Intelligent decision data processing method and system

Abstract

The invention relates to an intelligent decision data processing method and system in the technical field of computer information processing and artificial intelligence, and the method comprises the steps: constructing a causal graph, and training a dual machine learning model; obtaining a candidate strategy set, and for each candidate strategy in the candidate strategy set, executing anti-fact deduction by using the dual machine learning model; constructing a comprehensive risk function and a multi-objective utility function; and taking the comprehensive risk function as a penalty term to be fused with the multi-objective utility function, constructing an objective function, solving a strategy enabling the objective function to be maximized in the candidate strategy set, and determining the strategy as a global optimal target strategy. Compared with the prior art, the method provided by the invention can remarkably improve the judgment capability.

Description

Technical Field

[0001] This application relates to the fields of computer information processing and artificial intelligence technology, and in particular to an intelligent decision-making data processing method and system. Background Technology

[0002] In current big data processing and intelligent decision-making systems, computer systems need to continuously collect massive amounts of time-series data and interaction records from distributed network nodes (such as mobile terminals, server logs, and third-party API interfaces), and use algorithmic models to predict and intervene in the future state of the system. Existing data processing architectures typically rely on statistical dashboards or rule-based automated scripts to clean and aggregate historical data, and use preset numerical thresholds to trigger corresponding control commands (e.g., triggering an alert when the monitored traffic value is lower than a set value). To optimize the system's decision parameters, existing technologies often employ online comparative testing (such as A / B testing) to verify the effects of different parameter configurations through small-scale traffic allocation.

[0003] However, although existing data processing technologies are relatively mature in structured data storage and offline analysis, significant technical bottlenecks still exist when processing highly dynamic, non-stationary real-time data streams and constructing complex decision-making logic: First, in real-world network environments, the collected time-series signals (such as clickstreams and interaction frequencies) are often accompanied by significant amounts of random noise and aperiodic fluctuations. Existing monitoring systems mostly use moving averages or static statistical thresholds to determine signal status. These methods struggle to effectively separate effective signal components from background noise in the time-frequency domain, causing the system to be unable to respond at the millisecond or second level in the early stages of sudden changes in data distribution. It often requires waiting for data to accumulate to a certain magnitude before the trend can be confirmed, resulting in severe signal perception lag.

[0004] Second, due to the layered and isolated nature of network architecture, user interaction data across different applications or terminals is typically stored in independent databases (i.e., data silos). Existing data association technologies primarily rely on matching single identifiers. This prevents the system from linking scattered, fragmented interaction events into a complete temporal logical chain, thus hindering the support for complex cross-domain intent reasoning.

[0005] Third, most mainstream models currently rely on fitting the statistical correlation of historical observation data. However, such models are inherently unable to handle causal intervention problems. When it is necessary to evaluate an intervention strategy that has never appeared in the historical database (such as implementing a completely new parameter configuration or resource allocation scheme), it is impossible to simulate the potential impact of the intervention on the system state in a virtual environment. This forces the system to rely on trial and error when facing unknown scenarios, lacking predictive capabilities.

[0006] Fourth, in existing automated generation or decision-making systems, compliance, security, or specific rule verification typically employs a post-processing architecture. This involves the model generating candidate results first, followed by screening through filters. This mechanism not only wastes computing resources on generating a large number of invalid results, but more importantly, because the generation process itself lacks the guidance of constraints, it is highly susceptible to outputting instructions that do not conform to preset rules when filters fail or response times out, posing a risk to system operation. Summary of the Invention

[0007] To at least address the technical problem of existing decision-making methods lacking predictive capabilities, an intelligent decision-making data processing method and system are provided.

[0008] According to a first aspect of the present invention, an intelligent decision-making data processing method is provided, comprising: constructing a causal graph, including hybrid nodes containing contextual features, intervention nodes containing strategies, and result nodes containing observation indicators; and training a dual machine learning model, the dual machine learning model including a result predictor for fitting the mapping relationship between hybrid nodes and result nodes, and an intervention predictor for fitting the propensity relationship between hybrid nodes and intervention nodes; obtaining a set of candidate strategies; for each candidate strategy in the set of candidate strategies, performing counterfactual inference using the dual machine learning model to deduce the potential outcome expectation of the candidate strategy; constructing a comprehensive risk function, wherein at least a return volatility indicator is included; the return volatility indicator is calculated based on the difference distribution between the potential outcome expectation and the benchmark result output by the result predictor; constructing a multi-objective utility function, the multi-objective utility function being a weighted combination of immediate gain efficiency and long-term potential value calculated based on the potential outcome expectation; fusing the comprehensive risk function as a penalty term with the multi-objective utility function to construct an objective function; solving for the strategy in the set of candidate strategies that maximizes the objective function, and determining it as the globally optimal objective strategy.

