A Multi-Strategy Collaborative Intelligent Agent Risk Control Adaptation Method and System
By aligning and standardizing multi-source data with timestamps, a risk input package is generated, and the strategy weights are dynamically adjusted. This solves the stability problem of existing risk control solutions under the conditions of inconsistent multi-source data and system load fluctuations, and achieves stability and efficiency of risk control results.
Patent Information
- Application Number
- CN202611375871.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-07
AI Technical Summary
Existing risk control solutions lack dynamic adaptability when faced with real-world conditions such as missing multi-source data, inconsistent time series, model output differences exceeding preset ranges, and system load fluctuations. This results in risk control results failing to meet preset stability indicators, inaccurate risk assessments, conflicting handling instructions, system response delays, and excessive resource consumption.
By acquiring multi-source data and aligning it with timestamps, standardizing fields, and marking missing data based on the business request reception time, a risk input package is generated. Combining multi-dimensional risk characteristics and strategy combinations, the strategy weights are dynamically adjusted to generate a state machine and a decision rule tree. Risk control processing is then performed and the results are fed back, enabling dynamic evaluation and combination of strategies.
In situations of inconsistent multi-source data and fluctuating system load, improve the stability and interpretability of risk control results, avoid time-consuming strategy calls and resource waste, and ensure the rationality and effectiveness of risk control processing.
Smart Images

Figure CN122862916B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial risk control and intelligent decision-making, specifically to an intelligent agent risk control adaptation method and system based on multi-strategy collaboration. Background Technology
[0002] Existing risk control solutions typically involve jointly analyzing multi-source business data and using rule-based judgment, model scoring, or weighted evaluation methods to configure and determine execution strategies when making risk identification and disposal decisions.
[0003] However, when faced with real-world scenarios such as missing multi-source data, inconsistent time series, model output differences exceeding preset ranges, and fluctuating system loads, existing solutions rely excessively on fixed configurations or subjective settings for strategy selection and weight adjustment, lacking dynamic adaptability. This results in risk control results failing to meet preset stability indicators. Such rigid evaluation and invocation mechanisms are prone to causing inaccurate risk assessments and conflicting handling instructions. Furthermore, in high-concurrency business scenarios, this often leads to the unreasonable invocation of time-consuming strategies, resulting in system response delays exceeding preset latency thresholds, excessive resource consumption, and inconsistent compliance control execution. It fails to effectively balance risk control accuracy, real-time response requirements, and system resource constraints. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent agent risk control adaptation method and system based on multi-strategy collaboration. This avoids problems such as risk control results failing to meet preset stability indicators, inaccurate risk judgment, conflicting handling instructions, and unreasonable invocation of time-consuming strategies due to fixed configurations or subjective settings when there are missing multi-source data, inconsistent timing, model output differences exceeding preset ranges, and system load fluctuations. Furthermore, it is easier to control the dynamic evaluation combination and fixed node sequential execution of risk control strategies under limited latency budgets, resource budgets, and compliance constraints, thereby balancing risk identification capabilities, execution time, and resource consumption, and ensuring the rationality and effectiveness of the overall risk control process.
[0005] The objective of this invention can be achieved through the following technical solutions:
[0006] Acquire data related to business requests, including transactions, accounts, devices, networks, locations, behaviors, channels, systems, data graphs, historical policies, and preset blocking lists, as well as latency budgets, resource budgets, and compliance constraint parameters. The compliance constraint parameters include at least prohibited models, mandatory verification steps, and data processing or disposal instruction restrictions.
[0007] Assign risk event identifiers to business requests, and based on the time the business request is received, perform timestamp alignment, field standardization, and missing data marking on the above data within a preset time window to form a risk input package;
[0008] Based on the risk input package, multidimensional risk characteristics are generated, including traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, and business risk tolerance.
[0009] The risk interception effectiveness, false interception index, execution time and resource consumption of each strategy in the preset risk control strategy library are normalized and evaluated. Under the parameters of time delay budget, resource budget and compliance constraint, the strategies are screened, sorted and combined and the weights of strategies are adjusted. Conflicts are handled according to the preset disposal instruction priority to generate strategy combinations.
[0010] Based on the strategy combination, a state machine, threshold table, decision rule tree, risk instruction set, model call table and blocking control table are generated;
[0011] Risk control processing is executed in the order of fixed nodes in the decision rule tree, outputting risk results, business handling instructions and risk control execution instructions, and feeding back the execution results and the post-transaction results after the transaction or review is completed to the intelligent agent control module.
[0012] Preferably, the transaction data includes the amount, transaction type, and timestamp; the account data includes the account identifier, registration duration, number of successful transactions, and number of failed transactions; and the device data includes the device identifier, device fingerprint integrity, and device switching frequency.
[0013] Network data includes risk level, agent identifier, and network latency; location data includes location information, origin information, and location deviation distance.
[0014] Behavioral data includes the mean and variance of click intervals; channel data includes channel codes, feature missing rate, and proportion of outlier samples.
[0015] System data includes current throughput, latency, processor utilization, memory utilization, cache hit rate, and queue length;
[0016] The graph data includes the number of associated accounts, the number of associated devices, and the clustering coefficient; the historical policy data includes the number of rule hits, model scores, model credibility, historical secondary verification records, and historical handling results; the preset blocking list data includes at least one of account identifiers, device identifiers, and network node addresses.
[0017] Preferably, continuous features are truncated according to the corresponding preset lower and upper limits and normalized to the range of zero to one, and missing values are marked with missing tags.
[0018] The preset lower limit and preset upper limit are determined by the historical sample statistical range or configuration parameters; traffic pressure is generated based on the normalized current throughput, latency and queue length; resource pressure is generated based on processor utilization, memory utilization and cache hit rate; data reliability is generated based on feature missing rate, device fingerprint integrity and network latency; channel anomaly degree is generated based on the proportion of abnormal samples in the channel, location deviation distance and risk level; gang association degree is generated based on the number of associated accounts, the number of associated devices and the aggregation coefficient; and strategy probing degree is generated based on the number of rule hits, historical secondary verification failure rate and device switching frequency.
[0019] Preferably, the model output credibility is determined based on the model score, model confidence, input feature completeness, and consistency of multiple model outputs from the currently invoked base model, deep model, or other risk control model.
[0020] The business risk tolerance is determined based on the risk loss limit configured according to the business type, transaction amount, customer level, and business scenario.
[0021] The risk interception effectiveness, false interception indicators, execution time and resource consumption of each risk control strategy are normalized, and a comprehensive evaluation score is calculated according to preset weights to achieve a unified measurement of strategy effectiveness, computational load and resource consumption.
[0022] Preferably, the risk control strategies include deep model invocation strategy, basic model invocation strategy, basic rule execution strategy, missing data compensation strategy, secondary verification strategy, limit strategy, delayed review strategy, silent monitoring strategy, risk diffusion strategy and strategy blocking strategy;
[0023] Missing compensation strategies are used to call a specified data source or use preset alternative features; risk diffusion strategies are used to increase the risk level of associated subjects or expand the detection range of associated objects; and strategy blocking strategies are used to stop the invocation of a specified strategy or model.
[0024] Retain strategies whose execution time does not exceed the latency budget, whose resource consumption does not exceed the resource budget, and whose compliance constraints are not violated, and adjust the strategy weights according to the preset strategy-scenario mapping relationship; when multiple strategies output different handling instructions, select the instruction with the highest priority according to the pre-configured risk instruction priority table.
[0025] Preferably, the state machine includes a normal monitoring state, a high-concurrency degradation state, a low-reliability data state, a suspected group fraud state, a suspected rule probing state, a strong control state, and a trial operation monitoring state; when the throughput is greater than the throughput threshold and the processor utilization is greater than the processor utilization threshold and continues for a preset duration, it enters the high-concurrency degradation state; if it is less than the transaction processing per second threshold or less than the processor utilization threshold, it remains in the normal monitoring state.
[0026] When the feature missing rate is greater than or equal to the missing rate threshold or the channel anomaly degree is greater than or equal to the channel anomaly threshold, the data enters a low reliability state.
[0027] When the clustering coefficient is greater than or equal to the clustering coefficient threshold and the number of associated accounts is greater than or equal to the number of associated accounts threshold, the system enters a suspected group fraud state; when the strategy probing degree is greater than or equal to the probing threshold, the system enters a suspected rule probing state.
[0028] When a risk is hit on a high-risk preset blocking list or the risk score is greater than or equal to the strong control threshold, the system enters a strong control state; when the risk control strategy is under evaluation or de-weighting observation, the system enters a trial operation monitoring state; when multiple states are met simultaneously, the current state is determined according to the preset priority order.
[0029] Preferably, the judgment rule tree is executed in a fixed order: matching the preset blocking list, judging data reliability, judging initial risk score, judging gang association, judging model invocation, and executing instruction output;
[0030] The risk threshold is provided by the threshold table, and the initial risk score is determined according to the risk status index and the obtained model score according to the preset weight.
[0031] When the preset blocking list is hit or the initial risk score is greater than or equal to the risk threshold, the corresponding business processing instruction of rejection or strong control is output; when the preset blocking list is not hit and the initial risk score is less than the risk threshold, the corresponding business processing instruction of permission or silent monitoring is output, and model call nodes marked as high time consumption and unnecessary are skipped according to the model call table.
[0032] Business processing instructions include at least one of the following: allow, deny, SMS verification, facial verification, amount limit, delayed review, silent monitoring, and temporary freeze. Risk control execution instructions include at least one of the following: basic rule execution, basic model call, and deep model call.
