Mobile phone leasing risk control optimization method and system based on multi-dimensional AI model analysis
By analyzing multi-dimensional AI models, collecting and aligning static and dynamic data, constructing a risk control feature combination matrix, and using reinforcement learning algorithms to match the risk control model and dynamically adjust weights, the system solves the problems of data timeliness and model adaptation in the mobile phone rental risk control system, thereby improving the accuracy of risk assessment and the stability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING XIAOMA KUAI REN TECHNOLOGY CO LTD
- Filing Date
- 2025-06-20
- Publication Date
- 2026-06-23
AI Technical Summary
Existing mobile phone rental risk control systems have shortcomings in data timeliness alignment, model adaptation, and weight allocation, resulting in high risk assessment error rates, low fraud detection rates, and rapid decay of model accuracy, causing losses and compliance risks for operators.
By analyzing multi-dimensional AI models, static risk control data and dynamic leasing behavior data are collected and aligned to construct a risk control feature combination matrix. Reinforcement learning algorithms are used to match the risk control model, dynamically adjust the weights, and optimize risk control decisions by combining data security processing and edge computing.
It has achieved timeliness and consistency in risk assessment, improved fraud detection rate and stability of model accuracy, reduced bad debt rate and false rejection rate, and improved the accuracy and compliance of risk control decisions.
Smart Images
Figure CN120975537B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial technology using artificial intelligence, specifically to a method and system for optimizing risk control in mobile phone leasing based on multi-dimensional AI model analysis. Background Technology
[0002] Artificial intelligence refers to the intelligence exhibited by computer systems or machines created by humans. It means that these systems can perform tasks that usually require human intelligence to complete. The mobile phone rental risk control optimization system is a financial technology solution based on artificial intelligence technology, which aims to improve the efficiency and accuracy of risk management in mobile phone equipment rental business. The core module of the system consists of multi-dimensional data collection, real-time model matching and dynamic decision engine. By collecting static credit information and dynamic behavior data of users, it builds an intelligent risk control system to ensure the controllability of rental business risks.
[0003] Currently, there are three major technical bottlenecks in mobile phone rental risk control practices: First, due to the difference in the update frequency of credit data and the collection frequency of rental behavior data, existing systems cannot achieve time-series alignment. If the timeliness deviation of the data exceeds the threshold, it will lead to the distortion of user profiles and an increase in the error rate of risk assessment. Second, traditional rule engines rely on preset model selection logic, such as a fixed credit limit corresponding to a credit scoring model. They cannot dynamically match the optimal analysis model based on scenario characteristics such as device value and rental period. When encountering new fraud methods, the model adaptation error rate is as high as 52%. Finally, in the risk scoring dimension, the current system adopts static weight ensemble, such as a fixed RF / GBDT ratio of 6:4. It cannot automatically optimize the weight allocation based on the dynamic data of user performance, resulting in the continuous decline of model accuracy as business expands, decreasing by an average of 1.2 percentage points per month.
[0004] These problems directly led to a decline in industry operational efficiency: operators' bad debt rates, caused by inaccurate risk assessments, remained consistently above the industry warning line; and rising false rejection rates for high-end equipment leasing resulted in annual revenue losses exceeding [a certain threshold]. The figure of 2.6 million units per 10,000 units has also raised compliance concerns at the regulatory level—compliance complaints due to the lack of traceability of decision-making logic accounted for 67% of regulatory inquiries.
[0005] Therefore, we propose a method and system for optimizing risk control in mobile phone rental based on multi-dimensional AI model analysis to solve the above problems. Summary of the Invention
[0006] (a) Technical problems to be solved
[0007] To address the shortcomings of existing technologies, this invention provides a method and system for optimizing risk control in mobile phone rental based on multi-dimensional AI model analysis, thus solving the problems mentioned in the background section.
