A data risk prediction method and related products
By improving the GRPO reinforcement learning algorithm and introducing a dynamic asymmetric reward and punishment mechanism, the problem of existing models being unable to distinguish risk levels is solved, and the accurate identification of high-risk content and the reliability of the system are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING LEADSEC TECH
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-26
Smart Images
Figure CN122087802A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to a data risk prediction method and related products. Background Technology
[0002] In current large-scale content moderation operations, risk tags generally employ complex hierarchical structures to accurately characterize the nature and severity of violations. However, existing reinforcement learning-based moderation models (such as systems based on the GRPO algorithm) suffer from significant technical bottlenecks in the design and application of reward functions: their reward mechanisms typically rely solely on linear scoring based on the semantic path coverage depth, assigning similar or even identical weights to violations of different severity levels, such as illegal information and minor inappropriate remarks (vulgar language, crude vocabulary). This "one-size-fits-all" scoring method ignores the essential differences between risk levels.
[0003] Especially when the model misses detections at high levels, the existing reward function cannot apply orders-of-magnitude asymmetric penalties based on the preset level weights, resulting in weakened training signals and difficulty in effectively guiding the model to prioritize avoiding high-risk misclassifications. As a result, even when the overall accuracy of the model seems acceptable, it may still produce fatal missed detections of major security risks, severely undermining the reliability and security of the review system. Summary of the Invention
[0004] To address the aforementioned issues, this application provides a data risk prediction method and related products, aiming to improve the accuracy and reliability of data risk prediction.
[0005] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides a data risk prediction method, including: Obtain the target data; The target data is input into a risk prediction model to obtain the risk prediction result of the target data output by the risk prediction model. The risk prediction model is trained using a modified GRPO reinforcement learning method. The modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function. The dynamic asymmetric reward function includes a dynamically adjusted first weight combination. The first weight combination includes multiple first weight values. The first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The dynamic asymmetric penalty function includes a dynamically adjusted second weight combination. The second weight combination includes multiple second weight values. The second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
[0006] Optionally, the training process of the risk prediction model specifically includes: Obtain a training dataset; the training dataset includes various historical data and risk labels corresponding to each historical data; the risk labels include risk levels and risk discrimination paths corresponding to the risk levels; Build the initial model; The improved GRPO reinforcement learning is used to train the initial model on the training dataset to obtain a risk prediction model.
[0007] Optionally, the step of using the improved GRPO reinforcement learning to train the initial model based on the training dataset to obtain a risk prediction model specifically includes: Set the value of k to 1, and use the initial model as the model for the first iteration; Each type of historical data in the training dataset is input into the model in the k-th iteration, resulting in multiple risk prediction results corresponding to each type of historical data output by the model in the k-th iteration; the risk prediction results include the risk prediction level and the risk discrimination path corresponding to the risk prediction level; Based on multiple risk prediction results corresponding to each type of historical data and the third weight value corresponding to each type of historical data, a reweight value is obtained; the third weight value is related to the risk level in the risk label of the historical data. The weight parameters of the model in the k-th iteration are updated based on the reassigned weights to obtain the model in the (k+1)-th iteration. If the iteration termination condition is met, the iteration ends, and the model of the (k+1)th iteration is used as the risk prediction model. If the iteration termination condition is not met, increment the value of k by 1 and return to the step of inputting each type of historical data in the training dataset into the model of the kth iteration to obtain multiple risk prediction results corresponding to each type of historical data output by the model of the kth iteration.
[0008] Optionally, the step of obtaining the reweighted value based on multiple risk prediction results corresponding to each type of historical data and the third weight value corresponding to each type of historical data specifically includes: The reward value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric reward function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. The penalty value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric penalty function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. Based on the reward and penalty values corresponding to each risk prediction result for each type of historical data, the final reward value corresponding to each risk prediction result for each type of historical data is obtained; Based on the final reward value corresponding to each risk prediction result of each type of historical data, the advantage value of each type of historical data is obtained; Based on the advantage value of each historical data and the third weight value corresponding to each historical data, the reweight value is obtained.
[0009] Optionally, the dynamic asymmetric reward function is as follows: ; in, This represents the reward value corresponding to the Gth risk prediction result of the I-th historical data; This represents the first weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the preset maximum risk level; C represents the total number of nodes in the risk planning path corresponding to the risk prediction level in the Gth risk prediction result of the I-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk discrimination path corresponding to the risk level in the risk label of the i-th historical data.
