scale, while regular exposure applications can have a shorter attenuation scale.

[0064] When aggregating adjacent events, an account node uses a graph attention mechanism. The attention network concatenates event features, the account's current hidden state, behavior type adaptation results, and edge weights to generate an attention score, which is then normalized within the account's direct behavior neighborhood. (4) in, This represents the normalized attention weights that indicate the propagation of an event to account nodes; This indicates the event neighborhood of the account node in the direct behavior evidence view; Indicate the characteristics of the event; This represents the transformation matrix corresponding to the behavior type; This represents the direct behavioral characteristics after aggregation. The aggregation result is then input into a 64-dimensional GRU temporal memory layer, preserving the sequential relationship between consecutive consultations, follow-ups, and collections in a hidden state.

[0065] The relationship verification branch first assembles a sequence of the N most recent behavior types and their relative times for each account. The front end of the sequence uses a one-dimensional convolution with kernel widths of 3 and 7 to extract short-term action combinations, and then a bidirectional GRU or lightweight Transformer extracts longer-term interaction rhythms. The strength of the relationship between two accounts can be determined by sequence similarity, semantic content similarity, and shared interaction objects. (5) in, This indicates the strength of the relationship between account a and account b; Indicates the similarity of historical interaction sequences; Indicates the semantic similarity of content or consultation; This indicates the normalized overlap between shared operating entities and shared interaction objects; This represents a non-negative fusion coefficient. The relation graph attention layer multiplies this strength by the confidence level associated with the candidate hypothesis and then aggregates neighbor information.

[0066] The direct behavior branch and the relationship corroboration branch each stack two layers of graph attention updates, and residual connections are preserved between layers to avoid convergence of account representations after multiple layers of propagation. The state fusion unit performs attention pooling on the direct behavior features and relationship corroboration features of the account covered by the candidate hypothesis, and then concatenates them into a hypothesis-level representation, outputting a four-level trust state distribution: (6) in, This represents the L1 to L4 four-dimensional probability distribution corresponding to candidate hypothesis h; and These represent the hypothesis-level features of the direct action branch and the relational supporting branch, respectively; The bias is the state output layer parameter; the symbol "∥" indicates vector concatenation.

[0067] The model training samples are used to generate supervisory labels based on authorized and confirmed account associations, private domain access results, effective consultation depth, and car rental service feedback. Level 4 status labels can be initially generated by business rules, and then conflicting samples are reviewed by two people. The training objective simultaneously constrains state classification and dual-view... Figure 1 Consistency and correlation confidence calibration: (7) in, Indicates the total loss; This represents the cross-entropy of the fourth-level state; This item limits the output of two views to produce unfounded, excessively large differences; The association confidence level is calibrated using Brier loss or binary cross-entropy. λ represents the trainable parameters of the network; each λ represents a loss weight. This loss is only used for training, and online joint convergence is still updated round by round according to the candidate hypotheses.

[0068] Table 2 Examples of Two-Branch Graph Processing Networks and Migration Decision Parameters

[0069] Figure 6Corresponding to the network data flow described above, the upper direct action branch enters event encoding, type transformation, temporal gating, graph attention, and temporal memory layers from the action event edges; the lower relational evidence branch enters sequence encoding, relational strength calculation, and relational graph aggregation from the association edges, supporting edges, and historical sequences. Data quality values ​​are gated from the upper direct evidence, and association confidence is gated from the lower relational evidence. After the two branches meet in the state fusion unit, the L1 to L4 distribution is output by Softmax. The network can be implemented using a common graph tensor framework, with each layer's input and output being fixed-length tensors, facilitating batch execution on CPUs or GPUs.

[0070] In the joint convergence implementation, the first and second state distributions of each candidate hypothesis are obtained from the direct behavioral evidence view and the relational corroboration view, respectively. Local aggregation is then performed on the behavioral event edges connected to each account node to obtain the local account state distribution. Consistency differences are then generated based on the differences between the two view distributions and the differences between the local account distribution and the hypothesis-level distribution. The previous round of association confidence, association features, data quality, and consistency differences collectively generate the next round of association confidence, which is then updated and evidence propagation is re-executed.

[0071] The first state distribution is obtained by connecting the state output head separately to the direct action branch, and the second state distribution is obtained by connecting the isomorphic output head separately to the relational corroboration branch. The account local state distribution is calculated using only the account's own valid events after severing the aggregation edges between candidate user nodes and other accounts. To avoid inconsistencies in the scales of the two output heads, the same temperature parameter is used for probability calibration during offline training; a symmetric and bounded Jensen-Shannon divergence is used for online comparison. The consistency difference of the candidate hypotheses can be expressed as: (8) in, This indicates the consistency difference of candidate hypothesis h; and Let these represent the first state distribution and the second state distribution, respectively. This represents the local state distribution of the assumed account 'a'; This represents the account weight obtained by normalizing the number of valid events and the average data quality of the account. Used to balance differences between views and differences within an account.