[0009] Further, the deduction to obtain the potential outcome expectation of the candidate strategy includes: the outcome predictor outputting a baseline result based on the current context features; the intervention predictor outputting a historically predisposed strategy based on the current context features; calculating the outcome residual between the actual observation result and the baseline result output by the outcome predictor, and the action residual between the actual intervention action and the historically predisposed strategy output by the intervention predictor; performing correlation fitting calculation on the outcome residual and the action residual to obtain causal effect parameters; using the causal effect parameters, mapping the deviation of the current candidate strategy relative to the historically predisposed strategy into an outcome gain, and superimposing the outcome gain onto the baseline result, thereby synthesizing the potential outcome expectation.

[0010] Furthermore, the return volatility index is calculated as follows: the difference between the expected potential outcome and the benchmark outcome is calculated to obtain the incremental return of the strategy; a probability density function of the incremental return of the strategy is constructed, and the upper limit of the integral of the probability density function is solved, such that the definite integral of the probability density function in the interval from negative infinity to the upper limit of the integral is equal to a preset tail probability threshold; the upper limit of the integral is determined as the return volatility index.

[0011] Furthermore, the comprehensive risk function also includes a distribution stability index, which is calculated by calculating the divergence between the system state prediction distribution after the candidate strategy is implemented and the historical baseline distribution. The system state prediction distribution is the probability distribution of the predicted state indicators of the system after the candidate strategy is implemented, and the historical baseline distribution is the probability distribution of the state indicators under historical stable states.

[0012] Furthermore, the construction of the multi-objective utility function includes: calculating the difference ratio between the expected potential outcome and the execution cost of the candidate strategy as the immediate gain efficiency; using the expected potential outcome as the first term of the future time window value sequence, and summing the predicted values ​​of each future time step in combination with the time discount factor as the long-term potential value.

[0013] Furthermore, constructing the objective function includes: introducing a preset risk aversion coefficient to adjust the sensitivity of the decision-making process to risk; using the risk aversion coefficient to weight the comprehensive risk function to obtain a risk penalty term; and calculating the difference between the multi-objective utility function and the risk penalty term to construct the objective function.

[0014] Furthermore, after determining the globally optimal target strategy, the method further includes: determining whether the target strategy triggers a prohibited condition in a preset security constraint rule base; if it is triggered and the corresponding rule penalty weight exceeds a preset risk tolerance threshold, then the target strategy is determined to be non-compliant. Furthermore, after determining the globally optimal target strategy, the method further includes: Case 1: detecting whether the current target object's profile tag contains vulnerable features, and simultaneously detecting whether the target strategy contains exploitative attribute tags; Case 2: detecting whether the current target object's age attribute belongs to the category of minors, and simultaneously detecting whether the content classification of the target strategy involves restricted attributes; if the candidate strategy triggers any of the above cases, the ethical verification is deemed to have failed.

[0015] Furthermore, in response to the target strategy being non-compliant or failing ethical verification, an alternative strategy that has the smallest Euclidean distance to the target strategy and meets the compliance conditions is sought as the final target strategy to be executed.

[0016] According to another aspect of the present invention, an intelligent decision data processing system is provided, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the above-described method.

[0017] The beneficial effects of this invention include: This invention achieves a leap from traditional correlation fitting to causal counterfactual inference by constructing a causal graph and training a dual machine learning model. Unlike existing technologies that rely solely on historical data statistics, this application utilizes an outcome predictor and an intervention predictor to infer candidate strategies that have never appeared before, calculating their potential expected outcomes under mandatory enforcement conditions. This mechanism endows the system with the ability to predict unknown scenarios, effectively avoiding the resource waste caused by online trial and error, and significantly improving the accuracy and foresight of decision-making.