[0010] Optionally, the dynamic asymmetric penalty function is as follows: ; in, Let represent the penalty value corresponding to the Gth risk prediction result of the I-th historical data; This represents the number of nodes where the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the I-th historical data and the risk discrimination path corresponding to the risk level in the risk label of the I-th historical data first diverge. This represents the second weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk assessment path corresponding to the risk level in the risk label of the i-th historical data. This represents the confidence value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. To represent an asymmetric algorithm, only if > Triggered at time.
[0011] A second aspect of this application provides a data risk prediction device, comprising: The acquisition module is used to acquire target data; A prediction module is used to input the target data into a risk prediction model to obtain the risk prediction result of the target data output by the risk prediction model. The risk prediction model is trained using a modified GRPO reinforcement learning method. The modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function. The dynamic asymmetric reward function includes a dynamically adjusted first weight combination. The first weight combination includes multiple first weight values. The first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The dynamic asymmetric penalty function includes a dynamically adjusted second weight combination. The second weight combination includes multiple second weight values. The second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
[0012] A third aspect of this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the data risk prediction method provided in the first aspect.
[0013] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data risk prediction method provided in the first aspect.
[0014] The fifth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the data risk prediction method provided in the first aspect.
[0015] Compared with the prior art, this application has the following beneficial effects: This application includes acquiring target data; inputting the target data into a risk prediction model to obtain a risk prediction result of the target data output by the risk prediction model; the risk prediction model is trained using a modified GRPO reinforcement learning method; the modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function; the dynamic asymmetric reward function includes a dynamically adjusted first weight combination; the first weight combination includes multiple first weight values; the first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level; the dynamic asymmetric penalty function includes a dynamically adjusted second weight combination; the second weight combination includes multiple second weight values; the second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level; the risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
[0016] This application acquires the target data to be evaluated and inputs it into a risk prediction model, outputting risk prediction results, including risk levels and their corresponding risk discrimination paths. The core of the risk prediction model lies in employing an improved GRPO reinforcement learning algorithm. Its key innovation is the introduction of a dynamic asymmetric reward function and a dynamic asymmetric penalty function. The first weight value in the dynamic asymmetric reward function is dynamically determined based on the negative correlation between the risk prediction level and each node in the discrimination path: if a node has a strong indicative effect on the determination of a high-risk level, its negative correlation is weak, and its corresponding first weight value is set to a smaller value, thus giving a higher effective reward when the model correctly identifies high-risk content; conversely, if a node is semantically ambiguous and prone to false alarms, a larger first weight is assigned to suppress its interference with the reward signal.
[0017] Correspondingly, the second weight in the dynamic asymmetric penalty function is set based on the positive correlation between nodes and risk levels: when the model misses a report at a highly positively correlated high-level node, it automatically assigns a very large second weight, imposing a severe penalty across orders of magnitude; while for misjudgments of low-correlation or low-risk nodes, a lighter penalty is applied. Through this dynamic mechanism of "heavy penalty for high-risk, light penalty for low-risk, reward for high-confidence, and suppression for low-confidence," the problem of traditional auditing models treating "missed reports" and "misjudgments due to inappropriate wording" is effectively solved. During training, the model's sensitivity and interception capabilities for major security risks are significantly enhanced, greatly improving the accuracy and reliability of data risk prediction. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 A flowchart illustrating a data risk prediction method provided in this application embodiment; Figure 2 This is a structural diagram of a data risk prediction device provided in an embodiment of this application. Detailed Implementation
[0020] As described earlier, in current large-scale content moderation scenarios, risk labels typically employ a hierarchical structure to accurately reflect the type and severity of illegal content. However, existing moderation models based on reinforcement learning algorithms such as GRPO have significant flaws in their reward function design: their scoring mechanisms largely rely on linear scoring based on the coverage depth of semantic paths, treating illegal content the same as minor violations such as indecent language, ignoring the essential differences between risk levels. Especially when the model misses detections at higher levels, traditional reward functions cannot apply asymmetric penalties across orders of magnitude based on risk levels, resulting in ineffective feedback for critical errors and dilution of training signals. This means that while the model may perform reasonably well overall, it may fatally miss high-risk content, seriously threatening the reliability and robustness of the content security system.