[0072] If A1 and B1 in H1 consistently exhibit L3 behavior, while C1 only receives a small number of low-quality views, the account weight will limit C1's impact on the difference. However, if C1 has multiple high-quality rejection or complaint events that are locally concentrated in L1, the divergence between the local state and the assumed state increases significantly and cannot be simply overwhelmed by the number of events. Thus, consistency differences reflect both whether the interpretations of the two graph views support each other and whether there are accounts within the candidate combinations that should not be merged.

[0073] The association confidence scores of probabilistic candidate hypotheses are updated using damped updates to avoid confidence score jumps caused by a single round of distribution fluctuations. (9) in, Indicates the association confidence level in the r-th round; Indicates the update step size; This indicates account association characteristics that do not change with this round of dissemination; This represents the penalty coefficient for consistency differences; This indicates the consistency difference in round r. The clip operation limits the result to 0 to 1. For edges with deterministic associations, the hard_link constraint is maintained, and the corresponding edge weights are not updated.

[0074] The scheduler updates in a loop following the sequence of "graph propagation—distribution comparison—confidence update—repropagation". To improve efficiency, in the (r+1)th round, event encoding and node features unaffected by confidence are reused, and only relation gating, neighborhood aggregation, and state output are recalculated. Each round saves the iteration_id, the confidence level of each hypothesis, the four-dimensional state distribution, the difference components, and the time consumed. If the confidence level of a hypothesis is below 0.02, it can be put into dormancy but its record is retained, and it will be reactivated when new definitive evidence emerges.

[0075] The joint convergence criterion simultaneously constrains both the confidence level and the state distribution: (10) in, Indicates the threshold for changes in association confidence; This represents the threshold for the change in the norm of the confidence state distribution. `converged=true` is only written if the maximum changes of both valid candidate hypotheses are below the threshold; if the threshold is not met even after the maximum number of iterations, `converged=false` is written, and a non-convergence penalty is added to the independent confidence score, preventing the output of high-confidence transitions disguised as stable results.

[0076] Figure 7 A set of example iterations is given. The changes in association confidence and state distribution both decrease with each round, falling below 0.001 and 0.002 respectively in round 11, thus indicating joint convergence. Simultaneously, the association confidence of H1 increases from 0.56 and stabilizes at approximately 0.795, while H2 decreases from 0.31 and stabilizes at approximately 0.158. The curves in the figure illustrate that both types of variables must stabilize synchronously; one cannot only observe network loss or only observe account association scores. An online implementation can incorporate these two curves into monitoring metrics to detect prolonged oscillations caused by abnormal data.

[0077] During the state generation phase, the corresponding trust state distributions are weighted and aggregated based on the association confidence of each candidate user association hypothesis after joint convergence to obtain the user trust state distribution. Then, the independent confidence is determined based on the degree of association confidence concentration, consistency difference, platform and time coverage of effective behavioral events, data quality, and the degree of conflict between positive and negative behavioral evidence. The current user trust state data is determined by the user trust state distribution.

[0078] The candidate hypothesis weighted aggregation uses normalized association confidence scores. For the associated entity u to be analyzed, its candidate set is Hu, and the aggregation distribution can be expressed as: (11) in, This represents the distribution of user-level L1 to L4 trust states; This represents the set of valid candidate hypotheses corresponding to the user's associated entities; and Let represent the association confidence and hypothesis-level state distribution after joint convergence, respectively. If the sum of the confidence scores of all candidates is lower than the minimum value, the system degenerates into the single-account hypothesis with the highest confidence score and writes the low-confidence reason into the state data.

[0079] The concentration of association confidence can be calculated by inversely using the entropy of normalized candidate confidence: it approaches 1 when the main probability is concentrated on one hypothesis, and approaches 0 when multiple hypotheses are nearly uniform. Consistency difference is taken as the weighted average of valid candidates; platform coverage is calculated as the proportion of platforms with high-quality events to the total number of access platforms, and time coverage is calculated as the proportion of time buckets with valid events within the main window. The geometric average of the two forms the coverage value; data quality is taken as the weighted average of valid events after propagation.

[0080] Positive behavioral evidence includes complete broadcasts, favorites, continued following, proactive inquiries, private domain access, and confirmation of effective services. Negative behavioral evidence includes explicit rejection, unfollowing, complaints, denial of needs after inquiries, and verified invalid outreach. The system calculates the weighted sum of the two types of evidence separately, increasing the degree of conflict when both are high; the degree of conflict is low when only one type of evidence dominates. The independent confidence level can be generated by the following formula: (12) in, Indicates the degree of independence; Indicates the degree of concentration of confidence in candidate associations; This represents the weighted mean of consistency differences; Indicates the extent of platform and time coverage; Indicates the average data quality of valid events; It indicates the degree of conflict between the positive and negative evidence; each coefficient is determined by the confidence level calibration error on the validation set.

