A method and system for trusted association of events across data domains
By generating standardized events and reliability vectors, adaptive candidate windows, hierarchical candidate generation, and joint verification, the problem of event correlation under data domain isolation conditions is solved, achieving efficient and reliable cross-data domain event correlation, and reducing computational load and erroneous correlation.
Patent Information
- Application Number
- CN202611224815.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-17
AI Technical Summary
In two isolated data domains, events are subject to delays, missing events, duplication, out-of-order events, clock errors, network collisions, and anomaly risks, making it impossible to effectively share stable identifiers. Existing technologies struggle to establish reliable associations under these conditions, and the arrival of new events necessitates recalculating all historical events, increasing processor overhead and storage access volume.
By generating standardized events and reliability vectors, an adaptive candidate window is used to generate hierarchical candidate components. The reliability vector is used to adjust the candidate generation, probability calculation and decision-making process across stages. Only the affected conflicting connected components are updated, joint confirmation and rejection are performed, and a link number with a validity period is output.
It reduces the number of candidate combinations entering the association calculation, limits the scope of recalculation after the arrival of new events, eliminates conflicting candidate event chains, suppresses unreliable association results when there is insufficient evidence, and ensures the credibility and efficiency of the results.
Smart Images

Figure CN123114732A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer data processing technology, and in particular to a method and system for reliable association of events across data domains. Background Technology
[0002] In distributed computing systems, different events of the same technical process may originate in mutually isolated data domains. For example, the first data domain records event triggering, propagation, or arrival information, while the second data domain records installation, initial startup, session establishment, state changes, or subsequent behavior information. Due to factors such as data isolation, permission boundaries, and the inability to share identifiers, there is usually no stable user identifier or stable device identifier that can be directly linked between the two data domains.
[0003] Events in the two data domains may also use different time zones, time precisions, event type encodings, device environment classifications, and application version formats. Terminal clock skew, network transmission, queue congestion, batch reporting, and offline reporting can cause discrepancies between the event occurrence time and the reporting time, further resulting in delayed, out-of-order, and late events. Simply comparing two time points can easily eliminate true candidates or include irrelevant events in the same candidate window.
[0004] Shared networks, carrier-grade address translation, proxy networks, and frequent network switching can cause multiple terminals to exhibit the same or similar network characteristics within the same time period. Device model, system version, and device environment type may also have a high collision rate. If the consistency of network or device environment is directly regarded as strong matching evidence, erroneous associations can easily be formed in high-collision environments.
[0005] One approach is to set a fixed time window and perform combined comparisons of events from two domains within that window. However, a fixed window struggles to accommodate both low-latency and high-latency sources simultaneously. A window that is too small will miss late events, while a window that is too large will generate a large number of candidate combinations. Performing full pairwise comparisons on both event sets increases the number of candidates with the number of events, increasing processor computation, memory access, and real-time processing latency.
[0006] Another approach is to assign weights to time, network, device environment, or behavioral characteristics, and then select the candidate with the highest score. This method does not distinguish between states such as "field is true but has low discriminative power," "field is missing," "field is abnormal," and "field has expired," nor does it explain how the same quality information affects the candidate selection stage, probability calculation stage, and final decision-making stage. When multiple candidate scores are close, directly selecting the highest-scoring candidate may still result in unreliable associations.
[0007] Selecting candidates independently one by one also struggles to handle conflicts between candidate event chains. For example, the same application-side conversion event may simultaneously become the highest-scoring event in multiple candidate event chains, or the same temporary event may appear simultaneously in mutually exclusive confirmed chains. Without implementing consistency constraints at the candidate event chain set level, locally optimal results may conflict with each other.
[0008] In streaming event processing scenarios, new events, late events, or correction events continuously arrive. Recalculating the entire historical event chain each time a new event arrives involves repeatedly accessing data and constraints unrelated to the new event, increasing processor overhead and storage access volume. Using only a cache without explicitly defining the affected time buckets, candidate variables, adjacency constraints, and revision deadlines also fails to determine the boundaries of incremental updates.
[0009] Furthermore, hashing, bucketing, network prefixing, or local processing can only reduce the risk of direct identification; they do not necessarily eliminate the long-term linking capability resulting from different field combinations. If the output link number is stable over a long period, or if there is no retention period, version, or access control, the processing result may still be used for continuous tracing across tasks.
[0010] Therefore, an event association scheme is needed that can be implemented under conditions of missing stable identifiers and data domain isolation, so that quality information runs through candidate generation, probability calculation and result decision, limits the update range after the arrival of new events, uniformly resolves candidate event chain conflicts, and does not output forced association results when evidence is insufficient. Summary of the Invention
[0011] The technical problem to be solved by this invention is: when two mutually isolated data domains cannot share stable user identifiers or stable device identifiers, and events are subject to delays, missing, duplicates, out-of-order events, clock errors, network collisions, and anomaly risks, how to reduce the number of candidate combinations entering the association calculation, limit the scope of recalculation after the arrival of new events, eliminate conflicting candidate event chains, and suppress unreliable association results when there is insufficient evidence.
[0012] To address the aforementioned technical problems, this invention provides a cross-data domain event credible association method. This method first generates standardized events and reliability vectors in two isolated data domains; then, it generates candidate event chains, allowing the same reliability vector to adjust the generation of candidate event chains, evidence probability contributions, and confirmation thresholds across stages, while satisfying the monotonic constraint that a decrease in reliability does not make candidate event chains easier to confirm; when a new event, a late event, or a correction event arrives, only the affected conflicting connected component determined by the changed candidate event chain and its conflict relationships is updated, while the results outside this affected conflicting connected component remain unchanged; subsequently, a set of consistent links is determined within this component; finally, joint confirmation and rejection are performed based on the first candidate probability, candidate gap, uncertainty, conflict state, and risk gating.
[0013] The first and second data domains perform format standardization, time interval correction, field generalization, quality marking, and temporary numbering on the original events within their respective domains, generating standardized events that are allowed to enter the associated computation without requiring the exchange of names, mobile phone numbers, complete network addresses, or stable device identifiers.
