Resource constraint-oriented network transaction dynamic intelligent management and control method and system

Abstract

The invention discloses a resource constraint-oriented network transaction dynamic intelligent management and control method and a resource constraint-oriented network transaction dynamic intelligent management and control system. The method comprises the following steps: a state sensing step: acquiring monitoring data; a risk quantification step of generating a risk evaluation value; an intelligent decision-making step of determining a target management and control action based on a decision mapping model satisfying a mapping monotonicity constraint, the constraint ensuring that an expected system resource consumption level of the action does not decrease progressively along with a risk value; a resource scheduling step: dynamically regulating and controlling action execution according to expected consumption and a real-time resource state; and an action execution step. The system adopts a modular micro-service architecture to realize the method. According to the invention, by introducing the mapping monotonicity constraint and the dynamic resource scheduling, the problems of stability risk and resource mismatching caused by decoupling of the management and control action and the system resource state in the prior art are solved, and the dynamic optimal balance of the management and control efficiency and the resource consumption is realized.

Description

Technical Field

[0001] This invention relates to the fields of complex system resource management, data analysis, and automated decision-making. Specifically, it relates to an intelligent dynamic management and control method and system for managing and optimizing the operational status of a network trading ecosystem. This system, acting as a central upper-level management hub, can connect to and schedule multiple underlying functional modules, such as transaction execution, risk identification, and resource management, to achieve an optimal balance between system resource consumption and platform management effectiveness. Background Technology

[0002] Online transactions have become a core component of the digital economy, widely applied in diverse scenarios such as e-commerce retail, online services, and digital content delivery. Maintaining a healthy, efficient, and secure transaction environment, and preventing fraudulent transactions, fake orders, information abuse, and market manipulation, are core technological challenges that platform operators must address.

[0003] Existing technical solutions mostly approach the issue from a single perspective or with static rules, lacking a systematic and dynamically optimizable control perspective, and mainly suffer from the following technical defects.

[0004] Existing technical solutions typically have the following limitations in technical architecture and execution logic at the system implementation level, for example: 1. Decoupling of control rules from system resources leads to stability risks and resource waste: Existing solutions often employ static rules that trigger fixed actions based on fixed thresholds, implemented using hard-coded rules or simple rule engines. These rules have simplistic triggering logic and lack dynamic coupling with the underlying system resource status. For example, in high-concurrency scenarios such as large-scale promotions, when suspected fraudulent order behavior is detected, the system automatically triggers a resource-intensive operation (involving large-scale database queries, complex model calculations, and data synchronization) without any verification, performing a "deep order audit and full update of user profiles." This "one-size-fits-all" execution mode completely ignores the system's current CPU load, memory availability, and I / O bottlenecks, easily leading to systemic technical risks such as peak computing resource overload, sharp increases in service response latency, and even avalanche of core transaction links during high-traffic periods. This results in a severe imbalance between the execution cost (system resource overhead) of control operations and potential benefits.

[0005] 2. Lack of quantitative perception and closed-loop control technology for complex trading states: In the existing technical architecture, risk perception modules are mostly isolated detectors targeting single risk points (such as specific keyword matching or single behavioral thresholds). They lack a quantitative perception model capable of real-time collection, normalization, and fusion calculation of multi-dimensional and heterogeneous data such as "behavioral clustering," "abnormality of trading conditions," and "market liquidity imbalance." Consequently, the system cannot build a panoramic, digital diagnostic capability for the "sub-healthy" or "potentially failing" state of the trading environment. This results in a lack of precise data input for control decisions, relying solely on a retrospective, binary "intercept / non-intercept" logic, missing the optimal opportunity for closed-loop control through lightweight technical intervention in the early stages of risk.

[0006] 3. The control strategy engine lacks scenario-adaptive parameterized configuration technology: Within the same platform, there are often diverse business scenarios ranging from high-frequency, low-amount transactions to low-frequency, high-amount transactions, each with significantly different resource consumption patterns and risk characteristics. Existing technical solutions typically use a set of static rule parameters (such as fixed risk thresholds), failing to dynamically adjust the strength of control strategies and resource allocation weights based on the value attributes of the transaction object, historical risk data, or real-time system load through configurable parameter interfaces or adaptive algorithms. This rigid technical implementation prevents the system from implementing more prudent protection in high-value transaction scenarios and from downgrading control of low-risk transactions under high load, resulting in inappropriate allocation of technical resources.

[0007] Furthermore, existing risk control schemes based on fixed rules, known as "IF-THEN" schemes, typically simplify their decision-making logic to a step function based on a single threshold, and risk perception is mostly a binary "yes / no" judgment. Such schemes have inherent limitations in terms of the flexibility of their rule logic and their perception of system resource status. This invention constructs a systematic enhanced technical architecture by introducing a multi-dimensional risk quantification model, a resource constraint-aware scheduler, and a decision mapping model that can dynamically optimize parameters based on reinforcement learning. This aims to address the technical problems of decreased system robustness and low resource utilization efficiency caused by these limitations.

[0008] While existing technical literature includes independent "multi-dimensional risk quantification models" (e.g., for credit assessment) and "distributed system resource schedulers" (e.g., for microservice clusters), no technical solution has ever combined these two with the specific technical field of "real-time dynamic management and control of online trading platforms," ​​nor has it proposed a technical framework that "uses the quantified risk value of the trading environment as input and the system resource consumption level as a constraint to jointly determine the final selection and execution sequence of management and control actions." Simple combinations of existing technologies cannot solve the core technical optimization problem of "maximizing management and control effectiveness under limited and fluctuating system resources" in trading platform management. This invention, through a technical closed loop of 'perception-mapping-scheduling-reinforcement learning optimization,' achieves for the first time the unification and collaborative optimization of the three flows: environmental state data flow, management and control decision command flow, and system resource monitoring flow, possessing outstanding substantive technical characteristics.

