A cross-source data credibility dynamic evaluation method

Abstract

The application discloses a kind of cross-source data credibility dynamic evaluation method, it is related to data processing technical field.Data stream is obtained and field data is extracted to construct topological network, initial state data is configured;Capture its time series variance, generate time series credibility inertia data representing noise threshold in combination with historical data;When the abnormality of source node to be evaluated generates initial abnormal disturbance data and injects downstream propagation chain;Control disturbance conduction, scale subtraction comparison is carried out with time series inertia data at current edge, if it is not over threshold, then it is zero truncated, if it is over, then it is reduced to generate reconstruction state data, and it is updated into edge attribute to be used as next cycle calculation input;Finally, the reconstruction or initial state data of untruncated edge is aggregated to generate each path score and is packaged as evaluation result, according to which low-score link source end is executed weight reduction, fuse or recollects.The application realizes the adaptive blocking of abnormal disturbance and fine closed-loop routing control.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method for dynamic evaluation of cross-source data credibility. Background Technology

[0002] Big data governance and lineage analysis technologies represent important development directions in the fields of data element flow and trusted computing. They aim to achieve dynamic quality control and risk assessment of heterogeneous data by establishing complex data association networks and dependencies. In recent years, with advancements in distributed graph databases and data tracing algorithms, data credibility measurement schemes have gradually evolved towards a more end-to-end and dynamic approach.

[0003] Existing technologies typically configure static credibility weights on data sources or key field nodes, and combine these with topological relationships in the lineage graph for conflict detection and risk propagation to downstream levels, thereby achieving quality control and credibility fusion for downstream business tags. These methods can effectively intercept obvious errors at the source in conventional data cleaning and mapping scenarios, and can dynamically adjust the inherent weights of nodes based on their historical performance, demonstrating a certain degree of data governance adaptability.

[0004] However, when facing complex scenarios of cross-source propagation of heterogeneous data, the same data node often cascades along different lineage paths to multiple heterogeneous downstream business tags (such as simultaneously propagating to "loan amount tag" and "enterprise growth tag"). Because different paths use different intermediate operators and the historical communication frequency and business feedback verification stability of each path are naturally unevenly distributed, the "absorption capacity" and "temporal evolution pattern" of different paths for source-end abnormal noise fluctuate dramatically. Existing technologies fail to adjust the evaluation mechanism to address the inconsistent noise resistance rigidity of the same field propagating on different paths. When a source node experiences sudden, brief outages, update delays, or local conflicts, existing technologies can only indiscriminately propagate the risk to all downstream paths through that node. This can easily lead to the simultaneous and uniform downgrading of previously historically stable and robust downstream business tags, or the failure to promptly cut off risks in vulnerable, cold-start paths, increasing the risk of distorted data association evaluation and excessive diffusion of the impact of anomalies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a dynamic evaluation method for cross-source data credibility.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: This invention discloses a method for dynamic evaluation of cross-source data credibility, comprising the following steps: Acquire cross-source data streams and extract field data, establish a topology network containing nodes and directed edges according to a preset mapping relationship, and configure initial propagation state data containing historical propagation counts and historical conflict rates for each directed edge; The time series variance of the initial propagation state data is captured, and combined with the historical propagation count and historical conflict rate, damping correlation calculation is performed according to the preset damping function to generate time-series reliable inertial data that characterizes the noise immunity threshold of the directed edge. When the node residual of the source node to be evaluated exceeds the preset dynamic baseline, initial abnormal disturbance data is generated based on the amount and duration of the residual excess, and injected into all downstream propagation chains starting from the source node to be evaluated. The initial abnormal disturbance data is controlled to propagate step by step along each downstream propagation chain. At any current directed edge, the corresponding time-series reliable inertial data is retrieved, and the current-level disturbance input value is compared with the time-series reliable inertial data by subtraction under the unified reliability scale to generate penetrating disturbance data. If the penetration disturbance data does not exceed the preset threshold, the value is set to zero to cut off the propagation chain; If the initial propagation state data is reduced by using the penetration perturbation data, reconstructed state data is generated and written into the edge attribute of the corresponding directed edge to replace the original data. The reconstructed state data is then used as the input for calculating the time-series reliable inertial data in the next cycle. The reconstructed state data or the initial propagation state data that has not been downgraded for each uncut directed edge on each downstream propagation chain are aggregated across the entire chain. Multiple terminal path score data are generated for the same source node to be evaluated, and they are associated and encapsulated into an adaptive impact assessment result. Downweighting, circuit breaking, or re-collection are performed on the data source corresponding to the downstream propagation chain whose score is lower than the preset score threshold.

[0007] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention changes the limitation of traditional technology that regards credibility only as a static attribute of a node or a one-way linear probability. By dynamically fusing historical disturbance frequency and discrete variance, each directed edge in the network topology can form an adaptive noise barrier against occasional noise, thereby effectively filtering and absorbing background noise and weak interference in the early stage of data flow transmission.

[0008] 2. This invention transforms the abnormal transmission process into an adaptive adversarial evolution of energy dissipation, avoiding the drawbacks of traditional technologies where source-end abnormalities lead to the simultaneous degradation of all downstream business chains or the unlimited cascading spread of abnormalities. It achieves precise control over the scope of abnormal impact in the vertical dimension of the graph lineage.

[0009] 3. This invention enables the evaluation results to directly drive the physical control of the data link, completing the technical closed loop from "state awareness" to "proactive self-healing intervention". While ensuring the global robustness of the cross-source data flow graph, it improves the dynamic adaptive adjustment capability of complex networks to heterogeneous abnormal flows, and enhances the robustness and data security of the IoT platform layer data flow system. Attached Figure Description

[0010] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is a flowchart illustrating the working principle of the present invention; Figure 2 This is a timing flowchart of the present invention. Detailed Implementation

[0011] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0012] In existing technologies, the credibility verification of cross-source data streams often relies on static rule configuration or isolated index inversion of a single node, making it difficult to balance the real-time performance and accuracy of graph-level lineage propagation. When sudden anomalies or sporadic noise fluctuations occur at the source node, traditional methods can lead to disordered cascading propagation of abnormal features in the downstream dense topology network, increasing the evaluation error of unrelated downstream service nodes. Existing evaluation systems cannot synchronously perceive the impact of differences in the noise robustness of the topology propagation channel itself on the diffusion of disturbances. Especially under the pressure of concurrent data conflicts from complex heterogeneous data sources, single-attribute inversion evaluation models can exhibit systematic biases, failing to meet the requirements for fully automated, precise routing and control intervention.

