A method, system, and storage medium for multi-source conflict data fusion governance based on evaluation feedback and dynamic confidence.

By using dynamic confidence assessment of data sources and constructing a dual-confidence DS quality function, this paper solves the problems of dynamic changes in the confidence of data sources and conflict allocation in high-conflict scenarios in multi-source data fusion. It achieves unified classification results with high confidence and is applicable to application scenarios such as cybersecurity, financial risk control, medical diagnosis, and smart cities.

CN123116166APending Publication Date: 2026-11-17THE THIRD RES INST OF MIN OF PUBLIC SECURITY
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611471444.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies have failed to effectively address the issues of dynamic changes in data source credibility and the reasonable allocation of conflicting information in high-conflict scenarios during multi-source data fusion, resulting in unreliable decision-making quality.

Method used

By establishing a dynamic confidence assessment mechanism for data sources, constructing a dual-confidence DS quality function, and performing conflict backfilling in high-conflict scenarios, conflict evidence is allocated according to confidence ratios to achieve reasonable weighting and uncertainty absorption of conflict information.

Benefits of technology

It achieves dynamic and accurate evaluation of data source confidence, effectively suppresses the contribution of low-quality data sources, avoids the amplification of errors in weak consensus under high-conflict scenarios, and outputs unified classification results with high confidence.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116166A_ABST
    Figure CN123116166A_ABST
Patent Text Reader

Abstract

This invention discloses a method, system, and storage medium for multi-source conflict data fusion governance based on evaluation feedback and dynamic confidence, relating to the field of multi-source information fusion. The proposed solution periodically evaluates each data source based on a standardized test set. Dynamic confidence is obtained by weighting the current test accuracy and the historical average confidence of the sliding window using a time factor. A basic probability allocation function (DS) is constructed using the dynamic confidence and classification probability. Conflict coefficients are calculated; if the threshold is not exceeded, the Dempster combination rule is used; otherwise, conflict is backfilled—the conflict coefficient portion is assigned to the unknown category based on the uncertainty mean, and the remainder is assigned to the primary judgment category of each data source according to the dynamic confidence ratio. The category corresponding to the maximum value after fusion is taken as the result. This invention overcomes the counterintuitive fusion problem of DS theory under high conflict conditions, improving the accuracy and reliability of fusion.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, specifically to the fusion and governance technology of multi-source conflicting data, and more particularly to a multi-source conflicting data fusion and governance method, system, and storage medium based on evaluation feedback and dynamic confidence. Background Technology

[0002] With the deepening of digital transformation, the integration of conflicting data from multiple sources has become a core element in building high-quality knowledge bases, risk profiles, and decision support systems. In application scenarios such as cybersecurity, financial risk control, medical diagnosis, and smart cities, single data sources often suffer from incomplete coverage and a limited perspective, necessitating the fusion of multi-source data to improve the accuracy and robustness of decision-making.

[0003] Dempster-Shafer (DS) evidence theory, as a mathematical tool for handling uncertainty and incomplete information, has been successfully applied in fields such as target recognition, fault diagnosis, and information fusion. This theory, by constructing an identification framework and a basic probability assignment function (i.e., a basic quality function), can formally express the degree of evidence support and utilize Dempster's combination rules to achieve multi-source evidence fusion. However, classical DS theory suffers from the following prominent problems in practical applications.

[0004] First, classic DS theory assumes that all sources of evidence have equal credibility and does not consider the dynamic changes in the accuracy of the data sources themselves. When faced with input from low-reliability data sources, classic DS theory still assigns them excessive influence, seriously affecting the quality of the final decision.

[0005] Second, when there is high conflict among multiple pieces of evidence, Dempster's combination rule produces counterintuitive fusion results. Classical Dempster theory normalizes and amplifies the conflict coefficient K by a factor of 1 / (1-K). When K approaches 1, small amounts of consistent evidence are drastically amplified into false high-confidence conclusions, i.e., the Zadeh paradox.

[0006] To address the aforementioned issues, several improvement schemes have been proposed in the prior art. Patent application CN110009251B discloses a method for dynamic adjustment and fusion of multi-source farmland information, which determines whether to use the classic DS synthesis rule or the average evidence factor to replace the conflict factor based on whether the conflict coefficient is within a set threshold range. However, the "average evidence factor replacement" method used in this scheme essentially discards the original conflict information and replaces it with the average value, failing to distinguish the credibility differences between different data sources and failing to solve the problem of reasonable redistribution of conflict information.

[0007] In their article "A Partial Conflict Allocation Strategy in Evidential Reasoning," published in *Computer Applications Research*, Vol. 29, No. 9, 2012, Cao Jie and Meng Xing disclosed a method for handling conflicting evidence. This method sets a threshold and proportionally allocates local conflicts exceeding this threshold to the focal element that generated the conflict, then fuses the evidence using standard synthesis rules. While this method involves proportionally allocating conflicts to focal elements, its allocation logic is based on a retrospective allocation of the internal structure of the evidence—returning the conflict to the focal element that generated it—and assumes that all evidence has consistent credibility. This scheme fails to consider the credibility differences between data sources and does not propose a technical concept for directing the allocation of conflicts to the primary and unknown categories of the data source.

[0008] Florea et al., in their article "Dynamic estimation of evidence discounting rates based on information credibility" published in *RAIRO - Operations Research*, Volume 44, Issue 4, 2010, disclosed a technical approach for dynamically estimating evidence discounting rates based on information credibility. Furthermore, the article "credibility decay model in time-domain evidence fusion" published in *Systems Engineering and Electronics*, Issue 7, 2015, also involves a method for real-time reliability assessment of evidence based on credibility decay. However, both of these methods involve discounting the evidence sources before fusion (i.e., input-side correction), rather than targeted allocation of conflicts at the fusion result level.

[0009] In summary, there is an urgent need in the existing technology for a dynamic fusion method for multi-source conflicting data that can establish a systematic dynamic evaluation mechanism for data source confidence while preserving the sound mathematical foundation of DS theory, and can reasonably redistribute conflicting information in high-conflict scenarios—different from the "discard-replace" approach, the "input-side correction" approach, and the "returning the conflict to the focal element that generated the conflict" approach. Summary of the Invention

[0010] (a) Purpose of the invention

[0011] This invention aims to solve the technical problem of unreliable information normalization results caused by multi-source data conflicts in the prior art, and provides a multi-source conflict data fusion governance method, system and storage medium based on evaluation feedback and dynamic confidence.

[0012] This invention establishes a systematic dynamic confidence assessment mechanism for data sources, comprehensively considering the overall credibility of the data source and the classification confidence of individual data points, constructs a dual-confidence DS quality function, and achieves reasonable weighting and uncertainty absorption of conflicting evidence by performing conflict backfilling according to confidence ratio in high-confidence scenarios, thereby outputting a unified classification result with high confidence.

[0013] (II) Technical Solution

[0014] To address the aforementioned technical problems, this invention provides a multi-source conflict data fusion governance method based on evaluation feedback and dynamic confidence, comprising:

[0015] The dynamic confidence assessment step for data sources involves periodically evaluating multiple data sources based on a standardized test set, and calculating the dynamic confidence of each data source based on its current test accuracy on the standardized test set and the historical confidence average based on sliding window statistics.

[0016] The quality function construction step involves, for each data point to be fused, constructing a basic probability allocation function for the DS evidence theory based on the dynamic confidence level of the data source from which the data point originates, calculated in the data source dynamic confidence assessment step, and the classification probability distribution of the data point; and

[0017] The conflict fusion step involves calculating the conflict coefficients for multiple basic probability allocation functions, determining whether the conflict coefficients exceed a preset conflict threshold, and if they do not exceed the threshold, using the Dempster combination rule for fusion; if they exceed the threshold, initiating the conflict backfilling process.

[0018] The conflict backfilling process includes: allocating a portion of the conflict coefficient to the unknown category, and allocating the remaining portion of the conflict coefficient to the primary judgment category of each data source according to the relative proportion of the dynamic confidence of each data source; the primary judgment category is the category corresponding to the maximum value of the basic probability allocation function among the known categories.

[0019] Preferably, the dynamic confidence assessment step of the data source includes:

[0020] Test set maintenance sub-step: A standardized test set that is independently built and continuously maintained is used as an external benchmark. The standardized test set is built according to sample collection, labeling quality control and version management. The version management includes periodically and incrementally supplementing new business samples and eliminating expired samples that exceed the time limit.

[0021] Test execution sub-steps: Generate the current random test set by stratified random sampling from the standardized test set according to categories, call the processing interface of each data source to classify the current random test set, and calculate the current test accuracy of each data source;

[0022] The confidence score calculation sub-steps are as follows: Calculate volatility based on the historical confidence score sequences of each data source; adaptively adjust the length of the sliding window according to the volatility, with a shorter sliding window for higher volatility; calculate the mean of historical confidence scores within the sliding window as the historical confidence score mean; calculate a timeliness factor based on the most recent update time of the standardized test set; determine the weight of the current test accuracy based on the timeliness factor; and weight and sum the current test accuracy with the historical confidence score mean to obtain the dynamic confidence score.