[0014] For each standardized event and its comparable fields, a reliability vector is generated. The reliability vector can include at least source reliability, field completeness, data freshness, time reliability, network evidence reliability, device / environment distinguishability, and risk credibility. Each reliability component can be normalized to the range of 0-1.
[0015] Anomaly risks are first converted into risk confidence levels, which can be expressed as r_risk = 1 - normalize(Risk). Risk represents the anomaly risk before normalization, and normalize represents mapping the anomaly risk to the range of 0-1. The higher the anomaly risk, the lower r_risk. For events marked as critical risks, independent risk gating can also be set to prevent candidate event chains from directly entering the confirmation state.
[0016] An adaptive candidate window is generated based on the historical reported delay quantiles, historical delay fluctuations, and event time uncertainty for the corresponding source and event type. The adaptive candidate window can be represented as W = clamp(Qp + k1 × σd + k3 × Ut, Wmin, Wmax). Here, Qp is the historical reported delay quantile, σd is the historical delay fluctuation, Ut is the event time uncertainty, k1 and k3 are adjustment coefficients, Wmin and Wmax are the lower and upper limits of the window, respectively, and clamp indicates that the calculation results are restricted between the lower and upper limits of the window.
[0017] Network collision density is not included in the calculation of the adaptive candidate window, but is used to reverse-correct the reliability of network evidence. Network evidence reliability can be expressed as r_net = r_net_base × (1 - normalize(Cn)). Here, Cn is the network collision density, and r_net_base is the network evidence reliability before correction. Higher network collision density results in a lower probability contribution of network evidence, and correspondingly increases the independent evidence requirement or confirmation threshold.
[0018] Based on standardized events, adaptive candidate windows, and reliability vectors, hierarchical candidate generation is performed in order of increasing cost. The first layer filters based on task context, target object, and platform compatibility; the second layer filters based on corrected time intervals and adaptive candidate windows; the third layer excludes combinations that violate causal order based on event sequence; and the fourth layer forms soft evidence based on network type, device environment type, and version compatibility. For each event to be associated, no more than a maximum of K candidate sources are retained.
[0019] Incompatible task contexts, inconsistent target objects, and event order are unlikely to be hard constraints, and candidates that meet the hard constraint exclusion criteria can be directly deleted. Time distance, network type, device environment, version, and behavioral similarity are considered soft evidence. When a soft evidence field is missing, the candidate is not directly deleted, nor is the missing field recorded as having zero probability.
[0020] The candidate generation phase outputs candidate event chain records. These records may include the candidate event chain number, participating events and their order, reliability of each field, satisfied constraints, unsatisfied constraints, missing field markers, and intermediate features required for probability calculation.
[0021] Establish a reliability-aware probabilistic event chain for candidate event chains. The probabilistic event chain includes first data domain event nodes, second data domain event nodes, candidate correlation variables, time constraints, network collision correction constraints, device environment consistency constraints, context constraints, behavior order constraints, source reliability constraints, and missing field masks.
[0022] For a retained candidate event chain c corresponding to the same target event, the logarithmic score of the candidate event chain is represented as z_c = log(P0,c) + Σm mask_m,c × r_m,c × log(max(φ_m,c, ε)), and the probability of the candidate event chain is represented as p_c = exp(z_c) / Σj exp(z_j). Here, P0,c is the prior probability of the candidate event chain, mask_m,c is the missing field mask, r_m,c is the reliability of the m-th type of evidence, φ_m,c is the probability constraint of the m-th type of evidence, ε is a positive number to prevent undefined logarithms, and the summation index j iterates through all retained candidate event chains for the same target event.
[0023] The time probability constraint φ_time,c uses a Gaussian decay function of candidate delay relative to expected delay; the network probability constraint φ_net,c is assigned a value based on the generalized network category consistency, inconsistency, or missing state, where network collisions are only adjusted by the exponent through r_net; the device environment probability constraint φ_device,c uses a weighted consistency rate of system category, device category, language category, and hardware category; the behavior order probability constraint φ_order,c uses an exponential decay function of soft order deviation, and candidates that violate hard causal order are deleted during the candidate stage.
[0024] When a field is missing, the missing field mask causes the corresponding probability constraint to take a neutral value and increases the uncertainty of the candidate event chain. This distinguishes between "evidence does not support" and "no evidence," avoiding the direct setting of possible true candidates to zero due to missing fields.
[0025] The uncertainty of the candidate event chain is expressed as U_c = α × H_norm + β × M_c / q + γ × B_c / Wmax + δ × Var(R_c). Here, H_norm is the normalized entropy of the probability distribution of the candidate events for the same target event, which is 0 when there is only one candidate; M_c / q is the proportion of missing evidence; B_c is the total width of the correction time interval for each event in the candidate event chain; Var(R_c) is the dispersion of the reliability components of the candidate event chain; α, β, γ, and δ are all not less than 0 and their sum is 1.
[0026] The monotonic linkage of reliability vectors across stages refers to the simultaneous application of the same reliability vector across three stages: adjusting candidate generation intensity, candidate retention quantity, or independent evidence requirements in the candidate stage; adjusting the probabilistic contribution of various types of evidence in the probability stage; and adjusting the confirmation threshold in the decision stage. The confirmation threshold includes the confirmation probability threshold, candidate gap threshold, uncertainty cap, and risk gating. Each adjustment relationship satisfies monotonic constraints: a decrease in reliability can expand the candidate retention range, but this must be accompanied by an increase in independent evidence requirements, a higher confirmation probability threshold or candidate gap threshold, and a lower allowable uncertainty cap; a decrease in reliability does not make the candidate event chain easier to confirm.
[0027] When a new, late, or corrected event arrives, the hierarchical index is used to determine the affected time bucket. The candidate event chain added, deleted, or changed by this event is used as a seed node. The search proceeds along shared source events, shared target events, shared mutually exclusive temporary events, and other conflicting edges to obtain the affected conflicting connected components. Only the candidate related variables, probabilities, uncertainties, and conflict states within the affected conflicting connected components are updated, while the results outside the affected conflicting connected components remain unchanged.
[0028] The system sets a revision period for confirmed links. Late events within the revision period can trigger a partial recalculation and generate a new version; late events exceeding the revision period do not automatically rewrite confirmed links, but are written to the audit log or enter the manual review queue.
