A multi-target constraint revenue distribution method for data element circulation

Abstract

This invention discloses a multi-objective constrained benefit allocation method for data element circulation. The method includes: multi-dimensional evaluation of the contribution, data quality, reputation, and timeliness of each data provider; data quality employing a hybrid quality operator combining rule-based scoring and semantic scoring; semantic scoring supporting proxy scoring and interface with the real model, with fallback in case of anomalies; generating a comprehensive score by fusing multi-dimensional evaluation indicators based on a nonlinear coupling function; setting quality gating and reputation penalty mechanisms to filter low-quality or low-reputation participants; constructing a multi-objective utility function encompassing performance improvement, supply stability, and risk control; performing constraint optimization under budget conservation, non-negative allocation, and performance lower bound constraints to obtain a statically optimal allocation scheme; and synchronously outputting allocation fairness (Gini coefficient) as an independent monitoring indicator; further constructing a reinforcement learning environment, using multi-dimensional features as states and dynamic adjustment coefficients as actions, and learning an adaptive allocation strategy that adjusts with environmental changes through a policy gradient method. This invention can achieve a benefit allocation that balances efficiency, fairness, risk mitigation, and supply stability in multi-party data collaboration scenarios, providing reproducible and auditable evaluation results, and exhibiting significant advantages in comprehensive utility, supply stability, and risk control.

Description

Technical Field

[0001] This invention belongs to the technical field of market mechanism design and implementation of data trading platforms, specifically involving a multi-objective constrained revenue distribution method for the circulation of data elements. Background Technology

[0002] With the development of the data element market, multi-party data collaboration has become an important way to improve the performance and business value of artificial intelligence models. In a typical data element circulation scenario, the platform aggregates data resources from multiple data providers to serve downstream model training or task execution, and distributes the revenue generated from the tasks among the platform and the participating parties. How to design a reasonable revenue distribution mechanism is the core issue to ensure the healthy operation of the data element market.

[0003] The existing technology still has the following problems and shortcomings:

[0004] (1) The traditional method of allocating data based on contribution ratio only considers the data contribution of each participant and ignores the impact of factors such as data quality and reputation. When there are data providers with "low quality and high quantity", simply allocating data based on contribution will lead to low-quality data receiving unreasonable high returns, which will damage the quality environment of the data ecosystem.

[0005] (2) Although the uniform distribution method ensures formal fairness, it cannot distinguish between high-contribution and low-contribution participants, lacks incentive effect for the supply of high-quality data, and is difficult to maintain the long-term supply stability of the data market.

[0006] (3) Existing methods usually make allocation decisions only for a single optimization objective (such as maximizing efficiency or maximizing fairness), while real-world scenarios need to take into account multiple conflicting objectives such as performance improvement, supply stability, risk control and fair allocation.

[0007] (4) Existing static allocation rules cannot adapt to dynamic factors such as fluctuations in the quality of participants, changes in reputation, and fraudulent behavior in the data market, and lack the ability to adapt and adjust with changes in the environment.

[0008] (5) Existing methods lack a systematic risk control mechanism against malicious behaviors such as collusion attacks, repeated uploads, and delay attacks, which makes the allocation results easy to manipulate.

[0009] Therefore, there is an urgent need for a revenue distribution method that can comprehensively consider multiple factors, balance multiple optimization objectives, and support dynamic adaptive adjustment. Summary of the Invention

[0010] To address the aforementioned problems in existing technologies, this invention provides a multi-objective constrained benefit allocation method for data element circulation. By employing nonlinear coupled scoring, quality gating and reputation penalties, multi-objective constrained optimization, and reinforcement learning dynamic strategies, it achieves a benefit allocation method that balances efficiency, fairness, risk mitigation, and supply stability in multi-party data collaboration scenarios.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A multi-objective constrained benefit allocation method for data element circulation includes:

[0013] Step S1, Multi-dimensional feature collection and evaluation: Conduct multi-dimensional feature evaluation on the N data providers involved in the data element circulation, and obtain the contribution, data quality score, reputation score and timeliness score of each data provider respectively;

[0014] Step S2, Nonlinear Coupling Score Calculation: Based on the multidimensional evaluation features of each data provider, the comprehensive score is calculated through a nonlinear coupling function;

[0015] S3. Quality Gating and Reputation Penalty: Apply quality gating and reputation penalty processing to the comprehensive score to obtain a valid score; when the data provider's data quality score is lower than a preset quality threshold, the provider's valid score is set to zero; when the data provider's reputation score is lower than a preset reputation threshold, the provider's valid score is multiplied by a set penalty coefficient. ,in ;

[0016] S4. Weight Normalization and Budget Allocation: Based on the dynamic adjustment coefficient and effective score, calculate the allocation weight and final revenue of each data provider;

[0017] S5. Multi-objective constraint optimization solution: Construct a multi-objective utility function U, and solve for the optimal allocation scheme under budget conservation constraints, non-negativity constraints, performance lower bound constraints, and fairness constraints.

