Risk control policy optimization method, electronic device, and program product

Abstract

The specification provides a risk control strategy optimization method, an electronic device and a program product. The risk control strategy optimization method comprises: obtaining a to-be-optimized risk control strategy data set; wherein the to-be-optimized risk control strategy data set comprises a to-be-optimized risk control strategy composed of a plurality of to-be-optimized risk control rules, strategy evaluation index data and rule evaluation index data calculated under a preset risk control evaluation dimension; obtaining an optimization target for constraining the risk control evaluation dimension and a candidate rule set; the candidate rule set comprises a plurality of candidate risk control rules and rule evaluation index data calculated under the risk control evaluation dimension; based on the to-be-optimized risk control strategy data set, the optimization target and the candidate rule set, a strategy optimization prompt word is constructed; the strategy optimization prompt word is input into a language model to instruct the language model to optimize in the direction of achieving the optimization target, to reconstruct the to-be-optimized risk control strategy by referring to the candidate rule set, and to output an optimized risk control strategy and a corresponding optimization reason.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular to a risk control strategy optimization method, electronic device and program product. Background Technology

[0002] In the field of financial risk control, strategy optimization is a core link that runs through the entire life cycle of the risk control system. Its essence is to continuously iterate and optimize the risk control rules in the strategy so that the risk control system can adapt to the dynamically changing risk environment, thereby improving the risk prevention and control effect.

[0003] One strategy optimization method in related technologies relies heavily on expert experience, involving manual analysis of rule effects, design of adjustment schemes, and repeated testing and verification. This method has limitations such as low efficiency, long iteration cycles, and high dependence on individual experience, making it difficult to scale.

[0004] To address the shortcomings of manual tuning, another approach to strategy tuning in related technologies introduces automated optimization algorithms, such as genetic algorithms or greedy search. However, these methods typically treat risk control rules as purely mathematical combinations of parameters for optimization, often resulting in optimized rules that lack clear logical basis and fail to meet the stringent requirements of transparency and compliance audits in the risk control field. Summary of the Invention

[0005] In view of the above, one or more embodiments of this specification provide the following technical solutions: According to a first aspect of one or more embodiments of this specification, a risk control strategy optimization method is proposed, comprising: Obtain the dataset of risk control strategies to be optimized; wherein, the dataset of risk control strategies to be optimized includes: risk control strategies to be optimized consisting of multiple risk control rules to be optimized, strategy evaluation index data of the risk control strategies to be optimized calculated under a preset risk control evaluation dimension, and rule evaluation index data of each of the risk control rules to be optimized; Obtain the optimization target for constraining the risk control assessment dimension, and the alternative rule set; the alternative rule set includes: multiple alternative risk control rules, and rule assessment index data of each of the alternative risk control rules calculated under the risk control assessment dimension; Based on the dataset of risk control strategies to be optimized, the optimization target, and the set of alternative rules, construct strategy optimization prompt words; The strategy optimization prompts are input into the language model to instruct it to optimize towards the optimization goal. The risk control strategy to be optimized is reconstructed with reference to the alternative rule set, and the optimized risk control strategy and the corresponding optimization reasons are output.

[0006] According to a second aspect of the embodiments of this specification, an electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; Wherein, when the processor executes the executable instructions, it is used to implement the method described in the first aspect.

[0007] According to a third aspect of the embodiments of this specification, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in the first aspect.

[0008] According to a fourth aspect of the embodiments of this specification, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0009] As can be seen from the above embodiments, this specification describes the strategy evaluation index data and rule evaluation index data uniformly under the preset risk control evaluation dimension, and clarifies the optimization target as a constraint on the risk control evaluation dimension, so that the risk control strategy optimization process has a unified evaluation benchmark and optimization direction, thereby avoiding optimization deviations caused by inconsistent evaluation standards at different levels.

[0010] By constructing strategy optimization prompts based on the dataset of risk control strategies to be optimized, optimization objectives, and alternative rule sets, and inputting these prompts into a language model, the language model is guided to reconstruct the risk control strategy with reference to clear optimization objectives, relevant strategy evaluation index data, and rule evaluation index data. Finally, the optimized risk control strategy and the corresponding optimization reasons are output together, achieving traceable and explainable strategy optimization. The generated optimized risk control strategy has clear evaluation basis and rule selection logic, which helps to improve the interpretability and auditability of the strategy optimization process and meets the application requirements of the risk control system in terms of compliance and stability.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this specification. Attached Figure Description

[0012] Figure 1 This is a flowchart of a risk control strategy optimization method provided in an exemplary embodiment.

[0013] Figure 2 This is a timing diagram of a risk control strategy optimization process provided in an exemplary embodiment.

[0014] Figure 3 This is a flowchart of a risk control strategy optimization and iteration provided in an exemplary embodiment.

[0015] Figure 4 This is a timing diagram for determining the compliance of an initial risk control strategy, provided in an exemplary embodiment.

[0016] Figure 5 This is a timing diagram of a strategy simplification process provided in an exemplary embodiment.

[0017] Figure 6 This is a sequence diagram of a risk control strategy optimization report generation process provided in an exemplary embodiment.

[0018] Figure 7 This is a schematic diagram of the structure of a device provided in an exemplary embodiment. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this manual are all information and data authorized by the user or fully authorized by all parties. The collection, use and processing of related data shall comply with the relevant laws, regulations and standards of the relevant regions, and corresponding operation portals shall be provided for users to choose to authorize or refuse.

[0021] The following explains the relevant terms mentioned in the embodiments of this specification: (1) Language model.

[0022] In various embodiments of this specification, the language model is used for risk control strategy optimization, risk control strategy compliance determination, and generation of risk control strategy optimization reports.

[0023] Language models can be text processing models built based on machine learning or deep learning, such as sequence modeling models based on neural networks, models based on attention mechanisms, models based on encoder-decoder structures, or combinations of the above models. The embodiments in this specification do not limit the specific algorithm type, network structure, or parameter size of the language model.

[0024] Regarding model training and adaptation, one or more of the following methods can be used for training or optimization: ① Domain-adaptive training: Based on pre-training with general corpora, further pre-training or fine-tuning can be performed using textual data related to the risk control domain (including but not limited to feature variable description documents, rule configuration descriptions, and historical risk control strategy optimization experience) to enhance the model's ability to understand the semantics of the risk control domain. ② Prompt engineering and context learning: By carefully designing input prompt word templates and utilizing the model's context learning capabilities, the model can be guided to complete the risk control strategy optimization task without changing its parameters.

[0025] In terms of deployment and invocation, this language model can be deployed as a standalone service and invoked by the mining agent described in this manual via an API interface; it can also be integrated into the system as an embedded component.

[0026] (2) Risk control strategy.

[0027] It is a set of predefined logical criteria and action rules in a risk control system to achieve automated risk decision-making. Its core function is to evaluate the input risk-related data (such as user behavior, transaction information, credit records, etc.) in real time, and determine the risk level or type according to the predefined logic, thereby triggering corresponding risk handling actions.

[0028] A risk control strategy represents a complete solution that achieves a specific risk management objective, such as identifying fraudulent transactions or assessing credit risk. To achieve this objective, a risk control strategy must contain at least one risk control rule. These risk control rules can be executed in parallel, sequentially, or according to priority, and together they constitute the specific logical implementation of the strategy.

[0029] The components of a risk control rule include: ① Risk characteristic variables: These are the data inputs upon which the rules rely for logical judgment. They refer to quantitative or qualitative indicators extracted from users, transactions, devices, or historical behavior that can characterize the degree of risk, such as the amount of this transaction, the number of abnormal logins by a user in the past 30 days, or the distance between the application location and the user's usual location. Variables are the bridge connecting the raw data and the rule logic.

[0030] ② Judgment Condition: This is the core logical expression of the rule, specifying how to evaluate risk characteristic variables. It typically consists of an operator and a specific threshold (such as the numerical value 1000, or a list [location A, location B]). Operators include comparison operators (>, <, ≥, ≤, =) and logical connectors (& for AND, | for OR). For example, a transaction amount > 1000 or the application location not being in the frequently used location list. This condition determines when the rule will be triggered.

[0031] ③ Decision Result (Optional): This is the explicit action or output executed by the system after the rule is triggered (i.e., the judgment condition is true), representing the conclusion of this risk decision. Typical decision results include: rejecting the transaction, approving the application, transferring to manual review, sending an alert, or increasing the risk score by 10 points. It directly transforms logical judgment into concrete action. In the field of financial risk control, strategy optimization is a core link throughout the entire lifecycle of the risk control system. Its essence is to continuously iterate and optimize the risk control rules in the strategy, allowing the risk control system to adapt to the dynamically changing risk environment, thereby improving the effectiveness of risk prevention and control.

[0032] In various risk control scenarios such as financial lending, online transactions, and account security, risk control strategies are the core basis for identifying and intercepting various risky behaviors. Their effectiveness directly affects the control of risk losses and the accuracy of risk identification. As risk scenarios become increasingly complex and risk patterns rapidly evolve, risk control strategies need continuous optimization and iteration to adapt to the dynamically changing risk environment.