[0021] In view of the above problems, this application provides a data risk prediction generation method and related products. The method includes: acquiring target data; inputting the target data into a risk prediction model to obtain a risk prediction result of the target data output by the risk prediction model; the risk prediction model is trained using a modified GRPO reinforcement learning method; the modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function; the dynamic asymmetric reward function includes a dynamically adjusted first weight combination; the first weight combination includes multiple first weight values; the first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level; the dynamic asymmetric penalty function includes a dynamically adjusted second weight combination; the second weight combination includes multiple second weight values; the second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level; the risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
[0022] This application significantly improves the accuracy and security of high-risk content identification by introducing a dynamic asymmetric reward and penalty mechanism. The system inputs the target data to be evaluated into a risk prediction model and outputs a structured result containing the risk level and corresponding discrimination path. Specifically, the dynamic asymmetric reward function adjusts the first weight based on the negative correlation between nodes and risk levels: nodes with high discrimination power have a smaller weight to enhance effective rewards; fuzzy nodes have a larger weight to suppress false positives. The dynamic asymmetric penalty function sets the second weight based on positive correlation: imposing orders-of-magnitude penalties on missed detections and lighter penalties on low-risk false positives; effectively distinguishing the severity of risks, addressing the shortcomings of traditional models that treat "high-risk missed detections and low-risk false positives equally," and significantly improving the reliability and security boundaries of the review system.
[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.
[0024] Figure 1 A flowchart of a data risk prediction method provided in this application embodiment is shown below. Figure 1 As shown, a data risk prediction method includes: S101: Obtain target data.
[0025] This application does not limit the specific form and source of the target data; it can refer to any data to be reviewed that requires risk assessment. For example, target data includes, but is not limited to: user-uploaded text, images, audio, video, or multimodal content; posts, comments, and private messages published on social media platforms; real-time audio and video frames in live streams; press releases, forum posts, or self-media content, etc. Regardless of the data modality, it can serve as input to the risk prediction model in this application.
[0026] S102: Input the target data into the risk prediction model to obtain the risk prediction result of the target data output by the risk prediction model.
[0027] This application does not limit the construction method of the risk prediction model. For example, the risk prediction model is trained using the improved GRPO reinforcement learning. The core of the improved GRPO reinforcement learning lies in the introduction of a dynamic asymmetric reward function and a dynamic asymmetric penalty function, both of which can adaptively adjust according to the risk prediction level output by the model during training. Specifically, the risk prediction result includes the risk level and the risk discrimination path corresponding to the risk level (i.e., the complete classification path from the root node to the leaf node).
[0028] The dynamic asymmetric reward function includes a dynamically adjusted first weight combination; the first weight combination includes multiple first weight values; the first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction results obtained during training and the risk prediction level: if a node has a strong indicative effect on the high risk level (i.e., strong positive correlation), then its negative correlation is weak, and the corresponding first weight value is set to a smaller value, thereby amplifying the effective reward when the model correctly identifies high-risk content; conversely, nodes with ambiguous semantics or that are prone to false alarms are given a larger weight due to their high negative correlation, so as to suppress their interference with the reward signal.
[0029] The dynamic asymmetric penalty function includes a dynamically adjusted second weight combination; the second weight combination includes multiple second weight values; the second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level obtained during the training process and the risk prediction level: when the model misses a high-level node with high positive correlation, it automatically assigns a very large second weight and applies a severe penalty across orders of magnitude; while for misjudgments of low-correlation or low-risk nodes, a lighter penalty is applied.
[0030] This application abandons the binary feedback mechanism of "rewarding all correct answers and penalizing all incorrect ones" in traditional review models, and instead constructs an asymmetric, fine-grained feedback system. It no longer focuses solely on the correctness of the final judgment, but dynamically evaluates the model's cognitive alignment along the entire hierarchical risk labeling path: calculating the cognitive alignment score along the entire path from the root node to the leaf node, achieving "soft edge" feedback. Even if the model has minor deviations in end-level details (such as specific behavioral subcategories), as long as the main path (such as domain or topic level) is correct, it can still receive partial rewards, thereby encouraging the model to establish a stable and coherent reasoning chain.
[0031] During training, the system automatically locates the depth at which the model's predicted path first deviates from the standard path. If the deviation occurs at a high level, it is considered a fundamental logical error, and a severe penalty spanning orders of magnitude is immediately imposed. This penalty directly covers all subsequent sub-levels, preventing the error from being masked by local corrections. The severity of the penalty is positively correlated non-linearly with the "severity of underreporting risk" (e.g., amplified quadratically or exponentially). In particular, when the model incorrectly classifies high-risk content as low-risk with high confidence, the penalty magnitude increases exponentially, effectively curbing dangerous tendencies such as "minimizing serious cases" or "underreporting with high confidence."
[0032] This application precisely addresses the technical bottleneck of existing reinforcement learning review models being unable to implement differentiated feedback based on the severity of violations. It not only corrects the logical inconsistencies in multi-layered inference but also bridges the feedback gap between the initial screening and the final review stages, significantly improving the system's accuracy in identifying high-risk content, its security constraints, and its overall robustness.