[0033] Preferably, each business request is recorded with risk event identifier, multi-dimensional risk characteristics, traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, strategy selection, risk result, business handling instructions, risk control execution instructions, execution status and post-event results to form a strategy log;
[0034] Post-event results include at least one of the following: final transaction result, fraud confirmation result, review result, and customer complaint result;
[0035] When the preset evaluation period and minimum sample size conditions are met, the risk interception accuracy, risk interception effectiveness, false interception rate, average processing delay and resource consumption are calculated based on the strategy log.
[0036] When the false interception rate, average processing delay, or resource consumption exceeds the preset threshold, the corresponding risk control strategy will be disabled, downgraded, or transferred to trial operation monitoring mode; when none of them exceed the corresponding threshold, the current applicable status of the corresponding risk control strategy will be maintained.
[0037] When a strategy meets the preset re-activation conditions, demotion recovery conditions, or trial run end conditions, update the strategy weight, applicable status, threshold, or resource consumption limit in the strategy library, and add it to the strategy candidate set when the re-addment conditions are met.
[0038] A multi-strategy collaborative intelligent agent risk control adaptation system includes: a multi-source data acquisition module, used to acquire transaction, account, device, network, location, behavior, channel, system, graph, historical strategy and preset blocking list data, and to acquire latency budget, resource budget and compliance constraint parameters;
[0039] The risk feature construction module is used to perform timestamp alignment, field standardization, missing data marking, and feature normalization on the above data, using the business request reception time as a unified time benchmark.
[0040] The risk status analysis module is used to generate traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, and business risk tolerance.
[0041] The intelligent agent control module is used to evaluate, filter, sort, combine and adjust risk control strategies from the preset risk control strategy library, generate strategy combinations and send them to the strategy rule conversion module;
[0042] The strategy rule conversion module is used to generate state machines, threshold tables, decision rule trees, risk instruction sets, model call tables, and blocking control tables based on strategy combinations.
[0043] The risk control execution module is used to perform risk control processing according to the risk instruction priority table and judgment rule tree, and send business processing instructions to the transaction gateway, account management system, verification service or audit system through the communication interface;
[0044] The effect feedback and adjustment module is used to receive the execution status, handling results and post-event results returned by the transaction gateway, account management system, verification service or audit system, form a strategy log, and perform deactivation, demotion or trial operation of risk control strategies according to preset strategy adjustment conditions.
[0045] The beneficial effects of this invention are:
[0046] 1. This invention acquires transaction data, account data, device data, network data, location data, behavioral data, channel data, system data, graph data, historical strategy data, and preset blocking list data. It uses the business request reception time as a unified time benchmark to perform timestamp alignment, field standardization, missing data marking, and time benchmark correction. At the same time, it combines a logical state management mechanism of pending completion, computable, sealed, and timed-out closure. This enables multi-source data to form a computable risk input package within a preset time window under high concurrency conditions. It reduces misjudgment, cross-use, and main link blockage caused by multi-source data missingness, time sequence inconsistency, and clock drift, thereby improving the consistency, stability, and executability of risk control input.
[0047] 2. This invention truncates and normalizes continuous features, retains and marks missing facts, and generates traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, and business risk tolerance. Furthermore, it performs proportional redistribution or fixed default caliber processing on missing features, and performs fixed judgment tables or lookup table mapping on model output credibility and business risk tolerance. This gives risk status analysis a unified dimension, fixed branches, and clear output interface, which can reduce the impact of model output difference exceeding the preset range, insufficient input completeness, and system load fluctuations on the stability of risk judgment, thereby improving the interpretability and consistency of risk identification results.
[0048] 3. This invention, through an intelligent agent control module, normalizes, ranks, filters, combines, and adjusts the risk interception effectiveness, false interception indicators, execution time, and resource consumption of candidate risk control strategies under latency budget, resource budget, and compliance constraints. It also incorporates a risk instruction priority table, model call table, and blocking control table to constrain conflicting instructions and high-time-consuming strategies. This enables deep model call strategies, basic rule execution strategies, missing data compensation strategies, secondary verification strategies, quota strategies, delayed review strategies, silent monitoring strategies, risk propagation strategies, and strategy blocking strategies to collaboratively adapt to different business states. This avoids unreasonable invocation of high-time-consuming strategies, conflicting handling instructions, and inconsistent compliance control execution, thus balancing interception accuracy, false positive control, processing latency, and resource consumption. Attached Figure Description
[0049] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings, in which:
[0050] Figure 1 This is a schematic diagram of the process of an intelligent agent risk control adaptation method based on multi-strategy collaboration according to the present invention.
[0051] Figure 2The diagram shows a block diagram of an intelligent agent risk control adaptation system based on multi-strategy collaboration according to an embodiment of the present invention. Detailed Implementation
[0052] To make the above features and advantages of the present invention more apparent and understandable, exemplary embodiments are described below in detail with reference to the accompanying drawings.
[0053] Please see Figure 1 A method for adapting intelligent agent risk control based on multi-strategy collaboration includes: acquiring data related to business requests, such as transactions, accounts, devices, networks, locations, behaviors, channels, systems, graphs, historical strategies, and preset blocking lists, as well as latency budgets, resource budgets, and compliance constraint parameters. The compliance constraint parameters include at least the models that are prohibited from being called, the verification steps that must be performed, and restrictions on data processing or disposal instructions.
[0054] Assign risk event identifiers to business requests, and based on the time the business request is received, perform timestamp alignment, field standardization, and missing data marking on the above data within a preset time window to form a risk input package;
[0055] Based on the risk input package, multidimensional risk characteristics are generated, including traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, and business risk tolerance.
[0056] The risk interception effectiveness, false interception index, execution time and resource consumption of each strategy in the preset risk control strategy library are normalized and evaluated. Under the parameters of time delay budget, resource budget and compliance constraint, the strategies are screened, sorted and combined and the weights of strategies are adjusted. Conflicts are handled according to the preset disposal instruction priority to generate strategy combinations.
[0057] Based on the strategy combination, a state machine, threshold table, decision rule tree, risk instruction set, model call table and blocking control table are generated;
[0058] Risk control processing is executed in the order of fixed nodes in the judgment rule tree, outputting risk results, business handling instructions and risk control execution instructions, and feeding back the execution results and the post-transaction results after the transaction or review is completed to the intelligent agent control module;
[0059] In a preferred embodiment, the transaction data includes the amount, transaction type, and timestamp; the account data includes the account identifier, registration duration, number of successful transactions, and number of failed transactions.
[0060] Device data includes device identification, device fingerprint integrity, and device switching frequency; network data includes risk level, agent identification, and network latency.
[0061] Location data includes location information, place of origin information, and location deviation distance; behavioral data includes the mean and variance of click intervals.
[0062] Channel data includes channel code, feature missing rate, and proportion of abnormal samples; system data includes current throughput, latency, processor utilization, memory utilization, cache hit rate, and queue length calculated within a preset statistical window ending at the time the business request is received.
[0063] The graph data includes the number of associated accounts, the number of associated devices, and the clustering coefficient; the historical policy data includes the number of rule hits, model scores, model credibility, historical secondary verification records, and historical handling results; the preset blocking list data includes at least one of account identifiers, device identifiers, and network node addresses.
[0064] Continuous features in the multidimensional risk characteristics are truncated according to preset lower and upper limits for each feature, and then normalized to... Missing values are marked with missing tags; the preset lower limit and preset upper limit are determined by the historical sample statistical range or configuration parameters.
[0065] Traffic pressure is generated based on normalized current throughput, latency, and queue length; resource pressure is generated based on processor utilization, memory utilization, and cache hit rate; data reliability is generated based on feature missing rate, device fingerprint integrity, and network latency; channel anomaly is generated based on the proportion of abnormal samples in the channel, location deviation distance, and risk level; gang association degree is generated based on the number of associated accounts, the number of associated devices, and the aggregation coefficient; and strategy probing degree is generated based on the number of rule hits, the historical secondary verification failure rate calculated separately according to account identifier and device identifier, and the device switching frequency.
[0066] In this embodiment, the risk control server first truncates continuous features, and then uses linear normalization to map them to... The interval, specifically: when the original value is less than the preset lower limit, it is calculated according to the preset lower limit; when the original value is greater than the preset upper limit, it is calculated according to the preset upper limit. The normalized value is determined according to the relative position of the truncated value between the lower and upper limits, which can be expressed as:
[0067]
[0068] in, Indicates the original value According to the preset lower limit and preset upper limit Truncation is performed; and a lower limit is preset. Less than or equal to the preset upper limit When the preset lower limit coincides with the preset upper limit, the corresponding normalized value is directly set to 0 to avoid the denominator being 0, which would prevent calculation.
[0069] For missing values, the missing label is retained and they are not included in the numerical normalization of the feature. However, they are counted separately as missing facts in subsequent data reliability, model output credibility and strategy selection, so as to ensure the consistency of the correspondence between the normalization results, missing facts and subsequent judgment logic.
[0070] Based on this, traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree and strategy probing degree are all calculated based on the above normalized characteristics. Each indicator adopts the same dimension, which makes it easy to directly compare the relative risk levels of different dimensions when evaluating strategies in the future.
[0071] For similar features with multiple input sources, the latest and most valid data within the same preset time window is preferred; if multiple valid values exist for similar features, the average value or the feature value with the largest risk assessment value is taken according to the similar feature fusion rules configured for the business scenario, so as to ensure the stability of the input criteria.
[0072] To avoid logical inconsistencies caused by directly adding indicators from different directions, when generating the above risk status indicators, they are first uniformly converted into the same vector where a larger value indicates a higher risk or is less conducive to risk control.