[0081] The current trust state is the level with the highest probability in the user-level distribution, and the complete distribution is stored, not just the label. If the probability difference between L2 and L3 is less than 0.05, an AMBIGUOUS_BOUNDARY reason code can be added; this reason code does not change the distribution, but only allows the migration module to increase the persistence window requirement. Independent confidence and state probability are stored separately, for example, state_label=L3, state_probability=0.572, independent_confidence=0.810, to prevent "probability belonging to L3" from being mistakenly written as "overall reliability of the conclusion".

[0082] Figure 8 The diagram illustrates the correspondence between the three candidate hypotheses. H1 and H2 both have high probabilities at L3, but H1 has a correlation confidence of 0.78, while H2 is only 0.16. H3 is mainly distributed at L2 with a confidence of 0.06. After aggregating by correlation confidence, the user-level state still predominates at L3; combining candidate concentration, coverage, consistency, quality, and conflict, an independent confidence score of 0.81 is obtained. The heatmap, right-hand confidence bars, and bottom aggregation distribution represent the hypothesis-level conclusion, the credibility of account combinations, and the user-level conclusion, respectively, and these three are structurally independent.

[0083] During the state transition phase, the state change trend is determined based on historical user trust state data from multiple consecutive update cycles, and candidate transition directions are determined based on the change in the current state relative to historical states. When a candidate direction is consistent with the historical trend within a continuous window and the independent confidence level consistently reaches the corresponding threshold, a state upgrade or downgrade is executed. If no upgrade or downgrade is executed, and there is conflicting evidence, the process enters a pending confirmation phase; otherwise, the historical state is maintained. Upgrades and downgrades use different thresholds or continuous windows to create a transition hysteresis.

[0084] To analyze trends in continuous space, the fourth-level state distribution of period t is transformed into expected state values ​​between 0 and 1: (13) in, This represents the normalized state value in period t; This represents the probability that the current period belongs to the k-th trust level. L1, L2, L3, and L4 are mapped to 0, 1 / 3, 2 / 3, and 1, respectively. This continuous value is used for trend calculation, and the final output still retains the discrete levels and their complete probability distribution.

[0085] Historical state change trends are represented using a least-squares slope with decaying weights, with higher weights for recent periods. (14) in, Indicates the slope of the state changes in the history window; Indicates the length of the trend window; Indicates the time decay coefficient; and These represent the weighted time mean and the weighted state mean, respectively. A slope greater than a positive threshold indicates a continuous increase, less than a negative threshold indicates a continuous decrease, and the rest are considered stationary. For missing periods, the most recent state value is used, but the independent confidence level is set to 0 to avoid misclassifying missing values ​​as stable evidence.

[0086] The difference between the current state value and the weighted average of the historical windows forms the state change. An upward candidate direction is generated when the change is greater than the upgrade threshold, a downward candidate direction is generated when it is less than the negative downgrade threshold, otherwise a hold direction is generated. Upward candidate directions are counted in consecutive upgrade counts only when the historical slope is also positive; downward candidate directions are counted in consecutive downgrade counts only when the historical slope is negative. Direction interruption, insufficient confidence, or missing windows will reset the corresponding consecutive count to zero, but will not directly result in reverse migration.

[0087] In the example, the upgrade boundary is set to 0.68, the independent confidence threshold is 0.75, and a three-period consecutive update is required; the downgrade boundary is set to 0.34, the independent confidence threshold is 0.70, and a three-period consecutive update is also required. The upgrade threshold, downgrade threshold, and duration window can differ and are determined on the validation set by the costs of incorrect upgrades and incorrect downgrades. If the current state is L4, the upward direction switches to hold; if it is L1, the downward direction switches to hold. To prevent a single calculation from spanning multiple levels, by default, each state version migrates at most one level, and migration only occurs in subsequent periods if the conditions are still met.

[0088] The "pending confirmation" status is a migration strategy, not a new fifth trust level. When the weighted sum of both positive and negative evidence exceeds its respective threshold, or when the primary candidate hypothesis indicates a contrary migration direction, the system maintains the previous discrete level, sets `migration_action` to `PENDING`, records the conflicting evidence path, and the review validity period. New consultations or service feedback within the validity period can remove the "pending confirmation" status; if no new evidence is found by the expiration date, the status is recalculated, and "pending confirmation" is not automatically treated as an upgrade or downgrade.