[0009] Furthermore, the underlying computing infrastructure supporting the operation of these online trading platforms (such as cloud service platforms) also faces challenges in detecting internal workload anomalies, quantifying resource contention risks, and achieving efficient and stable scheduling under overall resource constraints. These two types of problems are essentially interconnected in the technical aspect of 'quantifying dynamic risks and thereby imposing resource-constrained controls'. Summary of the Invention

[0010] This invention aims to solve three core technical problems caused by the rigid and singular management mechanisms and the lack of resource coordination awareness in existing online trading platforms: 1. In risk control systems based on fixed rules, the triggering of control actions is decoupled from the real-time resource load status of the system. This can lead to systemic stability risks such as a surge in service response latency and database connection exhaustion when high-concurrency scenarios are triggered in a concentrated manner.

[0011] 2. The risk perception module of the existing system relies on a single or limited threshold judgment, which cannot perform integrated calculation and quantitative assessment of the complex trading environment state characterized by multiple dimensions such as behavioral clustering, condition anomaly degree, and market liquidity. As a result, the system cannot trigger early warning or light intervention at the risk initiation stage.

[0012] 3. The parameters of the control strategy engine (such as risk threshold and action intensity) are set globally and statically, and cannot be adaptively adjusted according to the value attributes of the transaction scenario, historical risk data flow or real-time system load, resulting in rigid technical resource allocation: insufficient protection in high-value transaction scenarios, while unnecessary excessive consumption of system resources in low-value high-concurrency scenarios.

[0013] To address the aforementioned issues, this invention proposes a dynamic intelligent management and control method for network transactions oriented towards resource constraints. Its core idea is to construct an intelligent closed-loop management and control system encompassing "perception-quantification-decision-scheduling-execution-evolution," treating the transaction environment as a complex dynamic system and implementing continuous monitoring, precise diagnosis, and adaptive intervention through technological means.

[0014] According to a first aspect of the present invention, a method for dynamic intelligent management and control of network transactions oriented towards resource constraints is provided, characterized in that the following steps are performed by a data processing system: Status awareness steps: Acquire monitoring data reflecting the status of the trading environment; Risk quantification step: Based on the monitoring data, generate a quantified risk assessment value; Intelligent decision-making steps: Input the risk assessment value into a pre-configured decision mapping model, and determine the target control action from a preset set of control actions based on mapping logic that satisfies the mapping monotonicity constraint; wherein, - Each control action in the preset control action set is associated with a preset expected system resource consumption level; - The mapping monotonicity constraint means that the decision mapping model is configured such that, for any two risk assessment values ​​R1 and R2, if R1>R2, then the expected system resource consumption level C1 of its output target control action is ≥C2. Resource scheduling steps: Based on the expected resource consumption level associated with the target control actions and the real-time resource status of the system, the execution timing, concurrency, or intensity of multiple target control actions are adaptively adjusted. This adjustment aims to ensure system stability and, while meeting real-time resource constraints, maintain as much as possible the matching relationship between risk and resource response levels established by the intelligent decision-making steps.

[0015] Action execution steps: Execute the controlled target actions after adjustment.

[0016] Furthermore, the risk quantification step includes: selecting monitoring data from at least two dimensions from the dimensions of abnormal behavior of trading entities, abnormal trading conditions, and abnormal market liquidity for fusion calculation.

[0017] Furthermore, the decision mapping model can be implemented through various technical forms, including but not limited to: structured lookup tables, encapsulated function libraries, rule scripts, and machine learning models. It is necessary to ensure that the mapping monotonicity constraints are met by using monotonic functions, adding monotonicity constraints during training, or configuring non-decreasing correspondences.

[0018] Furthermore, the resource scheduling steps include: real-time monitoring of system resource indicators (such as CPU, memory, network I / O); when the resource indicators exceed a safety threshold, taking at least one of the following control measures for target control actions with high expected resource consumption levels: delay, rate limiting, or degradation; and when the resource indicators recover to below the safety threshold, prioritizing the resumption of execution of the actions with high expected resource consumption levels.

[0019] Furthermore, the method also includes a dynamic optimization step: recording historical control event data and dynamically optimizing the parameter configuration of the decision mapping model based on a reinforcement learning algorithm to minimize long-term comprehensive costs (resource consumption costs and risks inadequate handling costs).

[0020] Furthermore, the method can be applied to specific scenarios such as online auction trading. In this scenario, risk quantification can be based on bid coordination monitoring data and transaction party correlation monitoring data; the decision mapping model is configured to output a control action (i.e., a high resource consumption control action) when the risk assessment value indicates the existence of coordinated bidding risk, with the expected resource consumption level not lower than a first threshold. The first threshold is a configurable system parameter used to define the minimum resource guarantee level required to cope with high-risk coordinated bidding behavior.

[0021] According to a second aspect of the present invention, a network transaction dynamic intelligent management and control system for implementing the above-described method is provided, characterized in that it adopts a modular microservice architecture, including a state perception module, a risk quantification module, an intelligent decision-making module, a resource monitoring module, a resource scheduling module, and an action execution module.