[0008] (II) Technical Solution
[0009] To achieve the above objectives, the present invention provides the following technical solution: a method and system for optimizing risk control in mobile phone rental based on multi-dimensional AI model analysis, wherein the method includes the following steps:
[0010] S1. Collect static risk control data and dynamic leasing behavior data from users;
[0011] S2. Perform data timeliness alignment processing on the static risk control data and dynamic leasing behavior data to generate timeliness-aligned static risk control correction data and dynamic leasing behavior benchmark data;
[0012] S3. Perform multi-dimensional feature combination processing on the static risk control correction data and the dynamic leasing behavior benchmark data to construct a risk control feature combination matrix;
[0013] S4. Based on the risk control feature combination matrix and the standard risk control feature matrix under the preset risk scenario, the risk control model type is matched by reinforcement learning algorithm.
[0014] S5. Call the risk control analysis program corresponding to the target risk control model identifier data, perform risk scoring calibration processing on the dynamic leasing behavior data, and generate an optimized leasing risk assessment result;
[0015] S6. Dynamically adjust the user's leasing permissions and quota parameters based on the leasing risk assessment results.
[0016] Preferably, S1 includes:
[0017] S11. Collect static risk control data of users through mobile SDK, including identity authentication information, historical credit records, and device hardware fingerprints;
[0018] S12. Obtain dynamic rental behavior data in real time through the rental platform API, including order fulfillment records, payment delay data, and abnormal equipment usage logs.
[0019] Preferably, S2 includes:
[0020] S21. Use a timestamp parsing algorithm to extract the data update frequency of the static risk control data. Data collection frequency of dynamic leasing behavior data ;
[0021] S22, when and When inconsistencies occur, the Lagrange interpolation algorithm is used to reconcile the static risk control data. Resampling generates static risk control correction data;
[0022] S23, when and When they are consistent, the original static risk control data is directly output as valid input.
[0023] Preferably, S3 includes:
[0024] S31. Stack the static risk control correction data and dynamic leasing behavior benchmark data vertically using user ID as the index to construct a risk control feature combination matrix:
[0025]
[0026] in This is data for static risk control correction. This serves as benchmark data for dynamic leasing behavior.
[0027] Preferably, S4 includes:
[0028] S41, Pre-built risk control model feature library:
[0029] ,
[0030] in The standard feature matrix corresponds to the fraud detection model, credit scoring model, equipment loss prediction model, performance capability model, and blacklist matching model;
[0031] S42. Matching using the reinforcement learning Q-Learning algorithm. and The optimal model:
[0032] Initialize the Q-value table and state space:
[0033]
[0034] The system state is defined as the combination matrix K of risk control features and the feature library W of the pre-set model. A model Cartesian product;
[0035] Select action at using an ε-greedy strategy;
[0036] Calculate the reward function:
[0037]
[0038] in, The similarity reward coefficient, This is the delay penalty coefficient. For model inference latency, Let Frobenius be the matrix norm.
[0039] Preferably, the reward function satisfies:
[0040]
[0041] in, The similarity reward coefficient, This is the delay penalty coefficient. For model inference latency, Let Frobenius be the matrix norm.
[0042] Preferably, S5 includes:
[0043] S51, according to Call the corresponding risk control model and input the risk control feature combination matrix. ;
[0044] S52. Output risk probability through an ensemble learning framework:
[0045]
[0046] in , For dynamic weights, For random forest, It is a gradient boosting tree;
[0047] S53, when When the threshold is greater than θ, a reduction in the lease limit and a freeze operation will be triggered, where θ is the industry risk threshold.
[0048] Preferably, the dynamic weights in S52 satisfy:
[0049]
[0050] in, These are the weights for the random forest model. To boost the weights of the tree model using gradient gradation, For random forest recall, For GBDT accuracy;
[0051] The model is dynamically adjusted based on its recall and precision over the past 30 days.
[0052] Preferably, after executing S4, a step of dynamically adjusting system resources is also included:
[0053] S71. Monitor the concurrent request volume of the model matching engine in real time. When the concurrency exceeds the preset threshold Q, dynamically increase the number of edge computing nodes according to the formula N=ceil(Q / 50).