[0018] Furthermore, the contribution level mentioned in step S1 The Shapley value method is used to evaluate the marginal contribution of each data provider to the performance of downstream tasks;

[0019] The data quality score mentioned in step S1 is calculated using a hybrid quality operator:

[0020]

[0021] in, For data quality score, For quality scoring based on semantic models, For rule-based quality scoring, The rule-based quality score is a mixed weighting. The result is calculated using a weighted combination of data missing rate, duplication rate, and time drift:

[0022]

[0023] in, For data missing rate, For data duplication rate, and This corresponds to the penalty coefficient;

[0024] The credit score mentioned in step S1 is calculated based on historical transaction records;

[0025] The timeliness score mentioned in step S1 is calculated by normalizing the mapping based on the data update delay.

[0026] Furthermore, the overall score in step S2 is:

[0027]

[0028] in, For the overall score, , , , These are the coupling indices corresponding to contribution, data quality, reputation, and timeliness, respectively.

[0029] Furthermore, the allocation weights and final benefits of each data provider in step S4 are as follows:

[0030]

[0031]

[0032] in, Assigning weights to each data provider For the ultimate benefit of each data provider, For a valid score, , This is a dynamic adjustment coefficient. For the allocable total budget, The total revenue generated from the circulation of data elements. This refers to the platform's commission rate.

[0033] Furthermore, the multi-objective utility function in step S5 is:

[0034]

[0035] in, For performance improvement items, As a supply stability term, As a risk control item, , , The weighting coefficients for the corresponding items; the performance improvement items The impact of the allocation scheme on the performance of downstream tasks is evaluated using a differentiable performance surrogate function, as follows:

[0036]

[0037]

[0038] in, For baseline performance, To influence the scale parameter, For normalized effective signals, The performance achievable by the current allocation scheme;

[0039] The constraints are: , , , This is the minimum acceptable performance threshold set.

[0040] Furthermore, the supply stability term A smoothed sigmoid function is used to approximate whether the payout of each participant has reached its minimum incentive threshold.

[0041]

[0042] in, Let be the minimum incentive threshold ratio for the i-th participant. is the kurtosis parameter of the sigmoid function. For smoothing scale parameters, This is the sigmoid function.

[0043] Furthermore, the aforementioned risk control items A weighted combination of basic risk and anti-fraud risk:

[0044]

[0045] in, To score the risk of collusion, To avoid repeatedly uploading risk scores, As a risk score for delayed attacks, , , This represents the corresponding anti-fraud weighting coefficient.

[0046] Furthermore, it also includes steps to monitor the fairness of the allocation results:

[0047] Calculate the Gini coefficient of the allocation scheme as a fairness monitoring indicator:

[0048]

[0049] in, The number after sorting in ascending order Each assigned value The total number of participants; the Gini coefficient is output synchronously with each allocation result for auditable recording and post-analysis, but does not directly participate in the calculation of the constrained optimization objective function.

[0050] Furthermore, it also includes step S6:

[0051] S6. Dynamic Policy Adjustment in Reinforcement Learning: Construct a reinforcement learning environment, using the multidimensional feature vectors of each data provider, the current budget ratio, the previous step utility value, and historical supply / risk statistics as the state space, and the dynamic adjustment coefficients of each participant. As an action space, with utility increment As an immediate reward, the training policy network learns to adaptively adjust the allocation policy according to changes in the environment; wherein, at each time step, random perturbations are applied to the quality, reputation, and anti-fraud characteristics of the data provider to simulate dynamic changes in the market, and the original action output by the policy network is mapped to a non-negative adjustment coefficient after a softplus transformation. .

[0052] Furthermore, it also includes a market sustainability assessment step: calculating long-term stability indicators based on the time-series allocation results, including supply volatility, rolling supply compliance rate, risk drift, participant retention rate, and market health score; wherein the market health score is a weighted comprehensive score of supply stability, risk control, and fairness.

[0053] A multi-objective constrained benefit distribution device for data element circulation includes:

[0054] The multi-dimensional feature acquisition and evaluation unit is used to conduct multi-dimensional feature evaluations on N data providers involved in the flow of data elements, and to obtain the contribution, data quality score, reputation score and timeliness score of each data provider respectively.

[0055] The nonlinear coupling score calculation unit calculates the comprehensive score based on the multidimensional evaluation features of each data provider through a nonlinear coupling function.

[0056] The quality gating and reputation penalty unit is used to apply quality gating and reputation penalty processing to the comprehensive score to obtain a valid score; when the data provider's data quality score is lower than a preset quality threshold, the provider's valid score is set to zero; when the data provider's reputation score is lower than a preset reputation threshold, the provider's valid score is multiplied by a penalty coefficient. ,in ;

[0057] The weighted normalization and budget allocation unit, based on dynamic adjustment coefficients and effective scores, is used to calculate the allocation weights and final revenues of each data provider.

[0058] The multi-objective constrained optimization solution unit is used to construct a multi-objective utility function U, solve for the optimal allocation scheme under budget conservation constraints, non-negativity constraints and performance lower bound constraints, and simultaneously output fairness monitoring indicators.

[0059] Compared with the prior art, the beneficial effects of the present invention include:

[0060] (1) Multi-dimensional comprehensive evaluation: By using a non-linear coupling function, contribution, quality, reputation and timeliness are unified into the same scoring framework, avoiding the limitations of single-dimensional evaluation. The non-linear power exponential form naturally has the "weakest link effect", so that serious deficiencies in any dimension will significantly lower the overall score.