[0013] To address the aforementioned issues, this study discovered a correlation between the historical conflict fluctuation variance of directed edges in the topological network and the spatial dissipation of anomalous disturbance energy. Error compensation was achieved by establishing a subtraction comparison model between temporal reliable inertia and anomalous disturbances. Further investigation revealed that directed edges with high-frequency propagation and low conflict characteristics are highly sensitive to the absorption of disturbance energy. While the reconstructed state data exhibits large single-cycle fluctuations, it demonstrates good boundary convergence stability for subsequent cycles. Therefore, a strategy of dynamically truncating or reducing the reconstruction propagation chain based on a penetration threshold was proposed. Further experimental verification demonstrated that the reconstructed state data is written into the edge attributes and inversely introduced into the mapping relationship of the temporal inertia calculation for the next cycle, thus incorporating a dynamic adjustment mechanism for model parameters and forming a closed-loop feedback system.

[0014] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] Example: like Figure 1 As shown, a cross-source data credibility dynamic assessment method falls under the category of data credibility management, platform robustness enhancement, and security protocol optimization technologies in the IoT security layer. It can be deployed in data platforms, edge computing nodes, or cloud data governance systems, serving key digital technology application scenarios such as industrial internet data fusion and blockchain distributed ledger trusted verification. The method includes the following steps: Acquire cross-source data streams and extract field data, establish a topology network containing nodes and directed edges according to a preset mapping relationship, and configure initial propagation state data containing historical propagation counts and historical conflict rates for each directed edge; The time series variance of the initial propagation state data is captured, and combined with the historical propagation count and historical conflict rate, damping correlation calculation is performed based on the preset damping function to generate time-series reliable inertial data that characterizes the noise immunity threshold of the directed edge. When the residual of the source node to be evaluated exceeds the preset dynamic baseline, initial abnormal disturbance data is generated based on the amount and duration of the residual excess, and injected into all downstream propagation chains starting from the source node to be evaluated. The system traverses all source nodes in the topology network with a fixed time window period, calculates the real-time residual of each node, and marks the node as the source node to be evaluated when the residual of a node exceeds its dynamic baseline. The initial abnormal disturbance data is propagated step by step along each downstream propagation chain. At any current directed edge, the corresponding time-series reliable inertial data is retrieved. Under a unified reliability scale, the current-level disturbance input value is subtracted from the time-series reliable inertial data for comparison, generating penetrating disturbance data. If the penetration disturbance data does not exceed the preset threshold, the value is set to zero to cut off the propagation chain; If the number of data exceeds the limit, the initial propagation state data is downgraded and reduced using the penetration perturbation data to generate reconstructed state data. This reconstructed state data is then written into the edge attributes of the corresponding directed edge to replace the original data. The reconstructed state data is then used as the input for calculating the time-series reliable inertial data in the next cycle. The system performs full-link aggregation on the reconstructed state data or the initial propagation state data that has not been downgraded for each uncut directed edge on each downstream propagation chain. It generates multiple terminal path score data for the same source node to be evaluated, and encapsulates them into an adaptive impact assessment result. It also performs weight reduction, circuit breaking, or re-collection on the data source corresponding to the downstream propagation chain whose score is lower than the preset score threshold, thereby achieving closed-loop control at the data security level and proactive defense at the platform security level.

[0016] like Figure 2 The diagram shown is a timing flowchart of the present invention. The working principle of this application is as follows: the data acquisition module continuously acquires cross-source data streams and extracts field data from them, registering each field entity as a node according to a preset mapping relationship. Register the lineage dependencies between fields as directed edges. Thus, a directed topology network is constructed. The graph computation engine configures initial propagation state data for each directed edge, which includes at least the number of historical propagations. Conflict rate with history Among them, the number of historical transmissions The historical conflict rate represents the cumulative number of times the directed edge is invoked and propagated through the bloodline path within a historical period. This characterizes the percentage of downstream business feedback verification failures in the historical propagation of this directed edge, with a value range of [value missing]. .

[0017] This application further proposes a credibility recovery mechanism to avoid the irreversible decay of directed edge credibility due to the monotonic accumulation of reconstruction degradation over multiple periods. When a directed edge is not subjected to a penetration perturbation in the current period (i.e., its source node is not marked as a source node to be evaluated, or the perturbation input value injected into the edge is zero), the graph computing engine performs exponential relaxation recovery on the current cell credibility of that edge, causing it to revert to the credibility baseline value recalculated based on the new observation data of the current period. Convergence, the restorative expression is: ; in, The current cell confidence level of the edge in the τth period (which may be the reconstructed value obtained from the downgrade of the previous period); The credibility baseline value for the τ+1th period is the result of a three-dimensional reweighted solution of spatial multi-source cross-consistency, temporal update decay freshness, and business feedback closed-loop correctness. To restore the relaxation coefficient, the value range is (0,1), and a preferred value is... =0.3. Thus, the directed edges that have been perturbed and downgraded will revert to their true three-dimensional credibility baseline exponentially after the threat is eliminated. Together with the downgrade reduction, this constitutes a two-way adjustment of "downgrading when overloaded and restoring when undisturbed", ensuring that credibility converges rather than monotonically declines under multi-cycle evolution.

[0018] This application further proposes the following specific implementation rules for establishing the topology network according to the preset mapping relationship: (1) Field-Node Mapping Rules: Based on their business semantic logic, the physical fields of each data source in the cross-source data stream are constructed using three factors: field name, data type, and enumeration value distribution range. Locality-sensitive hashing (such as MinHash / SimHash) is used to generate field fingerprints that retain similarity. The Hamming similarity or Jaccard similarity between fingerprints is compared. If the similarity exceeds a preset threshold (preferably 95%), the corresponding fields are merged into the same logical node v. If the similarity is below the threshold, they are registered as different logical nodes, thus completing the construction of the node set V. For non-enumerable continuous values ​​or free text fields, their value range statistics (quantiles, mean, and variance) are used instead of the enumeration value distribution to participate in fingerprint construction.