[0033] Traditional risk control strategies rely on expert experience for manual optimization, which is limited by low efficiency, long iteration cycles, and difficulty in scaling. While automated optimization algorithms improve this problem, they treat risk control rules as combinations of parameters for mathematical optimization, resulting in rules that lack logical basis and fail to meet the stringent requirements of transparency and compliance audits in the risk control field.

[0034] In recent years, large language models have demonstrated powerful analytical and generative capabilities, but their application in the field of risk control has been limited to polishing strategy texts or providing post-event explanations of decision results, and has not yet penetrated into the core decision-making process of strategy generation and iterative optimization.

[0035] Based on this, please refer to Figure 1 S100~S106 and Figure 2 Sections 21-27 of this specification provide a risk control strategy optimization method, which can be executed by a strategy optimization intelligent agent. The strategy optimization intelligent agent can be implemented as a software entity or logical function module running on an electronic device, including but not limited to servers, cloud computing platform nodes, terminal computing devices, and other hardware carriers with data processing and model invocation capabilities. The strategy optimization intelligent agent coordinates the entire process logic, including acquiring the dataset of the risk control strategy to be optimized, constructing prompt words, invoking the language model, and outputting the risk control strategy. Through centralized scheduling and automated execution of multi-step processing logic, it effectively reduces the degree of manual intervention, improves the stability and execution efficiency of the strategy optimization process, and facilitates deployment and expansion in different computing platforms or system environments.

[0036] Risk control strategy optimization methods include: In S100, the dataset of risk control strategies to be optimized is obtained; the dataset of risk control strategies to be optimized includes: risk control strategies to be optimized consisting of multiple risk control rules to be optimized, strategy evaluation index data of risk control strategies to be optimized calculated under the preset risk control evaluation dimension, and rule evaluation index data of each risk control rule to be optimized.

[0037] Among them, the risk control strategy to be optimized is a set of rules built for specific risk prevention and control scenarios. For example, in the credit risk prevention and control scenario, the risk control rules to be optimized may include: if the applicant has more than 3 overdue credit records, the application will be blocked; if the applicant's monthly income is less than 5,000 yuan, the application will be marked as a customer group to watch.

[0038] In one optional implementation, the acquisition of strategy evaluation metric data and rule evaluation metric data can be achieved by calling a strategy evaluation tool. Please refer to [link to relevant documentation]. Figure 2 The strategy optimization agent can call independently deployed strategy evaluation tools. The strategy evaluation tools are software modules with functions such as sample data loading, strategy simulation execution, and indicator quantification calculation. They can be deployed on the same electronic device as the strategy optimization agent or achieve data interaction through network communication.

[0039] The strategy optimization agent sends an evaluation request to the strategy evaluation tool. This request includes the risk control strategy to be optimized and pre-collected sample data. The pre-collected sample data is a set of historical risk-related samples, including samples marked as normal and those marked as risky. The sample data covers various typical situations within the evaluated scenario to ensure the objectivity and representativeness of the evaluation results. Upon receiving the evaluation request, the strategy evaluation tool, based on the pre-collected sample data, evaluates each optimization rule within the risk control strategy to be optimized, as well as the overall risk control strategy, within a pre-collected risk control evaluation dimension. Finally, it outputs the strategy evaluation index data for the risk control strategy to be optimized and the rule evaluation index data for each individual risk control rule to be optimized.

[0040] The rule evaluation index data is used to characterize the independent effect of each risk control rule in the risk control strategy to be optimized on the preset sample data. The evaluation process takes a single risk control rule as the evaluation object, counts the triggering of the rule in the preset sample data, and calculates the corresponding quantitative index under the risk control evaluation dimension. The rule evaluation index data can be used to characterize the ability and coverage of a single risk control rule to distinguish between risky and non-risky samples.

[0041] The strategy evaluation index data is used to characterize the overall effect of the risk control strategy to be optimized, which is formed by the combination of multiple risk control rules to be optimized, on the preset sample data. The evaluation process takes the complete risk control strategy as the evaluation object, and comprehensively considers the execution order, hit relationship and superposition effect among multiple risk control rules. The overall evaluation calculation of the risk control strategy is carried out under the risk control evaluation dimension, so as to obtain quantitative indicators reflecting the overall performance of the strategy.

[0042] In one optional implementation, strategy evaluation index data can be obtained by aggregating rule evaluation index data. However, this is not limited to a simple summation method. Instead, it involves a comprehensive calculation that combines the coverage relationship between rules, the execution order, and the sample hit overlap relationship. By simultaneously acquiring rule evaluation index data and strategy evaluation index data, the effectiveness of the risk control strategy to be optimized can be analyzed at both the rule level and the strategy level, providing a basis for subsequent strategy optimization, rule replacement, or rule simplification.

[0043] The preset risk control assessment dimensions include risk identification accuracy, customer coverage, and risk outcome distribution. The specific types and number of assessment dimensions can be configured according to the actual application scenario, and this manual does not limit this.

[0044] Among them, the risk identification accuracy dimension is used to evaluate the ability of the risk control strategy to distinguish between risky samples and normal samples, reflecting the effectiveness of the strategy in risk identification; the customer coverage dimension is used to evaluate the proportion of samples that are released after the implementation of the risk control strategy; the risk outcome distribution dimension is used to characterize the distribution of risky samples in the output results of the risk control strategy, which may include at least one of bad customer rate, risk proportion in passed samples, and risk proportion in intercepted samples, in order to evaluate the degree of concentration of the risk control strategy on risky samples from the perspective of outcome distribution.

[0045] The strategy evaluation metrics for the risk control strategy to be optimized include, but are not limited to, the LIFT metric and the pass rate. LIFT (Lift) is a commonly used quantitative metric in the risk control field, used to measure the effectiveness of a risk control strategy in differentiating the risk distribution of samples after execution. Its value reflects the strategy's ability to identify risks in the remaining sample set after execution; a higher LIFT value indicates better performance in terms of risk identification accuracy. The pass rate is the ratio of the number of samples approved by the risk control strategy to the total number of samples, used to characterize the strategy's performance in terms of customer coverage.

[0046] The rule evaluation metrics calculated for each risk control rule to be optimized include, but are not limited to, the LIFT and SUPPORT metrics corresponding to each risk control rule, used to measure the independent performance of a single risk control rule. Specifically, the rule-level LIFT metric is used to evaluate the risk discrimination capability of a single risk control rule within the hit sample set; the SUPPORT metric is used to evaluate the coverage ratio of a single risk control rule in the preset sample data, i.e., the ratio of the number of samples that meet the rule's judgment conditions to the total number of samples.

[0047] By simultaneously acquiring strategy evaluation index data and rule evaluation index data, risk control strategies can be analyzed and optimized from two levels—overall strategy effectiveness and individual rule performance—when subsequently reconstructing strategies based on language models. This avoids adjusting strategies solely based on overall strategy evaluation results while ignoring differences in coverage and risk identification capabilities of individual rules, thereby improving the controllability and stability of the strategy optimization process.

[0048] In S102, the optimization target used to constrain the risk control assessment dimension and the alternative rule set are obtained. The alternative rule set includes: multiple alternative risk control rules and rule assessment index data calculated for each alternative risk control rule under the risk control assessment dimension.

[0049] The optimization objective describes the constraints that are expected to be achieved in the risk control assessment dimension through the strategy optimization process. The optimization objective can be expressed in the form of threshold constraints, interval constraints, or a combination of constraints.

[0050] In one exemplary implementation, the optimization objectives must correspond one-to-one with the preset risk control assessment dimensions. For example, quantitative constraint standards can be set for the risk identification accuracy dimension and the customer coverage dimension. The optimization objectives may include: the remaining customer group LIFT is not higher than a preset threshold, and the pass rate is not lower than a preset proportion. By explicitly stating the optimization objectives as constraints on the risk control assessment dimensions, a clear and definite optimization direction can be provided for the subsequent strategy reconstruction process.

[0051] For example, suppose the optimization goals include: LIFT ≤ 1.2 for the remaining customer base and a pass rate ≥ 80%. This will not only keep the accuracy of the strategy's risk identification for the remaining customer base within a reasonable range, avoiding an excessively low pass rate due to an overemphasis on accuracy, but also ensure that the strategy has sufficient customer coverage to meet the basic requirements of the scenario for customer coverage.

[0052] The alternative rule set is a pre-built collection of candidate rules that can be used to replace or supplement rules in the risk control strategy to be optimized. It includes multiple alternative risk control rules and rule evaluation index data calculated for each alternative risk control rule under preset risk control evaluation dimensions. For example, the evaluation index data of the alternative rules also includes the LIFT index and SUPPORT index corresponding to each rule.

[0053] For example, a possible risk control rule could be: if an applicant's credit card usage exceeds 80% of their credit limit in the past 6 months, they are marked as a high-priority customer group, with a corresponding LIFT value of 1.1 and a SUPPORT value of 15%. These metrics clearly reflect the independent performance of this alternative rule, facilitating the language model's selection when reconstructing strategies. By introducing a set of alternative rules, a rich set of rules can be provided for reconstructing strategies to be optimized, avoiding the problem of insufficient optimization space caused by adjusting only the original strategy rules.

[0054] In S104, strategy optimization prompts are constructed based on the dataset of risk control strategies to be optimized, the optimization target, and the set of alternative rules.