[0054] S72. Allocate 3 times the standard computing resources to requests involving high-end equipment leasing;
[0055] S73. When the single model matching delay exceeds 200ms, automatically switch to the preset default credit scoring model.
[0056] S74. Update the latency penalty coefficient τ in the Q-Learning reward function based on the resource scheduling log. The update formula is:
[0057]
[0058] Where Q is the concurrent request threshold, N is the number of edge nodes to be expanded, and the update formula is an adaptive penalty coefficient rule.
[0059] Preferably, a data security preprocessing step is included before S1:
[0060] S01. End-to-end encryption is implemented on the collected static risk control data and dynamic leasing behavior data using the national cryptographic SM4 algorithm;
[0061] S02. Establish a secure channel using the SSL / TLS 1.3 protocol during data transmission;
[0062] S03. Verify the authenticity of the data source through a zero-knowledge proof protocol;
[0063] S04. Implement homomorphic encryption at the data storage layer;
[0064] S05. Generate data security audit logs and synchronize them to the blockchain evidence storage module;
[0065] S06. When data tampering is detected, a three-level security alarm mechanism is automatically triggered:
[0066] Level 1 Alert: Suspend the data acquisition process;
[0067] Level 2 Alert: Isolate suspicious data nodes;
[0068] Level 3 Alert: Reset the encryption key and report to the regulatory authority.
[0069] (III) Beneficial Effects
[0070] Compared with existing technologies, this invention provides a method and system for optimizing risk control in mobile phone rental based on multi-dimensional AI model analysis, which has the following beneficial effects:
[0071] 1. In this invention, by constructing a data timeliness central mechanism, during the risk control assessment of mobile phone leasing, the time-frequency deviation between static credit data and dynamic behavioral data is analyzed based on the Lagrange interpolation algorithm. Millisecond-level resampling processing is implemented on multi-source heterogeneous data, which can eliminate the problem of user profile distortion caused by data update delay in real time, ensuring the timeliness and consistency of risk assessment. Practical testing has verified that this mechanism reduces the error rate of risk assessment, improves the accuracy of user credit profiles, and enhances the precision of risk control decisions.
[0072] 2. In this invention, by deploying a reinforcement learning model adaptation engine, the similarity between the risk control feature matrix and the pre-set model library is calculated in real time based on the Q-Learning algorithm during the risk analysis process. The optimal model type is dynamically selected through an ε-greedy strategy. When encountering new fraud methods, the system automatically switches to the fraud detection model, solving the problem of rigidity in traditional system models. Actual business data shows that this mechanism improves the fraud detection rate of high-end equipment leasing and reduces the average annual loss.
[0073] 3. In this invention, by establishing a dynamic weight feedback closed loop, during the risk scoring output stage, the model weight coefficients α and β are dynamically allocated based on the real-time performance indicators of random forest and gradient boosting tree. By updating the model performance parameters daily and limiting the weight range, the concept drift phenomenon is continuously combated. This technology enables the system to maintain a lower model accuracy decay rate than the traditional system accuracy decay rate during business expansion, keep the bad debt rate stable within the industry benchmark value, and reduce the false rejection rate by a percentage point. Attached Figure Description
[0074] Figure 1 This is a schematic diagram of the method steps of the present invention;
[0075] Figure 2 This is a schematic diagram of the overall system architecture of the present invention. Detailed Implementation
[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0077] Please see Figure 1-2 The method and system for optimizing risk control in mobile phone rental based on multi-dimensional AI model analysis includes the following steps:
[0078] S1. Collect static risk control data and dynamic leasing behavior data from users;
[0079] S2. Perform timeliness alignment processing on static risk control data and dynamic leasing behavior data to generate timeliness-aligned static risk control correction data and dynamic leasing behavior benchmark data;
[0080] S3. Combine static risk control correction data with dynamic leasing behavior benchmark data through multi-dimensional feature processing to construct a risk control feature combination matrix.