[0019] (2) Directed edge-line determination rule: For any two logical nodes By parsing the Structured Query Language (SQL) lineage logs or ETL (Extract-Transform-Load) job configurations during the data flow process, if there is a source... to If the direct assignment, function computation dependency, or conditional aggregation dependency is found, then a direct lineage relationship is determined between the two, and a path is established accordingly. point to directed edges It is assigned to edge set E.

[0020] (3) Cross-source heterogeneous field disambiguation rules: For cross-source fields with different names but the same business meaning, semantic alignment is performed by using a preset data dictionary mapping table or domain knowledge graph. The aligned fields are uniformly pointed to the same logical node to ensure that cross-source data achieves uniqueness merging and disambiguation in the same topology network.

[0021] (4) Topological acyclic constraint rule: After completing node merging and edge construction, the graph computing engine performs directed acyclic verification on the edge set E (such as topological sorting or Tarjan strongly connected component detection); if a loop is detected, the directed edge causing back-pointing is disconnected according to the timestamp generated by the fields in the lineage log or the preset lineage priority, so that the final topological network G=(V,E) constitutes a directed acyclic graph, thereby ensuring that the downstream propagation chain is a directed path of finite length.

[0022] Specifically, the graph computing engine retrieves the initial propagation state data time series of the directed edge within a preset time window and calculates its discrete fluctuation variance. In conjunction with historical dissemination frequency Conflict rate with history (i.e., cumulative feedback conflict rate) Input a preset damping function to perform damping correlation calculation, and generate time-series reliable inertial data characterizing the noise immunity threshold of the directed edge. The design principle of the damping function is as follows: the greater the historical conflict rate and the time series variance, the more unstable the historical propagation of the directed edge is, and the weaker its ability to absorb occasional disturbances. Therefore, the reliable inertial data of the time series should decay accordingly. The greater the number of historical propagations, the more sufficient the statistical confidence of the edge is, and the more stable its inertia tends to be.

[0023] In one embodiment, the damping correlation calculation consists of two parts: One is the exponential original damping value based on historical conflict rate and time series variance. ; Secondly, it is based on the number of historical transmissions. Nonlinear exponential convergence factor It is used to characterize the accumulation of statistical confidence with the number of propagations.

[0024] Time-series reliable inertial data is obtained by combining the two: ; in, The current cell confidence level of the directed edge is used as the baseline amplitude before damping. Historical conflict rate; The time series variance of the initial propagation state data; and These are the conflict rate damping coefficient and the variance damping coefficient, respectively, used to adjust the relative contributions of conflict and fluctuation to inertial attenuation. Preferably, they are taken as follows: , It can be achieved through a limited number of trials. , In-interval calibration; The preferred convergence metric constant is denoted as . =20; To preset the lower bound threshold of confidence, the preferred value is... 0.05; exp(·) is the natural exponential function. ∈[0,1) Monotonically increasing and converging to 1, making When statistical confidence is sufficient, the damping value approaches the original value; when statistical confidence is insufficient, it is suppressed to the lower bound. Thus, the number of historical propagations, the historical conflict rate, and the time series variance all directly participate in the generation of reliable time-series inertial data.

[0025] When the source node to be evaluated When an anomaly is detected, the graph computation engine generates initial anomaly disturbance data based on the excess amount and duration of its node residuals. This initial anomalous disturbance data is then injected into all downstream propagation chains originating from the source node to be evaluated. Subsequently, the initial anomalous disturbance data is controlled to propagate sequentially along each downstream propagation chain: at any current directed edge... At this point, retrieve the time-series reliable inertial data corresponding to that edge. And under a unified confidence scale, the current level perturbation input value By subtracting and comparing with time-series reliable inertial data, penetration perturbation data is generated. The calculation formula is as follows: ; in, For the first Penetrating perturbation data with directed edges at each level; To be transmitted to the first The input perturbation data for a directed edge of order is for the direct downstream edge of the source node. ; For the first The temporal reliability inertial data of directed edges. The physical meaning of subtraction comparison is: the incident disturbance energy is rigidly canceled by the noise resistance threshold of the directed edge itself, and only when the incident disturbance exceeds the order of magnitude that the edge inertia can absorb, can the remaining energy penetrate to the downstream.

[0026] The graph computing engine will penetrate perturbation data and preset thresholds. The comparison is processed in a branched manner. If the penetration disturbance data does not exceed the preset threshold, then... This will set the penetrating disturbance data to zero. This is a "proactive immunity" mechanism at the data security level, selectively blocking abnormal disturbances along their propagation path, that is, making... This is to cut off the propagation chain, preventing the disturbance from being transmitted downstream; if the penetrating disturbance data exceeds a preset threshold, i.e. If the directed edge is determined to be overloaded by the disturbance, the initial propagation state data of the edge is downgraded and reduced using the penetrating disturbance data to generate reconstructed state data. In one embodiment, the downgrade reduction is performed on the current cell's credibility according to the following rules: ; in, and These are the current unit's credibility and historical conflict rate after reconstruction, respectively. Reduce the strength coefficient to improve credibility (preferably by taking) ); The conflict rate recharge coefficient (preferred value) ); This is used to constrain the conflict rate within a unified scale upper bound. The graph computing engine encapsulates the reconstructed current cell confidence level and historical conflict rate, along with the remaining fields of the original quintuple, into reconstructed state data. This reconstructed state data is then written into the edge attributes of the corresponding directed edge to replace the original initial propagation state data. This reconstructed state data is used as input for recalculating the temporal confidence inertia data of that edge in the next cycle, thus forming a cross-cycle state self-updating closed loop. Preset threshold. Used to determine whether a disturbance constitutes a valid penetration, its value is typically configured as follows: In this embodiment, .