[0055] Policy tuning cue words are used to provide the language model with the contextual information and constraints required for policy tuning in a structured or semi-structured form. By organizing and expressing the dataset of risk control strategies to be tuned, the tuning objectives, and the set of alternative rules, the language model can have clear input criteria and optimization directions when performing policy reconstruction.

[0056] For example, semantic information of risk feature variables corresponding to each risk control rule to be optimized in the risk control strategy can be loaded from a pre-set variable semantic dictionary to form a semantic information set. The strategy optimization prompt words can also contain this semantic information set, which enables the language model to accurately understand the actual meaning and risk indication of each risk feature variable when performing risk control strategy reconstruction. This reduces the probability of unreasonable rule combinations or semantic conflicts during the risk control strategy reconstruction process and helps generate risk control rules with semantic consistency and subsequent interpretable output results.

[0057] For example, preset strategy optimization knowledge can also be obtained and used as part of the strategy optimization prompts. This can provide clear domain constraints and optimization guidance for the strategy optimization process of the language model, so that it follows the established rule validity standards, threshold adjustment principles and rule combination logic when reconstructing risk control strategies, avoiding the generation of strategy results that deviate from existing risk control experience, thereby improving the stability, consistency and reusability of strategy optimization results.

[0058] In S106, strategy tuning prompts are input into the language model to indicate that the language model should optimize towards the tuning target. The risk control strategy to be tuned is reconstructed with reference to the alternative rule set, and the optimized risk control strategy and the corresponding optimization reasons are output.

[0059] The language model is a large language model with powerful natural language understanding, logical reasoning and text generation capabilities. It can be deployed on local electronic devices or cloud servers. The policy tuning agent interacts with the language model through application programming interfaces or network communication protocols.

[0060] In this embodiment, when the language model performs strategy reconstruction, it does not randomly generate risk control strategies. Instead, it adjusts the rule structure of the risk control strategy under the constraints of the optimization target, the strategy evaluation index data of the risk control strategy to be optimized, the rule evaluation index data of each risk control rule to be optimized, and the rule evaluation index data of each alternative risk control rule. Through this method, this embodiment can achieve automated optimization of risk control strategies under the constraints of a unified risk control evaluation dimension and clear optimization targets, thereby improving both the efficiency of strategy optimization and the controllability of the strategy generation process.

[0061] The language model outputs the generated optimized risk control strategies along with their corresponding optimization reasons, making the strategy tuning process traceable and interpretable. Specifically, the optimization reasons clearly describe the triggering basis for each strategy optimization operation, the changes in the referenced rule evaluation indicators, and the correspondence with the tuning objectives. This allows the strategy tuning results to be understood and verified by subsequent processing modules or manual review processes. By outputting the optimization reasons, they can also be reused as part of the historical tuning record when subsequent strategy evaluations fail to meet standards or require further tuning. This helps the language model understand past tuning approaches, avoiding the repeated execution of the same or conflicting optimization operations in multiple tuning rounds, thereby improving the stability and convergence efficiency of the multi-round strategy tuning process.

[0062] In some possible implementations, the strategy tuning agent organizes and populates the dataset of risk control strategies to be tuned, the tuning target, the alternative rule set, and the preset strategy tuning knowledge based on a preset strategy tuning prompt template, and generates strategy tuning prompt words to drive the language model to perform strategy tuning processing.

[0063] The strategy optimization knowledge can include general constraint rules related to risk control rule optimization, rule effectiveness judgment principles, and principles for handling relationships between rules. Examples include evaluation criteria for judging rule effectiveness and priority principles for handling rule coverage relationships. In this specification, strategy optimization knowledge, as an externally configurable knowledge source, works in conjunction with the parameterized knowledge of the language model itself, ensuring that the strategy optimization process is flexible while still being constrained by clear technical rules.

[0064] The strategy tuning prompt template can be predefined as a text template or structured data template with a fixed structure. It is divided into slots to carry different types of tuning information, so that the prompt information input to the language model has clear semantic boundaries and a consistent expression format.

[0065] For example, the strategy tuning prompt template may include slots for filling the dataset of risk control strategies to be tuned, slots for filling the alternative rule set, slots for filling the tuning target, slots for filling the strategy tuning knowledge, and instruction slots for guiding the language model to perform tuning behavior. By filling information from different sources and at different semantic levels into the corresponding slots, information mixing can be avoided, which could lead to the language model's unclear understanding of the tuning target or evaluation basis, thereby improving the stability and consistency of the strategy tuning results.

[0066] For example, the strategy tuning prompt template is configured with a set of instructions in the instruction slot to instruct the language model to perform strategy tuning. The instruction set can be represented as natural language instructions, structured control instructions, or a combination of both, and is used to define the behavioral boundaries and optimization direction of the language model when performing strategy reconstruction. For example, the instruction set instructs the language model, guided by strategy tuning knowledge, to perform at least one specified strategy optimization operation on the risk control strategy to be tuned, using the constraints of the tuning target on the risk control assessment dimension as the optimization basis, and referring to the alternative rule set, outputting the optimized risk control strategy and the corresponding optimization reason. Through the above method, this embodiment utilizes the instruction set in the strategy tuning prompt template to explicitly constrain and guide the strategy tuning behavior of the language model, so that the strategy tuning process no longer relies on the language model's free generation capability, but is transformed into a controlled strategy reconstruction process within a unified assessment dimension, explicit target constraints, and an existing rule space, thereby improving the controllability, consistency, and interpretability of the strategy tuning results.

[0067] Among these, at least some of the alternative risk control rules and at least some of the risk control rules to be optimized are comparable in rule structure. Specifically, the alternative risk control rules and the corresponding risk control rules to be optimized can have the same risk characteristic variables and the same condition judgment logic, differing only in the judgment threshold. By maintaining consistency in risk characteristic variables and judgment logic at the rule level, the evaluation results of different rules under the risk control assessment dimension can be compared, thus facilitating quantitative analysis and replacement selection of rule effects during strategy optimization.

[0068] For example, a risk control rule to be optimized can be represented as "txn_freq_3m > 100", where txn_freq_3m represents the transaction frequency over the past 3 months. Corresponding alternative risk control rules could include "txn_freq_3m > 102", "txn_freq_3m > 104", "txn_freq_3m > 106", etc. These alternative risk control rules maintain consistency in risk characteristic variables and conditional judgment logic, only adjusting the judgment threshold to create different rule instances, thus presenting different rule evaluation indicators in terms of sample coverage and risk identification capability.

[0069] Based on the above rule structure, at least one specified strategy optimization operation includes: (1) Deletion operation: Delete the risk control rules whose corresponding rule evaluation index values ​​do not contribute to the optimization goal from the risk control strategies to be optimized.

[0070] The language model can analyze the role of each risk control rule in the overall strategy by combining the rule evaluation index data of each rule to be optimized with the constraints of the optimization goal. When the rule evaluation index value of a certain risk control rule to be optimized indicates that it has no positive impact on the optimization goal, or its effect has been covered by other rules, the rule can be judged as a non-contributing rule and deleted during the strategy reconstruction process. By performing the deletion operation, the number of redundant or inefficient rules in the strategy can be reduced, thereby reducing the strategy complexity and avoiding interference from invalid rules on the overall strategy evaluation results.

[0071] The deletion operation is essentially rule pruning based on quantitative indicators and logical inclusion relationships. When performing this operation, the language model can identify and remove invalid rules based on the following two criteria: (1) Validity judgment based on marginal contribution: The language model calculates the marginal contribution value of each risk control rule to be optimized in the overall strategy. By comparing the correlation between the rule evaluation indicators (such as LIFT, KS value, PSI, etc.) of a single rule and the preset optimization target, rules that have no positive gain on the objective function are identified.

[0072] If a rule's discriminant metric (such as LIFT) approaches the baseline (e.g., LIFT≈1.0), it indicates that the rule lacks risk discrimination ability and is considered a noisy rule. If a rule has extremely low support and fails to hit any high-risk samples, it indicates insufficient statistical significance. Such rules are deemed "inefficient rules" and are deleted to reduce the strategy dimension, thereby reducing the risk of overfitting and computational overhead.

[0073] (2) Logical redundancy determination based on set inclusion relationship: The language model uses the semantic information of variables to construct the logical mapping of the rule-triggered sample set and analyzes the set relationship between rules.

[0074] Complete Coverage Decision: If the set of triggering conditions for rule R_A is a subset of the set of triggering conditions for rule R_B, and R_B already exists in the policy, then R_A constitutes logical redundancy.

[0075] Equivalence determination: If two rules are semantically equivalent or the overlap of the hit samples exceeds a preset threshold (e.g., Jaccard similarity coefficient > 0.95), the rule with the better evaluation index is retained and the other rule is deleted.

[0076] By eliminating logically redundant rules, the linear independence of the policy rule set is ensured, thereby optimizing the compactness of the policy structure.

[0077] (2) Replacement operation: For at least one risk control rule to be optimized, select a candidate risk control rule from the set of candidate risk control rules that has the same risk characteristic variables and condition judgment logic and has a better rule evaluation index value, and replace the risk control rule to be optimized.

[0078] A better rule evaluation index can be manifested as a higher risk identification capability or a more reasonable sample coverage while meeting the optimization objective constraints. Through rule replacement, the judgment strength of individual rules can be finely adjusted without changing the overall rule structure framework of the strategy, thereby causing the overall strategy evaluation index to converge towards the optimization objective.