[0081] S4. Based on the risk control feature combination matrix and the standard risk control feature matrix under the preset risk scenario, the risk control model type is matched by reinforcement learning algorithm.
[0082] S5. Call the risk control analysis program corresponding to the target risk control model identifier data, perform risk scoring calibration on the dynamic leasing behavior data, and generate optimized leasing risk assessment results.
[0083] S6. Dynamically adjust user leasing permissions and quota parameters based on leasing risk assessment results.
[0084] S11. Collect static risk control data of users through mobile SDK, including identity authentication information, historical credit records, and device hardware fingerprints;
[0085] S12. Obtain dynamic rental behavior data in real time through the rental platform API, including order fulfillment records, payment delay data, and abnormal equipment usage logs.
[0086] S21. Use a timestamp parsing algorithm to extract the data update frequency of static risk control data. Data collection frequency of dynamic leasing behavior data ;
[0087] S22, when and When inconsistencies occur, the Lagrange interpolation algorithm is used to reconcile the static risk control data. Resampling generates static risk control correction data;
[0088] S23, when and When they are consistent, the original static risk control data is directly output as valid input.
[0089] S31. Stack the static risk control correction data and the dynamic leasing behavior benchmark data vertically using user ID as the index to construct a risk control feature combination matrix:
[0090]
[0091] in This is data for static risk control correction. This serves as a benchmark for dynamic leasing behavior;
[0092] S41, Pre-built risk control model feature library:
[0093] ,
[0094] in The standard feature matrix corresponds to the fraud detection model, credit scoring model, equipment loss prediction model, performance capability model, and blacklist matching model;
[0095] S42. Matching using the reinforcement learning Q-Learning algorithm. and The optimal model:
[0096] Initialize the Q-value table and state space:
[0097]
[0098] The system state is defined as the combination matrix K of risk control features and the feature library W of the pre-set model. A model Cartesian product;
[0099] Select action at using an ε-greedy strategy;
[0100] Calculate the reward function:
[0101]
[0102] in, The similarity reward coefficient, This is the delay penalty coefficient. For model inference latency, Let Frobenius be the matrix norm.
[0103] The reward function satisfies:
[0104]
[0105] in, The similarity reward coefficient, This is the delay penalty coefficient. For model inference latency, The Frobenius norm of the matrix;
[0106] S51, according to Call the corresponding risk control model and input the risk control feature combination matrix. ;
[0107] S52. Output risk probability through an ensemble learning framework:
[0108]
[0109] in , For dynamic weights, For random forest, It is a gradient boosting tree;
[0110] S53, when When the threshold is >θ, a reduction in the lease limit and a freeze operation will be triggered, where θ is the industry risk threshold.
[0111] The dynamic weights in S52 satisfy the following:
[0112]
[0113] in, These are the weights for the random forest model. To boost the weights of the tree model using gradient gradation, For random forest recall, For GBDT accuracy;
[0114] The model is dynamically adjusted based on its recall and precision over the past 30 days.
[0115] The process of executing S4 also includes a step of dynamically adjusting system resources:
[0116] S71. Monitor the concurrent request volume of the model matching engine in real time. When the concurrency exceeds the preset threshold Q, dynamically increase the number of edge computing nodes according to the formula N=ceil(Q / 50).
[0117] S72. Allocate 3 times the standard computing resources to requests involving high-end equipment leasing;
[0118] S73. When the single model matching delay exceeds 200ms, automatically switch to the preset default credit scoring model.
[0119] S74. Update the latency penalty coefficient τ in the Q-Learning reward function based on the resource scheduling log. The update formula is:
[0120]
[0121] Where Q is the threshold for concurrent requests, N is the number of edge nodes to be expanded, and the update formula is an adaptive penalty coefficient rule;
[0122] Prior to S1, there is also a data security preprocessing step:
[0123] S01. End-to-end encryption is implemented on the collected static risk control data and dynamic leasing behavior data using the national cryptographic SM4 algorithm;
[0124] S02. Establish a secure channel using the SSL / TLS 1.3 protocol during data transmission;
[0125] S03. Verify the authenticity of the data source through a zero-knowledge proof protocol;
[0126] S04. Implement homomorphic encryption at the data storage layer;
[0127] S05. Generate data security audit logs and synchronize them to the blockchain evidence storage module;
[0128] S06. When data tampering is detected, a three-level security alarm mechanism is automatically triggered:
[0129] Level 1 Alert: Suspend the data acquisition process;
[0130] Level 2 Alert: Isolate suspicious data nodes;
[0131] Level 3 Alert: Reset the encryption key and report to the regulatory authority.