[0027] It should be noted that the replacement of edge attributes in the reconstructed state data and the three-dimensional weighted calculation of the current unit's credibility are triggered at different times: the former is triggered by the perturbation overload within the current cycle, which will downgrade the data. The edge attributes are written as initial values ​​for the next cycle's damped correlation calculation. The latter is triggered when a new round of cross-source data flows in and completes the re-collection of spatial consistency, timeliness, and business feedback accuracy, performing a full recalculation of the current unit's credibility. When both occur sequentially in the same cycle, the three-dimensional recalculation result is used as the standard, superimposed with the decay memory of the previous cycle's reconstruction degradation. That is, the initial value of the next cycle is a weighted average or the best of the three-dimensional recalculation value and the previous cycle's reconstruction value, which can be determined by the system configuration. This preserves the history of disturbances while reflecting the latest observations, ensuring unambiguous closed-loop updates.

[0028] After each downstream propagation chain completes its cascading propagation, the graph computing engine performs full-link aggregation on the reconstructed state data corresponding to the un-truncated directed edges in each downstream propagation chain, or the initial propagation state data that has not been downgraded. Multiple terminal path score data are generated for the same source node to be evaluated, and these score data, along with the anomaly source, damaged path, and risk impact range, are encapsulated into an adaptive impact assessment result. Finally, the routing control module determines the outcome based on the terminal path score data being below a preset score threshold. The data source corresponding to the downstream propagation chain implements closed-loop control: In one embodiment, when the score is slightly below the threshold, a weight reduction is implemented, that is, the weight of the data source is lowered in subsequent credibility fusion; when the score is significantly below the threshold and persists for multiple periods, a circuit breaker is implemented, that is, the transmission of the data source to downstream business tags is temporarily cut off; when the score is below the threshold and is determined by tracing to be a distortion at the collection layer, re-collection is implemented, that is, the data collection module is triggered to re-fetch and verify the data source. A preset scoring threshold is used. Normally configured as In this embodiment, .

[0029] By constructing time-series reliable inertial data for each directed edge that adaptively decays with historical conflicts and fluctuation variance, and by rigidly canceling incident disturbances step by step through subtraction comparison, each directed edge in the network topology can form an adaptive noise barrier against occasional noise. This allows for effective filtering and absorption of background noise and weak interference in the early stages of data flow transmission, thereby preventing the disorderly cascading diffusion of source anomalies in the downstream dense topology.

[0030] This application further proposes that the initial propagation state data be organized in the form of edge attribute quintuples of a directed graph. Specifically, the graph computation engine performs calculations for each directed edge. Maintain a quintuple edge attribute The 5-tuple includes: source node identifier Target node identifier Historical transmission count Cumulative feedback conflict rate (i.e., historical conflict rate) and current unit credibility The source node identifier and the target node identifier jointly and uniquely determine the position of the directed edge in the network topology; the cumulative feedback collision rate... That is, the historical conflict rate used in the aforementioned damping correlation calculation. Current unit credibility This is used to characterize the basic reliability of a single propagation of the directed edge, and participates in the damping correlation calculation as a component field for generating the initial propagation state data; that is, it serves as the reference amplitude before damping in the original damping value formula. Participate in the solution.

[0031] Preferably, the quintuple is stored in the graph computing engine in the form of an adjacency list or an edge index hash structure, which supports attribute retrieval and in-situ replacement writing with O(1) complexity by edge primary key, thus solving the storage bottleneck that traditional static mapping tables of lineage relationships cannot support dynamic reliability evolution.

[0032] This application uses a structured quintuple to uniformly organize the identifier, statistics, and credibility fields of each directed edge, so that all the inputs required for time-series reliable inertial calculation can be obtained directly by indexing edge attributes, and supports in-situ replacement of edge attributes by reconstructed state data, providing a standardized and efficiently addressable data base for subsequent step-by-step subtraction comparison and cross-cycle state self-update.

[0033] This application further proposes steps for obtaining the credibility of the current unit. Specifically, the graph computing engine obtains the spatial dimension multi-source cross-consistency of directed edges. Time dimension updates decay freshness and the accuracy of the business feedback loop The three dimensions mentioned above are then linearly weighted and summed according to preset weights to generate the credibility of the current unit. The calculation formula is as follows: ; in, The current unit's confidence level is determined by a value falling within the uniform confidence scale. Inside; The spatial dimension multi-source cross consistency is calculated by the consistency ratio of the values ​​of the field carried by the directed edge on multiple parallel data sources, which represents the lateral verification strength of the spatial dimension. The decaying freshness is updated over time and is obtained by mapping the time delay since the last valid update of this field through exponential decay. ,in To update latency, The preset freshness time constant characterizes the timeliness and reliability of the data; The accuracy of the business feedback loop is calculated by the success rate of the business feedback verification after the historical data is transmitted to the downstream business tags, which represents the empirical accuracy of the business loop. , , These are the linear weights for the three dimensions mentioned above, and their sum is normalized to... Preferably, take , , The relative weights of each dimension can be adjusted through a limited number of trials based on the characteristics of the data source.

[0034] Through the above technical solution, this application jointly characterizes the reliability of directed edges by three mutually orthogonal dimensions: spatial lateral verification, time-dependent decay, and business closed-loop empirical verification. These dimensions are then normalized and linearly weighted and fused into a single scalar current unit credibility. This not only ensures the multidimensional causality of credibility measurement but also ensures that its specification falls within a unified credibility scale, providing a physically interpretable benchmark amplitude input for damping correlation calculation.

[0035] This application further proposes a step for capturing the time series variance of the initial propagation state data. Specifically, the graph computing engine retrieves the historical data of the directed edge within a preset time window. Next to Current unit credibility First, calculate its time-series statistical mean, then calculate the variance of each historical value relative to this mean, which is used as the time-series variance. The calculation formula is as follows: ; in, For the directed edge within the preset time window, the first... The current cell reliability obtained from the next solution; For the above The time-series statistical mean of the current unit's credibility; The time series variance is used to quantitatively characterize the degree of discrete fluctuation in the credibility of the directed edge over time. The number of historical samples within a preset time window, which is greater than The integer, preferably taken as In this embodiment, The larger the variance of the time series, the more unstable the historical transmission of the directed edge is, and the more significant the decay of its original damping value.