[0073] Specifically, the current transaction processing volume per second, 99th percentile latency, queue length, processor utilization, memory utilization, feature missing rate, network latency, proportion of abnormal samples in the channel, location deviation distance, Internet Protocol address risk level, number of associated accounts, number of associated devices, clustering coefficient, number of rule hits, historical secondary verification failure rate, and device switching frequency are all calculated based on positive risk quantities.
[0074] Device fingerprint integrity and cache hit rate are positive guarantee quantities, which are first converted into negative risk quantities before being included in the calculation when participating in the synthesis.
[0075] Although data reliability is named after reliability, a low reliability risk intermediate value is first formed during the internal synthesis stage, and then converted into a reliability result consistent with the name during the output stage. The larger the reliability result value, the more complete and reliable the data is, and the smaller the value, the less reliable the data is.
[0076] For a single indicator composed of multiple normalized features, a weighted average is performed using preset weights, and the sum of each weight is 1. If a component feature is missing, the missing fact is included in the data reliability judgment through a missing label. At the same time, within the single indicator, only the features that are not missing and are allowed to participate in the calculation are re-normalized according to the remaining weights to avoid the distortion of the indicator caused by mechanical zeroing.
[0077] If all the constituent features of a certain single indicator are missing, the remaining weights will not be renormalized. Instead, they will be processed according to fixed default rules: the continuous values of traffic pressure, resource pressure, channel anomaly, group association degree and strategy probing degree will all be set to 0.5, and a flag indicating that all constituent features are missing will be set simultaneously.
[0078] The continuous values of data reliability are directly set to 0 and recorded as low reliability; this process avoids situations where the weight denominator is 0 or a single indicator has no definite value due to the absence of all inputs.
[0079] Furthermore, the data reliability output is stored according to the following fixed criteria: first, a low reliability risk intermediate value between 0 and 1 is formed based on the feature missing rate, the completeness of the reverse-processed device fingerprint, and the network latency; then, the data reliability value is obtained by subtracting this low reliability risk intermediate value from 1.
[0080] When it is necessary to participate in subsequent risk scoring and other unified inputs where the higher the value, the higher the risk, the data reliability value is then converted back to low reliability risk input, instead of directly adding the high reliability value to the other risk values.
[0081] After a business request is received by the transaction gateway, the risk control service first generates a risk event identifier for the request. The risk event identifier is used to match the collected data, judgment results and post-verification results associated with the same request, and is used to provide a basis for subsequent recording and review.
[0082] Within a preset time window based on the time the business request is received, the risk control service retrieves the corresponding fields from the transaction system, account management system, device fingerprint service, network monitoring component, location service, behavior tracking, channel service, system monitoring component, relationship graph service, historical strategy library, and preset blocking list library, and performs timestamp alignment, field standardization, and missing flags.
[0083] Among them, timestamp alignment refers to mapping data from various sources to the same request period to avoid the impact of inconsistent collection times on subsequent judgments.
[0084] Field standardization refers to organizing similar fields into a unified format, such as organizing fields like amount, delay, utilization rate, and frequency into comparable numerical forms; missing field marking refers to setting corresponding missing field flags for fields that have not been retrieved, so that the missing facts are retained in the risk input package instead of being directly discarded.
[0085] Furthermore, to ensure an executable data flow topology for multi-source data under high-concurrency scenarios, the risk control service manages data associated with the same risk event identifier using four logical states: pending completion, computable, archived, and timed out.
[0086] When a business request arrives, it first enters the pending completion state and continuously receives data fragments returned from various sources within the corresponding preset time window; when the minimum calculation condition is reached, it switches to the computable state, allowing the risk input package to be built first and then proceeding to the subsequent indicator calculation.
[0087] Once the decision rule tree completes the main process output, it enters the sealed state, allowing only the appending of execution results and post-processing results. When the waiting time for an external source to return exceeds the remaining latency budget of the request, it directly enters the timeout and closes state, preserving the missing facts for the unreturned source without blocking the main process.
[0088] Correspondingly, the data from each source is not accumulated indefinitely, but is temporarily stored within a sliding retention window centered on the time the business request is received; data fragments that exceed the window and are not referenced by the current risk event are removed in a first-come, first-served manner to avoid historical data fragments being incorrectly invoked by subsequent requests;
[0089] For multiple business requests that come in at the same time, the risk control service maintains independent request contexts according to the risk event identifiers and uses the remaining latency budget of the request as the priority basis for scheduling order.
[0090] When the system detects that the number of requests to be filled exceeds the preset concurrency limit, it first compresses the waiting time of low-priority sources, and then limits the high-time-consuming data filling actions to ensure that the main link forms a computable input within the budget.
[0091] Furthermore, to reduce alignment errors caused by clock skew between heterogeneous nodes, the data returned from each source undergoes a time base correction before being incorporated into the current request cycle:
[0092] If the absolute value of the deviation between the return time from the source side and the reception time of the transaction gateway does not exceed the preset drift tolerance, then the mapping will be directly performed based on the source time.
[0093] If the drift tolerance is exceeded, the average offset formed by multiple interactions of the source's nearest neighbors will be used for compensation before mapping. Data that still exceeds the preset time window after compensation will no longer be considered as valid values in the same period, but will be written into the timing mismatch flag and treated as missing.
[0094] Through the above time base correction, the multi-source data under the same risk event identifier has only three clear branches in the control flow: effective incorporation, missing retention, or timing mismatch removal, thus avoiding the cross-use of data from the same source between different requests due to clock drift.
[0095] After completing the above processing, the risk control service generates a risk input package and extracts multidimensional risk features from it. For continuous features, they are first truncated according to preset lower and upper limits, and then normalized. This is to prevent extreme values from causing non-linear shifts in the scoring;
[0096] The preset lower limit and preset upper limit are determined by the historical sample statistical range or configuration parameters, which facilitates the use of a unified calculation method in different business scenarios; for missing values, a missing mark is retained so that subsequent logic can identify whether the data is complete;
[0097] Based on this, the risk control service generates traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, and strategy probing degree respectively; traffic pressure is calculated by the normalized value of the current transaction processing per second, the 99th percentile latency, and the queue length, and is used to reflect whether there is queuing or timeout pressure in the current request link;
[0098] Resource pressure is calculated by processor utilization, memory utilization, and cache hit rate, and is used to reflect whether the running resources are close to the budget limit. Among them, cache hit rate is treated as having a reverse effect when participating in the resource pressure calculation, that is, the higher the cache hit rate, the weaker the effect on increasing resource pressure.
[0099] Data reliability is calculated by combining feature missing rate, device fingerprint integrity, and network latency, and is used to reflect whether the input data can be used for stability judgment; channel anomaly is calculated by combining the proportion of abnormal channel samples, location deviation distance, and Internet Protocol address risk level, and is used to reflect whether the channel source deviates from the preset channel feature standard;
[0100] The group association degree is calculated by the number of associated accounts, the number of associated devices, and the clustering coefficient, and is used to reflect whether there is a centralized association between multiple entities; the strategy probing degree is calculated by the number of rule hits, the historical secondary verification failure rate by account identifier and device identifier, and the device switching frequency, and is used to reflect whether there are probing behaviors that bypass the rules;
[0101] To avoid the aforementioned joint calculations only reflecting functional effects but lacking executable rules, in this embodiment, all the above indicators are generated by weighted average with preset weights, and the constituent features are first uniformly converted into the same vector where the larger the value, the more unfavorable it is for risk control processing, before participating in the synthesis.
[0102] Specifically, traffic pressure is obtained by multiplying the normalized values of the current transaction processing per second, the 99th percentile latency, and the queue length by their respective preset weights and then summing them up; resource pressure is obtained by multiplying the normalized values of processor utilization, memory utilization, and cache hit rate after reverse processing by their respective preset weights and then summing them up.
[0103] Internally, data reliability is first calculated by multiplying the feature missing rate, the completeness of the reverse-processed device fingerprint, and the normalized value of network latency by their respective preset weights, and then summing them to form a low reliability risk value. This value is then output as a data reliability numerical value according to a preset mapping relationship, whereby this data reliability value is uniformly mapped to... The range is defined, and the larger the value, the more reliable the data is and the more suitable it is to proceed directly to the next judgment.
[0104] Correspondingly, the larger the aforementioned low reliability risk value, the smaller the reliability value of the mapped data; the channel anomaly degree, gang association degree, and strategy probing degree are also obtained by weighted average of their constituent characteristics; the preset weights within each individual indicator are given by the configuration parameters and the weight sum is 1;
[0105] If a component feature is missing, that feature will not participate in the numerical synthesis of the current single indicator, and the weights of the remaining features that are not missing will be redistributed proportionally. At the same time, the missing fact will be reserved for data reliability judgment and subsequent strategy screening.
[0106] Furthermore, if all the constituent features of a particular metric are missing in the current request, the proportional reallocation process will not be performed; instead, the metric will be output according to a fixed default definition.
[0107] For traffic pressure, resource pressure, channel anomaly, group association degree, and strategy probing degree, set their continuous values to 0.5 and record them as medium, and write a mark indicating that all components of this single indicator are missing.
[0108] For data reliability, its continuous value is directly set to 0 and recorded as low reliability; by using the above default method, the situation where the weight denominator is 0 or the value of a single indicator is undefined is avoided when there is no valid input;
[0109] Furthermore, to ensure clear data flow and program branches, the above six indicators are all discretized and stored in three levels during output:
[0110] When the value of a single indicator is less than the first-level threshold, it is recorded as low-risk; when the value of a single indicator is not less than the first-level threshold and is less than the second-level threshold, it is recorded as medium-risk; when the value of a single indicator is not less than the second-level threshold, it is recorded as high-risk. The first-level threshold is less than the second-level threshold.