[0029] Perform consensus solving on the affected conflicting connected components. Assign a binary variable x_c to the candidate event chain c and determine the weight w_c based on the candidate event chain probability, average reliability, and uncertainty. Solve for max Σc w_c × x_c. Constraints include at least the variables corresponding to the same target event and a weight not greater than 1; candidate event chains sharing mutually exclusive temporary events cannot be selected simultaneously; and any pair of conflicting candidate event chains c and d satisfies x_c + x_d ≤ 1. This yields the set of consensus links with the maximum weight.
[0030] For the first candidate in the consistent link set, joint confirmation and rejection are performed based on the candidate probability, the candidate gap between the first and second candidates, the uncertainty of the candidate probability distribution, unresolved conflicts, and risk gating. A confirmation status is only output when the first candidate probability is not lower than the confirmation threshold, the candidate gap is not lower than the gap threshold, the uncertainty is not higher than the uncertainty threshold, there are no unresolved conflicts, and the risk gating is not triggered.
[0031] Candidate event chains that do not meet the confirmation criteria will be output as either pending, rejected, or manually reviewed, depending on the reason. A pending status indicates that partial recalculation is allowed within the revision period, triggered by a late event; a rejected status indicates insufficient evidence and no confirmation chain will be generated; and a manually reviewed status indicates the existence of anomalies, critical risks, or conflicts that cannot be automatically resolved.
[0032] The output uses randomly generated, rotated link numbers with expiration dates. Confirmed results may include event sequences, candidate probabilities or confidence levels, key supporting evidence, missing or degraded evidence, uncertainty, conflict resolution results, generation time, parameter versions, and retention periods. Unconfirmed results may include status, primary reason, whether recalculation is allowed, and the latest recalculation time.
[0033] The output does not include names, mobile phone numbers, complete network addresses, stable device identifiers, or link numbers that can be tracked across tasks for a long time. Long-term linkability of event-related results is limited through number validity periods, field generalization, access control, result retention periods, and deletion or rotation upon expiration.
[0034] Historical latency distribution, reliability parameters, probabilistic calibration relationships, and decision thresholds are determined through offline parameter calibration using confirmed matching samples, confirmed non-matching samples, and manually reviewed samples with clearly defined and authorized sources. Samples are grouped by source, source event type, and target event type, and further divided into parameter search sets, probabilistic calibration sets, and validation sets. Finite grid search or coordinate search is performed on k1, k3, and K. Under the condition that the true candidate recall rate is not lower than a preset lower limit, the parameter combination with the smaller target value is selected using J = λ1 × true candidate omissions + λ2 × average candidate count + λ3 × P95 processing time, where λ1, λ2, and λ3 are non-negative cost coefficients. The probabilistic calibration set is used to determine the probabilistic mapping, and the independent validation set is used to select T1, T2, and T3. During operation, changes in latency distribution, network collisions, field missing values, and review results are monitored; when changes exceed a set range, a recalibration task is generated, and the formal parameters are replaced after verification.
[0035] By using adaptive candidate windows, hierarchical indexes, hard constraints, and upper limits on the number of candidates, the number of candidate combinations entering the probability calculation stage can be reduced, avoiding unrestricted full combination comparisons between two sets of events.
[0036] By using the reliability vector to monotonically link across stages, the same reliability vector can simultaneously adjust the candidate generation strength, probabilistic evidence contribution, and confirmation threshold, and enforce the monotonous relationship that "a decrease in reliability does not make the candidate event chain easier to confirm." This can prevent low-discrimination or high-risk evidence from being interpreted inconsistently at multiple stages.
[0037] By using a missing field mask to separate "field missing" from "evidence not supported," uncertainty can be increased while retaining potential candidates, reducing erroneous exclusion and forced confirmation due to missing data.
[0038] By searching for the affected conflict-connected component along the conflict edges from the changed candidate event chain, and only updating and solving that component, the processing scope after a new event arrives can be related to the size of that component, rather than to the total number of historical events, thereby reducing redundant computations and irrelevant storage accesses.
[0039] By selecting a consistent set of links through conflict constraints, we can prevent the same target event or temporary event from being simultaneously confirmed by multiple mutually exclusive links. Through joint confirmation and rejection, we can prevent forced association results from being output when there is insufficient evidence, close candidates, high uncertainty, unresolved conflicts, or abnormal risks.
[0040] By using temporary link numbers with expiration dates, field generalization, access control, and retention periods, auditable event chains can be output without exchanging direct identity fields, and the ability of results to be linked across tasks for extended periods can be restricted. Attached Figure Description
[0041] Figure 1 This is a schematic diagram of the overall architecture of the cross-data domain event trusted association system according to an embodiment of the present invention.
[0042] Figure 2 This is a flowchart of the cross-data domain event trusted association method according to an embodiment of the present invention.
[0043] Figure 3 This is a schematic diagram of the cross-stage monotonic linkage of the reliability vector in an embodiment of the present invention.
[0044] Figure 4 This is a schematic diagram of the local update of the affected conflicting connected components according to an embodiment of the present invention.
[0045] Figure 5 This is a flowchart illustrating the conflict control, joint confirmation, and rejection process according to an embodiment of the present invention.
[0046] Figure 6 This is a schematic diagram illustrating the result output, auditing, and retention period control in an embodiment of the present invention.
[0047] Figure 7 This is a flowchart of the offline parameter calibration and drift monitoring process according to an embodiment of the present invention.
[0048] Figure 8 This is a schematic diagram illustrating the affected conflicting connected components and constraint solutions in an embodiment of the present invention. Detailed Implementation
[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention to specific data sources, service names, parameter values, or deployment forms. Those skilled in the art can make corresponding adjustments to the event types, index structures, probability calculation methods, and deployment locations without departing from the technical concept defined in the claims.
[0050] like Figure 1 As shown, the cross-data domain event trusted correlation system includes a first data domain 100, a first local processing module 110, a second data domain 200, a second local processing module 210, a trusted correlation engine 300, and a result output module 400. The trusted correlation engine 300 may include a candidate generation module 310, a probability chain module 320, a local update module 330, a conflict control and joint confirmation and rejection module 340, and an offline parameter calibration and drift monitoring module 350.