[0036] It should be noted that this is used to calculate the variance of time series. The current unit credibility sequence is taken from the baseline value recalculated in each period using spatial / temporal / business three-dimensional weighting. The reconstructed values ​​generated due to temporary degradation caused by disturbance overload are not included. This avoids the positive feedback formed by artificially increasing variance in a single abnormal event and further suppressing the reliable inertial data of the time series through the damping function, ensuring that the noise resistance threshold only reflects the inherent historical stability of that edge.

[0037] Through the above technical solution, this application quantifies the historical fluctuation rigidity of directed edges by the discrete variance of the current unit's credibility relative to its time series mean within the sliding window. This makes the time series credibility inertia data depend not only on the instantaneous amplitude of credibility, but also on its historical stability. As a result, historically stable edges obtain higher inertia, while historically jittery edges obtain lower inertia, thus achieving the distinguishable suppression of occasional noise and systemic instability.

[0038] This application further proposes that the step of generating reliable time-series inertial data also includes introducing a cold-start correction factor. Specifically, for directed edges with a small number of historical propagation times (such as newly established kinship relationships or cold-start paths), the statistical confidence of their historical conflict rate and time series variance is insufficient. If the original damping value is directly used as the inertia, it is easy to generate an artificially high noise immunity threshold due to small sample statistical noise, causing disturbances that should be truncated to be incorrectly absorbed. To address this, the graph computing engine uses the number of historical propagation times to solve the nonlinear exponential convergence function and generate a cold-start correction factor. The calculation formula is as follows: ; in, This is a cold start correction factor, with a value range of [value range missing]. With the number of times it has been spread throughout history Monotonically increasing and converging to ; The preset convergence metric constant (preferably taken as...) The graph calculation engine then utilizes a cold-start correction factor to adjust the original damping value calculated based on historical conflict rate and time series variance. Perform step scaling correction to address situations where the number of historical propagations is lower than a preset threshold. At that time, the time-series reliable inertial data will be forced to converge to a preset lower bound threshold of reliability under a unified reliability scale. Ultimately, reliable time-series inertial data is generated. The calculation formula is as follows: in, This is time-series reliable inertial data; To preset the lower bound threshold of confidence (preferably...) ); Set a preset number of times (preferably) The step part of the piecewise function in the above equation is composed of... and The comparison implementation, the scaling part is determined by the cold start correction factor. Implementation: When the number of historical propagations is less than At that time, the time-series reliable inertial data is forced to be compressed to a lower bound. This causes the cold-start edge to absorb almost no disturbances and tend to propagate risks downstream until sufficient statistics are available; when the number of historical propagations increases sufficiently... The time-series reliable inertial data smoothly converges to the original damping value. .

[0039] This application applies a step scaling correction to the original damping value through a nonlinear exponential convergence function driven by the number of historical propagations. This suppresses the risk of artificially high noise immunity thresholds caused by small sample statistical noise on cold-start edges, enabling time-series reliable inertial data to converge to a safe lower bound when statistical confidence is insufficient. This ensures that abnormal disturbances are not erroneously absorbed on newly created or low-frequency paths, and improves the robustness of the evaluation mechanism in scenarios of dynamic topology expansion.

[0040] This application further proposes the calculation steps for the node residuals of the source node to be evaluated and the preset dynamic baseline. Specifically, the graph computing engine uses a sliding time window to statistically analyze the historical average of the field data flowing through the source node to be evaluated. Compared with current observations The absolute deviation between the two is calculated as the nodal residual. The calculation formula is as follows: ; in, The node residuals of the source node to be evaluated; This is the current observation value; Within the sliding time window The mean of historical observations; The length of the sliding window (preferably) The graph calculation engine further retrieves the historical standard deviation of the absolute deviation within the sliding time window. And multiply the historical standard deviation of fluctuation by a preset gain factor. Generate dynamic baselines The calculation formula is as follows: ; in, For dynamic baseline; The historical standard deviation of the residuals at nodes within the window; This represents the mean of the residuals at the nodes within the window. This is a preset gain factor used to set the tolerance factor of the dynamic baseline relative to the residual fluctuation amplitude, preferably taken as... That is, using a statistical criterion of approximately three standard deviations, it is possible to... Calibration is performed within a finite number of trials within the interval. If and only if... An anomaly is detected in the source node to be evaluated. The graph calculation engine determines the excess of the node residual relative to the dynamic baseline. And the duration for which the node residuals consistently exceed the dynamic baseline. The data are input together into a preset normalized time series mapping function to generate initial anomalous perturbation data mapped to a unified confidence scale. The calculation formula is as follows: ; in, The initial anomalous disturbance data, after being compressed using the above normalized time-series mapping function, consistently falls within the unified confidence scale. Within the range; This is the normalized magnitude term of the residual excess relative to the dynamic baseline; The duration is relative to the reference duration. The normalized duration term; and These are magnitude weights and duration weights (preferred). , ); Preset reference duration (preferred) The more the residuals exceed the upper bound and the longer they persist, the closer the initial anomalous disturbance data will be to the upper bound. .

[0041] By constructing an adaptive dynamic baseline by multiplying the standard deviation of residual fluctuation within a sliding window by a gain coefficient, this application avoids misjudgment of fixed thresholds under different source node statistical characteristics. Furthermore, by simultaneously introducing a normalized time-series mapping of residual excess and duration, the initial abnormal disturbance data takes into account both the intensity and persistence of the anomaly, and strictly falls within a unified confidence scale, providing a dimensionally consistent disturbance injection source for subsequent step-by-step subtraction comparisons.

[0042] This application further proposes a step for controlling the propagation of initial anomalous disturbance data along each downstream propagation chain. Specifically, the graph computing engine compares the current-level directed edges to generate penetrating disturbance data. As input, multiply by a preset graph topology channel attenuation coefficient. This generates input perturbation data that is propagated to the next level's adjacent directed edge. The calculation formula is as follows: ; in, To be transmitted to the first Input perturbation data adjacent to the directed edge at the level; The attenuation coefficient of the graph topology channel has a value range of 1. Preferably, take This is used to simulate the spatial damping dissipation of perturbation energy propagating across edges in a graph structure. Where the penetrating perturbation data has been set to zero in the aforementioned subtraction comparison, i.e. From the above formula, we can see that the input disturbance data... By keeping it at zero, the truncated propagation chain remains constant at each downstream stage, thus achieving spatial damping dissipation of disturbance energy as it propagates on the graph structure.

