Cross-institution rural financial modeling method based on federated learning

Abstract

The invention relates to the technical field of federated learning and financial risk control, and discloses a cross-institution rural financial modeling method based on federated learning, and the method comprises the steps: a central server broadcasts a synchronous beacon containing a feature attention list according to an agricultural event calendar; and after each participant terminal receives the synchronization beacon, executing local model training according to the synchronization beacon, and performing information entropy analysis on the obtained initial gradient to screen out a high-value gradient subset and only uploading the high-value gradient subset to the central server. And the information value is taken as a transmission basis, and continuous data interaction in a weak network environment is avoided, so that the effectiveness of the risk control model is ensured, and the operation reliability and deployment feasibility of the whole collaborative modeling system in an actual rural scene are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a cross-institutional rural financial modeling method based on federated learning, belonging to the technical field of federated learning and financial risk control. BACKGROUND

[0002] In the current rural financial practice, in order to improve the accuracy of risk assessment, the introduction of federated learning to integrate discrete data among multiple institutions has become a technical trend. This approach follows the mature real-time modeling framework in urban financial technology based on continuous high-frequency communication, and its basic logic is to pursue higher model accuracy by increasing the interaction frequency and data dimension of model parameters.

[0003] However, when this technical framework is directly applied to the rural financial scene with significant industry specificity, the contradiction between its inherent design assumptions and the reality environment is highlighted. Agricultural production activities have strong periodicity, and financial demand and risk characteristics also fluctuate in a tidal pattern. For example, during the risk stable period of crop growth, continuous gradient exchange only consumes limited communication and terminal energy, but does not effectively improve the model's insight into potential risks. At critical nodes such as harvest or disaster, the instantaneous increase in data transmission demand often leads to congestion of existing communication networks, making the most critical model updates unable to be completed in time. This mismatch between the technical approach and the rhythm of agricultural finance leads to a structural contradiction between resource investment and model effectiveness.

[0004] Specifically, the existing technology mainly has the following deficiencies: 1. The operation mode of model collaboration cannot match the periodic rhythm of agricultural production, resulting in unnecessary resource consumption during the risk stable period, and communication bottlenecks during the risk high-incidence period, with low collaboration efficiency; 2. The continuous homogenization of data interaction makes the risk signals with high information value that appear in critical periods easily overwhelmed by the vast amount of low-value routine update data, and the model's timeliness in capturing critical risks is insufficient. Therefore, how to build a new method that can match the inherent periodic rhythm of agricultural production and achieve efficient and accurate collaborative modeling under limited communication and computing resources has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a cross-institutional rural financial modeling method based on federated learning, which mainly aims to solve the problem of contradiction between resource consumption and model timeliness caused by the inability of existing federated learning methods to adapt to the periodic rhythm of agriculture in rural scenarios while ensuring model effectiveness.

[0006] To achieve the above purpose, the present application provides a cross-institutional rural financial modeling method based on federated learning, comprising the following steps:

[0007] Step a, broadcasting a synchronization beacon containing event type code and feature attention list to multiple participant terminals by a central server according to an agricultural financial event calendar recording agricultural periodic events and corresponding feature attention list;

[0008] Step b, extracting corresponding features from local data according to feature attention list and performing local model training to obtain initial gradient after participant terminal receives synchronization beacon;

[0009] Step c, performing a gradient entropy screening step on initial gradient by participant terminal, screening step includes: performing information entropy analysis on each feature gradient specified by feature attention list in initial gradient, and determining gradient with information entropy higher than an entropy threshold as high-value gradient subset;

[0010] Step d, uploading high-value gradient subset to central server for aggregated update of global model by participant terminal.

[0011] Preferably, the step of broadcasting synchronization beacon in step a is specifically completed through at least one communication channel selected from wireless data system broadcast channel and short message service channel; and the method further comprises: a beacon daemon step performed by participant terminal, when synchronization beacon is not received through IP network channel within a time period before and after event time point recorded in agricultural financial event calendar, the beacon daemon step switches to receive through at least one communication channel.

[0012] Preferably, the beacon daemon step further comprises: adjusting actual execution time of local model training according to regional agricultural fine-tuning calendar recorded in terminal local storage and recording regional time deviation information after receiving synchronization beacon through at least one communication channel.