[0023] Preferably, in the confidence calculation sub-step, the volatility of each data source is calculated based on the mean squared error of the reference historical sequence; the length of the sliding window is adaptively adjusted according to the volatility in an inverse relationship, with a shorter window for higher volatility; the timeliness factor is calculated based on the time interval between the current moment and the most recent update moment of the standardized test set in an exponential decay manner; the weight of the current test accuracy is determined based on the timeliness factor, and this weight is between a preset minimum retention weight and a maximum weight; the dynamic confidence is the result of a weighted sum of the current test accuracy and the historical confidence mean according to the weight.

[0024] Preferably, in the quality function construction step, for any data source, the dynamic confidence level of the data source is used as the degree to which the output information of the data source is accepted; for each known category, the classification probability of each data to be fused is multiplied by the dynamic confidence level to obtain the basic probability allocation value of the data source for each known category; for unknown categories, the difference obtained by subtracting the dynamic confidence level from 1 is used as the quality of the unknown category to absorb the uncertainty of the data source.

[0025] Preferably, for two data sources, in the conflict fusion step, the sum of the products of the two quality functions on each focal element is first calculated, and the sum is subtracted from 1 to obtain the conflict coefficient; when the conflict coefficient is less than the preset conflict threshold, the Dempster combination rule is used for direct fusion; when the conflict coefficient is greater than or equal to the preset conflict threshold, the conflict backfilling process is initiated: a portion of the conflict coefficient is allocated to the unknown category according to the average uncertainty of the two data sources, and the remaining conflict coefficient is allocated to the respective primary judgment category of the two data sources according to the relative proportion of the dynamic confidence of the two data sources; the product of the two quality functions on each focal element is superimposed with the conflict coefficient component allocated to that focal element to obtain the fused quality function, and the category corresponding to the maximum value in the fused quality function is taken as the final fusion result label.

[0026] Preferably, when there are N data sources and N≥3, pairwise recursive combinations are performed according to the associative law of Dempster's combination rule, and the basic probability allocation function of the nth data source is combined with the current fusion result in turn, n = 1, 2, ..., N. Each combination step independently calculates the conflict coefficient of the current step and determines whether the conflict backfilling process is triggered; when the current fusion result is an intermediate fusion result, the representative confidence of the intermediate fusion result is taken as the arithmetic mean of the dynamic confidence of each original data source participating in the fusion.

[0027] To address the aforementioned technical problems, this invention also provides a multi-source conflict data fusion governance system based on evaluation feedback and dynamic confidence, comprising:

[0028] The standardized test set maintenance module is configured to independently build and continuously maintain standardized test sets, which are built according to sample collection, annotation quality control and version management.

[0029] The dynamic confidence assessment module is configured to periodically evaluate multiple data sources based on the standardized test set, and calculate the dynamic confidence of each data source based on the current test accuracy of each data source on the standardized test set and the historical confidence average based on sliding window statistics.

[0030] The quality function construction module is configured to, for each piece of data to be fused, construct a basic probability allocation function for the DS evidence theory based on the dynamic confidence level of the data source from which the data originates, calculated by the dynamic confidence level assessment module, and the classification probability distribution of the data.

[0031] The conflict fusion module is configured to calculate conflict coefficients for multiple basic probability allocation functions, determine whether the conflict coefficients exceed a preset conflict threshold, and if not, use the Dempster combination rule for fusion; if they exceed the threshold, initiate a conflict backfilling process. The conflict backfilling process includes: allocating a portion of the conflict coefficients to an unknown category, and allocating the remaining portion of the conflict coefficients to the primary judgment category of each data source according to the relative proportion of the dynamic confidence of each data source.

[0032] Preferably, the dynamic confidence assessment module includes:

[0033] The test execution unit is configured to generate a random test set for the current test by stratified random sampling from the standardized test set, call the processing interface of each of the data sources to classify the random test set for the current test, and calculate the current test accuracy of each of the data sources.

[0034] The confidence calculation unit is configured to calculate volatility based on the historical confidence sequence of each of the data sources, adaptively adjust the length of the sliding window according to the volatility, calculate the mean of historical confidence within the sliding window as the historical confidence mean, calculate a timeliness factor based on the most recent update time of the standardized test set, determine the weight of the current test accuracy based on the timeliness factor, and obtain the dynamic confidence by weighted summing of the current test accuracy and the historical confidence mean.

[0035] Preferably, when the conflict fusion module performs the conflict backfilling process for two data sources, it allocates a portion of the conflict coefficients to the unknown category according to the average uncertainty of the two data sources, and allocates the remaining conflict coefficients to the respective primary judgment categories of the two data sources according to the relative proportion of the dynamic confidence of the two data sources.

[0036] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method.

[0037] (III) Technical Effects

[0038] This invention achieves the following technical effects by establishing a complete technical chain: "dynamic evaluation of standardized test sets → construction of dual-confidence quality functions → judgment of conflict coefficient thresholds → backfilling according to confidence ratios":

[0039] First, this invention achieves dynamic and accurate evaluation of data source confidence through a standardized test set with versioned management and a confidence calculation mechanism based on a volatility-adaptive sliding window. Stable data sources are given a longer window to smooth out noise, while volatile data sources are given a shorter window to quickly respond to performance changes, effectively capturing performance degradation caused by model degradation, data distribution shifts, or temporary failures.

[0040] Secondly, this invention organically integrates the overall credibility of the data source with the classification confidence of a single data point through the construction of a dual-confidence DS quality function, so that even if a low-quality data source outputs a high-probability classification result, its overall contribution is effectively suppressed.

[0041] Third, this invention, through a conflict backfilling mechanism in high-conflict scenarios, backfills the conflict coefficients to the unknown category and the primary judgment category of each data source according to the confidence ratio of the data source, effectively avoiding the problem of the classic DS theory amplifying weak consensus errors in high-conflict scenarios.

[0042] Fourth, this invention can be applied to the fusion and governance of multi-source conflicting data in a homogeneous label space, and is suitable for various application scenarios such as network security, financial risk control, medical diagnosis, and smart cities. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0044] Figure 1 This is a flowchart illustrating the dynamic confidence assessment method for data sources provided in an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the structure of the dynamic confidence weighted calculation model provided in an embodiment of the present invention.

[0046] Figure 3 This is a flowchart illustrating the conflict fusion method provided in an embodiment of the present invention.

[0047] Figure 4 This is a structural block diagram of a multi-source conflict data fusion governance system based on evaluation feedback and dynamic confidence, provided in an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] To address the unreliable technical problem of multi-source conflicting data fusion in existing technologies, this invention makes the following technical innovations to form a multi-source conflicting data fusion governance method based on evaluation feedback and dynamic confidence.

[0050] First, this invention constructs a precise evaluation mechanism for the dynamic confidence level of data sources. Using independently constructed and version-maintained standardized test sets as external benchmarks, this invention generates current random test sets through stratified random sampling by category, and calculates the current test accuracy of each data source. Based on this, this invention introduces an adaptive sliding window based on historical confidence volatility (the larger the volatility, the shorter the window; the smaller the volatility, the longer the window), enabling stable data sources to obtain a smooth historical mean and volatile data sources to gain the ability to quickly respond to current changes. Simultaneously, this invention introduces a dynamic weighting mechanism based on the timeliness factor of test set updates; the newer the test set, the higher the weight of the current test accuracy; the older the test set, the higher the weight of the historical confidence mean.

[0051] Secondly, this invention constructs an innovative dual-confidence DS quality function. This invention organically combines the dynamic confidence η of the data source with the classification probability distribution p_i of a single data point, constructing a basic probability allocation function of m({S_{M+1}}) = 1-η and m({S_i}) = p_i ×η. This construction method systematically suppresses the contribution of low-confidence data sources to the overall quality function, even when they output high-probability classification results.

[0052] Furthermore, this invention constructs a conflict backfilling mechanism based on confidence ratio. In high-conflict scenarios (K ≥ K_th), this invention no longer employs the classic DS 1 / (1-K) normalization amplification. Instead, it allocates a portion of the conflict coefficient K to the unknown category based on the average of the uncertainties (1-η) of the two data sources, and backfills the remaining portion to their respective primary decision categories according to the relative proportions of the confidence η of the two data sources. This mechanism effectively avoids the problem of weak consensus being erroneously amplified in high-conflict scenarios.

[0053] Accordingly, in this invention, dynamic confidence assessment provides an accurate η value for the construction of the quality function, the construction of the quality function provides a reasonable input for conflict fusion, and the conflict backfilling mechanism performs directional allocation based on the η value when there is high conflict. The three work together to form a complete processing scheme from "assessment" to "construction" to "fusion".

[0054] The following is combined Figures 1 to 3 This paper provides a detailed description of the multi-source conflict data fusion governance method based on evaluation feedback and dynamic confidence provided by the present invention.

[0055] like Figure 1 and Figure 3 As shown, the method generally includes the following three steps:

[0056] Step S100: Dynamic confidence assessment of the data source.

[0057] This step is configured to use a standardized test suite that is independently built and continuously maintained as an external benchmark to periodically evaluate each data source and output the dynamic confidence level η of each data source.