[0079] The replacement operation is a local optimization of the rule parameter space while maintaining semantic consistency of the rules. This operation strictly adheres to isomorphism constraints to ensure the stability of the policy logic framework.

[0080] Mapping and filtering of isomorphic rule sets: The language model first establishes a mapping relationship between the rule to be tuned and the candidate rule set. A candidate rule is only included in the candidate set if it satisfies the following isomorphic condition: Consistency of feature variables: The risk feature variable identifiers referenced are the same; Consistency of logical operators: The comparison operators (such as >, <, =, \in>, <, =, ∈) and logical connectors used are the same; Semantic consistency: The risk types represented by the rules are consistent.

[0081] Under this constraint, the replacement operation is essentially a reselection of the threshold parameter or enumeration set of the same logical expression.

[0082] Indicator optimization mechanism under multi-objective constraints: In the isomorphic candidate set, the language model constructs a multi-objective optimization function based on the tuning objective and selects candidate rules.

[0083] Precision-oriented replacement: When the optimization goal focuses on improving the accuracy of risk identification, select alternative rules with higher LIFT or KS values, even if their coverage (SUPPORT) is slightly reduced.

[0084] Coverage-oriented replacement: When the optimization goal focuses on expanding customer coverage or improving pass rate, select alternative rules that have a higher support value or a lower false alarm rate while meeting the minimum risk discrimination constraint.

[0085] Global convergence verification: The language model predicts the impact of the replacement operation on the overall policy evaluation index vector, ensuring that the direction of index change after replacement is consistent with the gradient direction of the optimization target vector, thereby achieving convergence of policy performance towards the target space.

[0086] (3) Sorting operation: Adjust the execution order of multiple risk control rules in the risk control strategy to be optimized according to the rule evaluation index value of the risk control rule to be optimized.

[0087] In risk control strategies that execute multiple rules sequentially, the order in which different rules are executed affects the processing path of the samples and the overall evaluation results of the strategy. By prioritizing the execution of rules with better performance indicators or those that contribute more to the optimization goal, the overall performance of the strategy under limited execution resources can be improved, and the strategy evaluation results can more stably meet the constraints of the optimization goal.

[0088] The sorting operation is an optimization of the rule execution sequence by permuting and combining elements, taking advantage of the characteristics of sequential execution engines.

[0089] (1) Maximizing expected return based on conditional probability: In the sequential execution strategy, the subsequent evaluation is terminated once the sample hits the interception rule. The language model calculates the expected computational cost and expected risk interception return under different permutation orders based on the historical hit rate and risk discrimination of each rule.

[0090] In one implementation, rules with high LIFT values ​​(high signal-to-noise ratio) are placed at the beginning of the sequence to prioritize the interception of high-confidence risk samples, thereby reducing the base rate of subsequent rule processing and thus reducing the overall computational complexity.

[0091] In another implementation, for scenarios aimed at rapid passage, rules with high support values ​​and deemed safe are placed at the front to quickly divert normal samples and reduce the average processing latency of long-tail samples.

[0092] (2) Logical Dependency Topological Ordering: The language model analyzes the logical dependencies between rules and constructs a directed acyclic graph. If the effective execution of rule R_j depends on the preprocessing result or pre-filtering of rule R_i (e.g., R_j targets a specific customer segment, while R_i is used to identify that customer segment), then R_i is forced to be ranked before R_j.

[0093] By combining topological sorting with heuristic search, the language model determines an execution sequence that satisfies logical dependency constraints and statistically optimizes overall policy evaluation metrics (such as weighted accuracy and average response time). This operation does not alter the internal logic of the rules; it only optimizes the global performance of the policy by adjusting the execution order.

[0094] Through the above-mentioned optimization operations such as deletion, replacement, and sorting, this embodiment can reconstruct the risk control strategy structure in a targeted manner while maintaining the semantic consistency of risk control rules. This enables refined optimization of risk control strategies under the constraints of unified risk control assessment dimensions and clear optimization objectives, thereby improving the controllability of the strategy optimization process and the interpretability of the results.

[0095] In one exemplary embodiment, the policy tuning agent constructs policy tuning prompts, which includes: (1) Optimization targets: target LIFT, target pass rate, bad customer rate requirements; for example: target customer group LIFT≤1.2, target pass rate≥80%.

[0096] (2) Data set of risk control strategies to be optimized: multiple risk control rules to be optimized, strategy evaluation index data (such as remaining customer group LIFT, pass rate), and rule evaluation index data of each risk control rule to be optimized (such as LIFT / SUPPORT).

[0097] (3) Alternative rule set: multiple alternative risk control rules, and rule evaluation index data (such as LIFT / SUPPORT) for each alternative risk control rule.

[0098] (4) Strategy optimization knowledge: rule validity standards, threshold adjustment principles, and sequential optimization logic; for example, LIFT>1.7 is a valid rule, and high LIFT rules should be placed first.

[0099] (5) Instruction set.

[0100] (6) Output mandatory specifications: Optimized rule list + indicator to show whether the adjustment is complete + reason for optimization. This facilitates subsequent processing or review.

[0101] After inputting the strategy tuning prompts into the language model, the language model performs the following multi-step inference process; (1) Diagnose and optimize the status of risk control strategy: determine whether the remaining customer group LIFT exceeds the limit and whether the rules meet the validity criteria, such as LIFT>1.7, SUPPORT>0.005; (2) Reconstruct the risk control strategy to be optimized: ① Delete rules: Remove invalid rules with LIFT<1.5 or SUPPORT<0.005; ② Replace thresholds: Select candidate rules with higher LIFT (tightening strategy) or higher SUPPORT (relaxing strategy) from the set of candidate rules; ③ Adjust the order: Reorder the rules according to LIFT value, attribute (blacklist > model score > basic variable), and coverage; (3) Evaluate and optimize risk control strategies: Predict changes in LIFT and pass rate after the adjustment of the optimized risk control strategy to ensure that the optimization objectives are not violated.

[0102] In some embodiments, please refer to Figure 3 After generating the optimized risk control strategy, the strategy tuning agent can also evaluate the execution effect of the optimized risk control strategy to determine whether the optimized risk control strategy meets the preset tuning target.

[0103] In S300, the strategy optimization agent can call the strategy evaluation tool, so that the strategy evaluation tool can use preset sample data to evaluate the optimized risk control strategy under preset risk control evaluation dimensions, and obtain the strategy evaluation index data of the optimized risk control strategy and the rule evaluation index data of each optimized risk control rule in the optimized risk control strategy.

[0104] Strategy evaluation tools can be deployed as standalone evaluation modules or integrated into the strategy tuning agent. Internally, they can quantify the performance of risk control strategies across different evaluation dimensions through methods such as rule hit calculation and sample coverage statistics. By introducing strategy evaluation tools, optimized risk control strategies generated by language models can be transformed into quantifiable and verifiable evaluation results, thereby avoiding judgments based solely on the generated results and improving the controllability and stability of the strategy tuning process.

[0105] In S302, after obtaining the strategy evaluation index data for optimizing the risk control strategy, the strategy optimization agent can compare and analyze the strategy evaluation index data with the optimization target to determine whether the optimized risk control strategy meets the optimization target; and determine whether the current strategy optimization count has reached the preset count limit.

[0106] In S304, if the strategy evaluation index data of the optimized risk control strategy does not meet the optimization target and the current number of strategy optimizations has not reached the preset limit, the strategy optimization agent can determine the dataset of risk control strategies to be optimized in the next round of strategy optimization based on the optimized risk control strategy and its strategy evaluation index data, and the rule evaluation index data of each optimized risk control rule.

[0107] In this scenario, the optimized risk control strategy can be used as the risk control strategy to be optimized in the next round of strategy tuning, and together with the tuning objective and alternative rule set, it can be used to construct new strategy tuning prompts. Subsequently, the strategy tuning agent can return to the step of obtaining the dataset of risk control strategies to be tuned, thereby triggering a new round of strategy tuning process.

[0108] In S306, if the strategy evaluation index data for optimizing the risk control strategy does not meet the optimization target, and the current strategy optimization count reaches the preset limit, the strategy optimization process ends, and a prompt message indicating that no optimized risk control strategy can be generated is output.

[0109] In S308, if the strategy evaluation index data of the optimized risk control strategy meets the optimization target, the strategy optimization process ends, and the optimized risk control strategy obtained in the last round of strategy optimization is determined as the target risk control strategy.

[0110] Through the above methods, the risk control strategy optimization method in this embodiment can form an iterative optimization mechanism based on evaluation feedback, enabling the risk control strategy to gradually approach the preset optimization target during multiple rounds of optimization. On the one hand, it can avoid the problem that a single round of optimization is insufficient to meet the target constraints; on the other hand, by setting an upper limit on the number of strategy optimizations, it can effectively control the computational cost and execution complexity of the optimization process, preventing the strategy optimization process from falling into an unbounded loop.

[0111] In some embodiments, before obtaining the dataset of risk control strategies to be optimized, the strategy optimization agent may also perform initialization processing of the risk control strategies to obtain initial risk control strategies for subsequent strategy optimization.