[0111] The data reliability uses a storage method consistent with its name, that is, a high data reliability value is recorded as a high reliability level, a medium data reliability value is recorded as a medium reliability level, and a low data reliability value is recorded as a low reliability level. This level is stored as an independent field in the threshold table. When the state machine and the decision rule tree read this field, they will perform comparisons in the order of low reliability level, medium reliability level, and high reliability level, and it will not be mixed with the high and low risk levels of other risk indicators.
[0112] The threshold value is determined by the business scenario configuration or the statistical results of historical samples, and is written into the risk input package along with the corresponding continuous value. The subsequent state machine and decision rule tree can directly read the continuous value for comparison, or read the level for fast branching, so that there is a consistent input and output interface between feature construction, state judgment and policy execution.
[0113] The intelligent agent control module retrieves candidate strategies from the preset risk control strategy library, calculates the risk interception effectiveness, false interception index, execution time and resource consumption of each strategy, normalizes these indicators, and then calculates the comprehensive evaluation score according to the preset weights.
[0114] The comprehensive evaluation score is used to indicate whether the strategy is suitable for inclusion in the strategy portfolio under the current budget and compliance constraints. Strategies that do not meet the latency budget, resource budget, or compliance constraints are directly eliminated by the agent control module. Strategies that meet the conditions are sorted according to the comprehensive evaluation score and conflict is handled in combination with the preset handling instruction priority to form a strategy portfolio.
[0115] After generating the strategy combination, the strategy rule conversion module converts the combination into a state machine, threshold table, decision rule tree, risk instruction set, model call table and blocking control table;
[0116] The state machine describes the switching conditions between different operating states; the threshold table stores the threshold values for each judgment; and the judgment rule tree specifies the order of nodes in the risk control process.
[0117] The risk instruction set is used to store instructions such as allow, deny, verify, limit, delay audit, and silent monitoring; the model call table is used to specify which models can be called and which models should be skipped; the blocking control table is used to specify which policies should stop execution when timeout or resource limit is exceeded.
[0118] The risk control execution module executes the processing according to the fixed node order of the judgment rule tree. First, it judges the preset blocking list, then the data reliability, then the initial risk score, then the group association, then the model call, and outputs the risk result, business handling instructions and risk control execution instructions.
[0119] After execution, the risk control execution module returns the execution result to the feedback module; when the transaction is completed or the review is completed, the feedback module obtains the post-event result and sends the execution result and the post-event result back to the intelligent agent control module for subsequent adjustment of strategy weights, thresholds and applicable states;
[0120] In this embodiment, the above processes are all completed collaboratively by the risk control server, transaction gateway, account management system, verification service, and auditing system. The risk control server is responsible for calculating, filtering, and issuing instructions; the transaction gateway is responsible for receiving the processing results and controlling the transaction flow; the verification service is responsible for performing secondary verification; and the auditing system is responsible for post-verification processing of delayed auditing instructions.
[0121] As a preferred embodiment, the model output confidence is determined based on the model score, model confidence, input feature completeness, and consistency of multiple model outputs from the currently invoked base model, deep model, or other risk control model.
[0122] Business risk tolerance is determined based on the maximum risk loss configured according to the business type, transaction amount, customer level, and business scenario in the business request; the risk interception effectiveness, false interception indicators, execution time, and resource consumption of each risk control strategy are normalized, and a comprehensive evaluation score is calculated according to preset weights:
[0123]
[0124] in, Indicates the first The comprehensive evaluation score of each risk control strategy. , , and They represent the first The absolute values of risk interception effectiveness, false interception index, execution time, and resource consumption corresponding to each risk control strategy; before being substituted into the comprehensive evaluation score formula, the above items need to be converted into dimensionless normalized values through a minimax normalization function. , , , To ensure consistency of dimensions; risk interception effectiveness The calculation formula is:
[0125]
[0126] Risk interception effectiveness is expressed as the transaction amount covered by the strategy. Compared with the initial risk score The product; This indicates the summation of all transactions covered by the strategy; the original false positive indicator. The calculation formula is:
[0127]
[0128] in, The penalty coefficient for a single mis-blocking is preset for the business scenario; , , and These are preset weighting factors corresponding to risk interception effectiveness, false interception index, execution time, and resource consumption, respectively.
[0129] Risk control strategies include deep model invocation strategy, basic model invocation strategy, basic rule execution strategy, missing data compensation strategy, secondary verification strategy, limit strategy, delayed review strategy, silent monitoring strategy, risk diffusion strategy, and strategy blocking strategy.
[0130] Missing compensation strategies are used to call a specified data source or use preset alternative features; risk diffusion strategies are used to increase the risk level of associated subjects or expand the detection range of associated objects; and strategy blocking strategies are used to stop the invocation of a specified strategy or model.
[0131] Retain policies whose execution time does not exceed the latency budget, whose resource consumption does not exceed the resource budget, and whose compliance constraints are not violated, and adjust the policy weights according to the preset policy-scenario mapping relationship;
[0132] When multiple strategies output different handling instructions, the instruction with the highest priority is selected according to the pre-configured risk instruction priority table. The priority in the risk instruction priority table, from high to low, includes at least rejection, temporary freeze, SMS verification or face verification, amount limit, delayed review and allow.
[0133] The state machine includes normal monitoring state, high concurrency degradation state, low data reliability state, suspected group fraud state, suspected rule probing state, strong control state, and trial operation monitoring state;
[0134] When the current transaction volume per second is greater than or equal to the transaction volume per second threshold and the processor utilization is greater than or equal to the processor utilization threshold, and this continues for a preset duration, the system enters a high-concurrency degradation state; if it is less than the transaction volume per second threshold or the processor utilization threshold, the system remains in normal monitoring state.
[0135] When the feature missing rate is greater than or equal to the missing rate threshold or the channel anomaly degree is greater than or equal to the channel anomaly threshold, the system enters a low data reliability state; when the clustering coefficient is greater than or equal to the clustering coefficient threshold and the number of associated accounts is greater than or equal to the number of associated accounts threshold, the system enters a suspected group fraud state; when the strategy probing degree is greater than or equal to the probing threshold, the system enters a suspected rule probing state.
[0136] When a high-risk preset blocking list is hit or the risk score is greater than or equal to the strong control threshold, the system enters the strong control state; when the risk control strategy is under evaluation or de-weighting observation, the system enters the trial operation monitoring state; when multiple states are met simultaneously, the current state is determined in the following order: strong control state, suspected group fraud state, suspected rule probing state, low data reliability state, high concurrency degraded state, trial operation monitoring state, and normal monitoring state.
[0137] The thresholds for transaction volume per second, processor utilization, missing rate, channel anomaly, clustering coefficient, number of associated accounts, probing, and strong control are all provided by a preset threshold table or determined through historical sample statistical distribution.
[0138] In this implementation, the reliability of the model output does not depend solely on the output of a single model, but is determined jointly by the currently invoked base model, deep model, or other risk control models.
[0139] The specific model output steps include: the risk control server first reads the model score and model confidence of each model, maps the model confidence to three levels: high, medium and low, and then judges whether the current input is complete enough by combining the completeness of the input features and compares whether the outputs of multiple models are consistent.
[0140] If multiple models output in the same direction and the number of valid models reaches the preset number, and the input features are complete, the model output credibility is judged as high; if only a single model outputs valid and the model's confidence is greater than or equal to the preset confidence lower limit, it is judged as medium; if the output direction is opposite, or the input missing rate is greater than the preset threshold, the model output credibility is judged as low.
[0141] Model output consistency is judged by whether the direction of the output of each model is consistent under the same request, rather than requiring the values to be exactly the same, so as to take into account the differences in the scoring scale of different models.
[0142] To avoid direct comparisons between different scoring scales, before participating in consistency judgment and subsequent weighting, the scores of multiple models are first converted to a unified score range according to the historical statistical range or configuration range corresponding to each model.
[0143] If a model's score is missing, times out, or is prohibited from being called due to compliance constraints, then that model will be recorded as invalid output in this request, will not participate in consistency statistics, and will not participate in the weighted calculation of the current request.
[0144] The current call is limited to model outputs that have already returned results before entering the strategy evaluation and filtering process, as well as valid model results that can be directly read from historical strategy data and used for the judgment of this request. It does not include model outputs that have not yet been executed or returned in subsequent model call judgment nodes, so that the generation time of model output credibility is consistent with the execution time of subsequent rule trees.
[0145] Furthermore, to facilitate program implementation, the model output credibility is determined according to a fixed decision table: the completeness of the input features is first divided into three levels: high, medium, and low. The overall missing rate is not greater than 0.1 and is recorded as high, greater than 0.1 and not greater than 0.3 and is recorded as medium, and greater than 0.3 and is recorded as low.
[0146] The model confidence scores are then categorized into three levels: high, medium, and low, based on scores of at least 0.8, at least 0.6 but less than 0.8, and less than 0.6. The consistency of multiple model outputs is judged by whether the output directions are all high-risk or all low-risk. A consistency score of at least 0.8 is considered high, at least 0.5 but less than 0.8 is considered medium, and the rest are considered low. The confidence scores are then output according to the following rules:
[0147] The model output credibility is recorded as high when the input feature completeness is high, there are at least 2 valid models, at least 80% of the valid models have the same output direction, and at least one model among the models participating in the consensus judgment has a high confidence level.
[0148] When the completeness of the input features is not low and there is at least one valid model with a not low confidence level, the confidence level of the model output is denoted as medium.