[0089] Figure 9 The migration hysteresis is illustrated using 18 update cycles. The expected state value continuously exceeds the upgrade boundary and meets the independence confidence requirement from cycles 7 to 9, resulting in an upgrade in cycle 9. In cycle 10, although still in a high state range, the independence confidence decreases and conflicting evidence appears, leading to a migration result of pending confirmation. Thereafter, the state value gradually decreases, and downgrading is only performed when it continuously crosses the downgrading boundary and meets the downgrading confidence threshold from cycles 16 to 18. The gray area between the two boundaries is the hysteresis maintenance zone, which can suppress level fluctuations caused by single favorites, unfollows, or occasional inquiries.

[0090] The migration output includes previous_state, current_state, migration_action, trend_slope, candidate_direction, sustain_count, threshold_version, independent_confidence, and reason_codes. The state version consists of user_link_id, computation time, and an incrementing sequence number. Optimistic locking is used for writes: if a higher version has already been generated during the write cycle, the result of this round is moved to the recalculation queue to avoid historical chain forks caused by concurrent events.

[0091] During the feedback correction phase, a trust state package is generated based on the updated user trust state data. The state package includes user association identifiers, converged candidate hypotheses and association confidence levels, updated state data, independent confidence levels, valid evidence time windows, evidence path identifiers, state versions, and validity periods. Private domain access, consultation, or service feedback within the validity period is associated with user association identifiers and state versions and mapped to verification trust state data. A state error is generated based on the difference between the verified state and the predicted state and the independent confidence level. The source of the error is located along the candidate hypotheses, time windows, and evidence paths recorded in the state package, and the corresponding parameters are updated.

[0092] Trust state packets are transmitted using JSON or Protocol Buffers. A packet can contain user_link_id, state_version, issued_at, expire_at, a hypotheses array, state_distribution, state_label, independent_confidence, evidence_window, evidence_path_ids, migration_action, and model_versions. Elements in the hypotheses array store hypothesis_id, account_aliases, and converged_confidence; evidence_path_ids only stores controlled path numbers, and the original event text remains managed by the permission-isolated event library. The state user must return user_link_id and state_version exactly as is, not just the current account.

[0093] Private domain access feedback includes whether the redirect token has been verified, whether authorization has been completed, and whether an in-app account has been created; consultation feedback includes whether the request was initiated proactively, the number of effective dialogue rounds, whether the request was clearly stated, and whether a clear rejection was made; car rental service feedback includes results such as reservations, contracts, vehicle pickup, vehicle return, after-sales service, and complaints. The feedback mapper converts these discrete results into a verification state distribution according to versioning rules. For example, effective access and completion of multiple rounds of vehicle and price consultations can be mapped to a distribution mainly based on L3, fulfillment of the contract and follow-up visits in subsequent periods can be mapped to a distribution mainly based on L4, and a token not being opened only constitutes weak reverse evidence and is not directly mapped to L1.

[0094] Both the predicted state and the verification state are first converted into normalized state values ​​according to equation (13). To obtain a greater correction strength for high-confidence errors, the state error can be expressed as: (15) in, This represents the state error of state version t; This indicates the independent confidence level of this version; This represents the verification status value obtained from subsequent feedback mapping; This represents the original predicted state value. Low-confidence predictions, even if discrepancies exist, are used to supplement samples without excessively modifying the model; high-confidence predictions that are subsequently refuted by facts trigger the replay of key paths.

[0095] The error localization service replays four types of paths sequentially: candidate account combination path, direct behavior event path, relationship corroboration path, and migration determination path. For path j, the path is temporarily hidden or its weight is replaced with the baseline value, the state error is recalculated, and the path contribution can be normalized as follows: (16) in, This represents the normalized contribution of path j to the current error; This represents the error obtained after masking path j; This represents the set of valid evidence paths recorded in the state packet; the small value in the denominator is used to avoid division by zero when all contributions are zero. Paths with high contributions and traceable back to the original event are marked as major sources of error.

[0096] If the main contribution comes from a probabilistic account edge, the error source is marked as ACCOUNT_LINK, and the sample is sent to the calibration queue of the account association parameters; if it comes from a high-frequency browsing edge, it is marked as DIRECT_EVENT, and the adaptation weight or time decay scale for that behavior type is adjusted; if it comes from a cross-account semantically similar edge, it is marked as RELATION_SUPPORT, and the relationship strength model is updated; if the distribution of each state is reasonable but the transition is too early, it is marked as MIGRATION, and the threshold or persistence window is updated. Deterministic authorization relationships are not automatically deleted by a single business feedback, but rather a binding verification task is generated.