[0061] (2) Quality Assurance and Risk Control: The quality gating mechanism can effectively prevent low-quality data from contaminating the allocation results, and the reputation penalty mechanism reduces the allocation weight of low-reputation participants. Anti-fraud risk items (collusion, duplicate uploads, delayed attacks) provide systematic defense against malicious behavior.

[0062] (3) Multi-objective equilibrium optimization: By unifying the three objectives of performance improvement, supply stability and risk control through scalar utility function, the equilibrium trade-off of multiple objectives is achieved under the constrained optimization framework, which has a better comprehensive effect than single-objective optimization.

[0063] (4) Dynamic adaptive capability: The reinforcement learning strategy can adaptively adjust the allocation scheme according to the dynamic changes in the market environment (quality fluctuations, reputation changes, and attack behavior), overcoming the limitation that static rules cannot adapt to environmental changes.

[0064] (5) Explainable and auditable: All allocation processes can output complete intermediate calculation results (scores, weights, and utility items), supporting post-allocation auditing and interpretability analysis of allocation results. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the overall process architecture of the method of the present invention.

[0066] Figure 2 This is a schematic diagram of the algorithm code for the "multi-source quality scoring and backoff mechanism" of this invention.

[0067] Figure 3 This is a schematic diagram of the algorithm code for "nonlinear coupling, gating penalty and budget allocation" of the present invention.

[0068] Figure 4 This is a schematic diagram of the "Multi-Objective Constraint Optimization Solution (Ours-MOO)" algorithm code of the present invention.

[0069] Figure 5 This is a schematic diagram of the "Reinforcement Learning Dynamic Environment Allocation Stepping (Ours-RL)" algorithm code of the present invention.

[0070] Figure 6 This is a schematic diagram of the "NSGA-II algorithm for solving Pareto fronts" of the present invention.

[0071] Figure 7 This is a schematic diagram of the algorithm code for the "long-term sustainability assessment" of this invention.

[0072] Figure 8 A scatter plot showing the trade-off between utility and fairness for different methods.

[0073] Figure 9 Box plots comparing the health mean under different methods.

[0074] Figure 10 This is a graph showing the trend of the budget share of malicious groups changing over time in a risk attack scenario.

[0075] Figure 11 A comparison chart of supply trends for different resource allocation methods in various scenarios. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0077] Combination Figure 1 A multi-objective constrained benefit allocation method for data element circulation includes the following steps:

[0078] S1. Multidimensional Feature Acquisition and Evaluation:

[0079] There are N data providers participating in the data element circulation, with a total task revenue of R and a platform commission rate of p. The following multi-dimensional evaluations are conducted on each data provider:

[0080] (a) Contribution assessment: The contribution of each data provider was calculated using the Shapley value or an approximation thereof. Quantify the marginal contribution of its data to the performance of downstream tasks;

[0081] (b) Data quality assessment: A hybrid quality operator is used to fuse rule scoring and semantic scoring.

[0082]

[0083] in, Calculated based on data integrity metrics (missing rate, duplication rate, time drift). The semantic quality of data is evaluated based on a semantic model. The mixed weighting is used to balance the contributions of the two scoring methods. The process for multi-source quality scoring, semantic scoring API calls, and the fallback to rule-based scoring in case of API exceptions is as follows: Figure 2 As shown;

[0084] (c) Reputation assessment: Calculate the reputation score of each participant based on historical transaction records. ;

[0085] (d) Timeliness assessment: Calculate the timeliness score based on the novelty and time relevance of the data. .

[0086] S2, Nonlinear Coupling Score Calculation:

[0087] A comprehensive score is obtained by fusing multidimensional evaluation indicators using a nonlinear coupling function:

[0088]

[0089] in, , , , These are the coupling indices corresponding to contribution, quality, reputation, and timeliness, respectively. Compared to the traditional linear weighting method, the nonlinear coupling mechanism can better reflect the "weakest link effect" and interaction relationships among multiple factors. When any factor is at an extremely low level, the overall score will drop significantly, thus preventing the advantages of a single dimension from masking the shortcomings of other dimensions.

[0090] S3. Quality Gating and Reputation Penalties:

[0091] A quality gating and reputation penalty mechanism is introduced into the overall score to calculate the final effective score. :

[0092] (a) Quality gating: If If the score of the participant is zero, the participant will not participate in the distribution of profits, thus avoiding the pollution of the distribution results by low-quality data.

[0093] (b) Reputation penalty: If Then multiply the participant's valid score by the penalty coefficient. This reduces the revenue distribution weight of participants with low creditworthiness.

[0094] S4. Weight Normalization and Budget Allocation:

[0095] Based on dynamic adjustment coefficient and valid score Calculate the normalized weights and the share of revenue for each participant:

[0096]

[0097]

[0098] in, For an allocable budget, the budget conservation constraint must be satisfied. Dynamic adjustment coefficient The default value is 1 during the static optimization phase, and it is output by the policy network during the reinforcement learning phase.

[0099] The pseudocode for steps S2-S4 is as follows: Figure 3 As shown.

[0100] S5. Multi-objective constrained optimization solution:

[0101] Construct a multi-objective utility function that includes performance improvement, supply stability, and risk control:

[0102]

[0103] The definitions of each utility item are as follows:

[0104] (a) Performance Improvement Items Evaluate the impact of allocation schemes on downstream task performance using differentiable performance surrogate functions:

[0105]

[0106] in, For baseline performance, To influence the scale parameter, For normalized effective signals, ;

[0107] (b) Supply stability item The incentive achievement level of each participant is approximated using a smoothed sigmoid function.