[0051] The first data domain 100 and the second data domain 200 are isolated from each other in terms of permissions, storage, or network. The first local processing module 110 and the second local processing module 210 process the raw events in their respective data domains, providing only permitted standardized events, intermediate representations, and quality tags to the trusted association engine 300. The trusted association engine 300 can be deployed in an authorized federated computing environment or hosted by a controlled service on one side.
[0052] Events in the first data domain 100 may include trigger, exposure, click, share, transfer, call, or request events. Events in the second data domain 200 may include installation, first launch, session establishment, state change, response, or in-application behavior events. The event names above are for illustrative purposes only; in actual implementation, other event types that can express causal relationships may be used.
[0053] Each local processing module sequentially performs event type encoding standardization, time zone standardization, time precision standardization, platform and version format standardization, time interval correction, network and device field generalization, quality marking, and temporary number generation.
[0054] For event e, record the event occurrence time t_event and the reporting time t_report. Based on the source clock error, acquisition accuracy, and known queuing delay, generate a correction time interval I_e = [t_lower, t_upper]. When the terminal clock is uncalibrable, do not generate a pseudo-precise time point; instead, expand the time interval and increase the time uncertainty Ut.
[0055] The complete network address is locally converted into a network prefix, network category, or collision statistics for the corresponding time period. Device model, system version, language, and hardware environment are converted into categories with appropriate granularity. The original name, mobile phone number, stable device identifier, and complete network address are not included in the output results.
[0056] Each event uses a randomly generated temporary event number with an expiration date. The temporary event number is only valid within the specified task, version, and retention period and cannot be used as a long-term tracking identifier across tasks.
[0057] like Figure 3 As shown, the reliability vector 301 can be denoted as R_e=(r_source,r_complete,r_fresh,r_time,r_net,r_device,r_risk). Here, r_source represents source reliability, r_complete represents field completeness, r_fresh represents data freshness, r_time represents time reliability, r_net represents network evidence reliability, r_device represents device environment discrimination capability, and r_risk represents risk credibility.
[0058] Source reliability can be determined based on the integrity of the data collection chain, signature verification, duplication rate, and historical quality of the source. Field completeness can be determined based on the existence of the fields required for this association. Data freshness can be determined based on the interval between the reporting time and the processing time. Time reliability is negatively correlated with the width of the correction time interval.
[0059] The reliability of network evidence is first determined by r_net_base based on field validity and source quality, and then r_net = r_net_base × (1 - normalize(Cn)) is calculated based on the network collision density Cn for the current time period. When multiple events share the same network category, normalize(Cn) increases, and the reliability of network evidence decreases.
[0060] The ability to distinguish between different device environments can be determined based on the frequency of occurrence of a device category in the current candidate set. The more frequently a device environment category appears, the lower its ability to distinguish candidates independently. When the device environment field is missing, its reliability is not interpreted as reverse evidence.
[0061] Risk credibility is calculated as r_risk = 1 - normalize(Risk). Risk can be determined by considering factors such as abnormal source, recurring events, signs of forgery, abnormal frequency, and rule conflicts. For events that meet the preset key risk conditions, gate_risk = 0 is set to prevent the relevant candidate event chain from entering the confirmation state.
[0062] The specific method of cross-stage monotonically linked reliability vector is as follows: Reliability vector 301 is output to candidate generation intensity 311, evidence probability contribution 312, and confirmation threshold 313. The smaller value, geometric mean, or pre-calibrated conservative composite value of the corresponding reliability components of the events at both ends of the candidate event chain is taken to form the candidate event chain reliability R_c, and its mean of each available component is denoted as ρ_c. In the candidate stage, K can be increased or the minimum number of independent evidences can be increased as ρ_c decreases; in the probability stage, each r_m,c is used as the exponent of the corresponding probability constraint; in the decision stage, T1,c=T1,base+a1×(1-ρ_c), T2,c=T2,base+a2×(1-ρ_c), and T3,c=T3,base-a3×(1-ρ_c) can be used, where a1,a2, and a3 are not less than 0. Therefore, when reliability decreases, the confirmation probability threshold and candidate gap threshold do not decrease, and the upper limit of allowable uncertainty does not increase; even if the candidate range is expanded, the candidate event chain cannot be made more easily confirmed. Risk gating is independent of the aforementioned thresholds and has a priority blocking effect.
[0063] For source s, source event type a, and target event type b, the delay distribution is statistically reported from the authorized historical samples. A preset quantile Qp and delay standard deviation or robust volatility σd are taken, and combined with the current event time uncertainty Ut, W = clamp(Qp + k1 × σd + k3 × Ut, Wmin, Wmax).
[0064] When the source delay or time uncertainty increases, the candidate window W expands within its upper and lower limits; when the delay is stable and the time is reliable, the candidate window narrows. Network collision density does not participate in the adaptive candidate window, and Cn is not added to W to avoid further increasing the number of irrelevant candidates due to the shared network environment.
[0065] Parameters k1, k3, Wmin, and Wmax are configured according to the source and event type. A set of unrestricted ranges for initiating the offline search is as follows: k1 ranges from 0.5 to 3.0 with a step size of 0.25; k3 ranges from 0.5 to 2.0 with a step size of 0.25; K is an integer from 2 to 20; the latency quantile p is selected from 0.90, 0.95, and 0.99; Wmin is not less than the sum of the event acquisition accuracy and the minimum explainable latency; and Wmax is not greater than the business-allowed revision period. The above ranges are only used to generate a finite set of candidate parameters; the official values are determined by... Figure 7 The offline parameter calibration process shown is selected and the version is recorded with each candidate event chain result.
[0066] like Figure 2 As shown in S101 to S104, the system first obtains the standardized events and reliability vectors of the two data domains, then calculates the adaptive candidate window, and performs hierarchical candidate generation.
[0067] The candidate generation module 310 establishes a task context index, a time bucket index, an event sequence index, and an environment category index. The first layer filters based on task context, target object, and platform compatibility. The second layer determines whether two corrected time intervals can form an acceptable chronological relationship after being added to the candidate window. The third layer excludes combinations that violate causal order, such as the target event occurring earlier than all possible source events. The fourth layer calculates soft evidence such as network category, device environment category, version, and behavioral similarity.