[0112] The strategy optimization agent can first obtain a candidate rule set. The candidate rule set includes multiple candidate risk control rules, as well as rule evaluation index data calculated for each candidate risk control rule under preset risk control evaluation dimensions. The candidate risk control rules can come from historical rule bases, rule generation modules, or manually configured rule sets, and their corresponding rule evaluation index data are used to characterize the quantitative effect of each candidate risk control rule on the sample data.

[0113] After obtaining the candidate rule set, the strategy tuning agent can optimize the candidate rule set using a preset automated optimization algorithm, combined with the alternative risk control rule set, to generate an initial risk control strategy, with the optimization target as the optimization direction. The automated optimization algorithm can be a heuristic search algorithm, a rule combination optimization algorithm, or other automated processing algorithms suitable for rule selection and combination.

[0114] For example, rule combination optimization algorithms are used for rule optimization, with greedy algorithms being a typical application of this type of algorithm. During execution, the strategy optimization agent first prioritizes all candidate risk control rules in the candidate rule set based on their rule evaluation index data. The ranking criteria can be determined according to the priority of the optimization goal. For example, if the pass rate has a higher priority than LIFT in the optimization goal, the rules are prioritized in descending order of their contribution to improving the pass rate. Then, according to the ranking results, the strategy optimization agent sequentially selects rules to add to the initial rule combination, starting with the highest priority candidate risk control rule. After adding each candidate risk control rule, it calls the strategy evaluation tool to quickly calculate the strategy evaluation index corresponding to the current rule combination. If the current rule combination meets the basic constraints and the strategy evaluation index improves the fit with the optimization goal, the candidate risk control rule is retained. If adding a candidate risk control rule causes a logical conflict in the rule combination or the strategy evaluation index deviates from the optimization goal, the candidate risk control rule is removed, and the next candidate risk control rule is selected. During the selection of candidate risk control rules, the set of alternative risk control rules can serve as a supplementary rule pool. When the rules in the set of candidate rules cannot make the current rule combination meet the basic requirements of the optimization goal, the strategy optimization agent selects the candidate risk control rules that meet the conditions from the set of alternative risk control rules and adds them to the combination until a set of rules that meets all basic constraints and is as close as possible to the optimization goal is formed, which serves as the initial risk control strategy.

[0115] By introducing automated optimization algorithms, an initial risk control strategy that meets basic constraints can be quickly formed with low computational complexity. This avoids starting the subsequent strategy tuning process directly from the complete set of rules or an unordered set of rules, reducing the size of the strategy space and improving overall processing efficiency.

[0116] Please see Figure 4 In sections 41-43, after the initial risk control strategy is generated, the strategy optimization agent can call the strategy evaluation tool, which will then evaluate and calculate the initial risk control strategy based on preset sample data and preset risk control evaluation dimensions, to obtain the strategy evaluation index data of the initial risk control strategy and the rule evaluation index data of each initial risk control rule in the initial risk control strategy.

[0117] In this embodiment, through the aforementioned initialization process, the subsequent language model-based strategy tuning process does not start from an empty strategy or an unconstrained set of rules. Instead, it performs further refined optimization on an initial risk control strategy that already has a clear rule structure, quantifiable evaluation results, and meets basic constraints. This allows the language model tuning process to focus on local optimization issues such as rule combination relationships, parameter adjustments, and execution order, thereby improving the stability of multi-round tuning processes and reducing invalid or uncontrollable strategy generation results.

[0118] In some embodiments, after generating an initial risk control strategy through an automated optimization algorithm and completing the calculation of the corresponding strategy evaluation index data and rule evaluation index data, in order to avoid performing invalid optimization operations on the initial risk control strategy that has met the optimization target, and at the same time to provide targeted analysis basis for the subsequent optimization of strategies that have not met the target, the strategy optimization agent first executes the compliance determination process of the initial risk control strategy.

[0119] For example, the strategy tuning agent obtains an initial risk control strategy dataset, which includes: an initial risk control strategy consisting of multiple initial risk control rules, strategy evaluation index data calculated for the initial risk control strategy under the risk control evaluation dimension, and rule evaluation index data calculated for each initial risk control rule under the risk control evaluation dimension.

[0120] Please see Figure 4 In sections 44-47, the strategy optimization agent loads semantic information of risk characteristic variables corresponding to each initial risk control rule in the initial risk control strategy from a pre-built variable semantic dictionary to form the first semantic information set. The variable semantic dictionary is a pre-constructed feature knowledge base for the risk control domain, storing standardized semantic information of various risk characteristic variables in risk control scenarios. Specific content includes the unique identifier, definition, data source, value range, statistical scope, and correlation with the target risk type of the risk characteristic variable. For example, regarding the risk characteristic variable involved in the initial risk control rule: the number of credit card delinquencies in the past 6 months, the corresponding semantic information in the variable semantic dictionary includes: this variable represents the frequency of credit card repayment delinquencies by the applicant in the past 6 months; the data source is the financial credit reporting system; the value range is 0 to 12 times; the statistical scope is the cumulative number of delinquencies occurring within a calendar month; and it has a positive correlation with credit risk.

[0121] The strategy optimization agent constructs strategy compliance judgment prompts based on the initial risk control strategy dataset, the first semantic information set, and the optimization target. For example, the strategy optimization agent can populate a preset strategy compliance evaluation prompt template with the initial risk control strategy dataset, the first semantic information set, and the optimization target to generate strategy compliance evaluation prompts. The strategy compliance evaluation prompt template includes explicit instructions for the judgment task, requiring the language model to judge the compliance of the initial risk control strategy based on the optimization target and the first semantic information set, and output an evaluation conclusion including compliance status, indicator deviation analysis, and deviation cause reasoning.

[0122] The strategy optimization agent inputs strategy compliance assessment prompts into the language model to instruct the language model to analyze the strategy assessment index data of the initial risk control strategy based on the optimization target and combined with the first semantic information set, and output the compliance judgment conclusion of the initial risk control strategy.

[0123] After receiving the strategy compliance assessment prompts, the language model first performs multi-dimensional information parsing. On the one hand, it accurately extracts the rule structure and strategy assessment indicator values ​​of the initial risk control strategy. On the other hand, it combines the first semantic information set to understand the risk control logic behind each rule. Subsequently, the language model compares the strategy assessment indicator values ​​with the quantitative constraints described in the optimization target to determine whether the initial risk control strategy meets all constraints. If any indicators fail to meet the targets, the language model further combines the semantic information of the risk feature variables to analyze the causes of indicator deviations. For example, when the pass rate is lower than the optimization target threshold, the model can infer from the semantics of the risk feature variable "block if monthly income is lower than 5,000 yuan" in the rule that the support indicator of the rule is too high, causing a large number of normal customers to be mistakenly blocked, thereby lowering the overall pass rate of the strategy.

[0124] Finally, the language model outputs a structured compliance determination conclusion, which includes at least one of the following: first, a compliance status indicator, clarifying whether the initial risk control strategy meets the optimization target; second, details of indicator deviation, which quantifies the difference between each non-compliant indicator and the target threshold if the target is not met; and third, a deviation cause analysis, which explains the specific reasons for the non-compliance of indicators based on semantic information and rule logic.

[0125] If the compliance assessment result indicates that the standard is not met, the initial risk control strategy dataset will be identified as the risk control strategy dataset to be optimized, and will be directly transferred to subsequent datasets. Figure 1 and Figure 2 The strategy optimization process is shown below. For example, the deviation cause analysis results from the compliance determination conclusion can also be used as auxiliary information and incorporated into the construction process of subsequent strategy optimization prompts to improve the targeting of subsequent optimizations.

[0126] In some embodiments, please refer to Figure 5 In sections 51-53, when the compliance determination indicates that the initial risk control strategy has met the optimization target under the preset risk control evaluation dimension, the electronic device can further simplify the execution rules of the initial risk control strategy to reduce the complexity of the risk control strategy.

[0127] For example, the strategy optimization agent can construct simplified rule prompts based on the initial risk control strategy dataset, the first semantic information set, and the pre-defined redundant rule recognition knowledge.

[0128] The first semantic information set is used to characterize the semantic meaning of risk feature variables in each initial risk control rule and the relationship between them. By introducing the first semantic information set, the language model can analyze the relationship between rules not only based on the formal expression of the rules, but also by combining the semantic consistency or equivalence of risk feature variables, thereby improving the accuracy of rule coverage relationship identification.

[0129] Redundancy rule identification knowledge is used to describe the analytical principles used to determine rule redundancy, such as the criteria for determining rule coverage, the inclusion relationship of rule hit range, and the relative comparison method of rule effects.

[0130] After constructing the simplified rule suggestions, the strategy optimization agent inputs these suggestions into the language model. This instructs the language model, guided by knowledge of redundant rule identification, to analyze the coverage relationships among multiple initial risk control rules in the initial risk control strategy, in conjunction with the first semantic information set. During this process, the language model can analyze the interrelationships of different initial risk control rules in terms of risk characteristic variables, conditional judgment logic, and judgment thresholds to identify whether there are inclusion or high overlap relationships in the hit range between rules. When the hit range of an initial risk control rule is completely covered by another rule, or its contribution to the overall strategy evaluation index is lower than a preset standard, the initial risk control rule can be identified as a redundant rule.