[0097] For differentiable model parameters, the gradient update magnitude can be adjusted according to the path contribution: (17) in, Indicates the model parameters corresponding to error source j; This represents the feedback learning rate, which is subject to an upper limit. This represents the path contribution. For rule thresholds, gradients are not used directly. Instead, after accumulating to the minimum sample size, the grid is recalibrated based on the target error migration rate. Any new parameters are first validated in the shadow version, and only after passing the validation are model_versions generated and replaced with grayscale versions.

[0098] Figure 10 The path-level update architecture is illustrated. The left-hand state package fixes the predicted state, candidate hypothesis, evidence window, and path number within the same state version; the middle private domain access, consultation, and car rental service feedback are mapped to form verification trust state data; the right-hand side calculates the state error and performs evidence path replay; the bottom updates account association parameters, direct behavior evidence processing parameters, relationship corroboration processing parameters, or state transition parameters according to the error source. The updated parameters only apply to the next model version; the old state package is still interpreted according to its original model version, ensuring reproducibility.

[0099] For example, state version v102 predicts a related entity to be L4 with an independent confidence level of 0.88. Subsequent feedback indicates that platform B account fulfilled the obligation, but platform A account explicitly denied the related request. Path replay reveals that after blocking the probabilistic association edge between A1 and B1, the error decreased from 0.53 to 0.09. This path contributed the most, so the system attributes the error to account association rather than interpreting platform A's rejection as a decrease in user trust. Conversely, if all three accounts are confirmed to be bound and the events are consistent, and only the upgrade occurred prematurely, the error is located at the migration threshold or persistence window.

[0100] At the system implementation level, the cross-platform user trust state identification and migration system includes an event handling module, a candidate association module, a dual-view construction module, a joint convergence module, and a state generation and migration module. Each module exchanges data using a unified user_link_id, hypothesis_id, graph_batch_id, and state_version, and maintains idempotency through a state repository. Figure 2 The modular structure can be deployed as five service units of the same application, and can also be scaled up or down separately.

[0101] The event handling module includes a platform adapter, a pattern validator, a quality evaluator, and an event writer. The platform adapter is responsible for decoding the raw protocol, the pattern validator checks the type and enumeration based on the field dictionary version, the quality evaluator outputs `quality_info`, and the event writer writes standard events to the partition topic and event library. The candidate association module consists of a deterministic binding reader, an association feature service, a candidate combiner, and a confidence initializer. It outputs a set of candidate hypotheses and publishes a `HYPOTHESIS_READY` message to the graph construction service.

[0102] The dual-view construction module consists of a node resolver, a direct edge generator, a relation edge generator, and a graph batch processor. The node resolver unifies the node numbers for identical entities; the direct edge generator generates edge sequences based on event time; the relation edge generator generates associated and supporting edges based on candidate hypotheses; and the graph batch processor packages multiple candidate hypotheses according to the number of nodes and outputs sparse edge indices. The joint convergence module loads a version-managed bi-branch graph processing network, maintains a round cache and a convergence criterion; it can resume from the most recent complete round after an interruption, without using half-written confidence vectors.

[0103] The state generation and transition module includes a candidate aggregator, an independent confidence calculator, a trend analyzer, a hysteresis state machine, and a state packet generator. The candidate aggregator outputs user-level distributions; the independent confidence calculator summarizes correlation concentration, variance, coverage, quality, and conflict; the trend analyzer reads historical versions to generate change slopes; the hysteresis state machine executes upgrades, downgrades, maintains, or awaits confirmation; and the state packet generator writes the expiration date and evidence path. State packets can be used for hierarchical reach or service priority reference, but external applications do not write back state tags; they only return verifiable actual events.

[0104] Explicit version numbers are used in inter-module interfaces. Backward compatibility is maintained when field expansion occurs in standard event patterns; unknown fields are ignored in older models. A new `event_schema_version` is created when behavior type semantics change to avoid mixing enumerations with the same name. The model service returns `model_version` and `feature_dict_version`, and the state repository verifies their consistency with the computation task before writing. Duplicate message delivery is deduplicated using `idempotency_key`, and out-of-order message handling is achieved using the `event_time` watermark and a late arrival window. Events exceeding the late arrival window are moved to the next state version and marked as `late_event`.

[0105] System health monitoring records at least the platform receiving latency, pattern verification failure rate, proportion of low-quality events, number of candidate hypotheses per user, number of graph nodes and edges, number of joint convergence rounds, proportion of non-convergence events, independent confidence distribution, number of migrations, and number of state oscillations. When a platform interface is interrupted, events from other platforms can still form a state distribution, but the decrease in platform coverage will reduce the independent confidence. After recovery, late events will enter a new round of calculation based on their original occurrence time, without writing back historical state packets already used by the business.