[0149] All other cases are recorded as low; the aforementioned output direction adopts 0.5 in a unified score range as the direction dividing point in its implementation. A score greater than or equal to 0.5 is recorded as a high-risk direction, and a score less than 0.5 is recorded as a low-risk direction.
[0150] If the number of valid models is 0 when the current request enters the decision table, the aforementioned proportion of consistent models will not be calculated. Instead, consistency will be recorded as low and model output credibility will be recorded as low. The absence of valid model output will be written as an independent marker into the risk input package so that rule-based, verification-based, or missing compensation-based strategies can be prioritized during subsequent strategy screening.
[0151] Business risk tolerance is determined by the business type, transaction amount, customer level, and risk loss limit configured in the business request.
[0152] When processing a request, the risk control server first identifies the type of business the request belongs to, and then combines the transaction amount and customer level to perform a lookup mapping according to the risk loss limit table corresponding to the business type, amount range and customer level to determine the acceptable risk range of the request under the current scenario.
[0153] When the transaction amount is not greater than the first amount threshold and the customer level is the first preset level, the first risk tolerance is assigned accordingly; when the transaction amount is greater than the second amount threshold and the customer level is the second preset level or the initial state, the second risk tolerance is assigned accordingly; if the risk loss limit configured for the business scenario is less than the preset loss limit threshold, the business risk tolerance is adjusted down by one preset level.
[0154] This tolerance level is expressed in a tiered manner, that is, it is mapped to three control levels: low, medium and high, which correspond to the retention range of the first strategy, the retention range of the second strategy and the retention range of the third strategy, respectively, so as to be consistent with the strategy screening and disposal priority judgment.
[0155] Furthermore, the business risk tolerance is determined using a fixed table lookup logic, rather than relying on manual ad-hoc judgment: first, basic levels are divided according to business type, with transfers, withdrawals, and credit applications recorded as strict, consumer payments and top-ups recorded as standard, and inquiries or low-amount benefit receipts recorded as lenient;
[0156] The transaction amount is then divided into tiers, with amounts not exceeding the first threshold being classified as small, amounts exceeding the first threshold but not exceeding the second threshold being classified as medium, and amounts exceeding the second threshold being classified as large, where the first threshold is less than the second threshold.
[0157] Customers are further categorized into high-level, medium-level, and low-level customers. For strict business transactions, customers with large amounts or low-level status are directly classified as having low tolerance; customers with medium amounts and medium-level status are classified as having medium tolerance.
[0158] The rest are determined to be of medium or high tolerance; for standard business, if the customer is a large amount and low-level customer, the tolerance level is determined to be low; if the customer is a small amount and high-level customer, the tolerance level is determined to be high.
[0159] The rest are classified as medium tolerance; for lenient business, if the amount is small and the customer level is high or ordinary, it is classified as high tolerance; if the amount is large and the customer level is low, it is classified as medium tolerance.
[0160] The rest are determined to be of medium tolerance; the risk loss limit configured in the business scenario is treated as a correction item: when the limit is lower than the scenario preset lower limit, the above lookup result is lowered by one level; when the limit is higher than the scenario preset upper limit, the above lookup result is raised by one level; if it is already at the lowest or highest level, it remains unchanged;
[0161] Furthermore, to avoid thresholds and weights existing only in a preset abstract manner, in this embodiment, the model confidence level threshold, input completeness level threshold, amount threshold, and state switching threshold are all determined by either historical sample statistical results of the same business scenario or manually configured criteria.
[0162] When the effective sample size reaches the minimum sample size condition within the evaluation period, the threshold is generated by using the quantile results of historical samples first; when the effective sample size is insufficient or the business has just been launched, the scenario initialization configuration value is used; once the threshold takes effect in the current evaluation period, it remains fixed within the period and does not fluctuate in real time with a single request.
[0163] The entire threshold set is allowed to be replaced with a new version only when the next evaluation period arrives and the sample conditions are met again; if the difference between the old and new thresholds exceeds the preset change limit, the update will be performed step by step according to the change limit to avoid large fluctuations in the state machine and rule tree due to statistical fluctuations.
[0164] Correspondingly, the preset weights in the strategy evaluation also adopt a control method of being fixed within the period and updated when the period changes; if the sample size of a certain strategy is insufficient in the current period, the weight of the previous period will be used, and its influence will not be immediately amplified or reduced due to local sample anomalies.
[0165] In the strategy evaluation section, the risk control server calculates the risk interception effectiveness, false interception index, execution time and resource consumption for each risk control strategy, and performs normalization before sorting them according to the aforementioned comprehensive evaluation score.
[0166] The higher the score, the more suitable the strategy is to be included in the strategy portfolio in the current scenario; the weight is determined by historical sample statistics, business risk control objectives and strategy execution results, and is used to balance the relationship between risk identification capability and processing latency and resource consumption.
[0167] In cases where there are multiple candidate model invocation strategies, first compare the reliability of the model output, and then decide whether to retain the high-time-consuming model invocation strategy based on the business risk tolerance.
[0168] When the model output has low credibility and the business risk tolerance is low, rule-based and verification-based strategies should be retained first to avoid misjudgment caused by relying solely on the model output. The aforementioned risk interception effectiveness, false interception indicators, execution time and resource consumption are all normalized without dimension before being substituted into the comprehensive evaluation score. Therefore, each item in the comprehensive evaluation score can be directly linearly combined according to the weights, without directly adding or subtracting the original values of different dimensions.
[0169] To avoid merely describing the results when adjusting strategy weights based on scenarios, the pre-defined strategy-scenario mapping relationship uses a fixed weighting rule:
[0170] When the current state is a low-reliability data state, the weight coefficients of the missing data compensation strategy and the secondary verification strategy are each increased by a preset step, and the weight coefficient of the deep model calling strategy is decreased by a preset step.
[0171] When the current state is suspected of being a group fraud, the weight coefficients of the risk diffusion strategy, the basic rule execution strategy, and the delayed review strategy are each increased by one preset step; when the current state is suspected of being a rule probing state, the weight coefficients of the silent monitoring strategy, the secondary verification strategy, and the limit strategy are each increased by one preset step.
[0172] When the current state is high-concurrency degradation state, the weight coefficients of the deep model calling strategy and the high-time-consuming missing compensation strategy are each reduced by a preset step, while the weight coefficient of the basic rule execution strategy is increased by a preset step; when the current state is trial operation monitoring state, the weight coefficients of the strategies in the evaluation period are uniformly reduced by a preset step but not directly removed.
[0173] The preset step size is given by the configuration parameters, and the weight coefficients of each strategy are limited to the preset minimum and maximum values after a single adjustment. After the agent control module completes the above scenario correction, it multiplies the corrected weight coefficients by the original comprehensive evaluation score of the strategy as the final ranking basis for the strategy.
[0174] In terms of strategy types, the preset risk control strategy library includes at least the following: deep model calling strategy, basic model calling strategy, basic rule execution strategy, missing compensation strategy, secondary verification strategy, limit strategy, delayed review strategy, silent monitoring strategy, risk diffusion strategy, and strategy blocking strategy.
[0175] The missing data compensation strategy uses a specified data source or preset alternative features to fill in the missing items when the input fields are insufficient; the risk diffusion strategy increases the risk level of the associated entities or expands the detection scope of the associated objects when gang association or linkage risks are identified.
[0176] The policy blocking policy stops the invocation of a policy or its corresponding model when a policy exceeds its budget, triggers an exception, or fails to comply with compliance constraints.
[0177] When selecting strategies, the risk control server retains strategies whose execution time does not exceed the latency budget, whose resource consumption does not exceed the resource budget, and whose compliance constraints are not violated, and adjusts the strategy weights according to the preset strategy-scenario mapping relationship;
[0178] The strategy-scenario mapping relationship is used to explain which strategies are more suitable to be prioritized in different business scenarios. For example, when the data reliability is low, the weight of missing data compensation strategy and secondary verification strategy can be increased; when the strategy probing degree is high, the weight of risk diffusion strategy and silent monitoring strategy can be increased.
[0179] If multiple strategies output different handling instructions for the same request, the risk control server selects the instruction with the highest priority according to the pre-configured risk instruction priority table. The priority in the risk instruction priority table, from high to low, includes at least rejection, temporary freeze, SMS verification or face verification, amount limit, delayed review and allow, so as to ensure that there is only one definite output in the final handling result.
[0180] The state machine is operated by the risk control server, which switches between different states based on the current risk status indicators. The normal monitoring state is used for general requests; the high-concurrency degradation state is used to reduce some time-consuming processing when the system pressure is high; and the low-reliability data state is used to enhance validation when input fields are missing or channels are abnormal.
[0181] The suspected group fraud state is used to trigger corresponding strong control strategies when the number of associated accounts and devices is greater than or equal to a preset threshold; the suspected rule probing state is used to prevent probing when repeated rule boundaries are detected.
[0182] The strong control state is used to enforce control when a high-risk preset blocking list is hit or when the risk score is greater than or equal to the preset upper limit threshold; the trial operation monitoring state is used to record the operation results when the strategy is under evaluation or de-weighting observation.
[0183] When multiple states are satisfied simultaneously, the current state is determined according to the given priority order to avoid execution conflicts caused by entering multiple states at the same time;
[0184] The continuous preset duration is determined by the system monitoring component within a continuous monitoring window with the business request reception time as the endpoint. Only when the current transaction processing volume per second and processor utilization continuously exceed their respective thresholds within the continuous monitoring window will it switch to the high concurrency degradation state.