[0131] After identifying redundant initial risk control rules, the language model can simplify the initial risk control strategy, eliminating redundant rules and generating a target risk control strategy. The target risk control strategy has fewer rules than the initial risk control strategy, but it still maintains comparable or consistent strategy evaluation metrics performance under preset risk control evaluation dimensions.

[0132] For example, rule simplification prompts can be used to instruct the language model to output the simplification reasons along with the target risk control strategy, thus providing a clear and explainable basis for the rule simplification process. The simplification reasons explain the redundancy criteria for the removed risk control rules, including the coverage relationships between rules, the contribution analysis results of evaluation indicators to the overall strategy, and the substitution of retained rules for risk identification capabilities. This avoids rule simplification merely representing a change in the number of rules without a process explanation.

[0133] Through the above methods, this embodiment can reduce the number of rules in the risk control strategy without compromising its overall effectiveness, thereby reducing computational overhead during strategy execution and improving the clarity and maintainability of the strategy structure. Furthermore, since rule simplification is performed under the premise that the initial risk control strategy has already met the optimization objectives, it avoids the problem of strategy performance degradation due to excessive simplification.

[0134] In one exemplary embodiment, the strategy tuning agent can load a rule simplification prompt template to guide the language model to perform rule simplification and interpretation generation processing on the initial risk control strategy that has met the criteria.

[0135] The rule simplification prompt template has multiple slots configured to hold the structured input information required for the rule simplification process. These slots include at least the following: (1) Initial risk control strategy dataset. The initial risk control strategy dataset is used to describe the strategy structure to be simplified. Its contents include multiple initial risk control rules, strategy evaluation index data, and rule evaluation index data for each initial risk control rule.

[0136] (2) Constraint parameters are used to limit the boundary conditions of rule simplification, such as the maximum number of rules allowed and the minimum interpretation integrity requirement, so as to avoid the rule simplification process from oversimplifying the strategy structure or insufficient interpretation information.

[0137] The minimum interpretability completeness requirement is used to limit the completeness of the explanatory level of the risk control strategy after rule simplification, to prevent the simplification process from resulting in missing or incomprehensible explanatory information. For example, the minimum interpretability completeness requirement can be used to ensure that the explanatory document generated by the simplified risk control strategy includes at least a predetermined number of risk characteristic variable descriptions, rule judgment logic explanations, and risk effect explanations. Alternatively, the minimum interpretability completeness requirement can be used to constrain the number of rules or the scope of explanations retained in the simplified risk control strategy, ensuring that the simplified risk control strategy provides complete and consistent explanatory information while reducing the number of rules.

[0138] (3) Redundancy rule identification knowledge, used to describe the analytical criteria and interpretation generation specifications for determining rule redundancy. For example, it includes: the method of determining the coverage relationship of the rule-triggered sample set, and the priority retention principle based on the rule importance score.

[0139] (4) Output specifications: list of strategy rules + reasons for simplification.

[0140] After constructing the rule simplification prompts, the strategy tuning agent inputs the rule simplification prompts into the language model to instruct the language model to perform strategy simplification processing on the initial risk control strategy under the constraint of redundant rule recognition knowledge.

[0141] During strategy simplification, the language model can analyze multiple rules in the initial risk control strategy based on rule coverage relationships and rule importance scores. Rule coverage relationships can be calculated using a rule coverage matrix: when the set of trigger samples corresponding to one rule is completely contained within the set of trigger samples corresponding to another rule, the former is considered a covered rule. Rule importance scores can be quantified using the rule's LIFT value or the SHAP value calculated based on feature contribution to reflect the relative contribution of each rule to the overall strategy's effectiveness. After identifying redundant rules, the language model can remove those rules, generating a target risk control strategy with a reduced number of rules that still meets the constraint parameter requirements.

[0142] In an optional implementation, after rule simplification, the language model can also generate corresponding explanatory documents for the simplified risk control strategy based on the semantic information of risk feature variables and preset explanatory generation specifications. The explanatory documents can include explanations of the semantic meaning of the risk feature variables involved in each rule, descriptions of the logical relationships of the rule conditions, and explanations of the role of the rules in risk identification, and are output in a predefined format for subsequent use or archiving.

[0143] In this way, the analysis logic and interpretation generation requirements involved in the rule simplification process can be transformed into configurable prompt template constraints, so that the language model performs analysis and processing according to predetermined rules in the rule simplification stage, rather than performing unconstrained text generation, thereby improving the stability and consistency of the rule simplification results.

[0144] In some embodiments, please refer to Figure 6 In sections 61-67, after the final target risk control strategy is determined, the strategy tuning agent can also perform the summarization and recording of the strategy tuning results to generate a risk control strategy tuning report that describes the strategy tuning process and results.

[0145] The strategy tuning agent loads the semantic information of risk feature variables corresponding to each risk control rule in the target risk control strategy from a pre-set variable semantic dictionary to form a second semantic information set. The second semantic information set is used to characterize the semantic meaning and explanatory information of the risk feature variables involved in the target risk control strategy, thereby providing semantic support for the subsequent generation of strategy tuning reports.

[0146] The strategy optimization agent can call the strategy evaluation tool, which will calculate the target risk control strategy based on preset sample data and preset risk control evaluation dimensions. This will obtain the strategy evaluation index data of the target risk control strategy and the rule evaluation index data of each risk control rule in the target risk control strategy. The target risk control strategy and its corresponding strategy evaluation index data, as well as the rule evaluation index data of each initial risk control rule in the target risk control strategy, will form the target risk control strategy dataset.

[0147] The strategy tuning agent can construct summary prompts based on the target risk control strategy dataset, a second semantic information set, and historical data related to the target risk control strategy. The historical data can include information such as intermediate risk control strategies generated during the strategy tuning process, corresponding strategy evaluation metrics, types of strategy optimization operations performed, and the number of tuning iterations, reflecting the evolutionary path during strategy tuning. For example, historical data can be a tuning process log. By incorporating the target risk control strategy, strategy evaluation metrics, semantic information set, and historical data into the summary prompts, the language model can simultaneously consider the strategy structure, quantitative evaluation results, and the context of the tuning process when generating the tuning report, thus avoiding an isolated description of only the final strategy.

[0148] After constructing the summary prompts, the strategy optimization agent inputs these prompts into the language model to instruct it to generate a risk control strategy optimization report. The risk control strategy optimization report is used to structurally record and explain the strategy optimization process and its final results. Its content may include: the rule composition of the target risk control strategy, the evaluation index data of the target risk control strategy under the risk control assessment dimension, the optimization operations performed during the strategy optimization process, an explanation of the impact of various optimization operations on changes in the strategy evaluation index, a semantic explanation of each risk control rule in the target risk control strategy, and deployment recommendations, etc.

[0149] Through the above methods, this embodiment can generate an optimization report corresponding to the target risk control strategy after the optimization is completed, thereby providing a traceable record for subsequent strategy review, version comparison, or archive management. Furthermore, since the optimization report generation process is constrained by a pre-set semantic information set and historical data, it avoids inconsistencies between the generated results and the actual optimization process, improving the accuracy and consistency of the optimization result records.

[0150] In one exemplary embodiment, the scenario is set as credit approval risk control for a consumer finance company: The optimization goal is to minimize risk (i.e., reduce the bad debt rate) while maintaining a pass rate of ≥90%. The final strategy must meet the following requirements: LIFT of the remaining customer base ≤1.2 (i.e., the risk of unblocked users is no higher than 1.2 times the average risk of the entire customer base); and the number of strategy rules should be ≤10 to facilitate expert review.

[0151] The input data is prepared as shown in Table 1.

[0152] Table 1 Among them, all the alternative risk control rule indicators in the alternative rule set are derived from offline backtesting of X and y.

[0153] In one possible implementation, the following exemplary description illustrates the generated initial risk control strategy. If the target is met, further simplification will then be implemented: An initial risk control strategy S0 is generated, and a greedy algorithm is used to select rules round by round while satisfying the optimization objective: Objective function: Maximize weighted return = 0.7 × LIFT + 0.3 × (1 / Support) (favoring rules with high discrimination and low coverage to control the pass rate). Constraint check: Immediately simulate the pass rate and LIFT for each risk control rule added; skip if violated.

[0154] Greedy process: Round 1: Select Rule 3 (LIFT=3.0), current pass rate = 95%, LIFT=1.1 → compliant.

[0155] Round 2: Select Rule 1 (LIFT=2.5), pass rate = 92%, LIFT=1.15 → compliant.

[0156] Round 3: Try rule 2, after adding it the pass rate = 89% → violates "≥90%" → skip.

[0157] Round 3: Select Rule 4 (LIFT=1.8), pass rate = 90.5%, LIFT=1.18 → compliant.

[0158] Continuing to try other rules will result in a pass rate <90% or LIFT>1.2 → stop.

[0159] Output the initial risk control strategy S0: {RuleA: income_change<-0.3 AND loan_count_3m≥3;RuleB: overdue_days>30;RuleC: device_change_freq>10}.

[0160] The strategy evaluation tool returns the initial risk control strategy. The strategy evaluation metrics are as follows: pass rate = 90.5%; remaining customer base LIFT = 1.18 (satisfying ≤ 1.2).