[0043] In addition, to prevent propagation from not terminating due to residual loops in the graph, the graph computation engine maintains a set of visited directed edge labels as it propagates along any downstream propagation chain, and sets a preset maximum number of hops. (Preferred selection) When the propagation encounters a visited directed edge, or the current link hop count reaches a certain threshold. When this occurs, the chain's propagation ceases. This attenuation coefficient, combined with the subtraction comparison of each side, constitutes a dual energy dissipation mechanism along the path depth.

[0044] This application applies a stepwise multiplicative attenuation to the penetrating disturbance data through the graph topology channel attenuation coefficient, so that the disturbance energy is naturally dissipated with the propagation depth, and the truncated chain set to zero is always zero downstream. This mechanism ensures the monotonic convergence and precise control of the anomaly's influence range along the longitudinal dimension of the lineage, and avoids the unbounded spread of the anomaly in the deep path.

[0045] This application further proposes a cross-path topology self-healing repair step following the step of setting the value of the penetrating perturbation data to zero to truncate the propagation. Specifically, when the number of downstream propagation chains corresponding to the same source node to be evaluated that are truncated exceeds a preset proportion threshold... When, that is, satisfied (in For the number of truncated chains, (The total number of downstream propagation chains of the source node). The graph computing engine determines that the anomaly intensity of the source node to be evaluated is high enough to trigger truncation protection for most of its surrounding downstream links. At this time, a few sibling paths that have not been truncated or have generated reconstructed state data due to overload are experiencing concentrated residual disturbances. Therefore, a cross-path topology self-healing repair step is initiated to perform lateral compensation on the stressed overloaded sibling edges, with a preset proportion threshold. Preferred selection .

[0046] The self-healing repair steps include: the graph computing engine retrieves the historical average current cell confidence level of the currently truncated directed edge within the most recent complete cycle before truncation. and its time series variance Based on this, the path redundancy reserve coefficient, independent of the noise immunity threshold, is calculated. The calculation formula is as follows: ; in, The time-series statistical mean of the credibility of the current cell in the most recent complete cycle before the truncation of the truncated directed edge; This corresponds to the time series variance. The physical meaning of this formula is that the higher the historical confidence mean and the smaller the fluctuation variance of an edge, the more sufficient its statistical confidence is, and the stronger the "inertial redundancy reserve" released after it is truncated.

[0047] The graph computing engine combines the path redundancy reserve coefficient with a preset repair gain. Multiplying these together, we can solve for the lateral self-healing compensation operator Ψ, whose calculation formula is as follows: ; Where Ψ is the horizontal self-healing compensation operator, whose dimension is consistent with the current unit's reliability; The preset repair gain has a value range of (0,1), and is preferably set to [value missing]. =0.6.

[0048] The graph computing engine then searches the topology for sibling parallel edges that share a common source node or a common target node with the currently truncated directed edge. When it detects that a sibling parallel edge generates reconstructed state data due to perturbation overload, the lateral self-healing compensation operator is injected across the topology into the sibling parallel edge to compensate for the loss of current cell credibility due to perturbation reduction. The injection rules are as follows: ; in, The credibility of the current cell after reconstruction of the sibling parallel edges; This is used to constrain the credibility of the injected compensation to within the upper bound of a uniform scale.

[0049] It is important to clearly distinguish between: time-series reliable inertial data and perturbation input values. Current unit credibility With path redundancy reserve coefficient They are all normalized to the same unified credibility scale [0,1], but they play different functional roles: It is only used as the "upper limit of absorbable disturbance" in the longitudinal subtraction comparison (the difference between disturbance requirement and absorption threshold). and Both falling under the "credibility level / reserve" category, they only participate in horizontal additive compensation. This solution ensures consistency through additive constraints of the same type at the computational level. It avoids any credibility addition operations from beginning to end, thus preventing the logical error of mistakenly adding the "absorption threshold" as the "credibility amplitude".

[0050] When a source node causes a large-scale truncation of its surrounding area due to abnormally high intensity, and the residual risk concentrates on a few untruncation paths, this application converts the historical statistical confidence of the truncated edge into a horizontal self-healing compensation operator and injects it across the topology into the stressed sibling overloaded edge. This allows the local lineage subgraph to redistribute the confidence reserves between paths, alleviating the excessive concentration of risk on a few paths, preventing overloaded edges from being continuously reduced due to isolated stress, and maintaining the evaluation robustness of the local subgraph.

[0051] This application further proposes a step of performing full-link aggregation to generate multiple distinct terminal path score data for the same source node to be evaluated. Specifically, on any given downstream propagation chain, the graph computing engine performs a normalized percentage conversion on the reconstructed state data corresponding to each untruncated directed edge on that chain, or the unchanged initial propagation state data. To eliminate the cumulative scale bias caused by different propagation chain depths, this application adopts a path depth normalized chain multiplication rule, that is, multiplying the normalized confidence of each edge along the path, and then calculating the total number of directed edges for that path. Square root extraction generates terminal path score data for a specific downstream propagation chain label corresponding to the same source node to be evaluated. The calculation formula is as follows: ; in, Let e ​​be the confidence level of the directed edge e on the chain at the aggregation time. For edges that have undergone degradation reduction, take the confidence level of the current unit after reconstruction. For edges that have not been perturbed and reduced, their original current cell confidence level is taken. ; and These are the lower and upper bounds of the unified credibility scale, respectively (i.e.) and ); This represents the total number of directed edges (i.e., path depth) contained in this specific downstream propagation chain. This means that after performing cumulative multiplication according to the normalization multiplication law, the geometric mean of the path depth is then taken.

[0052] The terminal path score data corresponding to this specific downstream propagation chain, after being converted to a percentage system, falls within... In this application, the terminal scores of two links with identical confidence distributions converge to the same order of magnitude, regardless of their path length. This avoids the systematic bias of long links being naturally lower due to excessive multiplication factors under the traditional pure multiplication rule, ensuring that the terminal path score depends only on the unit average confidence of each edge on the path, accurately reflecting the true end-to-end impact of source anomalies on the path.