[0106] To verify the above implementation, a set of cross-platform simulation and anonymized mixed data without real identity information can be constructed. The example data covers 90 days, including 18,000 platform accounts, 320,000 behavioral events, and 2,400 sets of cross-platform associated entities confirmed by the authorization chain. L1 to L4 verification labels are generated based on subsequent private domain access, consultation, and car rental service results. The data is randomly divided into a 70% training set, a 15% validation set, and a 15% test set based on associated entities rather than events to prevent the same user's account from appearing in different sets and causing information leakage.

[0107] Five control groups were set up for verification: fixed association confidence level (meaning account association results do not accept reverse correction for state consistency); single direct view (meaning the relationship supporting branches are removed); no time gating (meaning the time gating value for all events is fixed at 1); no quality weighting (meaning the quality value of valid events is fixed at 1); and no migration hysteresis (meaning migration occurs when a single cycle exceeds the limit). The complete scheme uses the parameters in Table 2, with each setting trained with three random seeds. Table 3 lists the mean values. The F1 score for account association is calculated based on confirmed cross-platform association pairs, the F1 score for state macro-average is calculated based on four-level labels, the calibration error uses the expected calibration error of ten equal-width intervals, and the error migration rate represents the proportion of upgrades or downgrades that are opposite to the direction of the verification state to the total number of migrations.

[0108] Table 3. Exemplary simulation and ablation verification results

[0109] As shown in Table 3, when the association confidence level is fixed, the association F1, state macro F1, and calibration performance all decrease simultaneously, indicating that the inverse correction of candidate associations and state distribution helps to eliminate erroneous combinations. Retaining only the direct behavioral view is insufficient for corroborating cross-platform sparse accounts. Removing time gating or data quality weighting increases the impact of high-frequency old events, recurring events, and anomalous events. Removing hysteresis has little impact on single state identification, but the error transfer rate increases from 2.8% to 11.8%, indicating that historical trends, persistence windows, and independent confidence thresholds primarily improve the stability of state updates. The above values ​​are exemplary simulation results illustrating the verification methods; actual deployments should be retrained and calibrated using authorized data.

[0110] Further perturbation experiments were conducted: when 10% of the ordinary behavioral events were randomly removed from the test set, the average F1 score of the complete scheme's state macros decreased by no more than 0.018; when 8% of the probabilistic account edges were replaced with erroneous edges, joint convergence reduced the association confidence of approximately 71% of the erroneous edges to below 0.20; when 5% of the events were copied twice, idempotent deduplication reduced the mean L1 norm of the output state distribution change to below 0.003. For simulations of the same entity over 18 consecutive update cycles, the number of state oscillations after hysteresis decreased from an average of 3.6 times to 0.9 times. All perturbation data retained the injection markers, verifying that the program could repeatedly calculate each index.

[0111] Inference efficiency verification can be performed on an 8-core CPU and 32 GB memory environment. Taking 32 candidate hypotheses per graph batch and a maximum of 128 main window events per account as an example, the median time for a single round of graph propagation after an event encoding cache hit is approximately 41 milliseconds, and the average batch processing time is approximately 0.34 seconds when converging in 7.1 rounds. As the number of events increases, memory is kept under control by batching by hypothesis and using sparse edge indexing. If a GPU is available, graph batching and text encoding can be processed in parallel; if only a CPU is deployed, the semantic vector of the content is pre-cached without affecting the data relationships in the method.

[0112] Before the model goes live, out-of-time validation can be used: training for the first 60 days and testing for the next 30 days to check performance after content theme changes and platform behavior enumeration updates; temperature scaling is used to calibrate association confidence and state probability; statistics are performed separately by platform coverage, event sparsity, and personalized operating entities to avoid overall metrics masking single-platform failures. The state transition threshold is selected by weighting the business costs of error escalation and error degradation, and shadow runs are used to compare the state version differences between the old and new models on the same event stream.

[0113] In this implementation, the content contact dimension is formed by encoding events such as browsing duration, completion time, collection, and forwarding; the interaction dimension is formed by comments, private messages, and inquiries; the intention dimension is formed by inquiry topics such as car model, self-driving needs, and price; and the relationship dimension is formed by following, repeat visits, private domain access, and confirmed account relationships. These dimensions no longer directly output levels through a fixed total score, but instead are processed through direct behavior branches, relationship corroboration branches, joint convergence of candidate hypotheses, and independent confidence calculations. Ultimately, the four-level tags can still be read by the car rental service's tiered outreach, priority service, or exclusive benefits interface.

[0114] It should be noted that in actual implementation, batch processing or stream processing can be selected according to the platform interface conditions. Evidence propagation can be implemented using graph attention networks, graph convolutional networks, or models with equivalent neighborhood aggregation functions. Event memory can also be implemented using GRU, LSTM, or masked temporal attention layers. As long as the distinction between behavioral evidence view and relational corroboration view is maintained, the joint update of association confidence and state distribution is maintained, the independent confidence is output separately, and the state transition relationship with persistence conditions is maintained, the above data processing process can be completed.