[0108]

[0109] in, For the first The minimum incentive threshold ratio for each participant. This is the sigmoid steepness parameter. The scaling parameter is used for smoothing. This design ensures that when the participant's reward reaches its expected threshold, the supply contribution approaches 1, otherwise it approaches 0, and the entire function is continuously differentiable, making it suitable for gradient optimization.

[0110] (c) Risk control items Integrating basic data risks and anti-fraud risks:

[0111]

[0112] in, The assessment scores are for conspiracy risk, duplicate upload risk, and delayed attack risk, respectively. This corresponds to a weight. This item means that participants with higher assigned weights will incur greater risk penalties if they pose a higher risk.

[0113] (d) Fairness monitoring indicators: The Gini coefficient is used to measure the imbalance in the allocation results. It is output synchronously with each allocation result for auditable recording and post-analysis.

[0114]

[0115] in, The values ​​are the allocation values ​​sorted in ascending order. The Gini coefficient is not directly involved in the calculation of the objective function; instead, it serves as an independent monitoring indicator, providing auditable evidence of fairness for platform operators and regulators.

[0116] The constraints include: (1) Non-negativity constraints: (2) Budget conservation: (3) Performance lower bound: .

[0117] The constrained optimization problem is solved using a constrained numerical optimization algorithm. The preferred order is SLSQP for fast solution, followed by a fallback to Trust-Constrain for difficult samples. The specific Ours-MOO optimization solution process is as follows: Figure 4 As shown. During the solution process, an initial solution is first constructed by proportionally allocating effective scores, limiting the optimization variables to active participants (participants with effective scores greater than zero) to reduce the problem's dimensionality. If the solver fails to converge, it reverts to the initial proportional allocation scheme to ensure the method's robustness.

[0118] S6. Dynamic policy adjustment in reinforcement learning:

[0119] To address the dynamic characteristics of the data market, such as fluctuations in quality and changes in reputation, a reinforcement learning environment is further constructed to implement an adaptive allocation strategy.

[0120] (a) State space: composed of multidimensional feature vectors of each participant. It consists of the current budget ratio, platform commission ratio, previous step utility value, current step ratio, and historical supply / risk statistics;

[0121] (b) Action Space: The original adjustment signals of each participant output by the policy network are mapped to non-negative dynamic adjustment coefficients after gain scaling and softplus transformation. Substitute the weights into the weight formula of step S4 to calculate the final allocation.

[0122] (c) Reward function: based on utility increment As an immediate reward, the policy network learns allocation strategies that continuously improve utility.

[0123] (d) Environmental Dynamics: At each time step, random perturbations are applied to the quality, reputation, and anti-fraud characteristics of the participants to simulate supply fluctuations, quality drift, and changes in attack behavior in the real market. When the time series model is enabled, the timeliness score is also superimposed with periodic trend changes;

[0124] (e) The policy network is trained using the Proximal Policy Optimization (PPO) algorithm. The reinforcement learning environment completes state reading, action transformation, payoff allocation, utility calculation, perturbation update, and reward return within a single time step. Its Ours-RL environment stepping algorithm is as follows: Figure 5 As shown.

[0125] Optionally, the platform revenue sharing ratio p can also be incorporated into the action space, with p serving as an additional dimension of the action space, and mapped to it via a sigmoid transformation. The interval allows for joint optimization of platform commission and distribution among various parties.

[0126] This invention also includes a market sustainability assessment step:

[0127] Long-term stability indicators are calculated based on the time-series allocation results, including supply volatility, rolling supply compliance rate, risk drift, participant retention rate, and market health score. The market health score is a weighted composite score of supply stability, risk control, and fairness. Specifically, supply volatility, rolling supply compliance rate, risk drift, participant retention rate, and market health score are as follows:

[0128] (a) Supply volatility: the standard deviation of supply stability values ​​at each time step;

[0129] (b) Rolling supply compliance rate: The proportion of the average supply stability that reaches the threshold within the sliding window;

[0130] (c) Risk drift: the mean of the time differences of the risk series, a positive value indicates an upward trend in risk;

[0131] (d) Participant retention rate: The proportion of continuously active participants;

[0132] (e) Market health score: a weighted composite score of supply stability, risk control, and fairness.

[0133]

[0134] The code for the sustainability assessment algorithm is as follows: Figure 7 As shown, its output health, supply compliance rate, risk drift, and retention rate are used for long-term market performance comparisons in subsequent embodiments.

[0135] This embodiment also provides a multi-objective constrained benefit allocation device for data element circulation, including:

[0136] The multi-dimensional feature acquisition and evaluation unit is used to conduct multi-dimensional feature evaluations on N data providers involved in the flow of data elements, and to obtain the contribution, data quality score, reputation score and timeliness score of each data provider respectively.

[0137] The nonlinear coupling score calculation unit calculates the comprehensive score based on the multidimensional evaluation features of each data provider through a nonlinear coupling function.