[0132] The mobile phone rental risk control optimization system includes a data probe terminal, a timeliness hub terminal, a model adaptation terminal, a decision engine terminal, an optimization feedback terminal, and a blockchain evidence storage module.
[0133] The data probe is used to capture the hardware fingerprint of the device in real time through the mobile device's sensors and obtain the user credit data stream through the fintech platform's API gateway. It performs encryption and cleaning on multi-source heterogeneous data and triggers data source credibility verification in real time when abnormal device usage behavior and sudden changes in credit records are detected.
[0134] The timeliness center is used to analyze the deviation between the update timestamp of static risk control data and the collection frequency of dynamic behavior data through the Lagrange interpolation algorithm. It generates a feature matrix aligned with the time dimension according to the preset time-frequency calibration rules. When the timeliness deviation of the data is detected to exceed the threshold, the sliding window resampling mechanism is automatically started.
[0135] The model adaptation end is used to run the Q-Learning reinforcement learning algorithm, calculate the feature similarity between the risk control feature combination matrix and the pre-set model library in real time, dynamically select the optimal risk analysis model through the ε-greedy strategy, and automatically switch to the multi-model parallel verification mode when the model matching confidence is lower than the set standard.
[0136] The decision engine is used to perform dynamic weighted inference of random forest and gradient boosting tree, output user risk probability score in real time, generate differentiated quota control instructions based on equipment value coefficient and lease period parameters, and immediately activate the lease permission freeze protocol when the risk value exceeds the industry red line.
[0137] The optimized feedback end is used to continuously track the user's repayment trajectory. The deviation between the model's prediction results and the actual default behavior is analyzed through the confusion matrix. The integrated learning weight parameters are updated based on the dynamic ratio of recall and precision. When concept drift is detected, the model library retraining mechanism is triggered.
[0138] The blockchain evidence storage module is used to write the feature matrix, model selection basis, and quota adjustment instructions in the risk decision-making process into the smart contract in real time. The risk control logic chain is recorded immutably through the distributed ledger, and a complete and verifiable decision evidence package is automatically generated when the regulatory agency initiates an audit request.
[0139] The model matching engine is deployed on edge computing nodes, including:
[0140] Q-Learning processor, real-time calculation of state-action value matrix;
[0141] Feature similarity calculation unit, based on cosine similarity comparison;
[0142] The model cost evaluation unit dynamically monitors memory consumption and feeds it back to the reward function;
[0143] The feedback optimization module includes a blockchain-based evidence storage unit, which is used to record risk decision-making data in an immutable manner. Specific Implementation
[0145] Example 1: Data Timeliness Alignment and Feature Matrix Construction
[0146] In the actual deployment of the mobile phone rental risk control system, a certain operator uses this system for high-end mobile phone rental business. When user A initiates an iPhone 15 Pro Max rental application, the system collects the device's hardware fingerprint through the mobile SDK, including the device model, jailbreak status, and geolocation. At the same time, it retrieves the static risk control data from the credit reporting platform API, including 3 credit card overdue records, with the most recent update time being 48 hours ago. At this time, the dynamic rental behavior data stream is continuously input at a millisecond frequency. It is detected that user A repeatedly changed the delivery address 3 times within 10 seconds. The timeliness alignment module immediately starts the Lagrange interpolation algorithm: first, it parses the static data timestamp sequence, the 24-hour interval of the collection points, and the dynamic behavior data time axis to establish a time-frequency mapping function. Subsequently, based on dynamic data, static credit data is resampled to the same time granularity to generate static risk control correction data. The correction data and dynamic behavior baseline data, with abnormal address behavior marked as risk level II, are stacked vertically by user ID to finally construct a risk control feature combination matrix K. The matrix dimension is 12×150, that is, 12 feature dimensions and 150 time point data, among which the weight of high-risk features is increased to 0.78.