[0058] Furthermore, the confidence assessment in this step is not based on static weights or single test results, but rather integrates the current test accuracy of the data source on the standardized test set and the historical confidence average based on sliding window statistics. It also introduces a volatility-based window adaptive adjustment mechanism and a dynamic weighting mechanism based on the test set update time, thereby achieving accurate and dynamic confidence assessment.

[0059] Therefore, this step serves as the basis for subsequent quality function construction and conflict fusion, and its output dynamic confidence level will be used as an important input parameter for step S200.

[0060] Step S200: Construct the DS mass function.

[0061] This step is configured to construct the basic probability allocation function of the DS evidence theory for each piece of data to be fused, based on the dynamic confidence η of the data source from which the data originates and the classification probability distribution of the data.

[0062] Furthermore, this step specifically couples the overall confidence (η) of the data source with the classification confidence (probability distribution p) of a single data point in a multiplicative manner, constructing the form m({S_{M+1}})=1-η and m({S_i})=p_i×η, so that even if a low-confidence data source outputs a high-probability classification result, the contribution of its mass function is systematically suppressed.

[0063] During deployment configuration, the output of this step (the mass function of each data source) is the input of step S300.

[0064] Step S300, conflict fusion.

[0065] This step is configured to fuse multiple mass functions and output a unified classification label.

[0066] Furthermore, this step specifically calculates the conflict coefficient K and then determines whether K exceeds a preset conflict threshold K_th. If it does not exceed the threshold, the classic Dempster combination rule is used; if it does exceed the threshold, a conflict backfilling process is initiated—a portion of the conflict coefficient is allocated to the unknown category, and the remaining portion is allocated to the primary judgment category of each data source according to the relative proportion of the dynamic confidence levels of each data source. This step directly solves the counterintuitive technical problem of existing technologies in high-conflict scenarios.

[0067] Based on an understanding of the overall function and logical coordination of each step in the method of this invention, the specific implementation scheme of each step will be further explained from the perspective of technical implementation below.

[0068] As a specific example of an implementable form, step S100 in this method, when specifically implemented, further includes a test set maintenance sub-step S110, a test execution sub-step S120, and a confidence calculation sub-step S130.

[0069] (1) Test set maintenance sub-step S110:

[0070] This sub-step uses a standardized test suite that is independently built and continuously maintained as an external benchmark. The construction of the standardized test suite involves three phases:

[0071] Sample collection phase: Entity samples are extracted stratified by category from the business production environment, covering all known categories in the identification framework. The sample size ratio of each category is consistent with the business traffic distribution in the recent period to avoid undersampling of long-tail categories.

[0072] Annotation quality control phase: Each sample is independently annotated by no fewer than two domain experts. In case of discrepancies, a third party is consulted for arbitration, and only samples to which consensus is reached are included in the test set. Each sample must record at least the entity identifier, real label, annotation time, annotation source, and sample version number.

[0073] Versioning Management Phase: Test suites are managed using version numbers. Each update generates a new version and archives the old version, ensuring the reproducibility of historical evaluation results. Test suite maintenance includes both scheduled updates and triggered updates.

[0074] Regarding regular updates, new business samples are added incrementally on a periodic basis, while expired samples that have exceeded the timeliness threshold and have exited the business are eliminated.

[0075] More preferably, version management also includes a triggered update mechanism. In this mechanism, the sample proportion of each category in the production pipeline is continuously monitored. When the relative change in the proportion of a certain category exceeds a preset threshold, or when a completely new category appears in the production pipeline, the test set reconstruction process is immediately triggered, rather than waiting for the preset periodic maintenance cycle.

[0076] The test set reconstruction process described here includes: re-collecting samples from the business environment covering all known categories and new categories, performing annotation quality control, generating a new test set version number, and archiving the old version. This triggered update mechanism ensures that the test set can reflect changes in business distribution in a timely manner, making the confidence assessment always representative.

[0077] (2) Test execution sub-step S120:

[0078] This sub-step randomly samples test data from the data space corresponding to each label in the standardized test set to construct the current random test set.

[0079] Specifically, for category T, the number of samples is:

[0080] n_T = round(n_ref * ρ_T);

[0081] Where ρ_T is the percentage of samples of this category in the current test set, n_ref is the reference sample value, and round is the rounding function.

[0082] Subsequently, the processing interfaces of each data source are called, inputting the data from the current random test set to obtain the classification results from each data source. The classification results are then compared with the true labels of the test set to calculate the current test accuracy η_AC for each data source.

[0083] (3) Confidence calculation sub-step S130:

[0084] This sub-step calculates the dynamic confidence score based on the historical confidence score sequence of each data source.

[0085] For details, see Figure 2 In this confidence calculation sub-step, the dynamic confidence score is calculated in the following way:

[0086] First, the volatility of each data source is calculated based on the reference historical series. This volatility is used to quantify the stability of the historical performance of the data source. The higher the volatility, the less stable the recent accuracy of the data source is.

[0087] Then, the sliding window length used for statistical historical confidence is adaptively adjusted according to volatility. Stable data sources (low volatility) are given a longer window to smooth historical noise, while volatile data sources (high volatility) are given a shorter window to improve the response speed to performance changes.

[0088] Then, a timeliness factor is calculated based on the most recent update time of the standardized test set. This timeliness factor reflects the timeliness of the test set, and the newer the test set, the larger the timeliness factor. The weight of the current test accuracy in the final dynamic confidence is determined based on the timeliness factor, so that the newer the test set, the greater the contribution of the current test accuracy.

[0089] Finally, the current test accuracy and the historical confidence mean are weighted and summed according to the weights to obtain the dynamic confidence of the data source.

[0090] To further illustrate, the volatility α_A is calculated on a reference historical sequence of length W_ref:

[0091] α_A = sqrt( (1 / W_ref) * sum_{i=1}^{W_ref}(η_{A(ti)}-η_{Aavg})^2 );

[0092] In the formula, t is the current evaluation period number, η_{A(ti)} is the dynamic confidence level of data source A in the (ti)th evaluation period, and η_{Aavg} is the mean of the reference historical series. The larger α_A is, the more unstable the recent accuracy of the data source is.

[0093] Secondly, the sliding window length W_A is adaptively adjusted based on the volatility α_A:

[0094] W_A = max( W_min, min( W_max, round( W_max / (1 + λ * α_A) ) ) );

[0095] In the formula, W_min and W_max are the lower and upper bounds of the sliding window length, respectively, and λ is the sensitivity coefficient. This formula establishes a quantitative inverse relationship between volatility and window length: stable sources (smaller α) obtain longer windows to smooth noise, while volatile sources (larger α) obtain shorter windows to improve the response speed to changes.

[0096] Next, calculate the historical confidence mean η_AH:

[0097] η_AH = (1 / W_A) * sum_{i=1}^{W_A} η_{A(ti)};

[0098] Then, the timeliness factor is calculated based on the most recent update time of the standardized test set. Let t_upd be the most recent update time of the standardized test set, Δt = t - t_upd, and define the update timeliness factor β:

[0099] β = exp(-Δt / τ);

[0100] In the formula, τ is the time constant. As Δt increases, the timeliness of the test set deteriorates, and β approaches 0.

[0101] The weight γ for the current test accuracy is calculated based on the time factor β:

[0102] γ = γ_min + (γ_max - γ_min) * β;

[0103] In the formula, γ_min is the minimum retention weight for the current test accuracy, and γ_max is the maximum retention weight for the current test accuracy, with 0 ≤ γ_min < γ_max ≤ 1. This exponential decay model and the design of the minimum retention weight allow the timeliness of the test set to be quantified as a continuously adjustable factor, and even if the test set is severely outdated, the current test information is not completely discarded.

[0104] Finally, calculate the dynamic confidence level η_A of data source A:

[0105] η_A = γ * η_AC + (1 - γ) * η_AH.

[0106] As a specific example of an implementable form, step S200 in this method, in its actual implementation, involves constructing a basic probability allocation function of the DS evidence theory for each piece of data to be fused, based on the dynamic confidence level of the data source and the classification probability distribution of that piece of data.

[0107] Specifically, the basic probability assignment function is constructed in this step as follows:

[0108] For any data source A, its dynamic confidence level η_A is used as the degree to which the output information of that data source is accepted. For known categories S_1 to S_M, the classification probability p_i output by the data source is multiplied by the dynamic confidence level η_A to obtain the basic probability assignment value of the data source for each known category; for unknown category S_{M+1}, 1 is subtracted from the dynamic confidence level η_A to obtain the quality of the unknown category, which is used to absorb the uncertainty of the data source. This construction method ensures that even if a low-confidence data source outputs a high-probability classification result, its overall contribution is systematically suppressed.

[0109] As a further explanation, assume the complete set of data classification labels is Ψ = {S_1, S_2, …, S_M}, where M is the total number of known categories. Construct a recognition framework Θ = {S_1, S_2, …, S_M, S_{M+1}}, where S_{M+1} is the unknown category.

[0110] For any data source A, its dynamic confidence level is η_A. The classification probability distribution output by this data source for a certain data point is P({S_i}) = p_i, i ∈ [1, M], satisfying sum_{i=1}^{M} p_i = 1.