[0033] The above describes the main technical solution of this application. Further implementations of the main technical solution are now introduced. Details are as follows: Regarding the training process of the risk prediction model, this application provides an optional embodiment: Obtain the training dataset.
[0034] This application does not limit the method or specific form of obtaining the training dataset. For example, based on actual business scenarios, all violation samples in the dataset can be hierarchically risk-labeled according to preset business rules, thereby constructing a training dataset containing various historical data and their corresponding risk labels. The historical data can cover text, images, audio / video, or multimodal content, etc.; the risk labels not only include the final risk level but also the risk judgment path corresponding to the risk level, that is, a structured path that refines from the top-level domain (such as "violation information") to specific behavioral subclasses. This hierarchical labeling system provides the model with fine-grained supervision signals, supporting it in learning the logical chain and semantic hierarchy of risk judgment during training, thus laying the data foundation for the subsequent introduction of dynamic asymmetric reward and punishment mechanisms.
[0035] Build the initial model.
[0036] This application does not limit the construction method of the initial model, and can flexibly choose according to actual needs. For example, the initial risk prediction model can be built based on Transformer, BERT, ResNet, Graph Neural Network (GNN), or other deep learning architectures suitable for multimodal content understanding; alternatively, a pre-trained language model or a cross-modal large model can be used for transfer learning to leverage its powerful semantic representation capabilities. The initial model can take the form of a policy network, classifier, or sequence generation model, as long as it can receive target data input and output structured prediction results containing risk levels and corresponding risk discrimination paths. This design ensures the universality and scalability of this application under different technology stacks and business scenarios.
[0037] The improved GRPO reinforcement learning is used to train the initial model on the training dataset to obtain a risk prediction model.
[0038] This application does not limit the specific training process, but provides an optional embodiment: The improved GRPO reinforcement learning is used to train the initial model on the training dataset to obtain a risk prediction model, specifically including: Set the value of k to 1, and use the initial model as the model for the first iteration.
[0039] Each type of historical data in the training dataset is input into the model in the k-th iteration, resulting in multiple risk prediction results corresponding to each type of historical data output by the model in the k-th iteration.
[0040] This application does not limit the specific form of the risk prediction results, but to ensure interpretability, traceability and strategy consistency in practical applications, the risk prediction results should at least include the risk prediction level, the risk judgment path corresponding to the risk prediction level (a complete reasoning chain presented in a hierarchical structure, usually including nodes that are progressively refined such as "domain → topic → subcategory → specific behavior"), and the risk level score of each node in the path (i.e., the confidence or score of the model for the risk category at that node, used to quantify the strength of the judgment basis at each level).
[0041] The predicted path and its node scores can be compared longitudinally with a pre-set standard risk database. This not only verifies the compliance of the final conclusion but also confirms whether the intermediate reasoning process conforms to the business review logic. For example, if the model classifies content as "Level 1 Risk," but its scores at higher-level nodes such as "Level 2 Risk" or "Level 3 Risk" are significantly low, it may expose logical gaps or misjudgment risks. By introducing a node-level scoring mechanism, this application effectively ensures the consistency and reliability of review decisions at the deep reasoning level, avoiding problems such as "correct conclusion but incorrect path" or "high-confidence underreporting," thereby improving the transparency and security of the overall risk control system.
[0042] Based on the multiple risk prediction results corresponding to each type of historical data and the third weight value corresponding to each type of historical data, a reweight value is obtained; the third weight value is related to the risk level in the risk label of the historical data.
[0043] This application does not limit the specific calculation method for reassigning weights. Considering that the training dataset usually contains a large amount of historical data from which rich risk prediction results can be obtained, for example, 1% of the data can be extracted in real time as high-quality feedback samples within the sample range where the risk prediction results of the historical data are consistent with the corresponding risk labels. These samples can be directly used for the calculation of reassigning weights, and the model parameters can be dynamically updated based on these weights to achieve online optimization and adaptive adjustment of the policy network.
[0044] Furthermore, if numerous inconsistencies or misjudgments are detected in the risk prediction results within the monitoring window (such as frequent underreporting of high-risk content or excessive blocking of low-risk content), the system will be automatically classified as being in a "high escape risk" state. In this state, a targeted adjustment mechanism will be triggered: for each stage category (such as "first risk level" or "second risk level") appearing in the training dataset, the judgment threshold will be dynamically tightened or its node weight strengthened, enhancing the model's sensitivity and conservatism on the critical path, thereby quickly suppressing risk leakage and ensuring the security baseline of the review system. This mechanism achieves a closed-loop evolution from passive learning to proactive defense, significantly enhancing the system's robustness and responsiveness in complex adversarial environments.