[0013] Preferably, it further comprises a step: participant terminal locally caches and accumulates the remaining gradients which are not determined as high-value gradient subset in step c in the form of gradient residual, and adds accumulated gradient residual to newly generated initial gradient when performing local model training next time.

[0014] Preferably, the agricultural periodic events recorded in agricultural financial event calendar include one or more of spring ploughing and loan preparation period, crop growth period, harvest settlement period, agricultural product market price fluctuation period and natural disaster early warning period.

[0015] Preferably, it further comprises a feature confidence updating step performed after step d, the updating step includes: participant terminal calculates and updates a confidence score C i,t for each feature in a feature dictionary based on gradient residual of continuous multiple periods, and the updating rule of confidence score is: i,t C i,t-1 = (1-α)C + αδi,t wherein, C i,t is the confidence score of feature i in the current cycle, C i,t-1 is the confidence score of feature i in the previous cycle, a is a learning rate factor between 0 and 1, δ i,t is a state indicator, which is 1 when the gradient of feature i is determined as a high-value gradient subset in the current cycle, and 0 when the gradient of feature i is relegated to the gradient residual in the current cycle.

[0016] Preferably, the method further comprises a feature drift alerting step, the alerting step comprising: when the confidence score C i,t of a feature is lower than a drift determination threshold, generating a drift alert containing the index of the feature, and sending the drift alert to the central server together with the next uploaded high-value gradient subset; and the central server, upon receiving a number of drift alerts for the same feature index accumulated to a consensus number threshold, reducing the weight of the feature when performing the global model aggregation update.

[0017] Preferably, the method further comprises, before performing the local model training in step b, performing an event spatiotemporal fingerprint generating step, the generating step comprising: calling a satellite navigation module onboard the terminal to obtain a satellite time-stamped timestamp and a geographic grid code; and generating an event spatiotemporal fingerprint through a hash operation using the event type code parsed from the synchronization beacon, the obtained satellite time-stamped timestamp, and the geographic grid code; and in step d, uploading the generated event spatiotemporal fingerprint together with the high-value gradient subset.

[0018] Preferably, the method further comprises, before the central server performs the global model aggregation update, performing an event consensus arbitration step, the arbitration step comprising: performing consistency comparison on event spatiotemporal fingerprints received from a plurality of participant terminals in a geographically proximate location; and when an event spatiotemporal fingerprint of a certain terminal is identified to have a timestamp or a geographic grid code deviating from a spatiotemporal reference value calculated from event spatiotemporal fingerprints of more than one proximate terminal by more than a deviation threshold, determining the high-value gradient subset uploaded by the terminal this time as abnormal data and isolating it.

[0019] Preferably, the method further comprises, after the global model aggregation update, performing a post-decision arbitration step, the arbitration step comprising: after generating a preliminary credit decision using the updated global model, verifying the preliminary credit decision according to one or more financial risk control rules; the financial risk control rules including: verifying whether the preliminary credit decision is abnormally vetoed according to a whitelist of high-quality customers, and verifying whether the credit limit of the preliminary credit decision exceeds the upper limit constraint according to historical tax data of the applicant; and when the preliminary credit decision conflicts with any financial risk control rule, marking the decision and transferring it to a manual review process.

[0020] Compared with the prior art, the application has the following beneficial effects:

[0021] 1. The method provided by the application changes the operation mode of continuous high-frequency data interaction of the existing federated learning by triggering local model training of each participant terminal by the central server according to agricultural financial event history broadcast synchronization beacons, and combining a gradient entropy screening step to upload only a high-value gradient subset, the collaborative mechanism unified by the event rhythm and based on information value for transmission makes the model update action synchronized with the key nodes of agricultural production, avoids unnecessary communication and calculation in the risk stable period, and can efficiently capture key risk characteristics by focusing on high information entropy gradients in the risk concentration period, thereby establishing an effective balance between resource consumption and model efficiency under the limited rural network and energy conditions.

[0022] 2. The application avoids information loss caused by gradient sparsification by locally caching and accumulating the gradients that are not uploaded in the gradient entropy screening step, and superimposes them in subsequent training, thereby guaranteeing the long-term convergence of the model, and then uses the accumulated gradient residual to continuously diagnose the effectiveness of the feature dictionary, calculates and updates the confidence score of each feature, can timely identify the feature semantic drift caused by external environmental changes, and feeds back the change of this implicit state to the central server through a drift alarm, the design of changing the gradient residual from an information compensation element to a dictionary state probe enables the entire federated learning system to have a dynamic self-calibration capability, and can maintain its effectiveness on a longer time scale.