[0147] Example 2: Closed-Loop Feedback and System Optimization
[0148] User A completed the credit card delinquency processing and resubmitted the application within 72 hours after the application was rejected. The feedback optimization module tracked this behavior and updated the confusion matrix data: the original prediction of high risk was transformed into low risk. The system initiated a weight re-optimization program: recalculating the recall rate of the RF model in the most recent 30 days, and dynamically adjusting the weight α=0.84 / (0.84+0.81)=0.51→0.51. Because both increased simultaneously, the weight remained stable. At the same time, the blockchain evidence storage module wrote the two decision data, feature matrix K, model selection basis, and risk probability evolution curve into a smart contract to generate an auditable evidence chain. After 6 months of operation, the system achieved a bad debt rate of 7.9% in the high-end equipment leasing scenario, reduced the false rejection rate, and promoted the year-on-year revenue growth of this business line. The edge computing node automatically generates a model performance report every 24 hours. When the accuracy of the fraud recognition model is detected to be below 80% for 3 consecutive days, the online hot update mechanism of the model library is triggered.
Claims
1. A mobile phone rental risk control optimization method based on multi-dimensional AI model analysis, characterized by: The method includes the following steps: S1. Collect static risk control data and dynamic leasing behavior data from users; S2. Perform data timeliness alignment processing on the static risk control data and dynamic leasing behavior data to generate timeliness-aligned static risk control correction data and dynamic leasing behavior benchmark data; S3. Perform multi-dimensional feature combination processing on the static risk control correction data and the dynamic leasing behavior benchmark data to construct a risk control feature combination matrix; S4. Based on the risk control feature combination matrix and the standard risk control feature matrix under the preset risk scenario, the risk control model type is matched by the reinforcement learning algorithm to generate target risk control model identification data. S5. Call the risk control analysis program corresponding to the target risk control model identifier data, perform risk scoring calibration processing on the dynamic leasing behavior data, and generate an optimized leasing risk assessment result; S6. Dynamically adjust the user's leasing permissions and quota parameters based on the leasing risk assessment results; S1 includes: S11. Collect static risk control data of users through mobile SDK, including identity authentication information, historical credit records, and device hardware fingerprints; S12. Obtain dynamic rental behavior data in real time through the rental platform API, including order fulfillment records, payment delay data, and abnormal equipment usage logs; S3 includes: S31. Stack the static risk control correction data and dynamic leasing behavior benchmark data vertically using user ID as the index to construct a risk control feature combination matrix: ； wherein is static risk control correction data, is dynamic lease behavior benchmark data; S4 includes: S41, Pre-built risk control model feature library: ， ； wherein standard feature matrix corresponding to fraud detection model, credit scoring model, device attrition prediction model, performance capability model, black list matching model; S42, matching using reinforcement learning Q-Learning algorithm with optimal model: Initialize the Q-value table and state space: ； The system state representation is defined as the Cartesian product of the risk control feature combination matrix K and the preset model feature library W Select action at using an ε-greedy strategy; Calculate the reward function: ； wherein, is a similarity reward coefficient, is a time delay penalty coefficient, is a model inference time delay, is a matrix Frobenius norm; renew The value table continues to converge, outputting the highest value. The target risk control model identifier data corresponding to the value .
2. The mobile phone rental risk control optimization method and system based on multi-dimensional AI model analysis according to claim 1, characterized in that: S2 includes: S21. Use a timestamp parsing algorithm to extract the data update frequency of the static risk control data. Data collection frequency of dynamic leasing behavior data ; S22, when and When inconsistencies occur, the Lagrange interpolation algorithm is used to reconcile the static risk control data. Resampling generates static risk control correction data; S23, when and When they are consistent, the original static risk control data is directly output as valid input.