[0053] This embodiment uses a normalized geometric mean to composite the credibility of each side of the path: when the credibility of any side of the link approaches zero, the geometric mean also approaches zero, thus preserving the sensitivity to the link that has lost credibility; at the same time, because the square root normalization eliminates the cumulative bias caused by the path length, it makes two links with the same credibility distribution but different lengths converge to the same order of magnitude, ensuring that the terminal score reflects the unit average end-to-end credibility, rather than the path length.

[0054] This application further proposes an adaptive impact assessment result encapsulation and output step. Specifically, the graph computing engine uses the identification data of the source node to be assessed as the abnormal source label, the downstream propagation chain corresponding to the reconstructed state data as the damaged path label, and the terminal node label corresponding to the terminal path score data as the risk impact range. These are written together with multiple terminal path score data into a pre-structured association matrix for output. When two or more downstream propagation chains of the same source node to be assessed converge to the same terminal node, the graph computing engine takes the minimum value among the terminal path score data of these paths as the comprehensive risk score of the terminal node, and merges it into a single record in the association matrix using the terminal node as the index. This ensures that the risk characterization of the same terminal is unique and based on the worst path, avoiding duplicate or conflicting risk assessment results when multiple paths converge.

[0055] In one embodiment, the association matrix data Organized in the form of quadruples, each row of the record can be represented as: ; in, This is the label for the abnormal source, i.e., the identification data of the source node to be evaluated; The label for the damaged path, i.e., the first path in the reconstructed state data. A downstream transmission chain; The scope of risk impact refers to the terminal node label corresponding to the terminal path score data of this path; This provides the terminal path score data for this route. The routing control module uses this data to assign scores below a preset score threshold. The data source corresponding to the damaged path will be subject to closed-loop control such as weight reduction, circuit breaking, or re-collection.

[0056] By uniformly encapsulating and outputting anomaly sources, damaged paths, risk impact ranges, and multiple terminal path scoring data in the form of a structured correlation matrix, this application has completed the aggregation and expression from decentralized edge-by-edge state perception to a global impact map, enabling the assessment results to directly drive the physical control of the data link and realizing a technical closed loop from "state perception" to "proactive self-healing intervention".

[0057] The following is a specific implementation of a method for dynamically evaluating the credibility of cross-source data: In a multi-source data fusion scenario for enterprise credit reporting, the source nodes The "Enterprise Operation Flow Record" field contains two downstream bloodline transmission chains: Link 1 Through the "flow stability operator edge" "Credit score aggregation edge" Transmitted to the terminal "loan amount tag"; Link 2 via "growth mapping edge" "Public opinion correction edge" The propagation is transmitted to the terminal "Enterprise Growth Potential Tag". The initial propagation state data for each directed edge is as follows: of , , ; of , , ; of , , ; of , , .

[0058] The graph calculation engine retrieves the preset time window for each edge ( The current unit's confidence time series is calculated, and the time series variance is respectively... , , , According to public Formula (take) , Solve the time-series reliable inertial data to obtain , , ;and Due to historical transmission frequency Forced to converge to the lower bound, thus obtaining Among them, the side-level inertia Used for edge-by-edge adaptive absorption of incident perturbations, global threshold 0.10 is used to define whether a penetration disturbance constitutes a valid penetration; the two functions complement each other.

[0059] Source node to be evaluated via sliding window ( The statistical results of the nodal residuals are obtained. Dynamic baseline ,because Anomaly detected, duration According to the formula (Pick , , The initial anomalous disturbance data were obtained by solving the problem. ,injection and Two downstream transmission chains.

[0060] In Link 1 superior: Press Therefore, the penetration disturbance is set to zero, and the link is in It was immediately cut off. And the terminal's "loan amount tag" remains unaffected by disturbances, preserving its historically high robustness and stability. In link two... superior: Because of the extremely low inertia during cold starts, If overload is detected, the degradation reduction formula is applied (taking...). )have to Generate reconstructed state data and write it to the edge attributes; according to channel attenuation ( ) is transmitted to Input disturbance ,exist Place Overloaded and rebuilt again .

[0061] The graph computing engine performs full-link aggregation on the two links: the unbroken edges of link one retain their original values, and the terminal path score is obtained by multiplying them together. Link 2 is obtained by multiplying the reconstructed credibility values. .because The routing control module performs circuit breaking and re-collection on the "growth mapping" data source corresponding to link two, while maintaining the original weight of link one, which meets the score, without intervention.

[0062] Meanwhile, due to the truncated chain ratio downstream of the source node... This triggers cross-path topology self-healing. The truncated directed edge during propagation is... retrieval Historical average current cell confidence level within the most recent complete period before truncation (Based on the statistics from the aforementioned window) ) and its time series variance Calculate the path redundancy reserve coefficient .according to Solve for the lateral self-healing compensation operator. Search for and interact with the topology network. Sibling parallel edges with a common source node vs are obtained ;because Since the perturbation overload has already generated reconstructed state data, Ψ will be injected across the topology. ,according to Increase the credibility of its current unit.

[0063] Self-healing compensation is applied to edge attributes to stabilize the edge's inertial convergence in subsequent cycles. The terminal path score output in the current cycle still reflects the overload state, thereby alleviating the excessive concentration of residual risk on a few paths without reversing the evaluation conclusion of this cycle.

[0064] Ultimately, the graph computing engine labels the anomaly sources. Damaged path labels Scope of Risk Impact: "Enterprise Growth Potential Tag" and Terminal Path Score , Write the association matrix data together Output. This embodiment verifies that: based on the objective differences in the noise resistance rigidity of each path, this method can preserve the historically extremely stable Link 1 and promptly truncate and downgrade the vulnerable cold-start Link 2, achieving precise control and closed-loop intervention of the abnormal impact range in the longitudinal dimension of lineage.