[0068] For each target event, the candidate generation module 310 sorts the candidates according to the satisfaction of hard constraints, time compatibility, and available independent evidence, retaining no more than K candidate sources, where K is the upper limit of the number of candidates. K can be configured according to load and source quality. When reliability is low, more candidates can be retained for judgment, but this also increases the independent evidence requirement and the final confirmation threshold.
[0069] The candidate event chain record includes chain_id, event_ids, event_order, reliability_vector, satisfied_constraints, missing_evidence and features. chain_id is the link number with a limited validity period. The candidate record does not contain original direct identity fields.
[0070] The probability chain module 320 establishes event nodes, candidate association variables and probability constraints for the candidate event chain. For each candidate event chain c of the same target event, calculate z_c = log(P0,c) + Σm mask_m,c × r_m,c × log(max(φ_m,c, ε)), and then calculate p_c = exp(z_c) / Σj exp(z_j). Here, ε can take 10^-6, and j traverses all candidate event chains retained for the same target event. To avoid exponential overflow, in implementation, the maximum z value in the candidate set can be subtracted from each z_c before exponentiation; this transformation does not change the normalization result.
[0071] A specific form of the temporal probability constraint is φ_time,c = exp(-(d_c-μ_ab)^2 / (2×s_c^2)). Where d_c is the difference between the center of the target event time interval and the center of the source event time interval, μ_ab is the expected delay from source event type a to target event type b, s_c=max(s_min,σ_ab+h_A+h_B), σ_ab is the robust delay fluctuation of this type combination, h_A and h_B are respectively the half-width of the corrected time interval of the two events, and s_min is a positive minimum scale.
[0072] A specific form of the network probability constraint is: when two available generalized network categories are consistent, φ_net,c=q_same, when they are inconsistent, φ_net,c=q_diff, and when any network field is missing, mask_net,c=0; where 0<q_diff<q_same≤1, and the initial calibration candidate values can be 0.20 and 0.90 respectively. The network collision density Cn does not change the above category consistency judgment, nor does it expand the candidate window, but only reduces the exponent of the network constraint according to r_net = r_net_base × (1 - normalize(Cn)), and increases the independent evidence requirement or confirmation threshold through ρ_c.
[0073] One specific form of the device environment probability constraint is φ_device,c = max(ε,Σl ω_l×I(g_A,l=g_B,l)). Here, l iterates through the system category, device category, language category, and hardware category; ω_l is not less than 0 and sums to 1 after renormalization for available categories; and I is an indicator function that takes 1 when consistent and 0 when inconsistent. If a category is missing in any event, it does not participate in that weighting, and the corresponding missing value is included in M_c.
[0074] One specific form of the behavioral order probability constraint is φ_order,c = exp(-λo×v_c), where λo is greater than 0 and v_c is the soft order deviation number. If the target event occurs earlier than all possible times of the source event, or violates the pre-registered hard causal order, it is deleted during the candidate generation stage and not included in the probability calculation. Source evidence can be labeled with a probability φ_source,c from 0 to 1 based on the integrity of the acquisition chain and signature verification.
[0075] When the missing field mask mask_m,c is 1, it indicates that the m-th type of evidence is computable; when it is 0, it indicates that the field is missing. When mask_m,c is 0, the corresponding item in the log score is 0, which is equivalent to the neutral value of 1 in the probability product. The missing amount is still included in the uncertainty, thus distinguishing between "no evidence" and "evidence does not support it".
[0076] Uncertainty is calculated as U_c = α × H_norm + β × M_c / q + γ × B_c / Wmax + δ × Var(R_c). Where H_norm = -Σj p_j × log(p_j) / log(n), n is the number of candidates for the same target event (H_norm is 0 when n=1); M_c is the number of missing evidences, q is the preset total number of evidences; B_c is the sum of the correction time interval widths of each event in the candidate event chain; Var(R_c) is the variance of the reliability components involved in the calculation; α, β, γ, and δ are not less than 0 and their sum is 1. All terms are in the range of 0 to 1 or truncated to this range; the higher the uncertainty, the more difficult it is to confirm.
[0077] like Figure 4 and Figure 8 As shown, the local update module 330 maintains the time bucket 410, candidate related variables 420, adjacency constraints 430, and candidate event chain states. Each candidate event chain is used as a conflict graph node; when two candidate event chains share a source event, a target event, or a mutually exclusive temporary event, or cannot simultaneously satisfy hard causal order or other registered mutually exclusive resource constraints, a conflict edge is set between them.
[0078] When a new event arrives, the affected time bucket 410 is first located based on the event time interval and candidate window. Then, new candidates, candidates to be deleted, or candidates whose score input has changed due to the new event are generated, and these candidates are used as a seed set. A breadth-first search or depth-first search is performed on the seed set, and adjacent candidates are repeatedly added along the conflict edges until there are no unvisited adjacent candidates. The resulting set of nodes is the affected conflicting connected component 440.
[0079] Figure 4 In this context, A1, A2, and C1 all represent standardized events in the first data domain, and B1 represents the target standardized event in the second data domain. C1 is a newly added source event that is standardized late, labeled as late standardized event 415. A1, A2, and C1 are all candidates for the source event of B1. The candidate event chain variable 420 includes candidate event chains c1=(A1,B1), c2=(A2,B1), and c3=(C1,B1), which are in... Figure 4 The three candidate event chains are labeled 421, 422, and 423 respectively; they share the target event B1 and therefore have conflict edges set between them. The newly generated candidate event chain c3, generated by C1, serves as the seed node, and late events themselves are not directly used as nodes in the conflict graph.
[0080] The local update module 330 only adds, deletes, or modifies candidate variables and adjacency constraints in the affected conflicting connected components 440, renormalizes the candidate probabilities of the shared target event, recalculates the uncertainty, and performs conflict resolution. Historical chains 450 not connected to the components, along with their versions, probabilities, and selection states, remain unchanged. Since there must be conflicting edges between candidates of the shared target event, the candidates affected by probability normalization will not be missed outside the components.
[0081] If a new event is a late event, the system checks its occurrence time, arrival time, and the revision period for the confirmed link. A new version is allowed to be generated within the revision period, while the old version, the reason for the change, and the triggering event are retained; outside the revision period, only audit logs or manual review tasks are generated.