3. The mobile phone rental risk control optimization method based on multi-dimensional AI model analysis according to claim 1, characterized in that: S5 includes: S51, according to Call the corresponding risk control model and input the risk control feature combination matrix. ; S52. Output risk probability through an ensemble learning framework: ； in , For dynamic weights, For random forest, It is a gradient boosting tree; S53, when When the threshold is greater than θ, a reduction in the lease limit and a freeze operation will be triggered, where θ is the industry risk threshold.
4. The mobile phone rental risk control optimization method based on multi-dimensional AI model analysis according to claim 3, characterized in that: The dynamic weights in S52 satisfy the following: ； in, These are the weights for the random forest model. To boost the weights of the tree model using gradient gradation, For random forest recall, For GBDT accuracy; The model is dynamically adjusted based on its recall and precision over the past 30 days.
5. The mobile phone rental risk control optimization method based on multi-dimensional AI model analysis according to claim 1, characterized in that: The process of executing S4 also includes a step of dynamically adjusting system resources: S71. Monitor the concurrent request volume of the model matching engine in real time. When the concurrency exceeds the preset threshold Q, dynamically increase the number of edge computing nodes according to the formula N=ceil(Q / 50). S72. Allocate 3 times the standard computing resources to requests involving high-end equipment leasing; S73. When the single model matching delay exceeds 200ms, automatically switch to the preset default credit scoring model. S74. Update the latency penalty coefficient τ in the Q-Learning reward function based on the resource scheduling log. The update formula is: ； Where Q is the concurrent request threshold, N is the number of edge nodes to be expanded, and the update formula is an adaptive penalty coefficient rule.
6. A mobile phone rental risk control optimization system based on multi-dimensional AI model analysis, used to implement the mobile phone rental risk control optimization method based on multi-dimensional AI model analysis as described in any one of claims 1-5, characterized in that: The mobile phone rental risk control optimization system includes a data probe terminal, a timeliness hub terminal, a model adaptation terminal, a decision engine terminal, an optimization feedback terminal, and a blockchain evidence storage module. The data probe is used to capture the hardware fingerprint of the device in real time through the mobile device's sensors and obtain the user credit data stream through the fintech platform's API gateway. It performs encryption and cleaning on multi-source heterogeneous data and triggers data source credibility verification in real time when abnormal device usage behavior and sudden changes in credit records are detected. The timeliness center is used to analyze the deviation between the update timestamp of static risk control data and the collection frequency of dynamic behavior data through the Lagrange interpolation algorithm. It generates a feature matrix aligned with the time dimension according to the preset time-frequency calibration rules. When the timeliness deviation of the data is detected to exceed the threshold, the sliding window resampling mechanism is automatically started. The model adaptation end is used to run the Q-Learning reinforcement learning algorithm, calculate the feature similarity between the risk control feature combination matrix and the pre-set model library in real time, dynamically select the optimal risk analysis model through the ε-greedy strategy, and automatically switch to the multi-model parallel verification mode when the model matching confidence is lower than the set standard. The decision engine is used to perform dynamic weighted inference of random forest and gradient boosting tree, output user risk probability score in real time, generate differentiated quota control instructions based on equipment value coefficient and lease period parameters, and immediately activate the lease permission freeze protocol when the risk value exceeds the industry red line. The optimized feedback end is used to continuously track the user's repayment trajectory. The deviation between the model's prediction results and the actual default behavior is analyzed through the confusion matrix. The integrated learning weight parameters are updated based on the dynamic ratio of recall and precision. When concept drift is detected, the model library retraining mechanism is triggered. The blockchain evidence storage module is used to write the feature matrix, model selection basis, and quota adjustment instructions in the risk decision-making process into the smart contract in real time. The risk control logic chain is recorded immutably through the distributed ledger, and a complete and verifiable decision evidence package is automatically generated when the regulatory agency initiates an audit request.