[0065] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for dynamic evaluation of cross-source data credibility, characterized in that, Includes the following steps: Acquire cross-source data streams and extract field data, establish a topology network containing nodes and directed edges according to a preset mapping relationship, and configure initial propagation state data containing historical propagation counts and historical conflict rates for each directed edge; The time series variance of the initial propagation state data is captured, and combined with the historical propagation count and historical conflict rate, damping correlation calculation is performed according to the preset damping function to generate time-series reliable inertial data that characterizes the noise immunity threshold of the directed edge. When the node residual of the source node to be evaluated exceeds the preset dynamic baseline, initial abnormal disturbance data is generated based on the amount and duration of the residual excess, and injected into all downstream propagation chains starting from the source node to be evaluated. The initial abnormal disturbance data is controlled to propagate step by step along each downstream propagation chain. At any current directed edge, the corresponding time-series reliable inertial data is retrieved, and the current-level disturbance input value is compared with the time-series reliable inertial data by subtraction under the unified reliability scale to generate penetrating disturbance data. If the penetration disturbance data does not exceed the preset threshold, the value is set to zero to cut off the propagation chain; If the initial propagation state data is reduced by using the penetration perturbation data, reconstructed state data is generated and written into the edge attribute of the corresponding directed edge to replace the original data. The reconstructed state data is then used as the input for calculating the time-series reliable inertial data in the next cycle. The reconstructed state data or the initial propagation state data that has not been downgraded for each uncut directed edge on each downstream propagation chain are aggregated across the entire chain. Multiple terminal path score data are generated for the same source node to be evaluated, and they are associated and encapsulated into an adaptive impact assessment result. Downweighting, circuit breaking, or re-collection are performed on the data source corresponding to the downstream propagation chain whose score is lower than the preset score threshold.

2. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, The initial propagation state data is organized in the form of edge attribute quintuples of a directed graph. The quintuples include: source node identifier, target node identifier, historical propagation count, historical conflict rate, and current cell credibility. The current cell credibility is used to characterize the basic reliability of a single propagation of a directed edge and participates in the damping correlation calculation as a component field for generating the initial propagation state data.

3. The method for dynamic evaluation of cross-source data credibility according to claim 2, characterized in that, The steps for obtaining the current unit's credibility include: The spatial dimension multi-source cross consistency, temporal dimension update decay freshness, and business feedback closed-loop correctness of the directed edges are obtained. The spatial dimension multi-source cross consistency, temporal dimension update decay freshness, and business feedback closed-loop correctness are then linearly weighted and summed according to preset weights to generate the current unit's credibility.

4. The method for dynamic evaluation of cross-source data credibility according to claim 3, characterized in that, The steps for capturing the time series variance of the initial propagation state data include: Retrieve the historical confidence scores of the current unit from the 1st to the Nth time within a preset time window, and calculate the discrete variance of the historical confidence scores of the current unit from the 1st to the Nth time relative to its time series statistical mean, as the time series variance, where N is an integer greater than 1.

5. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, The steps for generating time-series reliable inertial data also include introducing a cold-start correction factor: The nonlinear exponential convergence function is calculated using the historical propagation count to generate the cold start correction factor. The cold start correction factor is then used to perform a step scaling correction on the original damping value calculated based on the historical conflict rate and the time series variance. This is to force the time series reliable inertial data to converge to the preset reliability lower bound threshold under the unified reliability scale when the historical propagation count is lower than a preset count value.

6. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, The calculation steps for the nodal residuals of the source node to be evaluated and the preset dynamic baseline include: By using a sliding time window, the historical mean and current observation of the field data flowing through the source node to be evaluated are statistically analyzed, and the absolute deviation between the two is calculated as the node residual. The historical standard deviation of the absolute deviation within the sliding time window is retrieved, and the historical standard deviation of the absolute deviation is multiplied by a preset gain coefficient to generate the dynamic baseline. The amount by which the node residual exceeds the dynamic baseline, and the duration for which the node residual continuously exceeds the dynamic baseline, are input together into a preset normalized time-series mapping function to generate the initial abnormal disturbance data mapped to the unified confidence scale.

7. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, The steps for controlling the propagation of the initial anomalous disturbance data along each of the downstream propagation chains include: The penetration perturbation data generated by comparing the directed edges of the current level is used as the input value and multiplied by a preset graph topology channel attenuation coefficient to generate input perturbation data that is propagated to the next level adjacent directed edge; wherein, if the penetration perturbation data has been set to zero, the input perturbation data remains at zero, so as to realize the spatial damping dissipation of perturbation energy when it propagates on the graph structure.

8. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, After setting the value of the penetration perturbation data to zero to truncate the propagation, when the number of downstream propagation chains corresponding to the same source node to be evaluated is truncated exceeds a preset proportion threshold, a cross-path topology self-healing repair step is initiated, including: Retrieve the historical average current unit confidence and its time series variance of the currently truncated directed edge in the most recent complete cycle before truncation, and calculate the path redundancy reserve coefficient accordingly; multiply the path redundancy reserve coefficient by the preset repair gain to calculate the lateral self-healing compensation operator. In the topology network, retrieve sibling parallel edges that share a common source node or a common target node with the currently truncated directed edge. When it is detected that the sibling parallel edge generates the reconstructed state data due to disturbance overload, the lateral self-healing compensation operator is injected across the topology into the sibling parallel edge to increase the credibility of the current cell corresponding to the reconstructed state data.

9. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, The steps of performing end-to-end aggregation to generate multiple distinct terminal path score data for the same source node to be evaluated include: On any specific downstream propagation chain, the reconstructed state data corresponding to each uncut directed edge on the chain, or the unchanged initial propagation state data, is normalized to a percentage scale and then sequentially multiplied and compounded, and the total number of directed edges contained in the path is squared to normalize, thereby generating the terminal path score data corresponding to the specific downstream propagation chain label of the same source node to be evaluated.

10. The method for dynamic evaluation of cross-source data credibility according to claim 1, characterized in that, The steps for encapsulating and outputting the adaptive impact assessment results include: The identification data of the source node to be evaluated is used as the abnormal source label, the downstream propagation chain corresponding to the reconstructed state data is used as the damaged path label, and the terminal node label corresponding to the terminal path score data is used as the risk impact range. These are written together with multiple terminal path score data into the pre-structured association matrix data for output.

Images