[0111] Based on this, the fundamental probability allocation function m_A constructed in this invention is:

[0112] m_A({S_{M+1}}) = 1 - η_A;

[0113] m_A({S_i}) = p_i * η_A, i ∈ [1, M];

[0114] This construction method differs from the additive discounting method used in existing technologies. This invention adopts a multiplicative coupling method, where the unreliability of the data source (1-η_A) directly determines the quality of the unknown category S_{M+1}—the more unreliable the data source, the larger the "undetermined" part in its output.

[0115] When a data source only outputs hard classification results (i.e., only a single category label is given, without outputting the probabilities of each category), a one-hot encoding method is used to construct the probability distribution: the probability of that category is 1, and the probabilities of the other categories are 0. In this case, the confidence level of a single data point does not provide discrimination, and the fusion result mainly relies on the dynamic confidence level η of the data source for adjustment.

[0116] As a specific example of an implementable form, step S300 in this method is used to perform conflict fusion on multiple basic probability allocation functions, including fusion of two data sources and fusion of multiple data sources (N≥3).

[0117] (1) For the case of merging two data sources;

[0118] For ease of description, we will define a specific data point D, where the identification result from data source A is represented by probability distribution P_A, and the identification result from data source B is represented by probability distribution P_B. Based on this, quality functions m_A and m_B are constructed respectively according to step S200.

[0119] For two data sources A and B, the specific calculation scheme for conflict fusion in this step is configured as follows:

[0120] First, calculate the sum of the products of the two quality functions on each focal element, and then subtract this sum from 1 to obtain the conflict coefficient K.

[0121] When the conflict coefficient K is less than the preset threshold, the classic Dempster combination rule is directly used for fusion.

[0122] When the conflict coefficient K is greater than or equal to a preset threshold, the conflict backfilling process is initiated: a portion of the conflict coefficient K is allocated to the unknown category according to the average uncertainty of the two data sources; the remaining conflict coefficients are then allocated to their respective primary judgment categories according to the relative proportion of the dynamic confidence of the two data sources; finally, the product of the original quality function is superimposed with each allocation to obtain the fused quality function.

[0123] Specifically, the conflict coefficient K is first calculated according to the following calculation model:

[0124] K = 1 - sum_{j=1}^{M+1} q(S_j);

[0125] Among them, q(S_j) = m_A(S_j) * m_B(S_j).

[0126] The conflict coefficient K reflects the degree of conflict between two sources of evidence, and its value ranges from [0, 1].

[0127] Then, determine whether the conflict coefficient K exceeds the preset conflict threshold K_th.

[0128] Case 1: K < K_th;

[0129] At this point, the conflict between the evidence sources is relatively low, so the Dempster combination rule is directly applied for fusion. The fused quality function m = m_A ⊕ m_B is defined as:

[0130] m(S_j) = q(S_j) / (1 - K), j ∈ [1, M+1];

[0131] The final fusion result is labeled as follows:

[0132] T_m = arg max_{j ∈ [1, M+1]} m(S_j);

[0133] Case 2: K ≥ K_th;

[0134] At this point, the sources of evidence are in a state of high conflict. If 1 / (1-K) is continued to be used for normalization amplification, the only weak consistency between the two sources will be amplified into a false high-confidence conclusion. Therefore, this invention initiates the conflict backfilling process.

[0135] First, determine the primary judgment category for both data sources. The primary judgment category is defined as the label corresponding to the maximum value of the quality function among the known categories:

[0136] T_A = arg max_{i ∈ [1, M]} m_A(S_i);

[0137] T_B = arg max_{i ∈ [1, M]} m_B(S_i);

[0138] Then, conflict backfill allocation is performed according to the following formula:

[0139] K_{M+1} = K * ( (1 - η_A) + (1 - η_B) ) / 2;

[0140] K_A = (K - K_{M+1}) * η_A / (η_A + η_B);

[0141] K_B = (K - K_{M+1}) * η_B / (η_A + η_B);

[0142] Wherein, K_{M+1} is the conflict coefficient component assigned to the unknown category S_{M+1}, K_A is the conflict coefficient component assigned to the primary judgment category T_A of data source A, and K_B is the conflict coefficient component assigned to the primary judgment category T_B of data source B.

[0143] The fused mass function is:

[0144] m(S_j) = q(S_j) + I(j = T_A) * K_A + I(j = T_B) * K_B, j ∈ [1, M];

[0145] m(S_{M+1}) = q(S_{M+1}) + K_{M+1};

[0146] Where I(·) is the indicator function.

[0147] The final fusion result is labeled as follows:

[0148] T_m = arg max_{j ∈ [1, M+1]} m(S_j);

[0149] Unlike existing technologies (such as the patent in publication number CN110009251B) that use the "average evidence factor replacement" method (which essentially discards the conflict coefficient and replaces it with the average value), this invention fully preserves all the information contained in the conflict coefficient K through a conflict backfilling mechanism and redistributes it in a structured manner to the unknown category and the main judgment category of each data source. Unlike the existing technology's "partial conflict allocation strategy," which assigns conflicts to all focal elements that generated the conflict (a retrospective allocation based on the conflict source), this invention directionally backfills the conflict coefficient to the main judgment category and the unknown category—the main judgment category represents the most confident judgment of each data source, and the unknown category represents the overall uncertainty of the data source (a forward-looking allocation based on the trust direction). The two differ fundamentally in their allocation logic. Unlike existing technologies such as dynamic discount rate estimation methods that perform overall discounting on evidence before fusion (input-side correction), the conflict backfilling of this invention directionally allocates conflicts at the fusion result level, which is an output-side processing, and the technologies involved are different.

[0150] As an example, consider the extreme case where η_A = 1 and η_B = 0: In this case, K_{M+1} = K / 2, K_A = K / 2, and K_B = 0. Half of the conflict coefficient is assigned to the unknown category, and the other half is assigned to the dominant category of A. When data source B is completely unreliable, half of the conflict between it and the completely reliable data source A should be attributed to the uncertainty of B, and the other half to the difference between the category claimed by A and the random output of B. Based on the above scheme, even in extreme cases, the conflict coefficient can be reasonably allocated.

[0151] As a further preferred setting, the preset conflict threshold K_th in this invention can be dynamically configured according to the security requirements of the application scenario. For high-security scenarios, a lower K_th value (e.g., 0.6 to 0.7) is used so that the system triggers the backfill mechanism when the conflict level is relatively low; for general application scenarios, a higher K_th value (e.g., 0.8 to 0.9) is used.

[0152] As a further preferred configuration, the optimization of the conflict threshold K_th is divided into two stages: offline optimization and online inference. In the offline optimization stage, a grid search is performed using the labeled historical dataset. Within a preset threshold candidate interval, each candidate threshold is traversed with a fixed step size. The accuracy of the historical fusion results relative to the true labels is calculated for each threshold, and the threshold that maximizes the accuracy is selected as the optimal K_th. In the online inference stage, the optimal K_th determined in the offline optimization stage is used for real-time fusion. When the standardized test set is updated or the data source performance changes significantly, the offline optimization is re-executed.

[0153] (2) For the case of merging multiple data sources (N ≥ 3);

[0154] When there are N data sources (N ≥ 3), since Dempster's combination rules satisfy the commutative and associative laws, they can be combined in a pairwise recursive manner:

[0155] m = ( ( ( ( m_1 ⊕ m_2 ) ⊕ m_3 ) ⊕ m_4 ) ⊕ … ) ⊕ m_N;

[0156] Each step of the combination independently calculates the conflict coefficient K_n for that step, and selects the fusion strategy (classic Dempster combination or conflict backfilling) according to the above conflict threshold judgment logic.

[0157] When N ≥ 3, the intermediate fusion result m_AB generated after each pairwise fusion step is treated as a whole and participates in subsequent fusions. The representative confidence level η_AB of the intermediate fusion result m_AB is defined as the arithmetic mean of the confidence levels of each original data source participating in the fusion:

[0158] η_AB = (1 / k) * sum_{i ∈ Ω} η_i;

[0159] Where Ω represents the set of original data sources participating in the fusion, and k represents the number of data sources in Ω. When the intermediate fusion result m_AB is fused with the next data source C, if conflict backfilling is triggered, m_AB is regarded as a virtual data source with a confidence level of η_AB, which, together with the confidence level η_C of data source C, participates in the calculation of the backfilling formula.

[0160] Preferably, when N ≥ 3, a confidence-based fusion order optimization strategy is adopted in the engineering implementation. Specifically, given that the dynamic confidence of each data source is known, the data sources participating in the fusion are arranged in descending order of confidence, and then the sorted data sources are recursively combined with the current fusion result in pairs. This strategy prioritizes the fusion of high-confidence data sources, ensuring that high-confidence information is fully preserved in the early stages of fusion and avoiding dilution by low-confidence information in later stages.

[0161] This invention provides a feasible implementation scheme for the proposed multi-source conflict data fusion governance method based on evaluation feedback and dynamic confidence.