[0018] Furthermore, by utilizing causal effect parameters in the calculation of potential outcome expectations, the unbiasedness of causal effect estimation can be ensured, thereby providing an accurate data foundation for subsequent risk assessment and solving the technical problem that traditional regression models are easily affected by confounding variables.

[0019] This invention constructs a comprehensive risk function that includes an indicator of return volatility and a multi-objective utility function that includes immediate gain efficiency and long-term potential value. Furthermore, by incorporating the long-term potential value indicator, the system can automatically suppress short-sighted behavior, balancing current conversion with full lifecycle retention.

[0020] Furthermore, this invention introduces mechanisms for compliance verification and ethical conflict verification. Unlike the post-processing interception in existing technologies, this application performs interception during the generation stage, which can significantly reduce the processing burden on the computer. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart of an intelligent decision-making data processing method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a verification and correction method according to an embodiment of the present invention. Detailed Implementation

[0022] 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 some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] First aspect of the invention like Figure 1 The intelligent decision-making data processing method shown includes: Step S1: Construct a causal graph, which includes hybrid nodes with contextual features, intervention nodes with strategies, and result nodes with observation indicators. Train a dual machine learning model, which includes a result predictor for fitting the mapping relationship between hybrid nodes and result nodes, and an intervention predictor for fitting the tendency relationship between hybrid nodes and intervention nodes. Step S2: Obtain a set of candidate strategies. For each candidate strategy in the set of candidate strategies, use the dual machine learning model to perform counterfactual inference to deduce the expected potential outcome of the candidate strategy. Step S3: Construct a comprehensive risk function, which includes at least a return volatility index; the return volatility index is calculated based on the distribution of the difference between the expected potential outcome and the benchmark outcome output by the outcome predictor. Step S4: Construct a multi-objective utility function, which is a weighted combination of immediate gain efficiency and long-term potential value calculated based on the expected potential outcome. Step S5 involves fusing the comprehensive risk function as a penalty term with the multi-objective utility function to construct an objective function. Then, the strategy that maximizes the objective function from the candidate strategy set is identified and determined as the globally optimal objective strategy. This will be explained in detail below.

[0025] Specifically, step S1 includes the following process: S101. Constructing a causal structure graph First, define the strategy space based on the business logic. (For example, specific actions such as price adjustments and channel allocation), and define a cause-and-effect graph. .

[0026] Among them, the node set It must contain at least intervention nodes, outcome nodes, and hybrid nodes. Intervention nodes include various strategies A. iExamples include price adjustments and channel changes. Result nodes include observed indicators, such as price, sales volume, and other observed results Y. i Hybrid nodes include contextual features X. i Edge set It represents the directed dependencies between nodes.

[0027] In one embodiment, context feature X i It includes object attribute components and environment state components. In this embodiment, the object attribute components are composed of discrete attribute tags of the user (such as user ID, membership level) and historical behavior statistics (such as historical average consumption amount, number of active days in the last 30 days); the environment state components are composed of timestamp features of the current decision time (such as weekday, hour), business resource status (such as current inventory) and external market indicators (such as competitor prices).

[0028] S102. Train the dual machine learning model.

[0029] To accurately estimate the net effect of the intervention strategy on the outcome, the system uses a dual machine learning framework to train two auxiliary prediction models, namely the intervention predictor and the outcome predictor, based on historical data.

[0030] Result Predictor Using machine learning models (such as random forests or neural networks) to fit confounding variables With observation results The mapping relationship between them. The outcome predictor is used to predict the baseline outcome that the system will naturally produce given the environment and object state. For example, using a regression model, it learns the mapping relationship between the given features X. i Below, the observation result Y generated by the target object i .

[0031] Intervention predictor Fitting confounding variables using classification or regression models Compared with historical intervention strategies The propensity relationship between them. This intervention predictor is used to estimate propensity scores. That is, the probability of adopting a certain strategy in a specific state. Learning from given features X... i Under what circumstances does the historical system tend to take what kind of intervention action? (A) i .

[0032] S103. Construct a causal effect estimator.