[0022] According to a third aspect of the present invention, a medium storing a computer program and an electronic device including a processor are provided for implementing the above-described method.

[0023] The beneficial effects of this invention are as follows: 1. Introduce "mapping monotonicity constraint": ensure that the resource consumption level of control actions matches the severity of environmental risks from the source of algorithm logic, and avoid system stability risks caused by resource mismatch.

[0024] 2. Achieve dynamic resource scheduling: A dual feedback mechanism of "environmental risk - system resources" has been established, which significantly improves the robustness of the system and the efficiency of resource utilization. Actual tests show that it can reduce system overload events caused by centralized control actions by more than 90%.

[0025] 3. Constructing multidimensional risk quantification and adaptive decision-making: It achieves accurate perception and gradient-based adaptive control of complex transaction states, which can reduce the occurrence rate of risk events by about 60%.

[0026] 4. Support for continuous evolution: Through reinforcement learning optimization based on historical data, the management strategy has moved from static configuration to data-driven, model-guided dynamic adaptation, achieving a technological leap.

[0027] 5. Platform-level versatility: It adopts a modular and plug-in microservice architecture, which can quickly adapt to various online transaction scenarios (such as e-commerce, bidding, service transactions, etc.) and new risk models. As the top-level control center, it works in conjunction with the lower-level functional modules to generate significant synergistic benefits.

[0028] Therefore, this invention not only provides a specific solution for governing online trading markets, but more importantly, it reveals and verifies a general intelligent management paradigm that couples quantitative risk state perception with resource-constrained response decisions. This provides an innovative technical approach for building large-scale software systems with adaptive capabilities (especially platform-level systems that need to coordinate multiple tasks and resource types). Attached Figure Description

[0029] Figure 1 is an overall flowchart of the dynamic intelligent management and control method for online transactions, showing the core process of "state perception - risk quantification - intelligent decision-making - resource scheduling - action execution". Among them, the intelligent decision-making step is marked with "mapping monotonicity constraint", the resource scheduling step is associated with the "system real-time resource status monitoring" module, and the arrow at the bottom indicates that the execution data is used for subsequent dynamic optimization of the model.

[0030] Figure 2 This is a schematic diagram of the risk quantification model, which shows the process of integrating and calculating multi-dimensional risk values. Sub-indicators are collected from three dimensions: abnormal behavior of trading entities, abnormal trading conditions, and abnormal market liquidity. After sub-indicator calculation, dynamic weight adjustment, and integration calculation of risk assessment values, a quantitative risk value is output.

[0031] Figure 3 It is a schematic diagram of decision-making logic under discrete mapping, which realizes discrete mapping decision-making based on risk level interval division; the "mapping monotonicity constraint" is marked to ensure that the expected resource consumption of control actions increases with the risk level.

[0032] Figure 4 It is a schematic diagram of decision-making logic under continuous mapping, which realizes dynamic adjustment of action intensity based on continuous function mapping; the "mapping monotonicity constraint" is marked to ensure that the action intensity and expected resource consumption increase continuously with the risk value.

[0033] Figure 5 This is a schematic diagram of the microservice architecture of the dynamic management and control system, showing the collaborative relationships between modules at each level and the continuous iterative optimization loop; it includes, in order, the perception layer, risk quantification layer, intelligent decision-making layer, resource scheduling layer, action execution layer, and data feedback module, with arrows at the bottom indicating the data feedback-driven continuous iterative optimization logic.

[0034] Figure 6 This is a timing diagram of dynamic resource scheduling, illustrating the dynamic scheduling logic of control actions under high load scenarios. By inputting the target control action and expected resource consumption, the system evaluates resources and dynamically allocates resources before outputting scheduling instructions.

[0035] Figure 7 It is a timeline diagram of the entire process of the dynamic control system, showing the complete closed loop from the triggering of transaction events to data feedback optimization; connecting all aspects of status perception, risk quantification, intelligent decision-making, resource scheduling, action execution and data feedback.

[0036] Figure 8 This is a schematic diagram of the risk level and control action mapping matrix, showing the mapping relationship between low, medium and high risks and corresponding control actions; the "Dynamic Adjustment" section at the bottom echoes the dynamic optimization steps of this invention.

[0037] Figure 9 This is a schematic diagram of the dynamic management and control system deployment architecture, showing the edge-cloud collaborative deployment of the system; it includes, in sequence, the perception node layer, the edge computing layer, the cloud service layer, the management console, the data storage layer, and the security protection layer, with each layer associated with a corresponding technical module and diagram. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of protection of the invention.

[0039] Key terms and principles 1. Risk Assessment Value: A comprehensive indicator generated based on monitoring data to quantify the probability and severity of adverse consequences such as fraud, abuse, and market imbalances in the trading environment. It is obtained through the fusion calculation of multi-dimensional sub-indicators or inference through machine learning models.

[0040] 2. Expected System Resource Consumption Level: This refers to a quantitative measure of the data processing system computing resources expected to be occupied or consumed when executing a control action. It can be calculated by combining the incremental CPU utilization (ΔCPU), incremental memory usage (ΔMem), incremental network I / O (ΔNet), and incremental disk I / O (ΔDisk), according to pre-configured weights. An example is as follows: Expected Resource Consumption Level = w1 × ΔCPU + w2 × ΔMem + w3 × ΔNet + w4 × ΔDisk, where the weight coefficients w1, w2, w3, and w4 satisfy w1 + w2 + w3 + w4 = 1. Initial configuration can be performed based on system architecture characteristics through benchmark testing or expert experience. The calculation results can be used for classification; for example, a score < 30 indicates a low resource consumption action, 30 ≤ score < 60 indicates a medium resource consumption action, and a score ≥ 60 indicates a high resource consumption action. The system maintains a "Control Action Resource Consumption Archive" to provide a data foundation for the estimation.