[0023] 3. The application introduces a step of generating an event space-time fingerprint before local model training, which adds a space-time reference derived from an external physical system to each collaborative training behavior triggered by a synchronization beacon by calling the timestamp and geographic coding obtained by a satellite navigation module, the central server performs consensus arbitration on these space-time fingerprints before aggregating the model, which can effectively identify and isolate invalid data caused by terminal clock out-of-step or signal anomaly, and ensure the consistency of the data on which the global model update is based in the space-time dimension, the combination of logical verification at the algorithm level and reference verification at the physical level builds a barrier against external disturbances and internal failures for the distributed modeling system, and improves the operation reliability of the financial risk control model.

[0024] 4、The application sets a decision post-arbitration step after global model updating, which utilizes independent verification logic based on deterministic financial risk control rules to review the preliminary credit decisions generated by the federated learning model; this mechanism places model decisions based on statistical probability within a verification process based on business hard constraints, which can effectively intercept and mark abnormal decisions that deviate from business common sense; in this way, while exerting the data mining capabilities of the federated learning model, it also provides the final guarantee of conforming to the real business logic for the final financial decision, improving the credibility and security of the entire risk control system in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1 The overall flowchart of the application is based on the federated learning cross-institutional rural financial modeling method;

[0026] Fig. 2 The application is a comparison chart of the method and the traditional federated learning method in the amount of communication data throughout the year;

[0027] Fig. 3 The time-space fingerprint generation and consensus arbitration process timing diagram of the application.

[0028] The purpose of the application, functional characteristics and advantages will be further described with reference to the accompanying drawings. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme of the application will be described in detail below. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0030] The technical scheme of the federated learning-based cross-institutional rural financial modeling method disclosed in the application relies on a central server as a coordination hub and a plurality of participant terminals deployed in agricultural-related financial institutions. The method replaces the continuous high-frequency interaction mode in traditional federated learning through a set of event-driven collaboration mechanisms. The overall process is composed of several core stages, including beacon broadcasting by the central server based on agricultural periodic events, local model calculation and gradient screening by each participant terminal, and global model aggregation update and multi-dimensional arbitration verification, so that data can flow asymmetrically between the central server and each participant terminal according to the inherent rhythm of agricultural production and the actual value of information. In a specific application scenario, for example, to improve the cross-institutional credit risk control capability of a plurality of township banks and rural credit cooperatives in a county, the business terminals of these institutions serve as participant terminals. At key nodes of agricultural production, such as the spring planting and loan preparation period, the demand for funds and credit risk increase synchronously. At this time, the conventional continuous modeling method is prone to cause communication resource shortage and key risk signal flooding in weak network environments in rural areas. To address this challenge and ensure reliable operation of the system under intermittent signal interruption and other extreme conditions, the collaborative action of the application scheme is uniformly scheduled by an agricultural financial event calendar preset in the central server. The event calendar is a structured data table that pre-recorded key events throughout the agricultural year, such as the spring planting and loan preparation period, the crop growth period, the harvest settlement period, and the natural disaster warning period. Each type of event is associated with a feature attention list. When the time point of a preset event in the agricultural financial event calendar approaches, the central server starts the broadcast process of the synchronization beacon. This process extracts the event type code and corresponding feature attention list of the current period from the event calendar and encapsulates them as a lightweight synchronization beacon data packet. To ensure reliable delivery of the beacon in complex rural network environments, the system is configured with a multi-channel redundant communication scheme. The preferred interaction path is the IP network channel. When the beacon daemon step embedded in the participant terminal fails to receive the synchronization beacon through the IP network channel within a preset time window before and after the event time point, the daemon step automatically switches to a backup communication channel, such as a wireless data system broadcast channel or a short message service channel, for reception. This design aims to proactively eliminate the risk of synchronization instruction failure due to single network path failure. It is worth noting that after receiving the synchronization beacon, the participant terminal will further adjust the actual execution time of model training based on the local stored regional agricultural calendar to make the modeling action more accurately match the real agricultural operation rhythm in the local area.