[0045] The weight parameters of the model in the k-th iteration are updated based on the reassigned weights to obtain the model in the (k+1)-th iteration.
[0046] This application does not limit the specific method for updating the weight parameters, and may employ methods including but not limited to: gradient descent (such as SGD, Adam, RMSProp, etc. optimizers), policy gradient update (applicable to reinforcement learning frameworks), weighted fine-tuning based on importance sampling, or a hybrid optimization strategy combining momentum and adaptive learning rate. As long as the update mechanism can effectively integrate the risk perception signal reflected by the reassigned weights and guide the model to more accurately align the risk judgment path with the business security strategy in subsequent iterations, it falls within the scope of protection of this application.
[0047] If the iteration termination condition is met, the iteration ends, and the model of the (k+1)th iteration is used as the risk prediction model.
[0048] This application does not impose restrictions on the iteration termination condition, which can be flexibly set according to actual training needs and system resources. For example, one or more of the following conditions can be used as termination criteria: To reach the preset maximum number of iterations (e.g., 100 times), prevent the training process from looping indefinitely. If the change in reassigned weights is below a certain convergence threshold (e.g., the weight adjustment range is less than 0.001 for several consecutive rounds), it indicates that the model has stabilized. When the reassigned weights reach or approach the preset target threshold, it reflects that the degree of risk alignment meets business requirements. The high-risk false negative rate or path consistency index on the validation set meets the predetermined performance standard.
[0049] The above mechanism ensures that the training process has good convergence and computational efficiency while guaranteeing the effectiveness, taking into account both model performance and engineering feasibility.
[0050] If the iteration termination condition is not met, increment the value of k by 1 and return to the step of inputting each type of historical data in the training dataset into the model of the kth iteration to obtain multiple risk prediction results corresponding to each type of historical data output by the model of the kth iteration.
[0051] To obtain reweighted values based on multiple risk prediction results corresponding to each type of historical data and the third weight value corresponding to each type of historical data, this application provides an optional embodiment: The reward value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric reward function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data.
[0052] This application does not limit the dynamic asymmetric reward function. For example, the dynamic asymmetric reward function is as follows: ; in, This represents the reward value corresponding to the Gth risk prediction result of the I-th historical data; This represents the first weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data (satisfying...). ), usually set to The more specific the behavior, the higher the requirement for risk awareness; This indicates the preset maximum risk level (e.g., level 4); C represents the total number of nodes in the risk planning path corresponding to the risk prediction level in the Gth risk prediction result of the I-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk discrimination path corresponding to the risk level in the risk label of the i-th historical data.
[0053] This embodiment implements "soft edge rewards". If the model makes a correct judgment at a high level (such as the first level) but deviates slightly at a low level, it can still receive a partial reward instead of being directly reduced to zero.
[0054] The penalty value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric penalty function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. This application does not limit the dynamic asymmetric penalty function. For example, the dynamic asymmetric penalty function is as follows: ; in, Let represent the penalty value corresponding to the Gth risk prediction result of the I-th historical data; This represents the number of nodes where the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the I-th historical data and the risk discrimination path corresponding to the risk level in the risk label of the I-th historical data first diverge. This represents the second weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. For high-risk nodes (such as Domain level). Set to an extremely high value to ensure that the model is severely penalized when it makes a source identification error; This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk assessment path corresponding to the risk level in the risk label of the i-th historical data. The confidence value (“compliant / low risk”) of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data plays a role in adjusting the severity of the penalty. To represent an asymmetric algorithm, only if > It is triggered at any time, and the cost of missing high-risk nodes is amplified by squaring the terms.
[0055] The dynamic asymmetric risk penalty mechanism constructed in this embodiment achieves 'early stopping penalty' by identifying the first fork in the path and combining it with hierarchical weights. If the model has a high-risk false negative in the first layer, the penalty will directly cover the entire path, effectively suppressing the model's tendency to "avoid the difficult and focus on the easy".
[0056] Based on the reward and penalty values corresponding to each risk prediction result for each type of historical data, the final reward value corresponding to each risk prediction result for each type of historical data is obtained.