[0041] 3. Implementation method for pre-setting the association between the control action set and the expected consumption level The preset set of control actions and their associated expected system resource consumption levels are maintained and mapped within the system through a configurable control action-resource archive. This archive can be a structured database table, a configuration file (such as JSON or YAML), or an in-memory hash mapping table. Each control action entry includes at least a unique action identifier, an action description, an expected system resource consumption level value, and optional resource consumption calculation parameters. In the intelligent decision-making step, the system queries this archive in real time based on the target control action identifier output by the decision mapping model to obtain the corresponding expected resource consumption level, which is then used by the resource scheduling step. A simplified JSON format example is as follows: json { "action_id":"ACT_AUDIT_DEEP", "action_name":"Deep Order Audit", "expected_resource_cost":65, "cost_calc_params":{ "base_cpu":30, "base_memory":20, "per_order_cpu":0.5 } } The initial data for the archive can be pre-configured through system benchmark testing, historical performance data analysis, or expert experience, and can be dynamically calibrated and updated based on actual consumption data during system operation.

[0042] 4. Decision Mapping Model: This is a configurable logical carrier that can be loaded, interpreted, and executed by the system. Its form can be a structured lookup table, rule script file, machine learning model file, or an executable library encapsulating mathematical functions. Its function is to receive risk assessment values ​​as input and output control actions (or action intensity) instructions.

[0043] 5. Core Technical Constraint (Mapping Monotonicity Constraint): It should be particularly noted that the mapping monotonicity constraint is the core technical condition of this invention and must be met. Regardless of the technical form in which the decision mapping model is implemented, it must be configured, designed, or trained to ensure that its internal logic conforms to this constraint.

[0044] Specifically, this constraint means that when configuring or training the decision mapping model, it must be ensured that its internal mapping logic satisfies the following: when the risk assessment value of the input increases, the expected system resource consumption level of the control action output by the model remains non-decreasing, while other input conditions remain unchanged.

[0045] This constraint ensures that the technical cost of implementing control measures is commensurate with the severity of the environmental risks they address. This is a key technical feature that distinguishes this invention from existing random or empirical mapping methods, enabling optimized matching of resources and risks.

[0046] Example 1: Implementation of the Intelligent Decision Mapping Method This embodiment illustrates the core design constraint: the decision mapping must satisfy the "mapping monotonicity".

[0047] 1.1 Implementation of Discrete Mapping Form Scenario: An e-commerce platform monitors store transaction behavior. Based on the monitoring data, the system calculates the current risk assessment value R=70 in real time.

[0048] Decision mapping process: The system loads a structured lookup table, and the configuration strictly follows the mapping monotonicity constraint. -R∈(0,50]: Output "Log only", resource consumption score 25 points.

[0049] -R∈(50,80]: Output "Restrict concurrent access to some high-risk functions and send an operation prompt", resource consumption score 55 points.

[0050] -R∈(80,100]: Output "Initiate deep audit and temporarily freeze account", resource consumption score 90 points.

[0051] Monotonicity verification: As the risk value range increases, the action resource consumption score of the mapped output increases strictly monotonically from 25 to 55 to 90.

[0052] 1.2 Implementation of Continuous Mapping Pattern Scenario: A content platform identifies abnormal propagation. The system calculates a risk assessment value of R=85.

[0053] Decision mapping process: The system loads an executable library that encapsulates continuous mathematical functions. The functions explicitly guarantee that the output parameters are monotonically non-decreasing with respect to the input risk value, for example: - Content weighting factor = min(1.0, 0.015 × R) / / The slope is positive and monotonically increasing. - Review queue priority adjustment = max(0, 10×R-500) / / The slope is positive and monotonically increasing. - Threshold for triggering manual review = (R>75)?1:0 / / Step function, overall non-decreasing Monotonicity analysis: The above functions / rules all satisfy the following: if R1>R2, then the calculated intensity of each action f(R1)≥f(R2).

[0054] Example 1.3: Implementation of a Decision Mapping Model Based on Constraint Training This embodiment provides a specific training method to ensure that a machine learning model satisfies the mapping monotonicity constraint. It takes using a neural network as the decision mapping model as an example.

[0055] Model and Data Preparation: Construct a neural network model whose input is the risk assessment value R, and whose output is the probability distribution of the selection of each action in the control action set or the direct output of the action intensity parameters. The training data is a historical control event sample set, and each sample contains: historical risk assessment value R_hist, the identifier of the final executed control action A_hist, and the actual resource consumption of the action C_hist.

[0056] Loss function design: The total loss function L_total consists of two parts: The main task loss L_task, such as cross-entropy loss (used for action classification) or mean squared error loss (used for intensity regression), is used to encourage the model to accurately fit historical decisions.

[0057] L_monotonic, a monotonicity constraint loss, is used to force the model to satisfy the mapping monotonicity constraint. Its design is as follows: Two pairs of samples (R_i, A_i) and (R_j, A_j) are randomly sampled from the training batch, where R_i > R_j.

[0058] Obtain the corresponding expected resource consumption levels C_i and C_j (from the control action - resource archive).