[0161] In the process of assessing the compliance of the initial risk control strategy, the initial risk control strategy It was deemed to have met the standard.

[0162] The policy tuning agent constructs the following simplified rule prompts: Input the simplified rule suggestions into the language model. The language model outputs the target risk control strategy {RuleA: income_change<-0.3 AND loan_count_3m≥3; RuleB: overdue_days>30} and the following simplification reason: Rule C (frequent device changes) is already covered by most users by Rules A and B, and its individual LIFT is low. It is recommended to delete Rule C and focus on the core risk logic.

[0163] After re-evaluating the target risk control strategy using the strategy evaluation tool, the corresponding strategy evaluation indicators include: pass rate = 91.2%, LIFT = 1.17 → still meets the optimization target.

[0164] The strategy tuning agent inputs the generated summary prompts into the language model, so that the language model outputs the following risk control strategy tuning report: In another possible implementation, the following exemplifies the generated initial risk control strategy. If the target is not met, then strategy optimization is performed: This is a completely new strategy generation process. The initial rule set is empty. The strategy optimization agent directly uses 50 candidate rules as the candidate rule set, taking the optimization target as the optimization direction, and selecting a greedy algorithm as the automated optimization algorithm. Basic constraints are set: support for a single rule ≥ 3%, no logical mutual exclusion between rules, and the number of rule combinations ≤ 10. Rule screening and combination are then carried out. Finally, the initial risk control strategy is generated. This includes the following six initial risk control strategies: 1. device_change_freq>10→LIFT=1.8,Support=15%; 2. call_fee_3m<50→LIFT=1.7,Support=14%; 3. trans_amount_std<1000→LIFT=1.6,Support=13%; 4. online_time_daily<2→LIFT=1.5,Support=11%; 5. card_use_freq<3→LIFT=1.4,Support=9%; 6. pay_delay_1m<2→LIFT=1.3,Support=8%.

[0165] Strategy evaluation metrics: pass rate = 85%, remaining customer base LIFT = 1.5, number of rules = 6.

[0166] In the process of assessing the compliance of the initial risk control strategy, the following strategy compliance assessment prompts are constructed and input into the language model: In the process of assessing the compliance of the initial risk control strategy, the initial risk control strategy The dataset was deemed unqualified. The strategy optimization agent directly identified the initial risk control strategy dataset as the dataset to be optimized. Based on the dataset, optimization target, and alternative rule set, it constructed targeted strategy optimization prompts. These prompts, while fully presenting all basic data, incorporated deletion, replacement, and sorting operations as core constraints. An example of a strategy optimization prompt is shown below: After receiving strategy optimization prompts, the language model follows a deletion → replacement → sorting logic to reconstruct the strategy. The language model generates an optimized risk control strategy and its corresponding reasons for optimization. The strategy optimization agent then calls a strategy evaluation tool to perform full-sample indicator verification calculations based on 80,000 historical application records. The final optimized risk control strategy and its verification indicators are as follows: 1. device_change_freq>10→LIFT=1.1,Support=14% (after replacement); 2. call_fee_3m<50→LIFT=1.2,Support=13% (after replacement); 3. pay_delay_1m<2→LIFT=1.3,Support=8% (original rule retained); 4. card_use_freq<3→LIFT=1.4,Support=9% (original rule retained); 5. trans_amount_std<1000→LIFT=1.6,Support=13% (original rule retained).

[0167] Strategy evaluation metrics: Pass rate = 93% (≥90%, meeting the optimization target, an improvement of 8 percentage points compared to the original strategy); Remaining customer group LIFT = 1.18 (≤1.2, meeting the optimization target, a decrease of 0.32 compared to the original strategy); Number of rules = 5 (≤10, meeting the optimization target).

[0168] Reasons for optimization: Deleting rules with no contribution directly reduced the overall LIFT base; the replacement operation replaced two high LIFT rules with low LIFT rules of the same logic, controlling the risk discrimination superposition effect from the source; the sorting method of ascending LIFT allows the strategy to first execute low-risk discrimination rules to gradually screen customer groups, avoiding the risk concentration of the remaining customer groups caused by a single high LIFT interception, and ultimately achieving the goal of meeting the optimization objectives for both core indicators.

[0169] The strategy optimization agent verified that all three evaluation indicators of the optimized risk control strategy met the preset optimization goals, and no secondary optimization was required. The optimization process of this risk control strategy was officially completed, and the strategy optimization agent output the optimized risk control strategy as the final target risk control strategy.

[0170] The various technical features in the above embodiments can be combined arbitrarily, as long as there is no conflict or contradiction between the combinations of features. However, due to space limitations, they are not described one by one. Therefore, the arbitrary combination of various technical features in the above embodiments is also within the scope of this specification.

[0171] Figure 7 This is a schematic structural diagram of a device provided in an exemplary embodiment. For example... Figure 7 As shown, device 700 mainly consists of a communication interface 702, a user interface 704, a processor 706, and a data storage 708. These components are interconnected and communicate with each other via a system bus, network, or other connection mechanism 710. The communication interface 702 enables device 700 to communicate with other devices, access networks, and transmission networks via analog or digital modulation. For example, the communication interface 702 may include an antenna and related processing devices for wireless communication with a radio access network or access point. Furthermore, the communication interface 702 can be a wired interface such as Ethernet, Token Ring, or a USB port, or a wireless interface such as Wi-Fi, Bluetooth, Global Positioning System (GPS), or a wide-area wireless interface (e.g., WiMAX or LTE). Of course, the communication interface 702 can also support other forms of physical layer interfaces and standard or proprietary communication protocols. The communication interface 702 may also include multiple physical communication interfaces, such as Wi-Fi, Bluetooth, and wide-area wireless interfaces.

[0172] User interface 704 includes receiving user input and providing output to the user. Therefore, user interface 704 may include input components such as a keypad, keyboard, touch-sensitive or presence-sensitive panel, computer mouse, trackball, joystick, microphone, still camera, and video camera, and output components such as a display screen (which may be combined with a touch-sensitive panel), CRT, LCD, LED, display using DLP technology, printer, and other similar devices known or developed in the future. User interface 704 may also generate auditory output via speakers, speaker jacks, audio output ports, audio output devices, headphones, and other similar devices known or developed in the future. In some embodiments, user interface 704 may include software, circuitry, or other forms of logic capable of transmitting and receiving data from external user input / output devices. Additionally or alternatively, device 700 may support remote access from other devices via communication interface 702 or another physical interface (not shown). User interface 704 may be configured to receive user input, the position and movement of which may be indicated by indicators or cursors described herein. User interface 704 may also be configured as a display device for rendering or displaying text fragments.

[0173] Processor 706 may contain one or more general-purpose processors and / or special-purpose processors.

[0174] Data storage 708 may include one or more volatile and / or non-volatile storage components and may be integrated wholly or partially with processor 706. Data storage 708 may include removable and non-removable components.

[0175] Processor 706 is capable of executing program instructions 718 (e.g., compiled or uncompiled program logic and / or machine code) stored in data storage 708 to perform the various functions described herein. Data storage 708 may contain a non-transitory computer-readable medium on which program instructions are stored, which, when executed by device 700, enable device 700 to perform any methods, processes, or functions disclosed in this specification and / or the accompanying drawings. Processor 706 executing program instructions 718 may result in processor 706 using data 712.

[0176] For example, program instructions 718 may include an operating system 722 (e.g., an operating system kernel, device drivers, and / or other modules) installed on device 700 and one or more applications 720 (e.g., a browser, social application, or game application). Similarly, data 712 may include operating system data 716 and application data 714. Operating system data 716 is primarily accessible to the operating system 722, while application data 714 is primarily accessible to one or more applications 720. Application data 714 may reside in a file system visible or hidden from the user of device 700.

[0177] Application 720 can communicate with operating system 722 through one or more application programming interfaces (APIs). These APIs help application 720 read and / or write application data 714, transmit or receive information via communication interface 702, receive or display information on user interface 704, etc.

[0178] In some terminology, application 720 may be simply referred to as "app". Furthermore, application 720 can be downloaded to device 700 through one or more online app stores or app markets. However, applications can also be installed on device 700 in other ways, such as through a web browser or a physical interface on device 700 (e.g., a USB port).

[0179] In some embodiments, the risk control strategy optimization device can be applied to, for example... Figure 7 The device shown implements the technical solution described in this specification. The risk control strategy optimization device may include: The acquisition module is used to acquire the dataset of risk control strategies to be optimized; wherein, the dataset of risk control strategies to be optimized includes: the risk control strategy to be optimized consisting of multiple risk control rules to be optimized, the strategy evaluation index data of the risk control strategy to be optimized calculated under the preset risk control evaluation dimension, and the rule evaluation index data of each risk control rule to be optimized. The acquisition module is also used to acquire the optimization target used to constrain the risk control assessment dimension, as well as the alternative rule set; the alternative rule set includes: multiple alternative risk control rules, and rule assessment index data of each alternative risk control rule calculated under the risk control assessment dimension; The prompt word construction module is used to construct strategy optimization prompt words based on the dataset of risk control strategies to be optimized, the optimization target, and the set of alternative rules; The optimization module is used to input strategy optimization prompts into the language model to indicate the optimization direction of the language model to achieve the optimization goal. It reconstructs the risk control strategy to be optimized with reference to the alternative rule set and outputs the optimized risk control strategy and the corresponding optimization reasons.