[0031] Before performing the local model training, to ensure the spatio-temporal consistency and auditability of the subsequent aggregated data, the terminal first performs an event spatio-temporal fingerprint generation step, which is configured to call the satellite navigation module on board the terminal, obtain satellite time stamp and geographic grid encoding originating from the external physical system, and generate a fingerprint capable of uniquely identifying the spatio-temporal background of the training event by a deterministic hash operation, combining the event type code parsed from the synchronization beacon with the two above. After generating the fingerprint, the participant terminal extracts the corresponding features from the local data according to the feature attention list, performs the local model training, and obtains the initial gradient. Since the transmission of the complete initial gradient in a weak network environment involves huge resource overhead, and the information value density contained in each dimension of the gradient is significantly different, the participant terminal performs a gradient entropy screening step after obtaining the initial gradient. The core of this step is to quantitatively analyze the information entropy of each feature gradient specified by the feature attention list in the initial gradient. The information entropy is used here as an indicator to measure the new information or unexpectedness carried by the gradient. If the information entropy of a feature gradient is higher than a preset entropy threshold, it is considered to reflect the risk dynamics deviating from the regular state with high value. The determination of the entropy threshold follows an objective calibration procedure, that is, by analyzing the statistical distribution of gradient information entropy in the system historical data, and setting a certain high percentile point, such as the 90th percentile, as the threshold, so that only the gradient part with information entropy higher than the threshold is identified as the high-value gradient subset and prepared for uploading. This mechanism enables the limited communication bandwidth to be accurately used to transmit the most critical risk change signals.

[0032] For the remaining gradients that are not selected in the gradient entropy screening step, the system does not discard them directly, but instead stores them locally in the form of gradient residuals and accumulates them across periods. In the next local model training, the accumulated gradient residuals are vectorially added to the newly generated initial gradient to compensate for the information loss that may be caused by gradient sparsification. In other words, the accumulated value of the gradient residual is further used to build a dynamic diagnostic mechanism for feature effectiveness. Based on the gradient residual status of consecutive periods, the terminal calculates and updates the confidence score C i,t of each feature in a locally maintained feature dictionary. The update rule follows the formula C i,t = (1 - a)C i,t-1 + aD i,t , where C i,t and C i,t-1 are the confidence scores of feature i in the current period and the previous period, respectively, a is a learning rate factor, and the value of the state indicator D i,t depends on whether the gradient of feature i is included in the high-value gradient subset in the current period. When the confidence score C i,tWhen the drift is below a preset drift determination threshold, the system generates a corresponding feature drift alarm; finally, the participant terminal uploads the screened high-value gradient subset, the generated event space-time fingerprint, and the possible triggered feature drift alarm to the central server. Before aggregating the data of each party, the server will first perform an event consensus arbitration step, which compares the consistency of the event space-time fingerprints received from multiple terminals in a geographical location. When the fingerprint of a terminal deviates significantly from the space-time reference value formed by the group of neighboring terminals, the gradient data uploaded by the terminal is determined to be abnormal and is isolated. At the same time, if the drift alarm for the same feature reaches a consensus quantity threshold, the server will correspondingly reduce the weight of the feature when performing global model aggregation. This multi-dimensional arbitration mechanism significantly improves the data quality of global model updating; after the global model is aggregated and updated, in order to ensure that the output of the model can match the existing financial business rules, the system also sets a decision post-arbitration step. After generating a preliminary decision on a credit application using the updated global model, the decision will be reviewed according to a set of deterministic financial risk control rules, which include but are not limited to: verifying whether there is an abnormal veto according to the whitelist of high-quality customers, and verifying whether the credit limit exceeds the upper limit of rigid constraints according to the historical financial data of the applicant. Any preliminary credit decision that conflicts with the financial risk control rules will be automatically marked by the system and transferred to the manual review process, thereby providing the entire automated risk control system with the final logical guarantee and safety redundancy.