[0059] Calculate the constraint violation: violation = max(0, C_j - C_i). If C_j > C_i, then the monotonicity is violated (higher risk corresponds to lower resource consumption).

[0060] The batch monotonicity loss is the mean of the violations of all sample pairs within the batch: L_monotonic = mean(violation).

[0061] Training process: The total loss is L_total = L_task + λ * L_monotonic, where λ is the constraint strength hyperparameter (e.g., set to 1.0). Model parameters are optimized through backpropagation and gradient descent. The larger the value of λ, the more strictly the model adheres to the monotonicity constraint.

[0062] The neural network decision mapping model trained by this method can learn decision-making patterns from data. At the same time, it is forced to have the monotonic characteristic that the higher the input risk value, the higher the expected resource consumption of the output action, thus reliably meeting the core technical constraints of this invention.

[0063] Example 2: Implementation of Dynamic Resource Scheduling Method like Figure 7 As shown, this demonstrates how the resource scheduling module dynamically coordinates multiple control actions when the system is under high load.

[0064] Scenario: During a major sales event, the system concurrently processes multiple transaction events that require management.

[0065] Scheduler input: -Task A (High-Value Merchant Audit): High resource consumption, predicted consumption score 85 points.

[0066] -Task B (Batch Order Rule Detection): Medium resource consumption, predicted consumption score 50 points.

[0067] -Task C (Send Warning Message): Low resource consumption, predicted consumption score 20 points.

[0068] -System resource status: CPU utilization 92%, database connection pool utilization 95%.

[0069] Scheduling decision and execution: 1. Pattern recognition: The scheduler determines that it is currently in "overload protection mode".

[0070] 2. Dynamic adjustment: For task A, temporarily store its "deep audit" sub-action and only execute the core detection logic; immediately execute task C; place the adjusted task A and task B into the delay queue and set task B to have a higher priority.

[0071] 3. Resilient Recovery: After 5 minutes, the CPU load drops to 75%. The scheduler retrieves and executes tasks B and A from the delay queue according to priority.

[0072] Example 3: Implementation of Risk Quantification Method This example demonstrates how to generate a comprehensive risk assessment value from multi-dimensional monitoring data.

[0073] 3.1 Sub-indicator Collection and Calculation The system acquires monitoring data through event-driven, timed polling, or streaming methods, and calculates sub-indicators for the following dimensions (baseline data is dynamically maintained by the system based on historical data): - Behavioral Anomaly Sub-Indicator: Reflects the clustering and regularity anomalies in the behavior of transaction entities. For example, it calculates the deviation of the frequency of transactions initiated by the same device or IP within a short period of time from the global benchmark. A specific calculation formula is as follows: The clustering score is calculated as: min(100, (observation frequency / historical benchmark frequency) × 100 × penalty factor), where the penalty factor is greater than 1 when the IPs are the same and the device fingerprints are similar.

[0074] -Conditional Anomaly Sub-Indicator: Reflects the dispersion or deviation of conditions such as transaction price and quantity. For example, calculating the coefficient of variation of a commodity's transaction price. Price Anomaly Score = (Price Coefficient of Variation / Category Benchmark Coefficient of Variation) × 100.

[0075] - Liquidity Anomaly Sub-Indicator: Reflects anomalies in transaction frequency and transfer rate. For example, calculating the sudden increase in transaction cancellation or refund requests within a specific period. Liquidity Score = min(100, (Sudden Increase Rate / Steady Period Baseline Increase Rate) × 100).

[0076] 3.2 Weight Determination and Dynamic Adjustment Initial weight configuration (example): Behavioral anomaly score 0.4, Conditional anomaly score 0.4, Liquidity anomaly score 0.2.

[0077] Dynamic adjustment rules (example): - When the target is a high-value commodity: the conditional anomaly weight is increased to 0.5.

[0078] -When the historical dispute rate is high: the weight of behavioral abnormality is increased to 0.5.

[0079] 3.3 Calculation of Comprehensive Risk Assessment Value Risk assessment value = Σ(sub-indicator score_i × weight_i).

[0080] Example 4: Dynamic Management of Anti-Collaborative Bidding Scenarios in Online Auction Transactions This embodiment demonstrates the application of the method of the present invention in the specific scenario of online competitive bidding, for dynamic management of collaborative bidding risks, so as to directly support claim 7.

[0081] Scenario and Problem Background: In an online engineering bidding or asset auction platform, multiple bidders may manipulate the bidding results through collusion (bid rigging), harming the interests of the bidding party. Identifying collusive bidding requires complex graph association analysis and time series pattern analysis, which is computationally intensive. Precise targeting is necessary under resource constraints.

[0082] System execution process Step 1: State Awareness and Risk Quantification Monitoring data collection: Price coordination monitoring data: Collect price sequences, price timestamps, and price correction behaviors of all bidders under the same bidding section.

[0083] Transaction party correlation monitoring data: Collects the registration information of bidders (such as contact information, IP address, device fingerprint), historical bidding symbiotic relationships, and equity correlation information.

[0084] Risk Quantification Calculation: Calculate two core sub-indicators: Quotation behavior coordination score: Based on the statistical distribution of quotations (such as abnormally low coefficient of variation) and the synchronicity of quotation corrections, the anomaly degree is calculated by clustering algorithm and normalized to 0-100 points.

[0085] Bidder Relationship Depth Score: Based on correlation monitoring data, a bidder relationship diagram is constructed, the degree of closeness of the largest group or community is calculated, and converted into a risk score.