[0082] like Figure 5 and Figure 8 As shown, the conflict control and joint confirmation and rejection module 340 receives a set 510 of candidate event chains within the affected conflict connectivity component and establishes conflict constraints 520. For each candidate event chain c, x_c∈{0,1} is set, where x_c=1 indicates selection of the chain. The weights of the candidate event chains can be calculated as w_c=max(0,ηp×p_c+ηr×mean(R_c)-ηu×U_c), where ηp, ηr, and ηu are non-negative coefficients, and mean(R_c) is the mean of the available reliability components.
[0083] The objective of the constraint solution is to max Σc w_c × x_c. For a candidate set C(b) of the same objective event, we set Σc∈C(b) x_c≤1; for a candidate set C(e) of shared mutually exclusive temporary events, we set Σc∈C(e) x_c≤1; for any pair of conflicting candidate event chains c and d, we set x_c + x_d ≤ 1. Candidates that violate the hard causal order are deleted before entering the solution process.
[0084] A directly implementable branch and bound solution involves the following steps: First, arrange the candidate event chains within the component from highest to lowest according to w_c; second, establish a search node containing the candidate position, the selected set, the current weight, and the upper bound corresponding to the sum of the remaining non-negative weights; third, prune if the upper bound is not greater than the current optimal weight; fourth, try the two branches "select without violating constraints" and "do not select" for the current candidate; fifth, update the optimal set with the current feasible set when the end of the arrangement is reached; sixth, output the optimal set 530 when the search stack is empty. This algorithm will inevitably terminate with a finite set of candidates. The Hungarian algorithm can only be used when the constraints degenerate into a one-to-one bipartite graph between the source and target events and there are no additional conflicting edges; otherwise, the aforementioned 0-1 constraint solution is still used.
[0085] For each target event in the consensus link set, obtain the first candidate probability p1, the second candidate probability p2, the candidate gap Δ = p1 - p2, and the uncertainty U. The joint confirmation and rejection conditions include p1 ≥ T1, Δ ≥ T2, U ≤ T3, no unresolved conflicts, and gate_risk not being 0.
[0086] Thresholds T1, T2, and T3 are probabilistically calibrated using authorized samples, and the threshold versions are recorded. If overall reliability decreases, T1 or T2 can be increased, the upper limit of permissible uncertainty can be decreased, or additional independent evidence can be required. Threshold adjustments do not alter the deterrent effect of critical risk gating.
[0087] Output "Confirmation 551" when all confirmation conditions are met; output "Pending 552" when evidence may be supplemented by a late event within the revision period; output "Reject Association 553" when evidence is insufficient and automatic recalculation is not allowed; output "Manual Review 554" when there are key risks, sources of anomalies, or conflicts that cannot be automatically resolved.
[0088] like Figure 6 As shown, the result output module 400 generates the link result 610, evidence description 620, version and generation time 630, retention period control 640, and audit log 650. When the retention period is reached or the task termination condition is met, the expired deletion or rotation 660 is executed.
[0089] Link Result 610 uses a random link number with an expiration date. Evidence Description 620 records the main supporting evidence, missing evidence, factors that reduce reliability, conflict resolution results, and reasons for rejection. Version and Generation Time 630 is used to reproduce the parameters, thresholds, and candidate states at that time.
[0090] Access control restricts the ability to read the subject and purpose of the link results. Retention period control (640) sets the retention periods for candidate records, confirmed results, pending results, and audit records. Records exceeding their retention periods are deleted, or given a new temporary number and their cross-version linking is terminated.
[0091] The following data are simulated computations used to demonstrate the feasibility of the computation process and do not represent real customer data, production test results, or performance commitments. Assume that domain A has source events A1 and A2, and domain B has a target event B1. The task contexts and target objects of all three are compatible and do not violate hard causality order. With parameters Qp=30 seconds, σd=8 seconds, Ut=4 seconds, k1=1.5, k3=1, Wmin=10 seconds, and Wmax=60 seconds, the uncrunted window is 30 + 1.5 × 8 + 1 × 4 = 46 seconds, therefore W = 46 seconds. The maximum number of candidates K=2, and the hierarchical index retains candidate event chains c1=A1-B1 and c2=A2-B1.
[0092] Assume the prior P0 of both candidates is 0.5, and the available evidence includes time, network, device environment, behavioral sequence, and source. For c1, set φ_time=0.78, φ_net=0.90, φ_device=0.95, φ_order=0.95, φ_source=0.90; for c2, set φ_time=0.88, φ_net=0.90, φ_device=0.20, φ_order=0.95, φ_source=0.90. The corresponding reliability for both candidates is set to r_time=0.85, r_net=0.25, r_device=0.80, r_order=0.90, r_source=0.90. A lower r_net indicates a higher collision rate in the current network category; this collision density does not participate in W, but only weakens the network evidence index and raises the confirmation threshold.
[0093] According to the probability formula, the unnormalized score of c1 is 0.5×0.78^0.85×0.90^0.25×0.95^0.80×0.95^0.90×0.90^0.90=0.3287; the unnormalized score of c2 is 0.5×0.88^0.85×0.90^0.25×0.20^0.80×0.95^0.90×0.90^0.90=0.1047. After normalization, p1=0.758, p2=0.242, and the candidate difference Δ=p1-p2=0.517.
[0094] The uncertainty coefficients are taken as α=0.4, β=0.3, γ=0.2, and δ=0.1. The normalized probability entropy of the two candidates is H_norm=0.798, the proportion of missing evidence in c1 is M_c / q=0, the proportion of total width of the time interval is B_c / Wmax=46 / 60=0.767, and the reliability discrete term Var(R_c)=0.05. Therefore, U=0.4×0.798+0.3×0+0.2×0.767+0.1×0.05=0.477.
[0095] Suppose that T1=0.70, T2=0.20, and T3=0.50 are selected by the independent validation set, and the risk gating corresponding to c1 is not triggered. c1 and c2 share the target event B1, thus constraining x_c1+x_c2≤1; the maximum weight consistent link set selects c1. At this time, p1=0.758 is not lower than T1, Δ=0.517 is not lower than T2, and U=0.477 is not higher than T3. There are no unresolved conflicts and the risk gating is not triggered, so the system outputs c1 as a confirmed link.