[0138] The quality gating and reputation penalty unit is used to apply quality gating and reputation penalty processing to the comprehensive score to obtain a valid score; when the data provider's data quality score is lower than a preset quality threshold, the provider's valid score is set to zero; when the data provider's reputation score is lower than a preset reputation threshold, the provider's valid score is multiplied by a penalty coefficient. ,in ;

[0139] The weighted normalization and budget allocation unit, based on dynamic adjustment coefficients and effective scores, is used to calculate the allocation weights and final revenues of each data provider.

[0140] The multi-objective constrained optimization solution unit is used to construct a multi-objective utility function U, solve for the optimal allocation scheme under budget conservation constraints, non-negativity constraints and performance lower bound constraints, and simultaneously output fairness monitoring indicators.

[0141] This invention can achieve a distribution of benefits that balances efficiency, fairness, risk mitigation, and supply stability in multi-party data collaboration scenarios, and provides reproducible and auditable evaluation results. It has significant advantages in terms of comprehensive utility, supply stability, and risk control.

[0142] Example 1

[0143] This embodiment uses the joint training of medical image-assisted diagnostic models as an application background: The buyer is a regional consortium of top-tier hospitals, which needs to train a classification model for early identification of benign and malignant pulmonary nodules across multiple data holders without aggregating original medical records; The seller has N=12 participants, including 9 medical institutions of different levels and sizes, 2 third-party data processing companies with medical image annotation qualifications, and 1 image AI company that provides synthetic negative sample augmentation; The platform is responsible for federated learning matching and settlement, model aggregation, contribution evaluation, and the final distribution and write-back of revenue.

[0144] In this embodiment, the parameters are shown in Table 1 below.

[0145] Table 1 Parameters and Definitions

[0146]

[0147] I. S1 Multidimensional Feature Acquisition and Normalization

[0148] First, the platform collects and evaluates the multi-dimensional characteristics of each participant, specifically including:

[0149] (1) Contribution The Shapley value approximation method is used to simulate the contributions of each participant based on a log-normal distribution. and normalization makes In this embodiment, the contribution vector of each participant is represented as follows:

[0150]

[0151] (2) Data quality A hybrid calculation method combining LLM semantic scoring and rule-based scoring is adopted, with the hybrid weight set as follows: In this embodiment, rule scoring Based on missing rate (evenly distributed in) ) and repetition rate (evenly distributed in) Simulation calculation:

[0152]

[0153] Semantic scoring Generated via a pluggable scorer; in offline experiments, it is generated by a random feature proxy of sigmoid transformation; in online scenarios, it can switch to calling the real model for scoring via an OpenAI-compatible interface, and automatically falls back to rule-based scoring in case of interface failure.

[0154]

[0155] (3) Credit score This result is derived from a comprehensive assessment of historical violation records and audit feedback. In the simulation calculations of this case, the following approach was adopted. The distribution is generated to reflect the actual distribution characteristics where most participants have high credibility and a few participants have low credibility.

[0156] (4) Timeliness score The result is obtained by normalization calculation based on data update latency. This embodiment uses... Distribution generates timeliness scores .

[0157] (5) Risk characteristics: including compliance deviation Delay in rule updates and delivery delays Three risk sub-items.

[0158] (6) Abnormal participant simulation: During the compliance audit phase, the system detects abnormal participants. and A security alert was triggered. In response to the aforementioned malicious behavior, the quality gating and reputation penalty mechanism of this invention was activated: based on a preset suspicion attenuation factor, [the system] penalizes / punishes [the individual / other individuals]. and Data quality score , and reputation score , Perform downward correction logic to reduce the quality to the preset quality threshold. The following; simultaneously, additional weighted penalties will be applied to the compliance deviations of these two participating parties, specifically manifested as follows: In addition, the participating parties Due to past reputational issues triggering severe penalties, its current reputation score has dropped to [a certain level]. If the score falls below the preset reputation penalty threshold, a corresponding reduction in allocation weight will be triggered in subsequent steps.

[0159] II. S2 Nonlinear Coupling Score

[0160] After numerically pruning each component to prevent numerical instability caused by zero values, the nonlinear coupling score is calculated:

[0161]

[0162] Calculate the overall score for each participant. The participant with the highest score... ,and and Due to a double decline in both quality and reputation, the score is extremely low. , This creates a score difference of approximately 71 times, effectively distinguishing between high-quality and suspicious participants.

[0163] III. S3 Quality Gating and Reputation Penalties

[0164] For quality score For participants whose scores are zero, their valid scores are completely excluded from allocation. Regarding reputation scores... For each participant, their effective score is multiplied by a penalty coefficient. This reduces their allocation weight. After gating and penalties, only participants who meet both quality and reputation standards are eligible for further optimization.

[0165] In this embodiment, ( ), ( If the quality score is below 0.25, gating is triggered, and the effective score is reset to zero. ( , Triggering a reputation penalty will halve the effective score. . to All are controlled through a double-layer gating system.

[0166] IV. S4 Weight Normalization and Initial Allocation

[0167] Calculate the initial allocation scheme based on the proportion of effective scores. The current total valid score for all participants is: Among them, the participating parties The highest initial allocation yield ( The return weighting is mainly concentrated in to Participants Only accounting for a small percentage of the total allocation ( ), and the participating parties and The initial profit allocation is then set to zero.