[0086] By assigning dynamic weights (such as focusing more on the degree of correlation), the integrated calculation yields a comprehensive risk assessment value of R=88, which clearly indicates a high degree of suspicion of coordinated pricing.

[0087] Step 2: Intelligent Decision Making The decision mapping model (a rule engine that satisfies monotonicity constraints) receives R=88.

[0088] According to the configuration, when the risk value is in the range of [80, 100], the target control action is output: "Initiate full historical transaction audit and real-time communication monitoring of the associated bidder group". The expected resource consumption level associated with this action is "high" (score 85 points), which meets the configuration requirement of "not lower than the first threshold" in claim 7.

[0089] Step 3: Resource Scheduling and Execution Resource monitoring shows that the current data analysis service cluster is under high load.

[0090] The resource scheduling module allocates a dedicated computing resource pool for this high-consumption action and limits its maximum number of concurrent threads to ensure that its execution does not affect the stability of the core bidding process.

[0091] The action execution module initiated an audit and successfully identified a bid-rigging group involving three bidders.

[0092] Technical effect This embodiment confirms the technical solution of claim 7: in the online bidding transaction scenario, risk is quantified based on the monitoring data of bid coordination and transaction party correlation, and high resource consumption control actions are output when the risk is high through the decision mapping model, and the whole process follows the dynamic resource scheduling constraints of the present invention.

[0093] Example 5: Dynamic Management of Anti-Fraud Scenarios on E-commerce Platforms This embodiment demonstrates the application of the method of the present invention in a typical non-competitive transaction scenario of anti-fraud on a traditional e-commerce platform. The flowchart is as follows: Figure 8 As shown.

[0094] 1. Scenario and Problem Background: During the "6.18" shopping festival, a major e-commerce platform needs to implement real-time controls to prevent fraudulent order practices targeting specific products (i.e., fake transactions to boost sales and positive reviews). Fraudulent order behaviors exhibit diverse characteristics, and the resource consumption of control measures (such as order interception, user profile updates, and model retraining) varies significantly. Traditional static rules are prone to causing system overload under high concurrency.

[0095] 2. System Execution Process Step 1: State Awareness and Risk Quantification - Monitoring Data Acquisition: The system collects multiple data streams in real time, including: - Order behavior sequence: Buyer's order time, frequency, device fingerprint, IP address, and similarity of delivery address.

[0096] - Transaction pattern data: the number of times the same buyer purchases the same product, the distribution of payment time, and the similarity of review content.

[0097] - Merchant-related data: historical complaint rate of merchants, validity of logistics tracking numbers.

[0098] -Risk quantification calculation: Based on the above data, calculate three sub-indicators: 1. Behavioral Clustering Score: Calculates outliers based on the degree of clustering of devices, IPs, and addresses.

[0099] 2. Trading Pattern Deviation Score: Compare the current trading time distribution and evaluation pattern with the normal benchmark.

[0100] 3. Merchant Risk Correlation Score: Assess current transaction risk based on merchant historical data.

[0101] - Based on the strategy during the promotional period (with a greater focus on behavioral clustering), dynamic weights are assigned: behavioral clustering 0.5, transaction pattern deviation 0.3, and merchant risk association 0.2.

[0102] - The combined calculation yields a comprehensive risk assessment value of R=78 for the current transaction flow.

[0103] Step 2: Intelligent Decision Making The decision mapping model receives R=78. This model is a pre-trained gradient boosting decision tree with monotonicity constraints added during training.

[0104] - The model outputs the target control action based on the mapping logic (which already satisfies monotonicity): "Initiate deep order auditing, mark relevant user profiles as high-risk, and temporarily refrain from performing full-model real-time retraining, which consumes extremely high resources." The expected resource consumption level associated with this action is "medium-high" (score 65).

[0105] Step 3: Resource Scheduling The resource monitoring module shows that the current system CPU usage is 88% and memory usage is 85%.

[0106] - The resource scheduling module determines that the current load is high. Based on the strategy, "medium-high" consumption actions are placed in a dynamic priority queue, and the number of concurrent threads for "deep order auditing" is limited to 50% of the normal value. At the same time, non-core parts of "user profile update" are temporarily suspended.

[0107] Step 4: Action Execution and Feedback - The action execution module executes control actions according to the scheduled parameters.

[0108] - After execution, the probability of the transaction being identified as fraudulent was 92%, and it was successfully intercepted.

[0109] - The data from this control event (monitoring data, actions, actual resource consumption, and interception results) was recorded and used for subsequent reinforcement learning optimization.

[0110] 3. Technical Effects - Resource efficiency: Through dynamic scheduling, effective management can still be performed during periods of high system load (88% CPU), avoiding service avalanche that may be caused by directly starting "full model retraining" (prediction consumes 90+ points).

[0111] -Precise control: Decisions are made based on multi-dimensional quantitative risk values, which is more accurate than a single threshold (such as "blocking more than 5 orders within 1 minute") and reduces the false positive rate.

[0112] - Adaptability: The recorded data can be used to optimize the decision mapping model, enabling it to make decisions faster or choose better actions in similar scenarios in the future.