[0033] Embodiment 1: In an agricultural county covering multiple townships, a sudden pest outbreak occurs due to continuously abnormally high temperature weather, which is classified as a natural disaster warning period in the agricultural financial event chronicle. For financial institutions that have a large amount of credit business with local farmers, an urgent challenge arises: how to dynamically and accurately reevaluate the credit risks of different farmers in the affected area within a short period of time to deal with possible concentrated defaults. The communication infrastructure in the affected area is highly uneven, with basic IP network coverage in the township centers, but extremely weak or missing network signals in the vast remote farmland areas. Under this working condition, the system deployed in the central server immediately broadcasts a synchronization beacon containing the event type code and the corresponding feature attention list of this natural disaster warning period event to all participant terminals in the county according to the updated agricultural financial event chronicle. The feature attention list explicitly points to feature dimensions highly related to this disaster, such as pesticide purchase amount, irrigation water amount, and social network information activity among farmers, etc. The participant terminals deployed in the township centers receive this synchronization beacon through the IP network channel, while the terminals in the remote farmland areas switch and finally receive this instruction through the short message service channel after failing to obtain the beacon through the IP network channel. The combination of event-driven beacon broadcasting mechanism and multi-channel redundant communication scheme provides a reliable path for the coordinated issuance of instructions across heterogeneous networks.

[0034] After each participant terminal performs local model training and generates initial gradients, the subsequent gradient entropy screening step of the terminal obtains higher analysis relevance due to the context provided by the synchronization beacon. The terminal focuses computing resources on analyzing the gradients of key features such as pesticide purchase quantity according to the feature attention list. In some farmer data that has taken countermeasures, the gradient of pesticide purchase quantity presents a very high information entropy, while the gradients of other regular features maintain a low entropy state. Therefore, the system can screen these high-entropy gradients into a high-value gradient subset, and cache a large amount of regular, low-value gradient information as gradient residuals. This mechanism converts the original model update problem that cannot be performed under a weak network into an asymmetric communication problem of transmitting a small amount of key information in a limited channel. The focus of the system shifts from maintaining continuous and homogeneous data interaction to efficiently capturing and transmitting high-information-entropy risk features at key nodes. Finally, the central server only receives the high-value gradient subset uploaded by each terminal, which has a very small data volume but a very high information density, and quickly completes the aggregation update of the global model according to the high-value gradient subset. The updated model has the ability to differentiate the risk levels of different farmers under the current disaster, and financial institutions can maintain or increase credit for farmers who show positive coping behavior according to the model output, while starting a risk warning and communication mechanism for farmers in the disaster area who have no coping behavior. The risk exposure of the entire financial system remains stable under the impact of disasters.

[0035] Embodiment 2: To objectively verify the running efficiency and model validity of the technical scheme of the present application in the resource-constrained and periodic fluctuation scenario, a comparative test is set up. The purpose of the test is to quantitatively evaluate the performance of the method of the present application relative to the traditional federated learning method of continuous high-frequency communication in terms of communication overhead, terminal computing load and key event response capability. For this purpose, a simulation test platform is constructed, which includes a central server and 50 participant terminals. The platform simulates a typical mixed network environment in rural areas through software-defined network, and loads a credit history and behavior simulation data set covering a complete twelve-month agricultural production cycle, which includes a clear spring farming loan preparation period, a risk stable period, a harvest settlement period and a sudden natural disaster warning period. The test sets up a test group using the scheme of the present application, and a control group using the traditional high-frequency, full-amount gradient interaction scheme. The federated learning update frequency of the control group is set to once an hour to simulate the high-frequency interaction strategy adopted by the prior art in pursuit of real-time model. The core parameter entropy threshold of the test group is set in accordance with a deterministic procedure. The procedure is to pre-analyze an independent validation set in the simulation data set, to statistically analyze the information entropy distribution of all feature gradients in different agricultural cycles, and to take the 90th percentile point as the threshold. The engineering consideration of the threshold is to balance the information fidelity required for model updating and the resource consumption of communication transmission. Based on this procedure, the entropy threshold used in this test is 0.85.

[0036] During the simulation test running process, the system monitors and records a plurality of performance indicators of the two groups in different agricultural cycles. In the risk stable period lasting for several months, the terminals of the control group continuously generate and upload complete initial gradients at a fixed frequency, generating a large amount of stable communication and computing overhead. The terminals of the test group upload only a small amount of high-value gradient subsets during this period because most of the gradient information entropy is below the entropy threshold, and their communication and computing activities remain at a low level. When the simulation enters the natural disaster warning period in November, the network state is adjusted to a high congestion and high packet loss rate mode. At this time, the communication delay of the control group increases, and a large number of gradient update packets are lost due to timeout, resulting in a decrease in the convergence speed of the global model. The terminals of the test group begin to focus on the gradient changes of disaster-related features and upload only high-value gradient subsets directly related to the disaster through the gradient entropy screening step due to the triggering of the synchronization beacon. The data packet size of the test group is much smaller than that of the control group, so it performs better in the congested network. Table 1 shows the comparison of key performance data collected during the test in two typical cycles.