[0168] V. S5 Multi-Objective Constraint Optimization

[0169] Construct the optimization problem using active participants (those with valid scores greater than zero) as the optimization variables:

[0170] Objective function:

[0171] Constraints:

[0172] Nonnegativity constraint: Established for all active participants

[0173] Budget conservation: (Apply linear equality constraints to the static model)

[0174] Performance lower bound:

[0175] Initial allocation of valid scores: .

[0176] The solution uses SciPy constraint optimization. By default, SLSQP is executed first, with maxiter not exceeding 300. ftol is configurable. If it fails, it rolls back to trust-constr with maxiter set to 800.

[0177] In this embodiment, the Sequential Quadratic Programming (SLSQP) algorithm is used with 10 active variables ( to The solution is obtained by solving on [the platform]. The convergent optimal solution indicates that the participants [are involved in the solution]. The profit distribution amount was significantly increased and adjusted to (52.25% of the allocable budget B); Participating parties to The profit distribution amount was compressed to slightly above their respective incentive thresholds; participating parties Excluded from actual allocation due to negative net gain (i.e. At this point, the system's accuracy metric meets the preset constraints, specifically manifested as follows: The overall utility value of the system is This represents a 16.3% improvement compared to the initial allocation plan, while also reducing the risk cost. It decreased by 13.0%.

[0178] VI. Experimental Results and Comparative Analysis

[0179] After 200 episodes of PPO training, the steady-state exogenous weights of the system exhibit the following distribution characteristics: for participants with moderate contributions ( to The corresponding allocation weight was slightly increased for the participants who triggered reputational penalties. Its weighting coefficient is set at 0.40 to retain a small amount of room for exploration; while for participants with serious compliance risks and malicious behavior... and Their weighting coefficients are all precisely assigned as... .

[0180] The methods of this invention, Ours-MOO (Multi-Objective Constrained Optimization Static Assignment) and Ours-RL (Reinforcement Learning Dynamic Adjustment), are compared with the following baselines:

[0181] (1) Uniform distribution: All participants receive the same adjustment coefficient. ;

[0182] (2) Shapley Proportional Allocation (Shapley_prop): The adjustment coefficient is proportional to the contribution. .

[0183] With 12 participants, a total revenue of 1000, and a platform commission of 15%, the performance metrics of each method are compared as follows:

[0184] Table 2 Comparison of Method Performance Indicators

[0185]

[0186] Experimental results show that:

[0187] Results Hard constraints and risk-weighted terms will To achieve the highest overall utility, the allocation is reduced to zero. With the lowest risk cost ; Allocate 30.72 to This leads to a significant increase in risk; Ours-RL uses the Gini penalty to distribute the over-concentrated budget across... to This approach improves fairness while slightly sacrificing utility. The two solvers complement each other: static Ours-MOO is suitable for fast solutions with known preferences, while dynamic Ours-RL is suitable for adaptive scenarios with multiple rounds of interaction. The trade-offs between utility and fairness for different methods are as follows: Figure 8 As shown.

[0188] VII. Market Sustainability Verification

[0189] In a typical volatile market scenario, multiple rounds of time series simulations were conducted to evaluate each method. The results show that the market health score of the method presented in this invention (Ours-MOO) is 0.586, higher than 0.520 for the Shapley proportional allocation and 0.544 for the uniform allocation. Box plots comparing the mean health scores of different methods after multiple experiments are shown below. Figure 9 As shown, the participant retention rate remained at 0.90%, the rolling supply compliance rate was 0.80%, the supply volatility was only 0.065, and the risk drift was a small positive value of 0.002, indicating that the allocation mechanism remained stable in long-term operation.

[0190] Example 2

[0191] This embodiment introduces two types of abnormal events based on the basic scenario of Embodiment 1 to verify the robustness of the method under dynamic disturbances. The total budget, gating parameters, etc., remain consistent with Embodiment 1.

[0192] Scenario A: Sudden Supply Shock

[0193] Participants The data quality has declined significantly due to upgrades in medical imaging equipment. Below the preset quality threshold ); Participants Some data was revoked due to previous violations. ); Participants Maintain the fake sample alarm status ( Since the quality scores of the above three samples are all below the gating threshold, the quality gating zeroing mechanism in step S3 is triggered simultaneously, and the set of samples excluded by the gating is changed from that in Example 1. Expand to In addition, the participating parties No alarms were triggered in this period, and the data quality improved to 0.4760, thus allowing it to proceed with the subsequent allocation process.

[0194] After solving the multi-objective constraint optimization problem in step S5, the results show that: [Participants] The profit distribution amount further increased to (57.21% of the allocable budget B); Participating parties Although a benefit of 9.23 was obtained in the initial allocation solution, it was ultimately zeroed out by the optimizer due to its high risk cost. At this point, the system's optimal utility value is... Compared to Example 1, only 5.4% was lost; meanwhile, since all high-risk samples were blocked through the gating mechanism, the system's risk cost was reduced. The efficiency was further reduced from 0.2158 to 0.1943. This demonstrates that, even in the extreme case of simultaneous failure of three key suppliers, the method of this invention can effectively control the system's utility loss to within 5.4% by automatically eliminating failed nodes and concentrating the budget on robust samples. The supply trends of each method in this scenario can be combined... Figure 11 Conduct observations.