[0113] It is understood that the above embodiments (especially embodiments 4, 5, and the subsequent embodiment 6) elaborate on the application of the present invention in business systems with clear business logic (such as transaction and content platforms). After understanding the core ideas of the present invention, those skilled in the art can apply it to another type of key software system—namely, the basic platform system responsible for resource supply and scheduling. In this type of system, 'monitoring data' can be converted into performance indicators and health status of the platform's internal workload, 'risk' can be understood as 'the risk of service quality decline or resource conflict', and 'control actions' correspond to dynamic scheduling strategies for basic resources such as computing, storage, and networks. Although the data objects and specific actions in the application scenarios are different, the core technical path of 'state perception → risk quantification → (under mapping constraints) decision-making → scheduling under resource constraints' remains consistent.

[0114] Example 6: Dynamic Management of Infringing Content on Digital Content Platforms This embodiment demonstrates the application of the method of the present invention in a non-transactional scenario but closely related to online content publishing, and its flowchart is as follows. Figure 9 As shown, this demonstrates the universality of its technical solution.

[0115] 1. Scenario and Problem Background: A video sharing platform needs to screen user-uploaded content for infringing content (such as pirated film clips) or other violations in real time. Traditional solutions involve either sampling a fixed percentage or reviewing all popular content. The former has a high miss rate, while the latter suffers from severe shortages of computing resources (especially GPU resources) during peak traffic periods.

[0116] 2. System Execution Process Step 1: State Awareness and Risk Quantification - Monitoring data collection: - Content Features: Extract keyframe features and audio fingerprints from the video using a pre-trained model, compare them with the copyright database, and generate an initial infringement confidence score.

[0117] -Uploader characteristics: uploader's historical violation records, account registration time, and upload frequency.

[0118] - Contextual characteristics: upload time (e.g., inactive periods), sensitivity of title / description text, and the overall review queue load of the current platform.

[0119] -Risk quantification calculation: - Calculate three sub-indicators: 1. Content infringement risk score: a normalized score calculated based on audio and video similarity.

[0120] 2. Uploader's historical risk score: A score based on the number and severity of historical violations.

[0121] 3. Context Anomaly Score: A comprehensive score combining abnormal upload time and text risk.

[0122] - Dynamic weighting: During peak traffic hours in the evening (when review resources are scarce), increase the weight of "uploader's historical risk score" (e.g., from 0.3 to 0.5) to make the system more inclined to process content from users with known risks.

[0123] - The comprehensive risk assessment value of the currently uploaded content is R=82 obtained by fusion calculation.

[0124] Step 2: Intelligent Decision Making - The decision mapping model (a neural network with monotonicity constraints) receives R=82.

[0125] - The model outputs the following control action: "Place this content in a high-priority review queue, review it using a high-precision but resource-intensive 'multimodal deep infringement detection model,' and delay its distribution." The expected resource consumption level for this action is "high" (score 80).

[0126] Step 3: Resource Scheduling - The resource monitoring module shows that the current GPU memory usage is 90%, and the audit service response latency has increased.

[0127] - The resource scheduling module initiates dynamic control: the number of concurrent instances of the "high-precision model" is limited to 2 (normally 5), and a high queue priority is set for this task to ensure that it can be executed as soon as possible, while not affecting the lightweight review pipeline of other low-to-medium risk content.

[0128] Step 4: Action Execution and Feedback - The action execution module calls a restricted high-precision model for review, confirms that the probability of infringement is 95%, executes the removal of the content and notifies the uploader.

[0129] - The data from this control was recorded and used to optimize model parameters and scheduling strategies.

[0130] 3. Technical Effects - This demonstrates the cross-domain versatility of the method: This embodiment shows how the same "perception-quantification-decision-scheduling" framework can be seamlessly migrated from "e-commerce transaction risk control" to "content security review", and the core constraints (mapping monotonicity) and technical advantages (dynamic resource adaptation) still hold.

[0131] - This highlights the value of resource scheduling: When specific hardware resources (GPU) become a bottleneck, this method can ensure the completion of the most critical task (high-risk infringing content review) through dynamic rate limiting and priority scheduling, thereby achieving optimal overall performance under resource constraints.

[0132] Example 7: Dynamic Optimization Based on Reinforcement Learning This embodiment uses reinforcement learning as an example to demonstrate how to optimize the parameters of a decision mapping model through dynamic learning. The control strategy optimization problem is modeled as a Markov decision process (MDP): - State space S: includes risk assessment value R, trading volume, time characteristics, current CPU / GPU utilization, etc.

[0133] - Action Space A: Adjustment operations for decision model parameters, such as adjusting the risk threshold range and adjusting the weights of each sub-indicator (adjustment step size ±0.05).

[0134] - Reward function: Designed to comprehensively consider both control effectiveness and resource costs. A specific example is: Reward = 0.7 × (Risk handling success rate) - 0.3 × (Standardized resource consumption rate) - 0.1 × (False alarm rate). Here, the standardized resource consumption rate is the ratio of actual consumption to expected consumption.

[0135] Online learning was conducted using a policy gradient-based reinforcement learning algorithm. The system collected a batch of historical control event data every hour, calculated the average reward, and updated the model parameters. After a week of learning, the optimized strategy showed in A / B testing that, while maintaining the same risk interception rate, the average resource consumption per control event decreased by approximately 18%, and the false positive rate decreased by 5%.

[0136] The above optimization process is carried out on the premise of ensuring that the decision mapping model always satisfies the mapping monotonicity constraint. For example, it can be done by setting a strict constraint penalty term in the reward function or by using a model structure that satisfies the monotonicity constraint.