[0033] After training the two predictors mentioned above, a causal effect estimator based on orthogonalized residuals is constructed. Causal effect parameters. Calculated using the following formula:

[0034] in: The number of samples; For the first Historical actual intervention actions for each sample; For the first Historical actual observation results for each sample; The above intervention predictor is based on confounding variables The output predicted value; The above result predictor is based on confounding variables The output is the predicted value. The intervention predictor is also used to calculate propensity score estimates. It is used to adjust the sample weights.

[0035] Through the above steps, using and The product of the two residuals cuts off the confounding variables. The resulting collinearity provides an unbiased computational basis for counterfactual inference.

[0036] Step S2 is used for counterfactual reasoning calculations, and the specific execution process is as follows: S201. Obtain the set of candidate strategies.

[0037] Generate a set of candidate strategies to be evaluated, denoted as: Each candidate strategy This represents a specific strategy or action vector, such as raising prices by 5%. These strategies 'a' are related to the aforementioned 'A'. i It's the same data, but it hasn't been executed yet at the current moment; it's a counterfactual hypothetical action.

[0038] S202. Perform counterfactual reasoning.

[0039] To evaluate each candidate strategy The effect of introducing causal inference into the system Operators intervene in the model. This differs from traditional prediction models that only calculate conditional probabilities. Unlike other steps, this step calculates the time required to perform the action. The expected outcome is as follows. Computationally, this means cutting the edge from the hybrid node X to the intervention node a in the causal graph, forcing the intervention node to be assigned the value a, while keeping the direct influence path of x on the outcome node Y unchanged.

[0040] Potential results The calculation formula is:

[0041] To solve for the above expectation, the system utilizes the intervention predictor from step S1. and outcome predictor Perform bias correction calculation. Because directly using... or Making predictions will include... The resulting regularization bias leads to the system employing the following orthogonalization method, the specific calculation formula of which is:

[0042] For example, the current confounding variables For the following scenario: The user is price-sensitive, and the current period is not a major promotional period. Candidate strategies: For: Issuing 20% ​​discount coupons (e.g., numerical discounting) Calculate the causal effect parameters according to step S1. for (That is, for every 1% increase in discount, the conversion rate increases by 0.5%).

[0043] Will Input two trained intervention predictors and outcome predictor Obtain the baseline results. Based on historical data, the conversion rate for this type of user in the current environment is approximately 15%. This leads to a historical bias strategy. Based on historical data, the system typically only issues 5% coupons to these types of users. This leads to the expected potential outcome. =0.225 means that if the 20% discount coupon policy is enforced now, the user's purchase conversion rate is expected to be 22.5%.

[0044] Based on the above steps, the expected potential results of each candidate strategy can be calculated.

[0045] Step S3 is used to construct the comprehensive risk function in order to quantify each candidate strategy. The risk includes at least a return volatility indicator; the return volatility indicator is calculated based on the distribution of the difference between the expected potential outcome and the benchmark outcome output by the outcome predictor.

[0046] The formula for calculating the comprehensive risk function is as follows:

[0047] According to the above formula, comprehensive risk includes indicators of return volatility and indicators of distribution stability.

[0048] Return volatility index is expressed as . The first weighting coefficient is preset and used to adjust the system's sensitivity to the uncertainty of the result.

[0049] Value at Risk (VaR) operator at a 95% confidence level. Traditional prediction models typically output only one expected value. The average value masks potential risks. In this embodiment, the aforementioned return volatility indicator indicates: if the strategy is executed... In a 5% scenario, how much would system utility (such as revenue and throughput) decrease? This is a downside risk control indicator. Specifically, it can be used to assess the strategy. Perform multiple Monte Carlo simulations or quantile-based regression to generate the probability density function of the predicted results. Solve for the upper limit of integration. , making 0.05 is the tail probability threshold. That is The corresponding risk value, if A negative number with a large absolute value indicates a significant potential risk of loss, and this value will increase significantly. This prevents candidate strategy a from being selected.

[0050] Distribution stability index is ,in The preset second weighting coefficient is used to adjust the system's aversion to drastic state shifts. It is a divergence operator. The posterior distribution of the strategy is the probability distribution of the predicted state indicators (such as user satisfaction scores and system load rate) after executing candidate strategy a. The baseline prior distribution is the probability distribution of the system's state indicators under historical steady-state conditions. The principle is that some strategies may greatly improve returns in the short term, but they can cause drastic distortions in the system's state distribution (e.g., extreme clickbait leading to abnormal user behavior patterns). By calculating the distance between the two probability distributions, we can measure whether a strategy will disrupt the system's steady state. The smaller the value, the more moderate the system state change brought about by candidate strategy a; the larger the value, the more likely the strategy may lead to uncontrollable system state changes.