[0037] Table 1: Comparison of performance data of test group and control group in different cycles.

[0038]

[0039] Referring to Table 1, during periods of stable risk, the experimental group maintained the same level of model effectiveness with significantly lower resource consumption than the control group. This phenomenon stems from the effective filtering of a large amount of redundant update information during this phase by the gradient entropy screening step. At the same time, the local accumulation compensation mechanism of the gradient residuals prevented significant degradation of model performance. During critical natural disaster warning periods, the model effectiveness of the experimental group was significantly higher than that of the control group, whose performance deteriorated due to network congestion. Moreover, the response latency for critical events was approximately one-seventh that of the control group. These results indicate that the mechanism combining event-driven synchronization beacons and gradient entropy screening can prioritize the transmission channels of critical risk information when network resources become a bottleneck, thereby ensuring that the model responds quickly to emergencies.

[0040] Example 3: This example combines Figs. 1 to 3 This section explains the cross-institutional rural finance modeling method based on federated learning, such as... Fig. 1 As shown in the diagram, this diagram starts with an agricultural financial event calendar, pre-setting a list of cyclical agricultural events and features of interest. Based on this calendar, the central server triggers a broadcast synchronization beacon at key nodes, sending collaborative instructions containing event codes and feature lists to each terminal. After receiving the synchronization beacon, participating terminals enter the local model training phase. Based on the feature list, they train from local data and generate initial gradients. Simultaneously, the terminals call the satellite navigation module to execute the event spatiotemporal fingerprint generation step, obtaining timestamps and geographic grid codes, which are used to construct the physical system spatiotemporal reference for training behavior, constituting the event spatiotemporal fingerprint generation step shown in the diagram. The initial gradients obtained after training then enter the gradient review step. The terminals perform information entropy analysis on the gradients, identifying and filtering high-value gradient subsets with information entropy exceeding a threshold; the remaining gradients are cached and accumulated locally as gradient residuals, constituting the content shown in the diagram of local gradient residual caching and accumulation. This step is used to compensate for gradients that were not uploaded. Following the information loss, in the data packet upload stage, participating terminals upload high-value gradient subsets, event spatiotemporal fingerprints, and feature drift alerts, if any. Drift alerts originate from the feature confidence update and drift alert steps. This step diagnoses feature validity based on cumulative residuals, identifies long-term invalid features, and triggers drift warnings. After upload, the central server executes the event consensus arbitration step. By comparing the event spatiotemporal fingerprints of each terminal, it identifies and isolates invalid data caused by spatiotemporal anomalies. Subsequently, it performs global model aggregation and update. The central server aggregates the valid gradients after arbitration and dynamically adjusts feature weights in conjunction with drift alerts to update the global model. After aggregation, the system enters the post-decision arbitration step. Based on financial risk control rules such as whitelists and credit limits, it reviews the preliminary credit decision output by the federated model, forming a risk control closed loop. Finally, it outputs the final credit decision, which can be manually reviewed or automatically executed to complete the entire closed loop.

[0041] likeFig. 2 As shown, the horizontal axis is the time month, marked as January to December, representing the monthly cycle of a complete agricultural year; the vertical axis is the communication data volume MB / terminal, representing the communication load size of each terminal in the corresponding month, and two curves are used in the figure to represent two methods, where the solid curve represents the traditional federated learning method, which maintains a high communication load of about 45 MB / terminal in each month; the dashed curve represents the method of the present application, which significantly reduces the communication data volume during the non-critical agricultural event months such as February to September, and only maintains below 1 MB / terminal, while moderately increasing the communication overhead during the critical event period such as January, October and December, but is still significantly lower than the traditional method. The method of the present application focuses on uploading high-value data only during the critical period of agricultural finance under the premise of ensuring the effectiveness of risk modeling, thereby realizing the on-demand allocation and significant optimization of communication load, which is suitable for risk modeling scenarios in rural weak network environments.