[0162] Specifically, in practical applications, corresponding software programs can be constructed based on the given methodologies to form a multi-source conflict data fusion and governance system based on evaluation feedback and dynamic confidence levels. When running, this software program will execute the aforementioned multi-source conflict data fusion and governance method steps and implement the corresponding functions, while simultaneously storing the data in a suitable storage medium for the processor to retrieve and execute.

[0163] like Figure 4 As shown, the multi-source conflict data fusion governance system based on evaluation feedback and dynamic confidence level proposed in this invention mainly includes four functional modules: a standardized test set maintenance module 101, a dynamic confidence level evaluation module 102, a quality function construction module 103, and a conflict fusion module 104.

[0164] The standardized test set maintenance module 101 is configured to independently build and continuously maintain a standardized test set, providing an external benchmark for dynamic confidence assessment.

[0165] Specifically, this module manages the test set by version number, generating a new version and archiving the old version with each update, making historical evaluation results reproducible.

[0166] During deployment, this module communicates with the dynamic confidence assessment module 102, providing it with standardized test set data as the input data source for the dynamic confidence assessment module 102.

[0167] The dynamic confidence assessment module 102 in this system is configured to periodically assess multiple data sources based on the standardized test set provided by the standardized test set maintenance module 101, and calculate the dynamic confidence of each data source.

[0168] Specifically, when performing confidence assessment, this module combines the current test accuracy of the data source on the test set with the historical confidence average based on adaptive sliding window statistics, and introduces a volatility-based window adaptive adjustment mechanism and a dynamic weighting mechanism based on the test set update time to achieve accurate and dynamic confidence assessment.

[0169] During deployment, the dynamic confidence assessment module 102 communicates with the standardized test set maintenance module 101 and the external data source respectively. It obtains test set data from the standardized test set maintenance module 101, sends test samples to the external data source and receives classification results. After calculation, it outputs the dynamic confidence of each data source to the quality function construction module 103 as the input parameter of the quality function construction module 103.

[0170] As an example, when the dynamic confidence assessment module 102 sends test samples to an external data source and receives classification results, it can obtain test results by calling the external data source interface.

[0171] The quality function construction module 103 in this system is configured to construct the basic probability allocation function of DS evidence theory for each piece of data to be fused, based on the dynamic confidence of the data source from which the data originates and the classification probability distribution of the data.

[0172] This module specifically couples the dynamic confidence of the data source with the classification probability of a single data point in a multiplicative manner, and directly defines the source unconfidence as the quality of the unknown category, so that even if a low-confidence data source outputs a high-probability classification result, the contribution of its mass function is systematically suppressed.

[0173] During deployment, the quality function construction module 103 is communicatively connected to the dynamic confidence assessment module 102 and the conflict fusion module 104, respectively. It obtains dynamic confidence from the dynamic confidence assessment module 102 and outputs the mass function to the conflict fusion module 104.

[0174] The conflict fusion module 104 in this system is configured to fuse multiple mass functions and output a unified classification label.

[0175] Specifically, the module is configured to calculate the conflict coefficient and then determine whether it exceeds the preset conflict threshold. If it does not exceed the threshold, the classic Dempster combination rule is used. If it exceeds the threshold, the unique conflict backfilling process is started—a portion of the conflict coefficient is allocated to the unknown category, and the remaining portion is allocated to the main judgment category of each data source according to the relative proportion of the dynamic confidence of each data source.

[0176] During deployment, the conflict fusion module 104 is connected to the quality function construction module 103, receives the mass function, calculates the conflict coefficient internally, selects a fusion strategy based on the threshold, and finally outputs the fusion result.

[0177] The multi-source conflict data fusion governance system, which is composed of the above four modules, can achieve complete data processing of "test set maintenance → confidence assessment → quality function construction → conflict fusion".

[0178] Similarly, based on an understanding of the overall functions and logical relationships of the modules in the above system, the following section will further elaborate on the specific configuration schemes of each component module from a technical implementation perspective.

[0179] As a specific example of an implementable form, the standardized test set maintenance module 101 in this system further includes a sample acquisition unit 1011, a labeling and quality control unit 1012, and a version management unit 1013.

[0180] The sample collection unit 1011 is configured to extract entity samples from the business production environment by category, covering all known categories in the recognition framework.

[0181] The sample size ratio of each category is kept consistent with the business traffic distribution of the recent period to avoid undersampling of long-tail categories.

[0182] The annotation quality control unit 1012 is configured to organize domain experts to conduct independent annotation and arbitration. Each sample is independently annotated by at least two domain experts; in case of disagreement, a third party is involved for arbitration, and only samples to which consensus is reached are included in the test set. Each sample must record at least the entity identifier, real label, annotation time, annotation source, and sample version number.

[0183] The version management unit 1013 is configured to manage test set version numbers, generating a new version and archiving the old version with each update, making historical evaluation results reproducible. Furthermore, this unit periodically adds new business samples incrementally and eliminates expired samples that have exceeded the timeliness threshold and have exited the business. This unit also continuously monitors changes in category distribution in the production flow, triggering test set reconstruction when the relative change in category proportion exceeds a preset threshold or when a new category appears.

[0184] Compared to existing technologies where test sets are mostly statically built or simply updated periodically, lacking version management and trigger-based refactoring mechanisms, the version management of this invention makes historical evaluation results reproducible, and trigger-based refactoring ensures that the test set can reflect changes in business distribution in a timely manner, making the confidence assessment always representative.

[0185] As a specific example of an implementable form, the dynamic confidence assessment module 102 in this system, when implemented, further includes a test execution unit 1021 and a confidence calculation unit 1022.

[0186] The test execution unit 1021 is configured to generate a random test set for the current test by stratified random sampling from the standardized test set according to categories.

[0187] Specifically, for category T, the number of samples drawn is n_T = round(n_ref * ρ_T), where ρ_T is the proportion of samples of this category in the current test set, and n_ref is the reference sampling value.

[0188] This unit calls the processing interface of each data source, inputs the data of the current random test set, obtains the classification results of each data source, compares the classification results with the true labels of the test set, and calculates the current test accuracy η_AC of each data source.

[0189] The confidence calculation unit 1022 is configured to calculate dynamic confidence based on the historical confidence sequence of each data source.

[0190] Specifically, this unit first calculates the volatility α_A on a reference historical sequence of length W_ref:

[0191] α_A = sqrt( (1 / W_ref) * sum_{i=1}^{W_ref}(η_{A(ti)}-η_{Aavg})^2 );

[0192] This volatility is used to quantify the stability of the historical performance of a data source; a higher volatility indicates that the recent accuracy of the data source is less stable. Unlike existing technologies that typically use fixed windows or simple moving averages to estimate the stability of a data source, this invention quantifies the dispersion of data source performance using volatility in the form of mean squared error, providing a precise mathematical basis for the adaptive adjustment of the window length.

[0193] Secondly, the unit adaptively adjusts the sliding window length W_A based on the volatility α_A:

[0194] W_A = max( W_min, min( W_max, round( W_max / (1 + λ * α_A) ) ) );

[0195] This formula establishes a quantitative inverse relationship between volatility and window length—stable sources (with smaller α) receive longer windows to smooth noise, while volatile sources (with larger α) receive shorter windows to improve response speed to changes. Unlike existing sliding window techniques that mostly use fixed lengths or simple equal-length sliding, this invention establishes a specific quantitative inverse relationship between volatility and window length and provides a complete mathematical formula and parameter configuration system, realizing a shift from "passive recording" to "active adaptation."

[0196] This unit is configured to calculate the historical confidence mean η_AH using the following calculation model:

[0197] η_AH =(1 / W_A)*sum_{i=1}^{W_A} η_{A(ti)}.

[0198] Then, this unit calculates the timeliness factor based on the most recent update time of the standardized test set. Let t_upd be the most recent update time of the standardized test set, and let Δt = t - t_upd be the timeliness interval. The update timeliness factor β is defined as β = exp(-Δt / τ). τ is a time constant; as Δt increases, the timeliness of the test set deteriorates, and β approaches 0. This unit differs from existing technologies by innovatively using the test set update time as a weighting factor for confidence assessment and employing an exponential decay model, thus achieving a continuous quantitative expression of the test set's timeliness.

[0199] Furthermore, this unit calculates the weight γ of the current test accuracy based on the time factor β:

[0200] γ = γ_min + (γ_max - γ_min) * β,

[0201] Where γ_min is the lowest retained weight of the current test accuracy, and γ_max is the highest retained weight of the current test accuracy, and 0 ≤ γ_min < γ_max ≤ 1.

[0202] By setting a minimum retention weight γ_min, even if the test set is severely outdated (β → 0), the minimum weight γ_min is still retained in the current evaluation, meaning that the latest test information will never be completely discarded.

[0203] Finally, the unit is configured to calculate the dynamic confidence level η_A of data source A using the following computational model:

[0204] η_A = γ * η_AC + (1 - γ) * η_AH.

[0205] As a specific example of an implementable form, the mass function construction module 103 in this system constructs a mass function for each piece of data to be fused, based on the dynamic confidence level η_A of the data source and the classification probability distribution p_i of that piece of data.