[0051] Specifically, the calculation is performed using the discrete form of the KL divergence formula:

[0052] in This represents the discrete value range of the state index.

[0053] In the above embodiments, the most critical indicator is the return volatility indicator, while the distribution stability indicator is an optional one.

[0054] In other embodiments, besides these two metrics, more metrics can be added, such as compliance. For example, a compliance indicator function can be constructed and weighted and summed with the return volatility and distribution stability terms in the above embodiments. The principle is that if a violation exists, a high penalty is directly imposed on the strategy.

[0055] Step S4 is used to construct a multi-objective utility function, which is a weighted combination of the immediate gain efficiency and long-term potential value calculated based on the expected potential outcomes. The specific execution process is as follows: S401. Construct a multi-objective utility function.

[0056] To comprehensively evaluate the payoff potential of candidate strategies, the system first constructs a multi-objective utility function. The calculation formula is as follows:

[0057] in: (i.e., the immediate gain efficiency after strategy execution). That is, the potential contribution of the strategy to the future lifecycle value of the target object; The preset weight balance coefficient ( ).

[0058] Representation strategy At the present moment The counterfactual gain efficiency. It directly depends on the calculation in step S2. The potential result output from step S2 is expected to be... Compare this to the resource consumption of the strategy itself. Specific calculation formula:

[0059] It is a strategy The inherent properties of [the policy / strategy]. For example: if the policy / strategy To issue a subsidy of 10 yuan, then If strategy In order to avoid intervention, (Minimum value).

[0060] Representation strategy Integral over the sequence of potential outcomes within a future time window. Specific calculation formula:

[0061] Indicates the execution strategy The next At each time step, the system predicts the potential future outcomes. : Preset time discount factor ( ). The preset future time window length.

[0062] Quantified the strategy The long-tail effect. Some strategies may lead to... Very high, but will lead to future It rapidly decays to 0. Introducing this term changes the multi-objective utility function. This will automatically punish such short-sighted, kill-the-goose-egg strategies, forcing the system to choose actions that can maintain long-term stability.

[0063] In the above embodiments, a weighted summation of immediate gain efficiency and long-term potential value was used. In other embodiments, geometric weighting or other types of weighting methods may also be used.

[0064] In step S5, the comprehensive risk function is fused with the multi-objective utility function as a penalty term to construct an objective function. The policy that maximizes the objective function in the candidate policy set is then determined as the globally optimal objective policy.

[0065] Specifically, the comprehensive risk value calculated in step S3 Construct a global objective function with a penalty term, and select the optimal objective policy by maximizing this objective function. The optimization equation is as follows:

[0066] The risk aversion coefficient is a system configuration parameter used to control the sensitivity of the decision-making process to risk (it can be freely configured by the enterprise).

[0067] This embodiment can achieve the pursuit Maximize while utilizing This is used as a regularization term to penalize high-risk strategies, thereby ensuring the selection of the target strategy. It is the solution that maximizes returns within the controllable risk boundary.

[0068] By applying the optimization equations described above, the globally optimal target policy can be found from multiple candidate policies. .

[0069] In one embodiment, the target strategy is determined. Then, it can be directly mapped into a set of control instructions that can be executed by a computer. If When numerical parameters are adjusted (such as price), the system generates a corresponding configuration update message; if For actions involving discrete actions (such as triggering notifications or blocking transactions), the system calls the corresponding API interface. Ultimately, the control command is sent to the execution terminal (such as an APP client, business server, or supply chain management system), thus completing the automated decision-making closed loop from data perception to physical execution.

[0070] In another embodiment, after the target policy is determined, it can be verified. If the verification passes, it can be mapped to computer-executable instructions for execution. Verification is divided into two types: compliance verification and ethical conflict verification.