[0042] As shown in Fig. 3 As shown, the participating terminal 1 and the participating terminal 2 first respectively initiate a request satellite time service operation to the satellite navigation module, and obtain the returned timestamp and geographic code, based on which event space-time fingerprint 1 and event space-time fingerprint 2 are respectively generated, then the two terminals respectively upload the gradient+space-time fingerprint 1 and gradient+space-time fingerprint 2 obtained by local training to the central server, after receiving the upload information, the central server triggers to start the consensus arbitration process, and calls the consensus arbitration module to perform space-time consistency judgment, the consensus arbitration module first compares the fingerprints of adjacent terminals, and calculates the space-time reference value according to the comparison, if it is found that the deviation of the fingerprint uploaded by a terminal exceeds the threshold-value, the terminal is identified as abnormal, and is marked as PT data abnormal, and then performs the operation of isolating abnormal data; if all fingerprint consistencies pass, it is judged that the data is credible, and the batch data enters the gradient aggregation calculation link, and is finally included in the global model update.

[0043] Embodiment 4: In a technical solution of the present application that has been running for multiple agricultural years, one long-term operation and maintenance problem is that the risk relevance of some data features may slowly drift in semantics with changes in market environment or planting structure, if such state changes cannot be effectively monitored, it may lead to the decay of the prediction ability of the global model, this embodiment aims to solve this problem Feature confidence update and feature drift alarm mechanism, its internal algorithm path and key parameter determination procedures are calculated and displayed; considering a feature with index i in the feature dictionary, its confidence score in the t-1 period is C i,t-1= 0.7, the update rate of the score is controlled by the learning rate factor a, the value of a is set according to the historical stability of the feature, for a feature with stable historical performance and solid risk connotation, the value of a can be set to 0.05 to enhance the anti-disturbance of the score, for a feature with emerging risk relevance in dynamic evolution, the value of a can be set to 0.2 to enhance the sensitivity of the score to new information patterns, in this case, feature i is considered as a new observation dimension, so the value of a is set to 0.2.

[0044] In the cooperative modeling of entering the t period, the initial gradient generated by the local model training of the participant terminal, the gradient component corresponding to feature i is analyzed by the gradient entropy screening step, the information entropy does not reach the entropy threshold, and is not included in the high value gradient subset, and is cached in the form of gradient residual, at this time, the value of the state indicator δ i,t is 0, the terminal triggers the update calculation of the feature confidence score, and the updated score is C i,t = (1- a) C i,t-1 + a δ i,t The decrease of this score quantifies the fact that the feature fails to provide high value information in this period; entering the t+1 period, after generating a new initial gradient, the terminal first performs vector superposition on the gradient residual of feature i cached in the t period and the newly generated feature t gradient component to form a modified gradient, the information entropy of the modified gradient is determined to be higher than the entropy threshold after the gradient entropy screening step, so it is included in the high value gradient subset and uploaded to the central server, at this time, the value of the state indicator δ i,t+1 is 1, the terminal updates the feature confidence score again, and the result is C i,t = 0.56, the rebound of the score reflects that the feature contributes effective high information update in this round; this continuous score update mechanism combined with a preset drift judgment threshold forms the trigger logic of the feature drift alarm, the setting of the drift judgment threshold follows a statistical analysis procedure, that is, by analyzing the confidence scores of a large number of stable features in the system in long-term operation, taking the position of two standard deviations below the mean value as the lower limit of the threshold, the lower limit can be set to 0.4, when the score C i,t of feature i continuously decreases and eventually falls below 0.4 in subsequent periods, the terminal generates a drift alarm containing the index of the feature, this whole closed loop operation from gradient residual accumulation to confidence dynamic score to alarm generation based on statistical threshold provides an automatic feature effectiveness self-diagnosis process for the whole federated learning system.

[0045] Embodiment 5: Before deploying the technical solution of the present application in a new cross-institutional rural financial cluster, in order to ensure the effectiveness of the event consensus arbitration step, a pre-emptive spatio-temporal benchmark calibration procedure is performed, which requires all newly added participant terminals in the cluster to respond to periodic calibration beacons sent by the central server within a pre-set 24-hour calibration period and continuously upload event spatio-temporal fingerprints generated by the respective call onboard satellite navigation modules. The central server collects all fingerprint data during this stage and, for the terminal cluster adjacent to the geographic location, calculates the fluctuation distribution of its timestamp and geographic grid code. Then, the deviation from the mean of this distribution that exceeds four standard deviations is set as the deviation threshold that triggers abnormal data isolation. This procedure establishes a judgment benchmark based on the local physical environment and device characteristics for subsequent real-time consensus arbitration.