[0185] If it is only a momentary spike, then the original state remains unchanged; the aforementioned risk score in the context of state machine switching corresponds to the current risk score formed based on the obtained risk state indicators and the previously available model scores, and is used to compare with the strong control threshold.
[0186] To avoid repeatedly switching between running states during the processing of a single request, the risk control server determines the current state of the same risk event identifier only once before entering the judgment rule tree;
[0187] This status remains unchanged until the request is completed, and subsequent requests will have their status reassessed based on the results of the new system monitoring window.
[0188] Furthermore, to ensure engineering stability during state transitions, in addition to meeting the entry conditions, the risk control server also implements minimum hold duration control for high-concurrency degraded states, low-reliability data states, and suspected rule-trial states:
[0189] If a request has been determined to a certain state before entering the decision rule tree, then even if a subsequent node reads a monitoring value that has fallen back briefly, it will not exit that state during the processing of the request.
[0190] When a new request is re-evaluated, the state will only exit if the corresponding exit condition is met consecutively a preset number of times. For the strong control state and the suspected group fraud state, no delayed entry condition is set. Once the condition is met, it will take effect immediately, but exiting still requires at least one full evaluation cycle of review and confirmation.
[0191] By using a state-reversal logic that prioritizes strict entry and lenient exit, we can avoid frequent oscillations in policy weights and model call permissions between adjacent requests caused by instantaneous fluctuations.
[0192] In a preferred embodiment, the judgment rule tree is executed in a fixed order: matching the preset blocking list, judging data reliability, judging initial risk score, judging gang association, judging model invocation, and executing instruction output.
[0193] Risk thresholds are provided by a threshold table. The initial risk score is determined based on traffic pressure, resource pressure, data reliability, channel anomaly, gang association, strategy probing, model output credibility, business risk tolerance, and the obtained model score according to preset weights.
[0194] When the preset blocking list is hit or the initial risk score is greater than or equal to the risk threshold, the corresponding business processing instruction of rejection or strong control is output; when the preset blocking list is not hit and the initial risk score is less than the risk threshold, the corresponding business processing instruction of permission or silent monitoring is output, and model call nodes marked as high time consumption and unnecessary are skipped according to the model call table.
[0195] Business processing instructions include at least one of the following: allow, deny, SMS verification, face verification, amount limit, delayed review, silent monitoring, and temporary freeze; risk control execution instructions include at least one of the following: basic rule execution, basic model call, and deep model call.
[0196] Record the risk event identifier, multi-dimensional risk characteristics, traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, strategy selection, risk result, business handling instructions, risk control execution instructions, execution status and post-event results for each business request to form a strategy log;
[0197] Post-event results include at least one of the following: final transaction result, fraud confirmation result, review result, and customer complaint result; when the preset evaluation period and minimum sample size conditions are met, the risk interception accuracy, risk interception effectiveness, false interception rate, average processing delay, and resource consumption are calculated based on the strategy log;
[0198] When the false interception rate, average processing delay, or resource consumption exceeds the preset threshold for false interception rate, processing delay, or resource consumption, the corresponding risk control strategy will be disabled, downgraded, or transferred to trial operation monitoring mode; when none of them exceed the corresponding threshold, the current applicable status of the corresponding risk control strategy will be maintained.
[0199] When a strategy meets the preset re-activation conditions, demotion recovery conditions, or trial run end conditions, update the strategy weight, applicable status, threshold, or strategy resource consumption limit in the strategy library, and add it to the strategy candidate set when the re-addment conditions are met.
[0200] Please see Figure 2 A multi-strategy collaborative intelligent agent risk control adaptation system includes: a multi-source data acquisition module, used to acquire transaction, account, device, network, location, behavior, channel, system, graph, historical strategy and preset blocking list data, and to acquire latency budget, resource budget and compliance constraint parameters;
[0201] The risk feature construction module is used to perform timestamp alignment, field standardization, missing data marking, and feature normalization on the above data, using the business request reception time as a unified time benchmark.
[0202] The risk status analysis module is used to generate traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, and business risk tolerance.
[0203] The intelligent agent control module is used to evaluate, filter, sort, combine and adjust risk control strategies from the preset risk control strategy library, generate strategy combinations and send them to the strategy rule conversion module;
[0204] The strategy rule conversion module is used to generate state machines, threshold tables, decision rule trees, risk instruction sets, model call tables, and blocking control tables based on strategy combinations.
[0205] The risk control execution module is used to perform risk control processing according to the risk instruction priority table and judgment rule tree, and send business processing instructions to the transaction gateway, account management system, verification service or audit system through the communication interface;
[0206] The effect feedback and adjustment module is used to receive the execution status, handling status and post-event results returned by the transaction gateway, account management system, verification service or audit system, form a strategy log, and perform deactivation, demotion or trial operation of risk control strategies according to preset strategy adjustment conditions.
[0207] In this implementation, the rule tree is executed in a fixed node order; the risk control server first determines whether the request matches the preset blocking list by matching the node; if it matches, the corresponding business processing instruction is directly output, and the subsequent time-consuming and unnecessary model call nodes can be skipped.
[0208] If the preset blocking list is not matched, proceed to the data reliability judgment node to check whether there are obvious missing or abnormalities in the current input;
[0209] Among them, the data reliability node uniformly reads the data reliability value and the corresponding reliability level in the risk input package. The higher the data reliability value, the more reliable the input. The lower the reliability level, the more reliable the input. The lower the reliability level, the more necessary it is to enhance the verification or reduce the policy calls to highly dependent complete inputs.
[0210] If the data reliability meets the requirements, proceed to the initial risk scoring judgment node. Calculate the initial risk score based on the risk status indicators and the obtained model score according to the preset weights, and compare it with the risk threshold in the threshold table. If the initial risk score is greater than the risk threshold, output the corresponding business processing instruction.
[0211] If the control is still not triggered, proceed to the group association judgment node to continue to judge whether there are any abnormalities in the number of associated accounts, the number of associated devices, and the aggregation coefficient;
[0212] If the group association is not triggered, the process proceeds to the model call judgment node, where the system decides whether to call the basic model or the deep model according to the model call table; then it proceeds to the instruction output node to generate the final risk control execution instruction and business processing instruction.
[0213] The risk thresholds in the above processing are provided by a threshold table, and the initial risk score is determined according to the risk status index and the obtained model score according to a preset weight.
[0214] Before scoring, the risk control server will first organize the current risk status indicators such as traffic pressure, resource pressure, data reliability, channel anomaly, gang association, strategy probing, model output credibility, and business risk tolerance into calculable inputs, and then form an initial risk score according to preset weights.
[0215] The obtained model scores are limited to valid scores that already exist in the risk input package before entering the model call judgment node. If multiple valid scores exist at the same time, they are first filtered according to model confidence and output consistency, and then the weighted average is taken as the current score input.
[0216] If only a single valid score exists, that score is used directly; if no score is available, the missing marker is retained and the missing score handling logic is used for scoring, so that the fixed node order is not disrupted due to missing scores.
[0217] To maintain temporal consistency, the obtained model scores are limited to the model results returned before entering the initial risk scoring judgment node, and do not include model outputs that have not yet been called or returned in subsequent model call judgment nodes;
[0218] Therefore, the initial risk score is calculated using the available prior information, while subsequent model call nodes only supplement new model results as the rule tree continues to be executed, without rewriting the judgments of the completed prior nodes in reverse.
[0219] Furthermore, to enable the initial risk score to be directly derived from the text, the risk control server employs a fixed input table and fixed output rules; the input vectors for the initial risk score are arranged in the following order:
[0220] Traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, business risk tolerance, and the total value of the model scores obtained;
[0221] Among them, traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, and the total value of the obtained model score are all input with values between 0 and 1. The model output credibility is mapped to 1, 0.5 and 0 for high, medium and low respectively, and the business risk tolerance is mapped to 0, 0.5 and 1 for high, medium and low respectively, so as to reflect that the lower the tolerance, the stricter the control.
[0222] The data reliability is directly adopted from the reliability value generated in the previous step. The larger the value, the more complete and reliable the data is. Therefore, before participating in the initial risk scoring, it is first converted into a low reliability risk input. The low reliability risk input is determined by subtracting the data reliability value from 1.
[0223] The obtained model score summary value is obtained by weighting the confidence of the effective model; if there is no effective model score, the summary value is set to 0.5 and the missing model score mark is retained for subsequent model call judgment nodes.
[0224] The initial risk score is not calculated using a random process, nor does it rely on online training results. Instead, it is formed by accumulating the above input vectors item by item according to preset weights. Each weight is synchronously saved by a threshold table, remains fixed within the same business scenario, and the total weight is 1.
[0225] When a preset blocking list is hit or the initial risk score is greater than the risk threshold, the risk control server outputs the corresponding business processing instruction and skips the model call nodes marked as high time consumption and unnecessary according to the model call table, so as to reduce unnecessary processing overhead.
[0226] Business processing instructions include at least one of the following: allow, deny, SMS verification, facial verification, amount limit, delayed review, silent monitoring, and temporary freeze. Temporary freeze is used to temporarily block and lock the current account or transaction that initiated the request. Risk control execution instructions include at least one of the following: basic rule execution, basic model call, and deep model call.
[0227] Upon receiving the instruction, the transaction gateway, account management system, verification service, or auditing system shall respectively execute the corresponding transaction release, account control, verification action, or auditing action;
[0228] To avoid the lack of clear mapping for the corresponding business processing instructions, the threshold table in this embodiment stores at least three types of thresholds: allow threshold, verification threshold and strong control threshold, and satisfies that the allow threshold is less than the verification threshold and the verification threshold is less than the strong control threshold.