[0057] This application does not limit the method for determining the final reward value. For example, the final reward value is determined using the following formula: ; in, This represents the final reward value corresponding to the Gth risk prediction result of the I-th historical data; express The weight value; The accuracy reward is indicated by continuing the reward logic in the Hi-Guard framework, reflecting the degree of matching between the model's prediction results and the true labels in the final risk level. For example, a positive reward is given for correct predictions, and a penalty is given for incorrect predictions. express The weight value; The format consistency reward continues the reward logic in the Hi-Guard framework. It is used to encourage the model to output risk discrimination paths that conform to the preset structure (such as domain → topic → subclass → specific behavior), ensuring that the output format is standardized and parsable. express The weight value; express The weight value.
[0058] Based on the final reward value corresponding to each risk prediction result of each type of historical data, the advantage value of each type of historical data is obtained.
[0059] This application does not limit the method for determining the dominance value. For example, the dominance value can be determined using the following dominance value calculation formula: ; in, The mean represents the advantage value of the i-th historical data point; mean() represents the total reward value for a set of data points. The mean of the total reward is calculated as the average of all historical samples. This is used to standardize the reward signal and eliminate global bias. std() represents the standard deviation of the total reward values for the same group. It reflects the dispersion of the reward distribution and is used for normalization to give the advantage value a uniform dimension and improve training stability.
[0060] Based on the advantage value of each historical data and the third weight value corresponding to each historical data, the reweight value is obtained.
[0061] For example, the reassignment weights are determined using the following reassignment weight calculation formula: ; in, Indicates reassignment of weights; This represents the third weight value corresponding to the i-th historical data point; weights are reassigned using the dominance value. The model weights are updated by reweighting the samples according to their lower-level risk levels when calculating the intra-group dominance value. This allows the gradients of high-risk samples to have a greater dominance in backpropagation, causing the update direction of the model parameters to shift towards the accuracy of high-risk samples.
[0062] This application is highly versatile and transferable. Its core idea, the "asymmetric reward and punishment mechanism," is not only applicable to content security auditing but can also be widely extended to other high-reliability scenarios that are risk-sensitive and have complex decision-making processes. In the field of financial risk control, a four-level risk assessment system can be constructed, namely "industry → company → financial status → personal credit". When the risk prediction model misclassifies a high-risk credit application (such as one with signs of fraud or default) as low-risk, the system automatically applies an asymmetric penalty that is much higher than that of ordinary misjudgments. This forces the model to prioritize the identification and interception of potential fraudulent activities, effectively reducing bad debts and systemic financial risks.
[0063] In medical auxiliary diagnosis scenarios, a hierarchical diagnostic path can be designed, consisting of "department → anatomical system → suspected disease category → specific symptom". If the risk prediction model misdiagnoses a high-risk disease (such as myocardial infarction or stroke) as a common symptom (such as a common cold or fatigue), a severe asymmetric penalty mechanism will be triggered, significantly amplifying the training cost of such missed diagnosis errors. This will drive the model to remain highly vigilant in the early screening stage, preventing serious medical accidents caused by misdiagnosis or missed diagnosis.
[0064] All of the above applications rely on the same core technology: by modeling the risk transmission logic through structured paths, and by dynamically adjusting the intensity of rewards and punishments based on the severity of the risk consequences, they achieve intelligent decision optimization of "minor mistakes can be tolerated, but major mistakes must be investigated", providing stronger security constraints and responsibility alignment capabilities for AI systems in high-risk fields.
[0065] This application overcomes the limitations of traditional hierarchical classification models that only focus on the accuracy of the final label by constructing an explicit multi-level risk assessment path and a dynamic asymmetric feedback mechanism. It not only requires the model to output the correct risk level, but also mandates that it develop a risk perception consistent with the business security strategy at each level, from "domain → topic → subcategory → specific behavior," ensuring vertical consistency in risk severity judgment throughout the entire reasoning path. It is particularly sensitive to misjudgments at higher levels (such as the domain or topic level). If high-risk content is incorrectly classified into compliant or low-risk categories, it is considered a "source failure," triggering severe penalties, thereby significantly reducing the risk of seriously violating content flowing into the platform.
[0066] The core innovation of this application is reflected in two dimensions: First, a node-weighted reasoning consistency reward mechanism. By assigning non-uniform weights to nodes at different depths (with higher weights for levels closer to high-risk sources), the model's risk scores at each level are evaluated in relation to the standard path, and this evaluation is incorporated into the reward function. This forces the model to calibrate risk perception layer by layer during the thought chain reasoning process, achieving deep logical self-consistency from root to leaf.