[0206] Specifically, for ease of description, we define the complete set of data classification labels as Ψ = {S_1, S_2, …, S_M}, where M is the total number of known categories, and the recognition frame Θ = {S_1, S_2, …, S_M, S_{M+1}}, where S_{M+1} is the unknown category.

[0207] For any data source A, its dynamic confidence level is η_A, and the classification probability distribution for a certain data output is P({S_i}) = p_i, i ∈ [1, M], satisfying sum_{i=1}^{M} p_i = 1.

[0208] Accordingly, the mass function construction module 103 is configured to construct the basic probability allocation function in the following manner:

[0209] m_A({S_{M+1}}) = 1 - η_A;

[0210] m_A({S_i}) = p_i * η_A, i ∈ [1, M];

[0211] Unlike existing weighted data structure (DS) methods that use additive discounting, this invention employs a multiplicative coupling approach. Specifically, it couples the data source confidence η_A with the probability p_i of a single data point in a multiplicative manner, directly defining the source unconfidence (1-η_A) as the quality of the unknown category S_{M+1}. The advantage of this approach is that the more unreliable the data source, the larger the "uncertain" portion of its output. Even if a low-quality data source outputs a high-probability classification result, its overall contribution is systematically suppressed.

[0212] Furthermore, the quality function construction module 103 is configured to construct a probability distribution using one-hot encoding when a data source only outputs hard classification results: the probability of that category is 1, and the probabilities of the other categories are 0. In this case, the confidence level of a single data point does not provide discrimination, and the fusion result mainly depends on the dynamic confidence level η of the data source for adjustment.

[0213] As a specific example of an implementable form, the conflict fusion module 104 in this system performs conflict fusion on multiple mass functions, specifically in two cases: fusion of two data sources and fusion of multiple data sources (N≥3).

[0214] For two data sources A and B, the conflict fusion module 104 is configured to first calculate the sum of the products of the two quality functions on each focal element, and then subtract this sum from 1 to obtain the conflict coefficient K.

[0215] K = 1 - sum_{j=1}^{M+1} q(S_j), where q(S_j) = m_A(S_j) * m_B(S_j);

[0216] The conflict coefficient K reflects the degree of conflict between two sources of evidence, and its value ranges from [0, 1].

[0217] Then, the conflict fusion module 104 determines whether the conflict coefficient K exceeds the preset conflict threshold K_th.

[0218] If K < K_th, the degree of conflict between evidence sources is low, and this module directly uses the Dempster combination rule for fusion:

[0219] m(S_j) = q(S_j) / (1 - K), j ∈ [1, M+1];

[0220] The final fusion result label is T_m = arg max_{j ∈ [1, M+1]} m(S_j).

[0221] If K ≥ K_th, the conflict fusion module 104 is set to initiate a conflict backfilling process. This conflict backfilling process fully preserves all the information contained in the conflict coefficient K and redistributes it in a structured manner. This is a "preservation-redistribution" rather than "discard-replacement" technical approach; secondly, this module backfills conflicts to the primary judgment category and the unknown category: the primary judgment category represents the most confident judgment of each data source, and the unknown category represents the overall uncertainty of the data source (a forward-looking allocation based on the trust direction); furthermore, this conflict backfilling is a targeted allocation of conflicts at the fusion result level, which belongs to output-side processing.

[0222] Specifically, the conflict fusion module 104 is configured to first determine the primary judgment category of the two data sources:

[0223] T_A = arg max_{i ∈ [1, M]} m_A(S_i);

[0224] T_B = arg max_{i ∈ [1, M]} m_B(S_i);

[0225] Then, conflict backfill allocation is performed according to the following formula:

[0226] K_{M+1} = K * ( (1 - η_A) + (1 - η_B) ) / 2;

[0227] K_A = (K - K_{M+1}) * η_A / (η_A + η_B);

[0228] K_B = (K - K_{M+1}) * η_B / (η_A + η_B);

[0229] Wherein, K_{M+1} is the conflict coefficient component assigned to the unknown category, K_A is the conflict coefficient component assigned to the primary judgment category of data source A, and K_B is the conflict coefficient component assigned to the primary judgment category of data source B.

[0230] The fused mass function is:

[0231] m(S_j) = q(S_j) + I(j = T_A) * K_A + I(j = T_B) * K_B, j ∈ [1, M];

[0232] m(S_{M+1}) = q(S_{M+1}) + K_{M+1};

[0233] The final fusion result label is T_m = arg max_{j ∈ [1, M+1]} m(S_j).

[0234] More preferably, the conflict fusion module 104 is configured to support dynamic configuration of the conflict threshold K_th: a lower K_th value (e.g., 0.6 to 0.7) is used for high-security scenarios, and a higher K_th value (e.g., 0.8 to 0.9) is used for general scenarios. The specific value of K_th is determined through offline grid search optimization.

[0235] When N ≥ 3, the conflict fusion module 104 is configured to perform pairwise recursive combinations according to the associative law of Dempster's combination rule. At each step, the conflict coefficient is calculated independently, and it is determined whether conflict backfilling is triggered. The representative confidence level of the intermediate fusion result is the arithmetic mean of the dynamic confidence levels of each original data source participating in the fusion.

[0236] As further explanation, the multi-source conflict data fusion governance system based on the aforementioned scheme, which is based on evaluation feedback and dynamic confidence, is also configured to execute the following operational logic:

[0237] After the system starts up, the standardized test set maintenance module 101 continuously maintains the standardized test set. This module incrementally collects new samples from the business environment and discards expired samples according to a preset cycle (such as weekly or monthly). When the category distribution changes significantly or a new category appears, the test set is reconstructed to ensure that the test set always maintains its timeliness and representativeness.

[0238] The dynamic confidence assessment module 102 retrieves standardized test set data from the standardized test set maintenance module 101 according to a preset assessment cycle (e.g., every 24 hours), and generates a random test set for the current period by stratified random sampling according to category. This module obtains classification results by calling the processing interfaces of each external data source and calculates the current test accuracy for each data source. Subsequently, this module reads the historical confidence sequence of each data source from the historical confidence database, calculates the volatility, adaptively adjusts the sliding window length based on the volatility, calculates the historical confidence mean, and calculates the timeliness factor and current accuracy weight according to the test set update timeliness. Finally, it weights and sums the current test accuracy with the historical confidence mean to obtain the dynamic confidence of each data source. This module stores the calculation results in the confidence database for subsequent fusion steps to query.

[0239] When an external system requests the fusion of a specific set of data, the source data source identifier and the corresponding classification probability distribution of that data are input into the system. The quality function construction module 103 queries the dynamic confidence of each data source from the confidence database based on the data source identifier, and combines it with the classification probability distribution of the data to construct the basic probability allocation function of each data source in a multiplicative coupling manner.

[0240] The conflict fusion module 104 receives each mass function, first calculates the conflict coefficient, and determines its relationship with a preset conflict threshold. If the conflict coefficient is below the threshold, it directly fuses the components using the classic Dempster combination rule; if the conflict coefficient reaches or exceeds the threshold, it initiates a conflict backfilling process. For multi-source scenarios with N ≥ 3, this module performs pairwise recursive combination according to the associative law of the Dempster combination rule, and can preferentially adopt a fusion order optimization strategy based on confidence ranking.

[0241] Finally, the module outputs the category label corresponding to the maximum value in the fused quality function as the final classification result.

[0242] Throughout the entire operation, the standardized test set maintenance module 101 and the dynamic confidence assessment module 102 run independently in the background to ensure that the data in the confidence database remains timely. The quality function construction module 103 and the conflict fusion module 104 respond immediately upon receiving an external fusion request, reading the latest dynamic confidence data from the confidence database to complete the fusion calculation. The two continuously running background modules and the two immediately responding front-end modules work together to form a complete "continuous assessment + immediate fusion" closed-loop system, which can both ensure the real-time updating of the data source confidence and respond quickly to fusion requests.

[0243] The following example, using the multi-source classification and fusion of fraudulent websites in the field of cybersecurity, illustrates in detail the application and implementation process of this invention.

[0244] (1) System deployment and configuration;

[0245] The system described in this invention was deployed at an anti-fraud center. The system is configured with three external data sources: Data Source A (a fraudulent website classification engine from an internet security vendor), Data Source B (a risk identification system from a financial institution), and Data Source C (a malicious information monitoring system from a telecommunications operator). The identification framework includes 10 known categories: S_1 = "Fake Investment and Wealth Management," S_2 = "Fake Online Loans," S_3 = "Fake Online Shopping with Rebates," S_4 = "Impersonating Customer Service," S_5 = "Fake Shopping," S_6 = "Fake Recruitment," S_7 = "Illegal Gambling," S_8 = "Dissemination of Malicious Information," S_9 = "Fake Prize Winning," S_10 = "Other Fraudulent Categories," and S_11 = "Unknown Category."

[0246] The preset conflict threshold K_th = 0.8. The dynamic confidence assessment parameters are set as follows: W_min = 3, W_max = 30, W_ref = 10, λ = 8, τ = 14 (days), γ_min = 0.30, γ_max = 0.80.