[0071] Specifically, compliance verification includes: Iterate through each prohibition rule in the preset security constraint rule base (including legal regulations and industry standards, which can be manually set), and perform the following conditional judgments for target strategy a in sequence: Condition 1: Determine whether the parameters of the target policy trigger the prohibition rule defined by the current security constraint rule; Condition 2: Determine whether the current rule's preset penalty weight is higher than the preset risk tolerance threshold (e.g., 0.7).

[0072] The judgment logic is as follows: If both conditions one and two are met simultaneously (i.e., the target strategy violates regulations and is a high-risk violation), then the target strategy is deemed non-compliant.

[0073] If only condition one is met but condition two is not met (i.e., the target strategy violates the rules but falls under a low-risk exemption), or if condition one is not met (i.e., the target strategy does not violate the rules), then skip the current rule and continue to check the next prohibited rule.

[0074] If none of the above non-compliance judgments are triggered after iterating through all prohibited rules, then the target policy is ultimately determined to be compliant.

[0075] Specifically, ethical conflict verification includes two scenarios: Scenario 1: The current user profile tags include vulnerable characteristics (e.g., elderly, cognitively impaired individuals, or high-risk users with a vulnerability score exceeding 0.85, the specific score can be generated using existing technologies); at the same time, the detection target strategy includes exploitative attribute tags (e.g., price discrimination based on big data, high-pressure inducement to pay, predatory pricing). Scenario 2: The current user's age falls into the category of minors; at the same time, the content classification of the target strategy involves restricted attributes (e.g., adult content, tobacco and alcohol marketing, gambling inducement). If a candidate strategy triggers any of the above conditions, the ethical verification is deemed to have failed. Otherwise (e.g., pushing restricted content to adult users, or pushing regular public welfare content to vulnerable users), the ethical verification is deemed to have passed, and the strategy is allowed to execute.

[0076] like Figure 2 As shown, if the compliance and ethical conflict checks fail, the target strategy can be modified. This involves performing a search in the instruction vector space A to find a safe alternative instruction that retains as much of the original instruction's intent as possible. The final alternative strategy is... (Alternative strategy is) The search process is a constrained optimization problem, with the objective function being to minimize the Euclidean distance between the substitute instruction and the original instruction.

[0077] The constraint is a substitute instruction. The above compliance checks must be met. Through this step, the system will remove the original high-risk instructions. Projection correction to safe instructions . This involves a strategy space for compliance. For example, strategy A offers a 50% discount, but this violates the rule that the maximum discount rate is 30%. A new strategy needs to be developed that keeps the discount rate below 30%. This makes the Euclidean distance ( (Minimum)

[0078] In one application scenario, after executing the finalized target strategy and collecting feedback data, the model can be updated. For example, when the terminal device executes the revised strategy... Subsequently, the system continuously monitors and records the actual business feedback that follows (denoted as the actual observed value Y). real ), thereby constructing instant sample pairs (X, ,Y real Subsequently, the system uses the instant sample pairs to calculate the prediction error of the local policy evaluation model, and transforms the error into the update gradient of the model parameters through the backpropagation algorithm, and uploads it to the server.

[0079] Specifically, in order to prevent the central server from using gradient information to deduce the user's private data, the terminal device uses a differential privacy-based privacy protection mechanism to perturb the gradient before uploading the local model gradient.

[0080] In one embodiment, the perturbation processing includes two processes: gradient clipping and noise injection. 1) Gradient clipping: the locally calculated gradient norm is limited to a preset threshold range to constrain the influence weight of individual samples on the overall model. 2) Noise injection: a random noise vector conforming to a specific probability distribution is superimposed on the clipped gradient. In this embodiment, the random noise vector can be generated using a Gaussian distribution, a Laplace distribution, or other statistical noise distributions that satisfy the differential privacy definition. After the above processing, an updated gradient that retains the statistical properties of the gradient while masking the privacy information of individual points is generated and sent to the server.

[0081] Second aspect of the invention This embodiment provides an intelligent decision-making data processing system. The system can be any electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The memory stores the computer program. The processor can be a central processing unit, a graphics processing unit, a neural network processor, or a combination thereof. The processor is configured to execute the computer program to at least implement steps S1 to S5 described above.