[0137] Example 8: System Architecture and Deployment The dynamic management and control system consists of the following independently deployed microservices: a state-aware service, a risk quantification service, an intelligent decision-making service, a resource scheduling service, an action execution service, and multiple fine-grained management and control action services (such as order service, user service, message service, and content moderation service). These services communicate asynchronously via RESTful APIs or message queues (such as Apache Kafka). The decision mapping model loaded by the intelligent decision-making service can be hot-updated through a configuration center. The resource monitoring service collects system metrics using tools such as Prometheus. This architecture supports high cohesion, low coupling, elastic scaling, and rapid iteration.

[0138] Example 9: Collaborative Work as the Top-Level Control Center This embodiment demonstrates the workflow of the system of the present invention, acting as a top-level control hub, collaborating with multiple lower-level professional functional modules (such as the order transaction engine, payment risk control module, content review module, and advertising anti-fraud module). All risk handling requests or self-inspection anomalies from lower-level modules are reported to the intelligent decision-making module of the present invention. The resource scheduling module has a global perspective, preventing various modules from simultaneously initiating high-consumption actions during peak traffic periods. Actual test data shows that this collaborative mechanism can reduce peak system resource consumption by 40%-70% under sudden risk events and reduce core business link response latency fluctuations by more than 50%, achieving a balance between global stability and local efficiency.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this invention, based on the technical solutions and inventive concepts of the present invention, should be covered within the protection scope of this invention.

Claims

1. A dynamic intelligent management and control method for network transactions oriented towards resource constraints, characterized in that, The data processing system performs the following steps: Status awareness steps: Acquire monitoring data reflecting the status of the trading environment; Risk quantification step: Based on the monitoring data, generate a quantified risk assessment value; Intelligent decision-making steps: Input the risk assessment value into a pre-configured decision mapping model, and determine the target control action from a preset set of control actions based on mapping logic that satisfies the mapping monotonicity constraint; wherein, - Each control action in the preset control action set is associated with a preset expected system resource consumption level; - The mapping monotonicity constraint means that for any two risk assessment values ​​R1 and R2, if R1>R2, then the expected system resource consumption level C1≥C2 of the target control action output by the model. Resource scheduling steps: Based on the expected resource consumption level of the target control actions and the real-time resource status of the system, the execution sequence, concurrency or intensity of multiple target control actions are adjusted. Action execution steps: Execute the controlled target actions after adjustment.

2. The method according to claim 1, characterized in that, The decision mapping model is implemented through at least one of the following technical forms, and ensures that the mapping monotonicity constraint is satisfied: (a) A structured lookup table storing the correspondence between risk level ranges and control actions and intensity parameters; (b) An executable function library that encapsulates monotonically increasing or non-decreasing mapping functions; (c) Rule scripts that are loaded and executed by the rule engine and whose rule configurations satisfy a non-decreasing correspondence; (d) Machine learning models with monotonicity regularization constraints added during training.

3. The method according to claim 1, characterized in that, The risk quantification steps include: The risk assessment value is calculated by collecting monitoring data from at least two of the following anomaly dimensions and then fusing the data: - Dimensions of abnormal behavior of trading entities: Abnormalities in the clustering, regularity, or coordination of the behavioral patterns of trading parties; - Dimensions of abnormal trading conditions: Abnormal dispersion or deviation based on trading price, quantity, or conditions; - Market liquidity anomaly dimension: anomalies based on transaction frequency, transfer rate, or participant activity.

4. The method according to claim 1, characterized in that, The resource scheduling steps include: Real-time monitoring of system resource metrics, including at least one of CPU utilization, memory usage, and network bandwidth usage; When resource indicators exceed preset safety thresholds, at least one of the following controls will be implemented for control actions where the expected resource consumption level is higher than the threshold: (a) Delay until resource indicators are restored; (b) Limit the number of concurrent executions; (c) Downgrade to an alternative action with lower resource consumption.

5. The method according to claim 1, characterized in that, It also includes dynamic optimization steps: Record historical control event data, including pre-control monitoring data, control actions performed, actual resource consumption, and post-control status changes; Based on the historical data, the parameter configuration of the decision mapping model is optimized through a reinforcement learning framework to minimize the long-term comprehensive cost, which includes resource consumption cost and risk management inadequacy cost.

6. The method according to claim 1, characterized in that, The data processing system is implemented using a microservice architecture. The state awareness step, risk quantification step, intelligent decision-making step, resource scheduling step, and action execution step are each executed by an independent, pluggable microservice module that communicates through a standardized interface.

7. The method according to any one of claims 1 to 6, characterized in that, The method is applied to online auction trading scenarios; The risk quantification step includes generating a risk assessment value based on price coordination monitoring data and transaction party correlation monitoring data; The decision mapping model is configured to output a control action whereby the expected resource consumption level is not lower than a first threshold when the risk assessment value indicates that there is a risk of collaborative bidding.

8. A dynamic intelligent control system for online transactions, used to implement the method as described in any one of claims 1 to 7, characterized in that, include: The status awareness module is used to acquire monitoring data; The risk quantification module, connected to the state perception module, is used to generate risk assessment values; The intelligent decision-making module, pre-configured with the decision mapping model, is connected to the risk quantification module and used to determine target control actions; The resource monitoring module is used to obtain the real-time resource status of the system; The resource scheduling module is connected to the intelligent decision-making module and the resource monitoring module respectively, and is used to perform adjustment actions based on expected resource consumption and real-time resource status. The action execution module is connected to the resource scheduling module and is used to perform control operations.

9. An electronic device, comprising a processor and a memory, characterized in that, When the processor executes a computer program stored in the memory, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images