[0096] In contrast, if the correction event indicates that the device context field of c1 is missing, and the device context probability constraint of c2 is corrected to 0.90, then the c1 device item will not participate in the scoring due to mask_device=0. The unnormalized score of c1 is 0.3424, and the unnormalized score of c2 is 0.3487. After normalization, the probabilities of the two candidates are approximately 0.495 and 0.505, and the candidate gap is reduced to 0.009. Since the missing proportion becomes 1 / 5, the normalized entropy is approximately 1.000, and U is approximately 0.618 using the same coefficient. At this time, Δ is lower than T2 and U is higher than T3, so the system does not force the selection of c2, which is ranked first, and outputs "pending". The two candidates caused by the new evidence and their conflicting edges constitute the affected conflicting connected components, and the local update does not change the historical links of other target events.
[0097] In shared networks or carrier-grade address translation scenarios, increasing network collision density and reducing network evidence contribution requires more independent evidence in terms of time, device environment, context, or behavioral sequence. For proxy networks or frequent network switching, network evidence should be flagged as anomalous, but inconsistencies in a single network should not directly exclude other candidates with sufficient evidence.
[0098] In cases of device clock skew, use the corrected time interval and increase Ut. In cases of reporting delays or out-of-order delivery, allow partial recalculation within the revision period and record the version. In cases of missing fields, enable missing field masking and increase uncertainty.
[0099] For repeated installations or reinstallations, deduplication or the establishment of new event cycles can be performed based on application instance, session, version, and time interval. For multiple users on the same device or multiple devices for the same user, the real user identity is not restored; if the evidence is insufficient, the association is rejected.
[0100] For duplicate events, deduplication is performed based on source, temporary number, time range, and event content, and the reliability of the abnormal source is reduced. For forged events or abnormal traffic, r_risk is reduced or critical risk gating is triggered, causing the corresponding candidates to undergo manual review instead of direct confirmation.
[0101] like Figure 7 As shown, the offline parameter calibration and drift monitoring module 350 uses confirmed matching samples, confirmed non-matching samples, and manually verified samples 710, all of which are from clearly identified and authorized sources. The samples are first grouped by source, source event type, and target event type 720; each group is then divided into a parameter search set, a probability calibration set, and an independent verification set according to time order or a random seed, to ensure that the same event does not occur repeatedly across sets.
[0102] For each true matching sample on the parameter search set, the reporting delay is calculated, and the candidate quantile Qp is used. The robust delay fluctuation σd can be 1.4826 multiplied by the absolute median difference of the delay, or the standard deviation after removing pre-registered outliers. Subsequently, a grid search or coordinate search is performed on a finite candidate set 730 with k1 ranging from 0.5 to 3.0, k3 ranging from 0.5 to 2.0, and K ranging from 2 to 20. For each combination, a candidate event chain is generated, and J = λ1 × true candidate omissions + λ2 × average candidate count + λ3 × P95 processing time is calculated. The combination with the smaller J is selected from those with a true candidate recall rate not lower than the preset lower limit R_min; if no combination satisfies R_min, the candidate range of Wmax is expanded first and the search is repeated, without directly lowering the confirmation threshold.
[0103] On the probabilistic calibration set 740, a finite search is performed on candidate function parameters for time, network, device environment, behavioral sequence, and source evidence, respectively, with the smaller negative logarithmic loss or Brier loss as the selection criterion, and ε, q_same, q_diff, λo, and each reliability mapping is fixed. On the independent validation set 750, T1, T2, and T3 are enumerated, first satisfying the preset upper limit of false confirmation rate and manual review capacity, and then selecting the combination with higher true match recall. The parameter search set must not be used again to report validation results.
[0104] The calibration output 760 includes the sample range and authorization record summary, Qp, σd, k1, k3, Wmin, Wmax, K, probability function parameters, α to δ, T1 to T3, reliability mapping version, random seed or time split point, effective time, and applicable source. During operation, latency distribution, network collision rate, field missing rate, and verification results are monitored according to a fixed period or sample size threshold 770; a recalibration task is generated when the preset variation range is exceeded. The new version is released after passing the same threshold on an independent validation set or controlled data 780, while the old version and its applicable results are retained in the audit log.
[0105] In one deployment, the two local processing modules send the allowed field only to the same controlled trusted association service. In another deployment, secure multi-party computation, private set intersection, or a trusted execution environment can be used to complete partial candidate intersection or probabilistic constraint calculations. The above secure computation methods are optional implementations and do not change the core relationships of cross-stage monotonic linkage of reliability vectors, local updates of affected conflicting connected components, conflict control, and joint confirmation and rejection.
[0106] This invention is not limited to the association between advertising events and application events, but can also be used in cross-data domain association scenarios such as device telemetry and server alarms, edge node events and central processing events, and call chain events and state change events, as long as the adopted solution contains the technical features and their interaction relationships as defined in the claims.
Claims
1. A method for reliable association of events across data domains, characterized in that, include: The original events are processed in two isolated data domains to generate standardized events and corresponding reliability vectors. Based on the standardized events, candidate event chains are generated, and the same reliability vector is used to adjust the generation, evidence probability contribution, and confirmation threshold of the candidate event chains across stages. Each adjustment relationship satisfies the monotonic constraint that a decrease in reliability does not make the candidate event chains easier to confirm. In response to the arrival of new, late, or corrective events, starting from the added, deleted, or changed candidate event chain, the affected conflicting connected component is determined along the conflict relationship between the candidate event chains. Only the state of the candidate event chain within the affected conflicting connected component is updated, while the results outside the affected conflicting connected component remain unchanged. Within the affected conflict connectivity component, a set of consistent links is determined based on the conflict relationship between candidate event chains. For the first candidate in the set of consistent links, joint confirmation and rejection are performed based on the first candidate probability, the candidate gap between the first candidate and the second candidate, uncertainty, conflict state, and risk gating. If all confirmation conditions are met, a confirmation state is output; otherwise, a pending state, a rejection association state, or a manual review state is output.