[0195] Scenario B: Cooperative Attack

[0196] Participants , , , The four parties simultaneously created data discrepancies across multiple dimensions, including quality, reputation, and compliance deviations, forming a distributed and covert collaborative attack pattern in an attempt to bypass single-point gating mechanisms and enter the final profit distribution list.

[0197] After the gating mechanism in step S3, the result shows: Participating parties , , Because the credit score is lower than the preset threshold And trigger the reputation penalty mechanism; participating parties , Because the quality score is lower than the preset threshold The score was reset to zero; the participating parties The credit score just meets the requirements. (This) temporarily passed the quality and reputation gate. Subsequently, the multi-objective constraint optimization in step S5 further constitutes the system's second line of defense: because the compliance offset risks of the above four attack participants meet the requirements... This results in a negative net gain per unit budget, and therefore the optimizer performs a complete zeroing process. At this point, the participants... The optimal profit distribution amount rises to (representing 68.11% of the allocable budget B); the optimal utility value is Compared to Example 1, the loss was 19.8%; however, the system's risk cost... The budget share of malicious groups decreased significantly from 0.2158 to 0.1666 (a reduction of 22.8%). This indicates that by focusing the budget on low-risk samples in optimization decisions, the system effectively reduced the risk cost when facing coordinated attacks. The trend of the budget share of malicious groups over time in high-risk attack scenarios is shown below. Figure 10 As shown, this further illustrates the sustained suppressive effect of the present invention on the budget share of malicious groups.

[0198] Example 3

[0199] This embodiment specifies the platform commission ratio. Incorporating these variables into the optimization framework verifies that the present invention can still provide efficient resource allocation for participants under the constraints of platform economy sustainability. The basic features in this embodiment are consistent with those in Embodiment 1.

[0200] Platform cost model:

[0201] The total cost of the system consists of fixed costs. and variable costs (in It consists of two parts. Therefore, the formula for calculating the platform's net profit is:

[0202]

[0203] Set the break-even constraint as follows Substituting the parameters into the above formula, we can derive the platform's break-even commission lower limit constraint:

[0204]

[0205] Joint optimization strategy:

[0206] This embodiment employs a two-layer joint optimization strategy: the outer layer optimizes in the decision space. The inner layer performs a grid search, while the inner layer optimization targets the revenue sharing ratio of each candidate platform. The Sequential Quadratic Programming (SLSQP) algorithm is called to solve for the corresponding optimal allocation scheme. Because in the feasible region Within, the optimal overall utility of the participants The trend is monotonically decreasing; therefore, the global optimal solution of the joint optimization occurs at the boundary where the cost-preservation constraint is exactly bound, i.e. At this point, the corresponding total budget that can be allocated... The platform's net profit is Set a fixed commission rate. The joint optimization strategy obtained above The comparison is shown in Table 3.

[0207] Table 3 Results of the Joint Optimization Strategy

[0208]

[0209] This joint optimization strategy resulted in a loss of only 2.18% in the optimal overall utility for all participants, enabling the platform to transition from a loss-making state to a break-even state, thus reconciling the interests of both the platform and the participants at minimal cost. Further sensitivity analysis showed that when the platform's fixed costs... When the commission rate increases from 80 to 120, the optimal platform commission rate is... Automatically increases to 0.30; when the variable cost factor... When the commission rate is reduced from 0.50 to 0.35, the optimal platform commission rate is... This will automatically decrease to 0.169, thus allocating more budget to the data providers. The platform commission calculation formula proposed in this invention is as follows:

[0210]

[0211] This constitutes a highly operable minimum sustainable revenue-sharing formula that can adaptively and dynamically adjust to changes in the platform economy environment.

[0212] Example 4

[0213] This embodiment illustrates a multi-objective Pareto front solution method based on NSGA-II.

[0214] When users need to obtain multiple sets of trade-offs between "performance-supply-risk" rather than a single scalar optimal solution, the NSGA-II algorithm can be used to solve the Pareto front.

[0215] The optimization objective is defined as three minimization objectives: (1) (Maximize performance), (2) (Maximize supply), (3) (Minimize risk) and impose penalties for constraint violations.

[0216] The population size was set to 64, the number of generations to 60, and the standard deviation of mutation to 0.12. Each individual was encoded as the weight allocation of active participants plus the platform's commission rate. Through non-dominated ranking and crowding distance selection, the first Pareto front was extracted after iterative evolution, providing decision-makers with multiple sets of non-dominated solutions to choose from. The algorithm flow for solving the Pareto front using NSGA-II is as follows: Figure 6 As shown.

[0217] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-objective constrained revenue allocation method for data element circulation, characterized in that, include: Step S1, Multi-dimensional feature collection and evaluation: Conduct multi-dimensional feature evaluation on the N data providers involved in the data element circulation, and obtain the contribution, data quality score, reputation score and timeliness score of each data provider respectively; Step S2, Nonlinear Coupling Score Calculation: Based on the multidimensional evaluation features of each data provider, the comprehensive score is calculated through a nonlinear coupling function; Step S3, Quality Gating and Reputation Penalty: Apply quality gating and reputation penalty processing to the comprehensive score to obtain the effective score; when the data provider's data quality score is lower than a preset quality threshold, set the provider's effective score to zero; when the data provider's reputation score is lower than a preset reputation threshold, multiply the provider's effective score by a set penalty coefficient. ,in ; Step S4, Weight Normalization and Budget Allocation: Based on the dynamic adjustment coefficient and effective score, calculate the allocation weight and final revenue of each data provider; Step S5: Multi-objective constraint optimization solution: Construct a multi-objective utility function U, and solve for the optimal allocation scheme under budget conservation constraints, non-negativity constraints and performance lower bound constraints, and output fairness monitoring indicators simultaneously.