[0229] The risk control server outputs the handling result according to the following fixed rules at the initial risk scoring judgment node: when it hits the high-risk preset blocking list, it directly outputs rejection or temporary freeze;
[0230] If the preset blocking list is not matched and the initial risk score is less than the allowed threshold, output "allow" or "silent monitoring"; if the initial risk score is not less than the allowed threshold but less than the verification threshold, output "SMS verification", "face verification" or "amount limit";
[0231] If the initial risk score is not less than the verification threshold and is less than the strong control threshold, output delayed review or temporary freeze; if the initial risk score is not less than the strong control threshold, output rejection or temporary freeze.
[0232] The specific instruction selected within the same scoring range is determined by the previously generated risk instruction priority table and the current business scenario configuration; if the scenario configuration restricts a certain handling instruction from being available, then the next priority instruction within the same range is selected.
[0233] Furthermore, the model call judgment node does not use a generalized expression, but is executed in a table-driven manner: the model call table contains at least five fields: model identifier, required flag, high time consumption flag, allowed state set, and timeout limit; after the risk control server enters the node, it first reads the current state machine state;
[0234] If the current state is not within the set of allowed states for a certain model, then skip that model; if a model is marked as high-time-consuming and the current request has already triggered verification, delayed review, rejection, or temporary freezing in the preceding node, then skip that model.
[0235] If a model is marked as necessary and is not prohibited from being called by compliance constraints, the model will continue to be executed even if the aforementioned handling results exist, and the output will only be used for logging, review support or subsequent policy feedback, without reversing the output of the previous handling conclusion; if the model execution exceeds its timeout limit, it will be recorded as timeout and subsequent calls to similar high-time-consuming models will be stopped according to the blocking control table.
[0236] Furthermore, in order to form a deterministic execution loop for the sending, receiving, and writing back processes, the risk control execution module adopts five execution states for the processing of each request: pending issuance, issued, confirmed, timed out and reissued, and closed and completed.
[0237] Specifically, after a business processing instruction is generated at the instruction output node, the instruction is first set to a pending state and a confirmation timer is started; after receiving the instruction, the transaction gateway, account management system, verification service, or auditing system returns the execution acceptance result and requests to change the status to issued.
[0238] When the external system confirms the action within the preset confirmation time, the system is switched to the confirmed state. If no confirmation is received after the confirmation time, the system enters the timeout resend state and only resends instructions with idempotency once. If no confirmation is received after resend, an execution exception flag is written according to the failure branch, but the output master decision is not rolled back.
[0239] When both the execution and processing statuses have stabilized, or when the preset maximum waiting time has been reached, the system will uniformly transition to the closed completion status. For instructions such as rejection, temporary freezing, and amount limits that directly affect the transaction flow, the risk control server will only consider the control to have been successfully implemented after receiving the acceptance result from the transaction gateway or account management system.
[0240] For subsequent action-oriented instructions such as SMS verification, facial verification, and delayed review, the intermediate completion conclusion will be output first after the verification service or review system accepts the request, and the subsequent results will be completed asynchronously.
[0241] Through the aforementioned state branches, even when the external system fails to respond in a timely manner, responds repeatedly, or responds late, the risk control link can still maintain a unique main conclusion and avoid multiple conflicting handling results corresponding to the same risk event identifier.
[0242] In the recording and retrieval section, the risk control execution module writes the risk event identifier, input characteristics, risk status indicators, strategy selection, risk result, business handling instructions, risk control execution instructions, execution status, and post-event results of each business request into the strategy log; the post-event results include at least one of the following: final transaction result, fraud confirmation result, review result, and customer appeal result;
[0243] After meeting the preset evaluation period and minimum sample size conditions, the effect feedback and adjustment module calculates the risk interception accuracy, risk interception effectiveness, false interception rate, average processing delay and resource consumption based on the strategy log, and uses these results to determine whether the strategy needs to be stopped, downgraded or transferred to trial operation monitoring state.
[0244] If the false interception rate, average processing delay, or resource consumption of a certain strategy exceeds the corresponding threshold, the strategy will be disabled, downgraded, or put into trial operation.
[0245] If the policy meets the conditions for reactivation, demotion recovery, or trial run termination, then update the policy weight, applicable status, threshold, or policy resource consumption limit in the policy library, and re-add it to the policy candidate set when the conditions for re-addment are met.
[0246] The execution status and handling status mentioned above can be written back immediately after the current request is processed, while the post-event results are written back asynchronously after the transaction is completed, the review is completed, or the appeal is processed. Both are associated through the same risk event identifier, but the post-event result feedback does not block the risk control decision output of the current business request.
[0247] Furthermore, to avoid the log write-back process slowing down the main decision-making process, the risk control execution module generates a read-only snapshot after the main process is completed and hands it over to the asynchronous recycling unit for processing:
[0248] The execution status, strategy selection, and current handling results are first added to the queue for collection in a sequential append manner. Then, the effect feedback and adjustment module merges the subsequent review results and appeal results returned for the same request according to the risk event identifier.
[0249] If the same risk event identifier is detected to be repeatedly transmitted during the asynchronous write-back phase, the intermediate state is overwritten with a valid state, but the state flipping order is retained for subsequent policy evaluation.
[0250] For requests that have not received a result after the preset collection time limit, instead of waiting indefinitely, they will be entered into the evaluation buffer as a result pending completion; they will be incorporated into the corresponding evaluation cycle when the result is completed later.
[0251] By separating the main process from the feedback process, the risk control decision output of the current request and the subsequent strategy effect statistics can be decoupled in time, while maintaining a consistent closed loop of risk event identification throughout the collection, decision-making, execution and post-testing stages.
[0252] In this embodiment, the system consists of a multi-source data acquisition module, a risk feature construction module, a risk status analysis module, an intelligent agent control module, a strategy rule conversion module, a risk control execution module, and an effect feedback and adjustment module. Each module works together through an internal communication interface and a business system interface.
[0253] The multi-source data acquisition module is responsible for sending data from the trading system, account management system, device fingerprint system, network monitoring component, location service, behavior tracking, channel service, system monitoring component, graph service, historical strategy library, and preset blocking list library into the risk control processing link;
[0254] The risk control execution module is responsible for sending business processing instructions to the transaction gateway, account management system, verification service, or auditing system; the effect feedback and adjustment module is responsible for writing the execution status, processing status, and post-event results back to the strategy log and updating the strategy library accordingly.
[0255] The technical principles of the present invention have been described above with reference to specific embodiments. These descriptions are merely for explaining the principles of the invention and should not be construed as limiting the scope of protection of the invention in any way. Based on this explanation, those skilled in the art can readily conceive of other specific embodiments of the invention without inventive effort, and these equivalent variations or substitutions are all included within the scope defined by the claims of this application.
Claims
1. An agent risk control adaptation method based on multi-strategy cooperation, characterized in that, include: Acquire data related to business requests, including transactions, accounts, devices, networks, locations, behaviors, channels, systems, data graphs, historical policies, and preset blocking lists, as well as latency budgets, resource budgets, and compliance constraint parameters. The compliance constraint parameters include at least prohibited models, mandatory verification steps, and data processing or disposal instruction restrictions. Assign a risk event identifier to the business request, and based on the receipt time of the business request, perform timestamp alignment, field standardization, and missing data marking on the above data within a preset time window to form a risk input package; Based on the aforementioned risk input package, multidimensional risk characteristics are generated, including traffic pressure, resource pressure, data reliability, channel anomaly, gang association, strategy probing, model output credibility, and business risk tolerance. The risk interception effectiveness, false interception index, execution time and resource consumption of each strategy in the preset risk control strategy library are normalized and evaluated. Under the time delay budget, the resource budget and the compliance constraint parameters, the strategy weights are screened, sorted and combined and adjusted. Conflicts are handled according to the preset handling instruction priority to generate strategy combinations. Based on the strategy combination, a state machine, threshold table, decision rule tree, risk instruction set, model call table, and blocking control table are generated. Risk control processing is executed according to the fixed node sequence of the judgment rule tree, outputting risk results, business handling instructions, and risk control execution instructions, and the execution results and post-transaction results after the transaction or review are fed back to the intelligent agent control module; in the fixed node sequence of the judgment rule tree, the initial risk score judgment node is located before the model call judgment node; the initial risk score is determined only based on the model results obtained before entering the initial risk score judgment node, and the model results obtained by the model call judgment node are not used to modify the judgments of the completed preceding nodes; The current state is determined only once for the same risk event identifier before entering the judgment rule tree, and remains unchanged until the business request is processed. The model call table contains at least the required flags and allowed state set for each model. When the current state is not in the allowed state set of the corresponding model, the call to that model is skipped. If a model is marked as required and is not prohibited from being called by compliance constraints, the model will continue to be executed even if the aforementioned processing results exist, without reversing the output of the previous processing conclusion.
2. The agent risk control adaptation method based on multi-strategy cooperation according to claim 1, characterized in that, The transaction includes the amount, transaction type, and timestamp; The account includes account identifier, registration duration, number of successful transactions, and number of failed transactions. The device includes device identification, device fingerprint integrity, and device switching frequency; The network includes risk level, agent identifier, and network latency; The location includes location information, place of origin information, and location deviation distance; The behavior includes the mean and variance of the click interval; The channel includes channel coding, feature missing rate, and proportion of abnormal samples; The system includes current throughput, latency, processor utilization, memory utilization, cache hit rate, and queue length calculated within a preset statistical window ending at the time the service request is received. The graph includes the number of associated accounts, the number of associated devices, and the clustering coefficient; The historical strategy includes rule hit count, model score, model credibility, historical secondary validation records, and historical handling results; The preset blocking list data includes at least one of account identifier, device identifier, and network node address.