[0082] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for intelligent decision-making data processing, characterized in that, include: Construct a causal graph, which includes hybrid nodes with contextual features, intervention nodes with strategies, and result nodes with observation indicators. Train a dual machine learning model, which includes a result predictor for fitting the mapping relationship between hybrid nodes and result nodes, and an intervention predictor for fitting the tendency relationship between hybrid nodes and intervention nodes. Obtain a set of candidate strategies, and for each candidate strategy in the set, perform counterfactual deduction using the dual machine learning model to deduce the expected potential outcome of the candidate strategy. Construct a comprehensive risk function, which includes at least a return volatility index; the return volatility index is calculated based on the distribution of the difference between the expected potential outcome and the benchmark outcome output by the outcome predictor; Construct a multi-objective utility function, which is a weighted combination of immediate gain efficiency and long-term potential value calculated based on the expected potential outcome; The comprehensive risk function is fused with the multi-objective utility function as a penalty term to construct an objective function. The policy that maximizes the objective function in the candidate policy set is then determined as the globally optimal objective policy.

2. The method according to claim 1, characterized in that, The deduction yields the expected potential results of the candidate strategy, including: The result predictor outputs a baseline result based on the current context features; The intervention predictor outputs a historical tendency strategy based on the current context features; Calculate the outcome residuals between the actual observation results and the baseline results output by the outcome predictor, and the action residuals between the actual intervention actions and the historical tendency strategies output by the intervention predictor; perform correlation fitting calculations on the outcome residuals and the action residuals to obtain causal effect parameters; Using the causal effect parameter, the deviation of the current candidate policy relative to the historical tendency policy is mapped to an outcome gain, and this outcome gain is superimposed on the baseline result to synthesize the potential outcome expectation.

3. The method according to claim 1, characterized in that, The return volatility index is calculated as follows: the difference between the expected potential outcome and the benchmark outcome is calculated to obtain the incremental return of the strategy; a probability density function of the incremental return of the strategy is constructed, and the upper limit of the integral of the probability density function is solved, such that the definite integral of the probability density function in the interval from negative infinity to the upper limit of the integral is equal to a preset tail probability threshold; the upper limit of the integral is determined as the return volatility index.

4. The method according to claim 3, characterized in that, The comprehensive risk function also includes a distribution stability index, which is calculated by calculating the divergence between the system state prediction distribution after the candidate strategy is implemented and the historical baseline distribution. The system state prediction distribution is the probability distribution of the predicted state indicators of the system after the candidate strategy is implemented, and the historical baseline distribution is the probability distribution of the state indicators under historical stable conditions.

5. The method according to claim 1, characterized in that, The construction of the multi-objective utility function includes: calculating the ratio of the difference between the expected potential outcome and the execution cost of the candidate strategy as the immediate gain efficiency; using the expected potential outcome as the first term of the future time window value sequence, and summing the predicted values ​​of each future time step in combination with the time discount factor as the long-term potential value.

6. The method according to claim 1, characterized in that, Construct the objective function, including: A preset risk aversion coefficient is introduced to adjust the sensitivity of the decision-making process to risk; the comprehensive risk function is weighted using the risk aversion coefficient to obtain a risk penalty term; the difference between the multi-objective utility function and the risk penalty term is calculated to construct the objective function.

7. The method according to claim 1, characterized in that, After determining the globally optimal target strategy, the method further includes: determining whether the target strategy triggers a prohibited condition in a preset security constraint rule base; if it is triggered and the corresponding rule penalty weight exceeds a preset risk tolerance threshold, then the target strategy is determined to be non-compliant.

8. The method according to claim 1, characterized in that, After determining the globally optimal target strategy, the method further includes: Scenario 1: Detect whether the profile tag of the current target object contains vulnerable features, and at the same time detect whether the target strategy contains exploitative attribute tags; Scenario 2: Detect whether the age attribute of the current target object falls into the category of minors, and at the same time detect whether the content classification of the target strategy involves restricted attributes; If a candidate strategy triggers any of the above conditions, the ethical verification is deemed to have failed.

9. The method according to claim 7 or 8, characterized in that, In response to the non-compliance or failure of ethical verification of the target strategy, an alternative strategy that has the smallest Euclidean distance to the target strategy and meets the compliance conditions is found and used as the final target strategy to be executed.

10. An intelligent decision-making data processing system, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method of any one of claims 1 to 9.

Images