[0180] For ease of description, the above devices are described by dividing them into various modules or units based on their functions. Of course, when implementing one or more of these specifications, the functions of each module or unit can be implemented in the same or different software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0181] Based on the same concept as the methods described above, this specification also provides an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein the processor performs the steps of the method as described in any of the above embodiments by executing the executable instructions.

[0182] Based on the same concept as the methods described above, this specification also provides a computer-readable storage medium having computer instructions stored thereon that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0183] Based on the same concept as the methods described above, this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the methods as described in any of the above embodiments.

[0184] What those skilled in the art will understand is: In this specification, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of additional identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded.

[0185] In this specification, “a,” “an,” and “the” do not specifically refer to the singular, but may also include the plural.

[0186] In this specification, ordinal numbers such as "first," "second," etc., do not necessarily indicate order; they are often used to distinguish between objects. For example, "first server" and "second server" usually refer to two servers. To differentiate between these two servers, they are described as "first server" and "second server." Of course, sometimes these two servers may be the same server.

[0187] In this specification, unless explicitly stated otherwise, "receiving and sending data" does not necessarily mean direct receiving and sending; it can also mean indirect receiving and sending. For example, A receiving data sent by B can be understood as A directly receiving the data sent by B, or it can be understood as A indirectly receiving the data sent by B through other entities such as C. Similarly, B sending data to A can be understood as B sending the data directly to A, or it can be understood as B indirectly sending the data to A through other entities such as C. Here, C can be one entity, or it can be two or more entities.

[0188] In this specification, unless explicitly stated otherwise, the relationships between structures can be direct or indirect. For example, when describing "A is connected to B," unless it is explicitly stated that A and B are directly connected, it should be understood that A can be directly connected to B or indirectly connected to B. Similarly, when describing "A is on top of B," unless it is explicitly stated that A is directly above B (AB is adjacent and A is above B), it should be understood that A can be directly above B or indirectly above B (AB is separated by other elements, and A is above B). And so on.

[0189] This specification uses specific terms to describe embodiments thereof. Terms such as "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that references to "an embodiment," "one embodiment," or "an alternative embodiment" in different locations throughout this specification do not necessarily refer to the same embodiment. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described herein, as well as the features of those different embodiments or examples, without contradiction.

[0190] Although one or more embodiments of this specification provide method steps as described in the embodiments or flowcharts, it is understood that the order of steps listed in the embodiments or flowcharts is only one of many possible execution orders and does not represent the only execution order. Therefore, when the claims involve method steps, any changes or adjustments to the order of such steps, or the parallelism between steps, are also within the scope of protection of the claims.

Claims

1. A method for optimizing risk control strategies, comprising: Obtain the dataset of risk control strategies to be optimized; wherein, the dataset of risk control strategies to be optimized includes: risk control strategies to be optimized consisting of multiple risk control rules to be optimized, strategy evaluation index data of the risk control strategies to be optimized calculated under a preset risk control evaluation dimension, and rule evaluation index data of each of the risk control rules to be optimized; Obtain the optimization target for constraining the risk control assessment dimension, and the alternative rule set; the alternative rule set includes: multiple alternative risk control rules, and rule assessment index data of each of the alternative risk control rules calculated under the risk control assessment dimension; Based on the dataset of risk control strategies to be optimized, the optimization target, and the set of alternative rules, construct strategy optimization prompt words; The strategy optimization prompts are input into the language model to instruct it to optimize towards the optimization goal. The risk control strategy to be optimized is reconstructed with reference to the alternative rule set, and the optimized risk control strategy and the corresponding optimization reasons are output.

2. The method according to claim 1, wherein constructing strategy optimization prompt words based on the dataset of risk control strategies to be optimized, the optimization target, and the candidate rule set includes: The dataset of risk control strategies to be optimized, the optimization target, the alternative rule set, and the preset strategy optimization knowledge are filled into the strategy optimization prompt template to generate the strategy optimization prompt words; The strategy tuning prompt template is configured with a set of instructions for instructing the language model to perform strategy tuning. The set of instructions instructs the language model, under the guidance of strategy tuning knowledge, to perform at least one specified strategy optimization operation on the risk control strategy to be tuned, based on the constraints of the tuning target on the risk control evaluation dimension, and with reference to the alternative rule set, and output the optimized risk control strategy and the corresponding optimization reason.

3. The method according to claim 2, wherein at least some of the candidate risk control rules and at least some of the risk control rules to be optimized have the same risk characteristic variables and condition judgment logic, and have different judgment thresholds; The at least one specified strategy optimization operation includes: Deletion operation: Delete risk control rules whose corresponding rule evaluation index values ​​do not contribute to the optimization target from the risk control strategies to be optimized; Replacement operation: For at least one risk control rule to be optimized, select a candidate risk control rule from the set of candidate risk control rules that has the same risk characteristic variables and condition judgment logic and has a better rule evaluation index value, and replace the risk control rule to be optimized. Sorting operation: Based on the rule evaluation index value of the risk control rule to be optimized, adjust the execution order of multiple risk control rules in the risk control strategy to be optimized.

4. The method according to claim 1, further comprising: Invoke the strategy evaluation tool so that the strategy evaluation tool uses preset sample data to evaluate the optimized risk control strategy under the risk control evaluation dimension, and obtain the strategy evaluation index data of the optimized risk control strategy and the rule evaluation index data of each optimized risk control rule in the optimized risk control strategy. If the strategy evaluation index data of the optimized risk control strategy does not meet the optimization target, and the current number of strategy optimizations has not reached the preset limit, based on the optimized risk control strategy, the corresponding strategy evaluation index data and rule evaluation index data, determine the risk control strategy dataset to be optimized in the next round of strategy optimization, and return to the step of obtaining the risk control strategy dataset to be optimized. If the strategy evaluation index data of the optimized risk control strategy meets the optimization target, the strategy optimization process ends, and the optimized risk control strategy obtained in the last round of strategy optimization is determined as the target risk control strategy.

5. The method according to claim 1, further comprising, before obtaining the dataset of risk control strategies to be optimized: Obtain an initial risk control strategy dataset, which includes: an initial risk control strategy consisting of multiple initial risk control rules, strategy evaluation index data of the initial risk control strategy calculated under the risk control evaluation dimension, and rule evaluation index data of each of the initial risk control rules; From the preset variable semantic dictionary, load the semantic information of the risk feature variables corresponding to each initial risk control rule in the initial risk control strategy to form a first semantic information set; Based on the initial risk control strategy dataset, the first semantic information set, and the optimization target, construct strategy compliance judgment prompt words; The strategy compliance judgment prompt is input into the language model to instruct the language model to analyze the strategy evaluation index data of the initial risk control strategy based on the optimization target and combined with the first semantic information set, and output the compliance judgment conclusion of the initial risk control strategy. If the compliance determination result indicates that the standard is not met, the initial risk control strategy dataset will be determined as the risk control strategy dataset to be optimized.

6. The method according to claim 5, further comprising: If the compliance determination conclusion indicates that the initial risk control strategy meets the requirements, a simplified rule prompt word is constructed based on the initial risk control strategy dataset, the first semantic information set, and the preset redundant rule recognition knowledge. The simplified rule prompts are input into the language model to instruct the language model, under the guidance of the redundant rule identification knowledge, to analyze the coverage relationship between each initial risk control rule in the initial risk control strategy in combination with the first semantic information set, and to remove redundant initial risk control rules based on the coverage relationship, and output a target risk control strategy with fewer rules than the initial risk control strategy.

7. The method according to claim 5, wherein obtaining the initial risk control strategy dataset includes: Obtain a candidate rule set, which includes: multiple candidate risk control rules, and rule evaluation index data calculated for each candidate risk control rule under the risk control evaluation dimension; Taking the optimization target as the optimization direction, the candidate rule set is optimized using a preset automated optimization algorithm and in combination with the alternative risk control rule set to generate the initial risk control strategy; The strategy evaluation tool is invoked so that it uses preset sample data to calculate the initial risk control strategy under the risk control evaluation dimension, thereby obtaining the strategy evaluation index data of the initial risk control strategy and the rule evaluation index data of each initial risk control rule in the initial risk control strategy.

8. The method according to claim 4 or 5, further comprising: From the pre-set variable semantic dictionary, load the semantic information of the risk feature variables corresponding to each risk control rule in the target risk control strategy to form a second semantic information set; The strategy evaluation tool is invoked to perform calculations on the target risk control strategy based on preset sample data and under preset risk control evaluation dimensions. This results in the strategy evaluation index data of the target risk control strategy and the rule evaluation index data of each risk control rule in the target risk control strategy, which together form a target risk control strategy dataset. Based on the target risk control strategy dataset, the second semantic information set, and historical data related to the target risk control strategy, a summary prompt word is constructed; The summary prompts are input into the language model to instruct the language model to generate a risk control strategy optimization report, which is used to record and explain the process and final results of strategy optimization.

9. An electronic device, comprising: processor; A memory for storing processor-executable instructions; wherein the processor implements the steps of the method as described in any one of claims 1-8 by executing the executable instructions.

10. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method as claimed in any one of claims 1-8.

Images