2. The method according to claim 1, characterized in that, The reliability vector includes source reliability, field completeness, data freshness, time reliability, network evidence reliability, device environment distinguishability, and risk credibility. The reliability of the candidate event chain is formed by taking the smaller value, geometric mean, or pre-calibrated conservative composite value of the reliability components of the events at both ends of the candidate event chain. The risk credibility is determined according to r_risk = 1 - normalize(Risk), where Risk represents abnormal risk, and the risk gating is triggered by the key risk marker.
3. The method according to claim 2, characterized in that, During the candidate phase, the upper limit of the number of candidates or the minimum number of independent evidences is adjusted based on the reliability of the candidate event chain; when the reliability of the candidate event chain decreases, the candidate retention range is expanded and the minimum number of independent evidences is increased, so that expanding the candidate retention range does not reduce the evidence requirements required for the candidate event chain to enter the confirmation state.
4. The method according to claim 3, characterized in that, In the probability phase, a missing field mask mask_m,c and a reliability of the m-th type of evidence r_m,c are set for the candidate event chain c; the log score of the candidate event chain c satisfies z_c = log(P0,c)+ Σm mask_m,c × r_m,c × log(max(φ_m,c, ε)), and the probability of the candidate event chain satisfies p_c = exp(z_c) / Σj exp(z_j); when the m-th type of evidence field is missing, mask_m,c is 0.
5. The method according to claim 4, characterized in that, The dynamic confirmation threshold for the decision-making stage is determined by the mean ρ_c of the available components of the reliability of the candidate event chain; the confirmation probability threshold satisfies T1,c=T1,base+a1×(1-ρ_c), the candidate gap threshold satisfies T2,c=T2,base+a2×(1-ρ_c), and the upper limit of uncertainty satisfies T3,c=T3,base-a3×(1-ρ_c), where a1, a2, and a3 are all not less than 0; when the reliability of the candidate event chain decreases, the confirmation probability threshold and the candidate gap threshold do not decrease, and the upper limit of uncertainty does not increase.
6. The method according to claim 1, characterized in that, The adaptive candidate window is determined according to W = clamp(Qp + k1× σd + k3 × Ut, Wmin, Wmax); Where Qp is the historical reporting delay quantile, σd is the robust historical delay fluctuation, and Ut is the event time uncertainty; the network collision density Cn does not participate in the calculation of the adaptive candidate window, and the reliability of network evidence is corrected in reverse according to r_net = r_net_base × (1 - normalize(Cn)).
7. The method according to claim 6, characterized in that, Hierarchical indexing is performed based on the compatibility of task context and target object, corrected time interval, causal order of events, and compatibility of network category, device environment category, and version. Task context incompatibility, target object inconsistency, and impossible event order are set as hard constraints, while time distance, network category, device environment category, version, and behavioral similarity are set as soft evidence. For each event to be associated, no more than the upper limit K of candidate sources is retained.
8. The method according to claim 1, characterized in that, Determining the affected conflicting connected components includes: using hierarchical indexing to locate time buckets that new events may affect; using candidate event chains that are added, deleted, or changed and are related to the new event, late event, or correction event within the time bucket as seed nodes; searching along shared source events, shared target events, shared mutually exclusive temporary events, or other conflicting edges to obtain the affected conflicting connected components; recalculating only the probability, uncertainty, and conflict state of candidate event chains within the affected conflicting connected components, while keeping the results outside the affected conflicting connected components unchanged.
9. The method according to claim 8, characterized in that, Set a binary selection variable x_c for the candidate event chain c within the affected conflicting connected component; with the goal of maximizing the candidate event chain weights and Σc w_c×x_c, set a variable and a constraint of no more than 1 for the candidate event chains of the same target event, and set x_c+x_d≤1 for any conflicting candidate event chains c and d; use branch and bound search to obtain the set of consistent links.
10. The method according to claim 1, characterized in that, T1 is the confirmation threshold, T2 is the candidate gap threshold, and T3 is the uncertainty threshold. The first candidate is output as the confirmation state only when p1 is not lower than T1, Δ is not lower than T2, U is not higher than T3, there are no unresolved conflicts, and the risk gating is not triggered. The uncertainty satisfies U_c = α × H_norm + β × M_c / q + γ × B_c / Wmax + δ × Var(R_c), where α, β, γ, and δ are all not less than 0 and their sum is 1. Otherwise, the pending state, the rejection state, or the manual review state is output according to the unmet conditions.
11. The method according to claim 10, characterized in that, Offline parameter calibration is performed using confirmed matching samples, confirmed non-matching samples, and manually verified samples with clearly identified and authorized sources; samples are grouped by source, source event type, and target event type, and divided into parameter search set, probability calibration set, and independent verification set; Under the condition that the actual candidate recall rate is not lower than the preset lower limit, the parameter combination is selected by J = λ1 × actual candidate omissions + λ2 × average candidate count + λ3 × P95 processing time, and a new parameter version is generated based on the running drift.
12. The method according to claim 1, characterized in that, A revision period is set for the event chain results of the confirmed status; a late event triggers the recalculation of the affected conflicting connected components and generates a new version of the link within the revision period, and enters the audit record or manual review queue when the revision period is exceeded; the event chain results use the validity-limited link number bound to the associated task and parameter version, and a retention period is set for the candidate record, the confirmed result and the audit record respectively.
13. A cross-data domain event trusted association system, characterized in that, include: The event handling module is used to generate standardized events and reliability vectors in mutually isolated data domains; The trusted association module is used to generate candidate event chains and adjust the generation of candidate event chains, evidence probability contribution, and confirmation threshold across stages using the same reliability vector, and the decrease in reliability does not make the candidate event chains easier to confirm; the local update module is used to determine the affected conflicting connected components using the changed candidate event chains caused by new events, late events, or correction events as seeds, and only updates the state within the component while keeping the results outside the component unchanged. The conflict control module is used to determine the set of consistent links based on the conflict relationships within the component; the joint confirmation and rejection module is used to perform joint confirmation and rejection based on the probability, candidate gap, uncertainty, conflict state and risk gating of the first candidate in the set of consistent links. When all confirmation conditions are met, the confirmation status is output; otherwise, the pending status, rejection status or manual review status is output.
14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 12.
15. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 12.