[0247] (2) Construction and trigger-based maintenance of standardized test sets;

[0248] System maintenance personnel extracted samples from historical fraudulent website data, categorized into 10 levels, with the sample size ratio for each category consistent with the business traffic distribution over the past quarter. Each sample was independently annotated by three domain experts, and any discrepancies in annotations were resolved through collective discussion and arbitration. The test set was released as version v1.0, containing 2000 annotated samples.

[0249] During system operation, the standardized test set maintenance module 101 continuously monitors the sample proportion of each category in the production flow. During operation, the system detects the emergence of a new category, "deepfake risk," and immediately triggers the test set reconstruction process: new samples covering all 11 categories (including the new category) are re-collected from the business environment, labeled and quality controlled, a v2.0 version test set is generated, and the v1.0 version is archived.

[0250] (3) Dynamic confidence assessment of data sources

[0251] The system performs a dynamic confidence assessment task every 24 hours. Taking a specific assessment as an example:

[0252] Stratified random sampling by category is performed from the standardized test set v2.3 (last updated 3 days ago), with n_ref = 200, to generate the current random test set.

[0253] The interfaces of three data sources are called for classification processing. The current test accuracy of data source A is η_AC = 0.960, the current test accuracy of data source B is η_BC = 0.890, and the current test accuracy of data source C is η_CC = 0.920.

[0254] Query the historical confidence database to calculate the volatility of each data source. Data source A has relatively low historical confidence series volatility (α_A = 0.04), classifying it as a stable source; data source B has relatively high historical confidence series volatility (α_B = 0.18), classifying it as a volatile source; data source C has moderate volatility (α_C = 0.10).

[0255] Calculate the sliding window length for each data source:

[0256] W_A = round(30 / (1 + 8 * 0.04)) = round(22.73) = 23;

[0257] W_B = round(30 / (1 + 8 * 0.18)) = round(12.30) = 12;

[0258] W_C = round(30 / (1 + 8 * 0.10)) = round(16.67) = 17.

[0259] Calculate the historical confidence mean for each data source (based on their respective sliding windows):

[0260] η_AH = 0.950;

[0261] η_BH = 0.930;

[0262] η_CH = 0.940.

[0263] Calculate the time factor (test set updated 3 days ago, τ = 14):

[0264] β = exp(-3 / 14) = 0.807.

[0265] Calculate the current test accuracy weights:

[0266] γ = 0.30 + (0.80 - 0.30) * 0.807 = 0.704.

[0267] Calculate the dynamic confidence scores for each data source:

[0268] η_A = 0.704 * 0.960 + 0.296 * 0.950 = 0.957;

[0269] η_B = 0.704 * 0.890 + 0.296 * 0.930 = 0.902;

[0270] η_C = 0.704 * 0.920 + 0.296 * 0.940 = 0.926.

[0271] (4) Conflict resolution and integration;

[0272] We now need to process a fraudulent website data point D that needs to be integrated. This website has been identified as belonging to different categories by three different data sources:

[0273] Data source A is identified as S_1 (fraudulent investment and financial management), with the following probability distributions: P(S_1) = 0.80, P(S_2) = 0.01, P(S_3) = 0.09, and the total probability of the remaining known categories is 0.10.

[0274] Data source B is identified as S_2 (fake online loan), with the following probability distributions: P(S_1) = 0.01, P(S_2) = 0.80, P(S_3) = 0.09, and the total probability of the remaining known categories is 0.10.

[0275] Data source C is identified as S_3 (brushing rebate), with the following probability distributions: P(S_1) = 0.05, P(S_2) = 0.05, P(S_3) = 0.80, and the total probability of the other known categories is 0.10.

[0276] Construct the quality function for each data source according to step S200 (with S_11 as the unknown category):

[0277] Data source A (η_A = 0.957):

[0278] m_A(S_1) = 0.80 * 0.957 = 0.7656;

[0279] m_A(S_2) = 0.01 * 0.957 = 0.00957;

[0280] m_A(S_3) = 0.09 * 0.957 = 0.08613;

[0281] m_A(S_11) = 0.043;

[0282] m_A (the rest total) = 0.0957.

[0283] Data source B (η_B = 0.902):

[0284] m_B(S_1) = 0.01 * 0.902 = 0.00902;

[0285] m_B(S_2) = 0.80 * 0.902 = 0.7216;

[0286] m_B(S_3) = 0.09 * 0.902 = 0.08118;

[0287] m_B(S_11) = 0.098;

[0288] m_B (the rest total) = 0.0902.

[0289] Data source C (η_C = 0.926):

[0290] m_C(S_1) = 0.05 * 0.926 = 0.0463;

[0291] m_C(S_2) = 0.05 * 0.926 = 0.0463;

[0292] m_C(S_3) = 0.80 * 0.926 = 0.7408;

[0293] m_C(S_11) = 0.074;

[0294] m_C (the rest total) = 0.0926.

[0295] According to the confidence ranking fusion strategy, the confidence ranking of the data sources is A (0.957) > C (0.926) > B (0.902). Therefore, A and C are fused first, and then the result is fused with B.

[0296] First, merge data source A and data source C. Calculate the q value:

[0297] q(S_1) = 0.7656 * 0.0463 = 0.03545;

[0298] q(S_2) = 0.00957 * 0.0463 = 0.00044;

[0299] q(S_3) = 0.08613 * 0.7408 = 0.06380;

[0300] q(S_11) = 0.043 * 0.074 = 0.00318;

[0301] The conflict coefficient K_AC = 1 - (0.03545 + 0.00044 + 0.06380 + 0.00318) = 1 - 0.10287 = 0.89713;

[0302] K_AC = 0.89713 > K_th = 0.8, triggering conflict backfilling.

[0303] Determine the primary judgment category: T_A = S_1, T_C = S_3.

[0304] Calculate backfill allocation:

[0305] K_11 = 0.89713 * ((1 - 0.957) + (1 - 0.926)) / 2 = 0.89713 * 0.0585 =0.05248;

[0306] K_A = (0.89713 - 0.05248) * 0.957 / (0.957 + 0.926) = 0.84465 *0.5082 = 0.42925;

[0307] K_C = (0.89713 - 0.05248) * 0.926 / (0.957 + 0.926) = 0.84465 *0.4918 = 0.41540.

[0308] Mass function after fusion:

[0309] m_AC(S_1) = q(S_1) + K_A = 0.03545 + 0.42925 = 0.46470;

[0310] m_AC(S_3) = q(S_3) + K_C = 0.06380 + 0.41540 = 0.47920;

[0311] m_AC(S_2) = q(S_2) = 0.00044;

[0312] m_AC(S_11) = q(S_11) + K_11 = 0.00318 + 0.05248 = 0.05566;

[0313] At this point, m_AC(S_3) = 0.47920 is slightly larger than m_AC(S_1) = 0.46470, indicating that after A and C are merged, the value tends to be S_3.

[0314] Then the fusion result m_AC is fused with data source B. The representative confidence level η_AC of the intermediate fusion result m_AC = (η_A + η_C) / 2 = (0.957 + 0.926) / 2 = 0.9415.

[0315] Calculate the conflict coefficient between m_AC and m_B. First, calculate the q value:

[0316] q(S_1) = 0.46470 * 0.00902 = 0.00419;

[0317] q(S_2) = 0.00044 * 0.7216 = 0.00032;

[0318] q(S_3) = 0.47920 * 0.08118 = 0.03890;

[0319] q(S_11) = 0.05566 * 0.098 = 0.00545;

[0320] The conflict coefficient K_ACB = 1 - (0.00419 + 0.00032 + 0.03890 + 0.00545) = 1 -0.04886 = 0.95114;

[0321] K_ACB = 0.95114 > K_th = 0.8, triggering conflict backfilling again.

[0322] Determine the primary judgment category: T_AC = S_3, T_B = S_2.

[0323] Backfill allocation:

[0324] K_11 = 0.95114 * ((1 - 0.9415) + (1 - 0.902)) / 2 = 0.95114 * 0.07825= 0.07442;

[0325] K_AC = (0.95114 - 0.07442) * 0.9415 / (0.9415 + 0.902) = 0.87672 *0.5107 = 0.44770;

[0326] K_B = (0.95114 - 0.07442) * 0.902 / (0.9415 + 0.902) = 0.87672 *0.4893 = 0.42898.

[0327] Final mass function:

[0328] m(S_3) = q(S_3) + K_AC = 0.03890 + 0.44770 = 0.48660;

[0329] m(S_2) = q(S_2) + K_B = 0.00032 + 0.42898 = 0.42930;

[0330] m(S_1) = q(S_1) = 0.00419;

[0331] m(S_11) = q(S_11) + K_11 = 0.00545 + 0.07442 = 0.07987

[0332] The final value is m(S_3) = 0.48660, and the output category label is S_3 (brushing orders for rebates).

[0333] (5) Technical effect analysis

[0334] In the above example, the three data sources support three different categories (S_1, S_2, S_3) with high confidence, which is a typical high-conflict multi-source fusion scenario.