[0067] Second, a quadratic asymmetric penalty mechanism based on path bifurcation points. This mechanism can accurately pinpoint the hierarchical location where the model-predicted path first deviates from the standard path. If the bifurcation point is located in a high-risk area, the penalty intensity is amplified quadratically or even exponentially, combining the severity of the risk at that node with the model's prediction confidence. For example, classifying high-risk content as ordinary entertainment with high confidence results in a penalty far exceeding linear growth, effectively curbing the dangerous tendency of "downplaying the severity" or "overconfident underreporting."
[0068] Figure 2 A structural diagram of a data risk prediction device provided in an embodiment of this application is shown below. Figure 2 As shown, based on the data risk prediction method provided in the preceding embodiments, this application also provides a data risk prediction device, including: The acquisition module is used to acquire target data.
[0069] A prediction module is used to input the target data into a risk prediction model to obtain the risk prediction result of the target data output by the risk prediction model. The risk prediction model is trained using a modified GRPO reinforcement learning method. The modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function. The dynamic asymmetric reward function includes a dynamically adjusted first weight combination. The first weight combination includes multiple first weight values. The first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The dynamic asymmetric penalty function includes a dynamically adjusted second weight combination. The second weight combination includes multiple second weight values. The second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
[0070] As an optional embodiment, the apparatus further includes: The acquisition module is also used to acquire a training dataset; the training dataset includes various historical data and risk labels corresponding to each type of historical data; the risk labels include risk levels and risk discrimination paths corresponding to the risk levels.
[0071] Build modules are used to build the initial model.
[0072] The training module is used to train the initial model based on the training dataset using the reinforcement learning of the improved GRPO to obtain a risk prediction model.
[0073] As an optional embodiment, the training module specifically includes: The setting unit is used to set the value of k to 1 and to use the initial model as the model for the first iteration.
[0074] The prediction unit is used to input each type of historical data in the training dataset into the model of the kth iteration, and obtain multiple risk prediction results corresponding to each type of historical data output by the model of the kth iteration; the risk prediction results include risk prediction level and risk discrimination path corresponding to the risk prediction level.
[0075] The calculation unit is used to obtain a reweight value based on multiple risk prediction results corresponding to each type of historical data and a third weight value corresponding to each type of historical data; the third weight value is related to the risk level in the risk label of the historical data.
[0076] The update unit is used to update the weight parameters of the model in the k-th iteration based on the reassigned weights to obtain the model in the (k+1)-th iteration.
[0077] The judgment unit is used to terminate the iteration if the iteration termination condition is met, and use the model of the (k+1)th iteration as the risk prediction model; if the iteration termination condition is not met, increment the value of k by 1, and return to the step of inputting each type of historical data in the training dataset into the model of the kth iteration to obtain multiple risk prediction results corresponding to each type of historical data output by the model of the kth iteration.
[0078] As an optional embodiment, the computing unit is specifically used for: The reward value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric reward function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. The penalty value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric penalty function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. Based on the reward and penalty values corresponding to each risk prediction result for each type of historical data, the final reward value corresponding to each risk prediction result for each type of historical data is obtained; Based on the final reward value corresponding to each risk prediction result of each type of historical data, the advantage value of each type of historical data is obtained; Based on the advantage value of each historical data and the third weight value corresponding to each historical data, the reweight value is obtained.
[0079] As an optional embodiment, the dynamic asymmetric reward function is as follows: ; in, This represents the reward value corresponding to the Gth risk prediction result of the I-th historical data; This represents the first weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the preset maximum risk level; C represents the total number of nodes in the risk planning path corresponding to the risk prediction level in the Gth risk prediction result of the I-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk discrimination path corresponding to the risk level in the risk label of the i-th historical data.
[0080] As an optional embodiment, the dynamic asymmetric penalty function is as follows: ; in, Let represent the penalty value corresponding to the Gth risk prediction result of the I-th historical data; This represents the number of nodes where the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the I-th historical data and the risk discrimination path corresponding to the risk level in the risk label of the I-th historical data first diverge. This represents the second weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk assessment path corresponding to the risk level in the risk label of the i-th historical data. This represents the confidence value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. To represent an asymmetric algorithm, only if > Triggered at time.
[0081] This application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a data risk prediction method.
[0082] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a data risk prediction method.
[0083] This application provides a computer program product, including a computer program that, when executed by a processor, implements a data risk prediction method.