[0115] The above are merely specific implementations of the present invention. For those skilled in the art, equivalent adjustments can be made to the data fields, model parameters, node types, time windows, and deployment methods without departing from the technical concept of the present invention; all such equivalent adjustments should be considered to fall within the protection scope of the present invention.

Claims

1. A cross-platform user trust state identification and migration method, characterized in that, include: Acquire heterogeneous cross-platform behavioral data and standardize it into behavioral event data with account identifiers; A set of candidate user association hypotheses and the association confidence of each hypothesis are generated based on account association characteristics and cross-platform behavior consistency. Each hypothesis in the set of candidate user association hypotheses represents the association relationship between cross-platform accounts belonging to the same user. Based on the behavioral event data of the accounts included in each hypothesis, a dual-view time-series correlation diagram is constructed, which includes a view of direct behavioral evidence and a view of relationship corroboration. The evidence propagation of the dual-view temporal correlation graph is controlled by the correlation confidence to obtain the trust state distribution of each hypothesis. The correlation confidence is updated according to the consistency difference between the trust state distributions, and the trust state distribution is recalculated with the updated correlation confidence until the correlation confidence and the trust state distribution converge together. The current user's trust status data and independent confidence are determined based on the converged trust status distribution and the associated confidence. The state transition strategy is determined based on historical and current user trust status data and independent confidence levels. The strategy is then executed and updated user trust status data is output.

2. The cross-platform user trust state identification and migration method according to claim 1, characterized in that, The standardization includes: mapping fields and normalizing behavior types for the cross-platform heterogeneous behavioral data; determining data quality information based on at least two of the following: field completeness, duplication, anomaly, and source platform reliability; obtaining behavioral event data containing platform identifier, account identifier, behavior object identifier, behavior type, behavior time, and the aforementioned data quality information, wherein the behavior object identifier is a content identifier or an operating entity identifier; and adjusting the propagation weight of the evidence formed by the behavioral event data in the dual-view time-series correlation graph based on the aforementioned data quality information and under the aforementioned association confidence control.

3. The cross-platform user trust state identification and migration method according to claim 1, characterized in that, The generation of the candidate user association hypothesis set and the association confidence of each hypothesis includes: determining the deterministic association relationship between cross-platform accounts based on account authorization relationship or account binding relationship; determining the probabilistic association relationship between cross-platform accounts based on at least two of behavior time continuity, content semantic similarity, common operating entity, and historical interaction sequence similarity; combining cross-platform accounts based on the deterministic and probabilistic association relationships to form multiple candidate user association hypotheses, and determining the corresponding initial association confidence based on the deterministic and probabilistic association relationships contained in each hypothesis; maintaining the deterministic association relationship unchanged during the joint convergence process, and updating the association confidence of each hypothesis based on the consistency difference between the trust state distributions.

4. The cross-platform user trust state identification and migration method according to claim 3, characterized in that, The dual-view time-series association diagram includes candidate user nodes, platform nodes, platform account nodes, content nodes, and operating entity nodes corresponding to each hypothesis, and the platform account node is connected to its respective platform node; The direct behavioral evidence view consists of behavioral event edges between platform account nodes and content nodes or operating entity nodes. Each behavioral event edge carries behavioral type, behavioral time, and data quality information, forming a temporal relationship according to the behavioral time. The data quality information represents the data quality of the corresponding behavioral event data, and the propagation weight of the behavioral event edge is determined based on the data quality information. The relationship corroboration view consists of association edges between candidate user nodes and platform account nodes, as well as behavioral corroboration edges between different platform account nodes. Each association edge is established based on the deterministic or probabilistic association relationship upon which the corresponding hypothesis is based, and is weighted by the association confidence level of the corresponding hypothesis. Each behavioral corroboration edge represents the continuity of behavioral time, semantic similarity of content, or relationship of common interaction objects.

5. The cross-platform user trust state identification and migration method according to claim 4, characterized in that, The evidence propagation is performed using a bi-branch graph processing network, including: encoding the behavior type, behavior time, and data quality information of the behavior event edges to obtain behavior event feature data; inputting the behavior event feature data into the direct behavior processing branch, performing differential transformation according to the behavior type, generating a time gating value based on the time interval between adjacent behavior events, and adjusting the propagation weight of the behavior event edges with the time gating value to obtain direct behavior features; inputting the association edge and the behavior corroboration edge into the relationship corroboration processing branch, determining the relationship strength based on the historical interaction sequence of cross-platform accounts, aggregating relationship information based on the relationship strength and the association confidence of the corresponding hypothesis to obtain relationship corroboration features; and fusing the direct behavior features and the relationship corroboration features through a state fusion unit to obtain the trust state distribution of each candidate user association hypothesis.