3. The intelligent agent risk control adaptation method based on multi-strategy collaboration according to claim 2, characterized in that, cutting off the continuous features in the multi-dimensional risk features according to preset lower limits and preset upper limits corresponding to each feature, and normalizing to setting a missing mark for the missing values; The preset lower limit and preset upper limit are determined by the historical sample statistical range or configuration parameters; traffic pressure is generated based on the normalized current throughput, latency and queue length; resource pressure is generated based on processor utilization, memory utilization and cache hit rate; data reliability is generated based on feature missing rate, device fingerprint integrity and network latency; channel anomaly degree is generated based on the proportion of abnormal samples in the channel, location deviation distance and risk level; gang association degree is generated based on the number of associated accounts, the number of associated devices and the aggregation coefficient; and strategy probing degree is generated based on the number of rule hits, the historical secondary verification failure rate and device switching frequency statistically calculated according to account identifier and device identifier respectively.
4. The agent risk control adaptation method based on multi-strategy cooperation according to claim 1, characterized in that, The reliability of the model output is determined based on the model score, model confidence, input feature completeness, and consistency of the outputs of multiple models from the currently invoked base model, deep model, or other risk control models. The business risk tolerance is determined based on the business type, transaction amount, customer level, and risk loss limit configured in the business scenario of the business request; the risk interception effectiveness, false interception index, execution time, and resource consumption of each risk control strategy are normalized, and a comprehensive evaluation score is calculated according to preset weights. wherein, represents the comprehensive evaluation score of the risk control strategy, , , and respectively represent the risk interception efficiency absolute value, the false interception index absolute value, the execution time consumption absolute value and the resource occupation absolute value corresponding to the risk control strategy. Each of the above needs to be converted into a dimensionless normalized value by the maximum-minimum normalization function before being substituted into the comprehensive evaluation score formula , , , to ensure dimensional consistency, wherein , , and are preset weight factors corresponding to the risk interception effectiveness, the false interception index, the execution time consumption, and the resource occupation, respectively.
5. The intelligent agent risk control adaptation method based on multi-strategy collaboration according to claim 1, characterized in that, The risk control strategies include deep model invocation strategy, basic model invocation strategy, basic rule execution strategy, missing data compensation strategy, secondary verification strategy, limit strategy, delayed review strategy, silent monitoring strategy, risk diffusion strategy, and strategy blocking strategy. The missing data compensation strategy is used to call a specified data source or adopt a preset alternative feature; the risk diffusion strategy is used to increase the risk level of the associated subject or expand the detection range of the associated object; and the strategy blocking strategy is used to stop the call of a specified strategy or model. Retain strategies that do not exceed the latency budget, do not exceed the resource budget, and do not violate compliance constraints, and adjust the strategy weights according to the preset strategy-scenario mapping relationship; When multiple strategies output different handling instructions, the instruction with the highest priority is selected according to the pre-configured risk instruction priority table. The priority in the risk instruction priority table, from high to low, includes at least rejection, temporary freeze, SMS verification or face verification, amount limit, delayed review, and permission.
6. The intelligent agent risk control adaptation method based on multi-strategy collaboration according to claim 2, characterized in that, The state machine includes a normal monitoring state, a high-concurrency degradation state, a low-reliability data state, a suspected group fraud state, a suspected rule probing state, a strong control state, and a trial operation monitoring state; When the current transaction volume per second is greater than or equal to the transaction volume per second threshold and the processor utilization is greater than or equal to the processor utilization threshold, and this continues for a preset duration, the system enters a high-concurrency degradation state. If the transaction volume per second threshold or the processor utilization threshold is less than the threshold, then normal monitoring status is maintained. When the feature missing rate is greater than or equal to the missing rate threshold or the channel anomaly degree is greater than or equal to the channel anomaly threshold, the data enters a low reliability state. When the clustering coefficient is greater than or equal to the clustering coefficient threshold and the number of associated accounts is greater than or equal to the number of associated accounts threshold, the system enters a suspected group fraud state. When the strategy probing degree is greater than or equal to the probing threshold, it enters the suspected rule probing state; when it hits the high-risk preset blocking list or the risk score is greater than or equal to the strong control threshold, it enters the strong control state. When the risk control strategy is under evaluation or demotion observation, it enters the trial operation monitoring state; when multiple states are met at the same time, the current state is determined in the following order: strong control state, suspected group fraud state, suspected rule probing state, low data reliability state, high concurrency demotion state, trial operation monitoring state, and normal monitoring state. The thresholds for transaction processing per second, processor utilization, missing rate, channel anomaly, clustering coefficient, number of associated accounts, trial threshold, and strong control threshold are all provided by a preset threshold table or determined through statistical distribution of historical samples.
7. The intelligent agent risk control adaptation method based on multi-strategy collaboration according to claim 1, characterized in that, The judgment rule tree is executed in a fixed order: matching the preset blocking list, judging data reliability, judging initial risk score, judging gang association, judging model invocation, and executing instruction output. The risk threshold is provided by the threshold table. The initial risk score is determined according to the traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility and business risk tolerance, and the obtained model score according to the preset weights. When the preset blocking list is hit or the initial risk score is greater than or equal to the risk threshold, the corresponding business processing instruction of rejection or strong control is output; when the preset blocking list is not hit and the initial risk score is less than the risk threshold, the corresponding business processing instruction of permission or silent monitoring is output, and model call nodes marked as high time consumption and unnecessary are skipped according to the model call table. The business processing instructions include at least one of the following: allow, deny, SMS verification, face verification, amount limit, delayed review, silent monitoring, and temporary freeze. The risk control execution instructions include at least one of the following: basic rule execution, basic model call, and deep model call.
8. The intelligent agent risk control adaptation method based on multi-strategy collaboration according to claim 1, characterized in that, Record the risk event identifier, multi-dimensional risk characteristics, traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, strategy selection, risk result, business handling instructions, risk control execution instructions, execution status and post-event results for each business request to form a strategy log; The post-event results include at least one of the following: final transaction result, fraud confirmation result, review result, and customer complaint result; When the preset evaluation period and minimum sample size conditions are met, the risk interception accuracy, risk interception effectiveness, false interception rate, average processing delay and resource consumption are calculated based on the strategy log. When the false interception rate, average processing delay, or resource consumption exceeds the preset threshold for false interception rate, processing delay, or resource consumption, the corresponding risk control strategy is disabled, downgraded, or transferred to trial operation monitoring mode; when none of the thresholds are exceeded, the current applicable state of the corresponding risk control strategy is maintained. When a strategy meets the preset re-activation conditions, demotion recovery conditions, or trial run end conditions, update the strategy weight, applicable status, threshold, or strategy resource consumption limit in the strategy library, and add it to the strategy candidate set when the re-addment conditions are met.
9. A multi-strategy collaborative intelligent agent risk control adaptation system, employing the multi-strategy collaborative intelligent agent risk control adaptation method according to any one of claims 1-8, characterized in that, include: The multi-source data acquisition module is used to acquire transaction, account, device, network, location, behavior, channel, system, graph, historical strategy and preset blocking list data associated with business requests, as well as latency budget, resource budget and compliance constraint parameters. The compliance constraint parameters include at least the models that are prohibited from being called, the verification steps that must be performed, and restrictions on data processing or disposal instructions. The risk feature construction module is used to assign risk event identifiers to the business requests. Based on the receipt time of the business requests, the module performs timestamp alignment, field standardization, and missing data marking on the data within a preset time window to form a risk input package. The risk status analysis module is used to generate multi-dimensional risk characteristics and traffic pressure, resource pressure, data reliability, channel anomaly, gang association degree, strategy probing degree, model output credibility, and business risk tolerance based on the risk input package. The intelligent agent control module is used to perform normalized evaluation of the risk interception effectiveness, false interception index, execution time and resource consumption of each strategy in the preset risk control strategy library. Under the time delay budget, the resource budget and the compliance constraint parameters, the module filters, sorts, combines and adjusts the strategy weights, and handles conflicts according to the preset handling instruction priority to generate strategy combinations. The strategy rule conversion module is used to generate a state machine, a threshold table, a decision rule tree, a risk instruction set, a model call table, and a blocking control table based on the strategy combination. The risk control execution module is used to execute risk control processing according to the fixed node sequence of the judgment rule tree, and output risk results, business handling instructions, and risk control execution instructions. In the fixed node sequence of the judgment rule tree, the initial risk score judgment node is located before the model call judgment node. The initial risk score is determined only based on the model results obtained before entering the initial risk score judgment node, and the model results obtained by the model call judgment node are not used to modify the judgments of completed preceding nodes. The current state identifier for the same risk event is determined only once before entering the judgment rule tree and remains unchanged until the business request is processed. The model call table contains at least the required flags and allowed state set for each model. When the current state is not in the allowed state set of the corresponding model, the call to that model is skipped. If a model is marked as required and is not prohibited from being called by compliance constraints, the model will continue to be executed even if the aforementioned processing results exist, without reversing the output of the previous processing conclusions; The effect feedback and adjustment module is used to receive the execution status, processing results and post-event results returned by the transaction gateway, account management system, verification service or audit system, form a strategy log, and perform deactivation, demotion or trial operation of risk control strategy according to preset strategy adjustment conditions, and feed back the execution results of the risk control execution module and the post-event results after the transaction or review are completed to the intelligent agent control module.
Citation Information
Patent Citations
Risk control model prediction result assessment method and device
CN114048992A
Multi-source model score consistency calibration method and system
CN121614367A