[0084] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0085] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A data risk prediction method, characterized in that, The method includes: Obtain the target data; The target data is input into a risk prediction model to obtain the risk prediction result of the target data output by the risk prediction model. The risk prediction model is trained using a modified GRPO reinforcement learning method. The modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function. The dynamic asymmetric reward function includes a dynamically adjusted first weight combination. The first weight combination includes multiple first weight values. The first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The dynamic asymmetric penalty function includes a dynamically adjusted second weight combination. The second weight combination includes multiple second weight values. The second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
2. The data risk prediction method according to claim 1, characterized in that, The training process of the risk prediction model specifically includes: Obtain a training dataset; the training dataset includes various historical data and risk labels corresponding to each historical data; the risk labels include risk levels and risk discrimination paths corresponding to the risk levels; Build the initial model; The improved GRPO reinforcement learning is used to train the initial model on the training dataset to obtain a risk prediction model.
3. The data risk prediction method according to claim 2, characterized in that, The reinforcement learning method using the improved GRPO to train the initial model based on the training dataset to obtain the risk prediction model specifically includes: Set the value of k to 1, and use the initial model as the model for the first iteration; Each type of historical data in the training dataset is input into the model in the k-th iteration, resulting in multiple risk prediction results corresponding to each type of historical data output by the model in the k-th iteration; the risk prediction results include the risk prediction level and the risk discrimination path corresponding to the risk prediction level; Based on multiple risk prediction results corresponding to each type of historical data and the third weight value corresponding to each type of historical data, a reweight value is obtained; the third weight value is related to the risk level in the risk label of the historical data. The weight parameters of the model in the k-th iteration are updated based on the reassigned weights to obtain the model in the (k+1)-th iteration. If the iteration termination condition is met, the iteration ends, and the model of the (k+1)th iteration is used as the risk prediction model. If the iteration termination condition is not met, increment the value of k by 1 and return to the step of inputting each type of historical data in the training dataset into the model of the kth iteration to obtain multiple risk prediction results corresponding to each type of historical data output by the model of the kth iteration.
4. The data risk prediction method according to claim 3, characterized in that, The reassignment of weights based on multiple risk prediction results corresponding to each type of historical data and a third weight value corresponding to each type of historical data specifically includes: The reward value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric reward function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. The penalty value corresponding to each risk prediction result of each historical data is obtained by using the dynamic asymmetric penalty function based on each risk prediction result corresponding to each historical data and the risk label corresponding to each historical data. Based on the reward and penalty values corresponding to each risk prediction result for each type of historical data, the final reward value corresponding to each risk prediction result for each type of historical data is obtained; Based on the final reward value corresponding to each risk prediction result of each type of historical data, the advantage value of each type of historical data is obtained; Based on the advantage value of each historical data and the third weight value corresponding to each historical data, the reweight value is obtained.
5. The data risk prediction method according to claim 1, characterized in that, The dynamic asymmetric reward function is as follows: ； in, This represents the reward value corresponding to the Gth risk prediction result of the I-th historical data; This represents the first weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the preset maximum risk level; C represents the total number of nodes in the risk planning path corresponding to the risk prediction level in the Gth risk prediction result of the I-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk discrimination path corresponding to the risk level in the risk label of the i-th historical data.
6. The data risk prediction method according to claim 1, characterized in that, The dynamic asymmetric penalty function is as follows: ； in, Let represent the penalty value corresponding to the Gth risk prediction result of the I-th historical data; This represents the number of nodes where the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the I-th historical data and the risk discrimination path corresponding to the risk level in the risk label of the I-th historical data first diverge. This represents the second weight value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk prediction level of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. This represents the risk level of the i-th node in the risk assessment path corresponding to the risk level in the risk label of the i-th historical data. This represents the confidence value of the i-th node in the risk planning path corresponding to the risk prediction level in the G-th risk prediction result of the i-th historical data. To represent an asymmetric algorithm, only if > Triggered at any time.
7. A data risk prediction device, characterized in that, The data risk prediction device includes: The acquisition module is used to acquire target data; A prediction module is used to input the target data into a risk prediction model to obtain the risk prediction result of the target data output by the risk prediction model. The risk prediction model is trained using a modified GRPO reinforcement learning method. The modified GRPO reinforcement learning method includes a dynamic asymmetric reward function and a dynamic asymmetric penalty function. The dynamic asymmetric reward function includes a dynamically adjusted first weight combination. The first weight combination includes multiple first weight values. The first weight values are determined based on the negative correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The dynamic asymmetric penalty function includes a dynamically adjusted second weight combination. The second weight combination includes multiple second weight values. The second weight values are determined based on the positive correlation between each node in the risk discrimination path corresponding to the risk prediction level in the risk prediction result obtained during training and the risk prediction level. The risk prediction result includes a risk level and the risk discrimination path corresponding to the risk level.
8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the data risk prediction method according to any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the data risk prediction method according to any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the data risk prediction method according to any one of claims 1-6.