2. The multi-objective constrained benefit allocation method according to claim 1, characterized in that, The contribution level mentioned in step S1 The Shapley value method is used to evaluate the marginal contribution of each data provider to the performance of downstream tasks; The data quality score mentioned in step S1 is calculated using a hybrid quality operator: in, For data quality score, For quality scoring based on semantic models, For rule-based quality scoring, The rule-based quality score uses a mixed weighting system. The result is calculated using a weighted combination of data missing rate, duplication rate, and time drift: in, For data missing rate, For data duplication rate, and This corresponds to the penalty coefficient; The credit score mentioned in step S1 is calculated based on historical transaction records; The timeliness score mentioned in step S1 is calculated by normalizing the mapping based on the data update delay.

3. The multi-objective constrained benefit allocation method according to claim 1, characterized in that, The overall score in step S2 is: in, For the overall score, , , , These are the coupling indices corresponding to contribution, data quality, reputation, and timeliness, respectively.

4. The multi-objective constrained benefit allocation method according to claim 1, characterized in that, The allocation weights and final revenues of each data provider in step S4 are as follows: in, Assigning weights to each data provider For the ultimate benefit of each data provider, For a valid score, , This is a dynamic adjustment coefficient. For the allocable total budget, The total revenue generated from the circulation of data elements. This refers to the platform's commission rate.

5. The multi-objective constrained benefit allocation method according to claim 4, characterized in that, The multi-objective utility function in step S5 is: in, For performance improvement items, As a supply stability term, As a risk control item, , , The weighting coefficients for the corresponding items; the performance improvement items The impact of the allocation scheme on the performance of downstream tasks is evaluated using a differentiable performance surrogate function, as follows: in, For baseline performance, To influence the scale parameter, For normalized effective signals, The performance achievable by the current allocation scheme; The constraints are: , , , This is the minimum acceptable performance threshold set.

6. The multi-objective constrained benefit allocation method according to claim 5, characterized in that, The supply stability item A smoothed sigmoid function is used to approximate whether the payout of each participant has reached its minimum incentive threshold. in, Let be the minimum incentive threshold ratio for the i-th participant. is the kurtosis parameter of the sigmoid function. For smoothing scale parameters, This is the sigmoid function.

7. The multi-objective constrained benefit allocation method according to claim 6, characterized in that, The aforementioned risk control items A weighted combination of basic risk and anti-fraud risk: in, To score the risk of collusion, To avoid repeatedly uploading risk scores, As a risk score for delayed attacks, , , This represents the corresponding anti-fraud weighting coefficient.

8. The multi-objective constrained benefit allocation method according to claim 1, characterized in that, It also includes steps to monitor the fairness of the allocation results: Calculate the Gini coefficient of the allocation scheme as a fairness monitoring indicator: in, The number after sorting in ascending order Each assigned value The total number of participants; the Gini coefficient is output synchronously with each allocation result for auditable recording and post-analysis, but does not directly participate in the calculation of the constrained optimization objective function.

9. The multi-objective constrained benefit allocation method according to claim 4, characterized in that, It also includes step S6: S6. Dynamic Policy Adjustment in Reinforcement Learning: Construct a reinforcement learning environment, using the multidimensional feature vectors of each data provider, the current budget ratio, the previous step utility value, and historical supply / risk statistics as the state space, and the dynamic adjustment coefficients of each participant. As an action space, with utility increment As an immediate reward, the training policy network learns to adaptively adjust the allocation policy according to changes in the environment; wherein, at each time step, random perturbations are applied to the quality, reputation, and anti-fraud characteristics of the data provider to simulate dynamic changes in the market, and the original action output by the policy network is mapped to a non-negative adjustment coefficient after a softplus transformation. .

10. A multi-objective constrained revenue allocation device for data element circulation based on the multi-objective constrained revenue allocation method according to any one of claims 1-9, characterized in that, include: The multi-dimensional feature acquisition and evaluation unit is used to conduct multi-dimensional feature evaluations on N data providers involved in the flow of data elements, and to obtain the contribution, data quality score, reputation score and timeliness score of each data provider respectively. The nonlinear coupling score calculation unit calculates the comprehensive score based on the multidimensional evaluation features of each data provider through a nonlinear coupling function. The quality gating and reputation penalty unit is used to apply quality gating and reputation penalty processing to the comprehensive score to obtain a valid score; when the data quality score of the data provider is lower than a preset quality threshold, the valid score of the provider is set to zero; when the reputation score of the data provider is lower than a preset reputation threshold, the valid score of the provider is multiplied by a penalty coefficient. ,in ; The weighted normalization and budget allocation unit, based on dynamic adjustment coefficients and effective scores, is used to calculate the allocation weights and final revenues of each data provider. The multi-objective constrained optimization solution unit is used to construct a multi-objective utility function U, solve for the optimal allocation scheme under budget conservation constraints, non-negativity constraints and performance lower bound constraints, and simultaneously output fairness monitoring indicators.

Images