[0335] If the three data sources are directly fused using the classic DS theory, since the three data sources all have a certain probability (0.09, 0.09, 0.80) on S_3, while S_1 and S_2 are only supported by a single data source with high confidence, the classic DS fusion will amplify the error of S_3 into the final result, which reflects the Zadeh paradox.

[0336] Furthermore, the present invention ensures the rationality of the fusion result through the following mechanism:

[0337] First, the dynamic confidence assessment accurately reflects the historical performance of each data source—data source A (η_A = 0.957) has historically outperformed data source B (η_B = 0.902) and data source C (η_C = 0.926), thus gaining corresponding influence in conflict backfilling.

[0338] Second, the construction of the dual-confidence quality function systematically suppresses the contribution of low-confidence data sources.

[0339] Third, the conflict backfilling mechanism backfills the conflict coefficients to their respective primary judgment categories according to the confidence ratio during each high-conflict fusion step, rather than amplifying weak consensus, thus avoiding counterintuitive fusion results.

[0340] Fourth, the confidence-based fusion order strategy allows high-confidence data sources A and C to be fused first, forming a relatively reliable intermediate result in the early stages, thus avoiding the introduction of too much noise from low-confidence data source B in the early stages.

[0341] In summary, this invention effectively solves the counterintuitive problem of multi-source data fusion in high-conflict scenarios by using a complete technical chain of "dynamic evaluation - dual confidence construction - conflict backfilling - confidence ranking fusion", and outputs a unified classification result with high confidence.

[0342] As a supplementary explanation, in relation to the provided multi-source conflict data fusion governance method based on evaluation feedback and dynamic confidence, this embodiment of the invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the steps of the above-mentioned multi-source conflict data fusion governance method.

[0343] This invention also provides a processor for running a program, wherein the program executes the steps of the above-described multi-source conflict data fusion and governance method during runtime.

[0344] This invention also provides a terminal device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. The program code is loaded and executed by the processor to implement the steps of the above-described multi-source conflict data fusion governance method.

[0345] The present invention also provides a computer program product, which, when executed on a data processing device, is adapted to perform the steps of the above-described multi-source conflict data fusion and governance method.

[0346] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0347] Finally, it should be noted that those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0348] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0349] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0350] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0351] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0352] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0353] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0354] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0355] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-source conflict data fusion governance method based on evaluation feedback and dynamic confidence, characterized in that, include: The dynamic confidence assessment step for data sources involves periodically evaluating multiple data sources based on a standardized test set, and calculating the dynamic confidence of each data source based on its current test accuracy on the standardized test set and the historical confidence average based on sliding window statistics. The quality function construction step involves constructing a basic probability allocation function for each piece of data to be fused, based on the dynamic confidence of the data source from which the data originates and the classification probability distribution of the data, calculated in the dynamic confidence assessment step. as well as The conflict fusion step involves calculating the conflict coefficients for multiple basic probability allocation functions, determining whether the conflict coefficients exceed a preset conflict threshold, and if they do not exceed the threshold, using the Dempster combination rule for fusion; if they exceed the threshold, initiating the conflict backfilling process. The conflict backfilling process includes: allocating a portion of the conflict coefficient to the unknown category, and allocating the remaining portion of the conflict coefficient to the primary judgment category of each data source according to the relative proportion of the dynamic confidence of each data source; the primary judgment category is the category corresponding to the maximum value of the basic probability allocation function among the known categories.

2. The method according to claim 1, characterized in that, The dynamic confidence assessment steps for the data source include: Test set maintenance sub-step: A standardized test set that is independently built and continuously maintained is used as an external benchmark. The standardized test set is built according to sample collection, labeling quality control and version management. The version management includes periodically and incrementally supplementing new business samples and eliminating expired samples that exceed the time limit. Test execution sub-steps: Generate the current random test set by stratified random sampling from the standardized test set according to categories, call the processing interface of each data source to classify the current random test set, and calculate the current test accuracy of each data source; The confidence score calculation sub-steps are as follows: Calculate volatility based on the historical confidence score sequences of each data source; adaptively adjust the length of the sliding window according to the volatility, with a shorter sliding window for higher volatility; calculate the mean of historical confidence scores within the sliding window as the historical confidence score mean; calculate a timeliness factor based on the most recent update time of the standardized test set; determine the weight of the current test accuracy based on the timeliness factor; and weight and sum the current test accuracy with the historical confidence score mean to obtain the dynamic confidence score.

3. The method according to claim 2, characterized in that, In the confidence calculation sub-step, the volatility of each data source is calculated based on the mean squared error of the reference historical sequence; the length of the sliding window is adaptively adjusted according to the volatility in an inverse relationship, with the window being shorter as the volatility increases. The timeliness factor is calculated based on the time interval between the current time and the most recent update time of the standardized test set in an exponential decay manner; the weight of the current test accuracy is determined based on the timeliness factor, and the weight is between the preset minimum retention weight and the maximum weight; The dynamic confidence level is the result of a weighted sum of the current test accuracy and the historical confidence average, weighted according to the specified weights.

4. The method according to claim 1, characterized in that, In the quality function construction step, for any data source, the dynamic confidence level of the data source is used as the degree to which the output information of the data source is accepted; for each known category, the classification probability of each data to be fused is multiplied by the dynamic confidence level to obtain the basic probability allocation value of the data source for each known category; for unknown categories, the difference obtained by subtracting the dynamic confidence level from 1 is used as the quality of the unknown category to absorb the uncertainty of the data source.

5. The method according to claim 1, characterized in that, For two data sources, in the conflict fusion step, the sum of the products of the two quality functions on each focal element is first calculated, and the sum is subtracted from 1 to obtain the conflict coefficient. When the conflict coefficient is less than the preset conflict threshold, the Dempster combination rule is used for direct fusion. When the conflict coefficient is greater than or equal to the preset conflict threshold, the conflict backfilling process is initiated: a portion of the conflict coefficient is allocated to the unknown category according to the average uncertainty of the two data sources, and the remaining conflict coefficient is allocated to the respective primary judgment category of the two data sources according to the relative proportion of the dynamic confidence of the two data sources. The product of the two quality functions on each focal element is superimposed with the conflict coefficient component allocated to that focal element to obtain the fused quality function, and the category corresponding to the maximum value in the fused quality function is taken as the final fusion result label.

6. The method according to claim 5, characterized in that, When there are N data sources and N≥3, the pairwise recursive combination is performed according to the associative law of Dempster's combination rule. The basic probability allocation function of the nth data source is combined with the current fusion result in turn, n=1,2,…,N. Each combination step independently calculates the conflict coefficient of the current step and determines whether the conflict backfilling process is triggered. When the current fusion result is an intermediate fusion result, the representative confidence level of the intermediate fusion result is taken as the arithmetic mean of the dynamic confidence levels of each original data source participating in the fusion.

7. A multi-source conflict data fusion governance system based on evaluation feedback and dynamic confidence, characterized in that, include: The standardized test set maintenance module (101) is configured to independently build and continuously maintain a standardized test set, which is built according to sample collection, annotation quality control and version management. The dynamic confidence assessment module (102) is configured to periodically evaluate multiple data sources based on the standardized test set, and calculate the dynamic confidence of each data source based on the current test accuracy of each data source on the standardized test set and the historical confidence average based on the sliding window statistics. The quality function construction module (103) is configured to construct the basic probability allocation function of DS evidence theory for each piece of data to be fused, based on the dynamic confidence of the data source from which the data originates and the classification probability distribution of the data, calculated by the dynamic confidence evaluation module (102). as well as The conflict fusion module (104) is configured to calculate the conflict coefficient for multiple basic probability allocation functions, determine whether the conflict coefficient exceeds the preset conflict threshold, and if it does not exceed the threshold, use the Dempster combination rule for fusion; if it exceeds the threshold, start the conflict backfilling process. The conflict backfilling process includes: allocating a portion of the conflict coefficient to the unknown category, and allocating the remaining portion of the conflict coefficient to the primary judgment category of each data source according to the relative proportion of the dynamic confidence of each data source.

8. The system according to claim 7, characterized in that, The dynamic confidence assessment module (102) includes: The test execution unit (1021) is configured to generate a random test set for the current test by stratified random sampling from the standardized test set, call the processing interface of each of the data sources to classify the random test set for the current test, and calculate the current test accuracy of each of the data sources. The confidence calculation unit (1022) is configured to calculate volatility based on the historical confidence sequence of each of the data sources, adaptively adjust the length of the sliding window according to the volatility, calculate the mean of historical confidence within the sliding window as the historical confidence mean, calculate a timeliness factor based on the most recent update time of the standardized test set, determine the weight of the current test accuracy based on the timeliness factor, and obtain the dynamic confidence by weighted summing of the current test accuracy and the historical confidence mean.

9. The system according to claim 7, characterized in that, When the conflict fusion module (104) performs the conflict backfilling process for two data sources, it allocates a portion of the conflict coefficient to the unknown category according to the average uncertainty of the two data sources, and allocates the remaining conflict coefficient to the respective main judgment category of the two data sources according to the relative proportion of the dynamic confidence of the two data sources.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • A method and system for dynamic adjustment and fusion of multi-source information in farmland

    CN110009251B