6. The cross-platform user trust state identification and migration method according to claim 4, characterized in that, The generation of consistency differences between the trust state distributions includes: obtaining the first state distribution and the second state distribution of each hypothesis based on the direct behavioral evidence view and the relational corroboration view, aggregating the behavioral event edges connected to each platform account node to obtain the local state distribution of the account; and generating consistency differences between the trust state distributions of each hypothesis based on the differences between the first state distribution and the second state distribution and the differences between the local state distribution of each account and the trust state distribution of the corresponding hypothesis. The process of updating the association confidence and determining joint convergence includes: obtaining the association confidence of the current iteration based on the consistency difference between the association confidence, account association features, data quality information, and trust state distribution of the previous iteration, and recalculating the trust state distribution of each hypothesis based on the association confidence of the current iteration; when the changes in association confidence and trust state distribution of adjacent iterations satisfy the corresponding preset convergence conditions, it is determined that the association confidence and trust state distribution have reached joint convergence.

7. The cross-platform user trust state identification and migration method according to claim 6, characterized in that, The determination of the current user trust state data and independent confidence scores includes: weighting and aggregating the corresponding trust state distributions based on the association confidence scores of each candidate user association hypothesis after joint convergence to obtain a user trust state distribution; determining the independent confidence scores based on the concentration of each association confidence score, the consistency difference between the trust state distributions, the coverage of effective behavioral events on the platform and time window, data quality information, and the degree of conflict between positive and negative behavioral evidence; determining the current user trust state data based on the user trust state distribution, wherein the independent confidence scores characterize the credibility of the current user trust state data and are output corresponding to the current user trust state data.

8. The cross-platform user trust state identification and migration method according to claim 7, characterized in that, The determination and execution of the state transition strategy includes: determining the state change trend based on historical user trust state data from multiple consecutive update cycles; determining candidate transition directions based on the change in current user trust state data relative to the historical user trust state data; when the candidate transition direction consistently aligns with the state change trend and the independent confidence level consistently reaches the corresponding transition threshold within a preset duration window, performing state upgrade or downgrade according to the candidate transition direction; if the state upgrade or downgrade is not performed, and if the positive behavioral evidence conflicts with the negative behavioral evidence, the state transition strategy is set to pending confirmation; otherwise, the trust state represented by the historical user trust state data is maintained; the state upgrade and state downgrade employ different transition thresholds or duration windows to create state transition hysteresis, and updated user trust state data is generated based on the execution result of the state transition strategy.

9. The cross-platform user trust state identification and migration method according to claim 1, characterized in that, Also includes: A trust status package is generated based on the updated user trust status data. The trust status package includes user association identifier, association hypotheses of each candidate user after joint convergence and corresponding association confidence, the updated user trust status data, the independent confidence, valid evidence time window, evidence path identifier, status version and validity period. Based on the user association identifier and the status version, the subsequent private domain access feedback, consultation feedback, or service feedback obtained within the validity period are associated with the trust status package, and the associated feedback is mapped to a status to obtain verified trust status data; the status error is determined based on the status difference between the verified trust status data and the updated user trust status data and the independent confidence level, and the error source of the status error is located based on the candidate user association hypothesis and the corresponding association confidence level, valid evidence time window, and evidence path identifier in the trust status package; the corresponding account association parameters, direct behavior evidence processing parameters, relationship corroboration processing parameters, or status transition parameters are updated based on the error source.

10. A cross-platform user trust state identification and migration system, characterized in that, include: The event handling module is used to acquire cross-platform heterogeneous behavioral data and standardize it into behavioral event data with account identifiers; The candidate association module is used to generate a set of candidate user association hypotheses and the association confidence of each hypothesis based on account association characteristics and cross-platform behavior consistency. Each hypothesis in the candidate user association hypothesis set represents the association relationship between cross-platform accounts belonging to the same user. The dual-view construction module is used to construct a dual-view time-series correlation diagram, which includes a view of direct behavioral evidence and a view of relationship corroboration, based on the behavioral event data of the accounts contained in each hypothesis. The joint convergence module is used to control the evidence propagation of the dual-view temporal correlation graph with the correlation confidence, obtain the trust state distribution of each hypothesis, update the correlation confidence based on the consistency difference between the trust state distributions, and recalculate the trust state distribution with the updated correlation confidence until the correlation confidence and the trust state distribution jointly converge. The state generation and transition module is used to determine the current user trust state data and independent confidence based on the converged trust state distribution and associated confidence, determine the state transition strategy based on the historical and current user trust state data and independent confidence, execute the state transition strategy and output the updated user trust state data.