[0046] After the cluster is formally put into operation, one of the participant terminals has a local clock failure, resulting in a two-hour forward timestamp offset in the event spatio-temporal fingerprint generated when responding to a spring farming loan period synchronization beacon, and the high-value gradient subset calculated by the terminal is uploaded together. After receiving the data, the central server performs the event consensus arbitration step and compares the spatio-temporal fingerprint of the terminal with the fingerprints uploaded by the remaining terminals adjacent to its geographic location for consistency. Since the timestamp deviation of the terminal is much larger than the deviation threshold set by the pre-calibration procedure, the system determines that the high-value gradient subset uploaded by the terminal this time is abnormal data and isolates it, excluding it from the global model aggregation update in this round. This mechanism avoids the impact on the stability of the global model that may be caused by a single terminal physical layer failure.

[0047] Embodiment 6: When a newly established rural financial cooperative needs to join an already stable running federated learning network of the present application, in order to achieve smooth access of new participants and ensure dynamic adaptation of global system parameters, a standardized new member access and parameter initialization procedure is performed. This procedure first requires the new participant terminal to download the current version of the global model from the central server. After downloading, the terminal enters a seven-day local model warm-up period during which the terminal only uses the global model to infer its local historical data, without uploading any gradient. This step aims to align the local environment of the new terminal with the existing federated system and complete the initialization and construction of the local feature dictionary.

[0048] After the preheating period, the new participant generates and submits rule parameters according to its internal data for the financial risk control rules required by the post-decision configuration arbitration step. For the whitelist of high-quality customers, the generation procedure is to filter out a list of customers with no overdue records in the past 36 months and the highest credit rating, and then hash the list and submit it. For the credit limit upper bound constraint, the rule is quantified as a calculation method, that is, the credit limit is not more than 3.5 times the average tax-paid income of the applicant in the past two years. At the same time, the central server dynamically resets the consensus quantity threshold according to the joining of new members. The setting procedure of the threshold is 30% of the total number of active participants in the current federated network, and the result is rounded up. After completing the above procedures, the new participant terminal can formally start responding to the synchronization beacon broadcast by the central server and participate in the subsequent local model training and upload process of the high-value gradient subset. This whole set of standardized access and calibration process provides a scalable and risk-controllable member expansion mechanism for the federated learning network, ensuring that the joining of new members will not disturb the stability of the existing global model, and enabling the risk control rule system and key arbitration parameters of the whole system to adaptively adjust with the evolution of the network.

[0049] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0050] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application.

Claims

1. A cross-institutional rural finance modeling method based on federated learning, characterized in that, Includes the following steps: Step a: A central server broadcasts a synchronization beacon containing event type codes and feature attention lists to multiple participating terminals based on an agricultural financial event calendar that records agricultural cyclical events and corresponding feature attention lists. Step b: After receiving the synchronization beacon, the participating terminal extracts the corresponding features from the local data according to the feature attention list and performs local model training to obtain the initial gradient. Step c, the participating terminal performs a gradient entropy screening step on the initial gradient. The screening step includes: performing information entropy analysis on each feature gradient specified by the feature interest list in the initial gradient, and determining the gradients with information entropy higher than the entropy threshold as a high-value gradient subset. In step d, the participating terminals upload a subset of high-value gradients to the central server for aggregation and updating of the global model.

2. The cross-institutional rural finance modeling method based on federated learning according to claim 1, characterized in that, The step of broadcasting the synchronization beacon in step a is specifically completed through at least one communication channel selected from the wireless data system broadcast channel and the short message service channel; and the method further includes: the participating terminal performs a beacon guarding step, and when the synchronization beacon is not received through the IP network channel within a time period before and after the event time point recorded in the agricultural financial event calendar, the beacon guarding step switches to at least one communication channel for reception.

3. The cross-institutional rural finance modeling method based on federated learning according to claim 2, characterized in that, The beacon guarding process also includes: after receiving a synchronization beacon through at least one communication channel, adjusting the actual execution time of local model training based on a regional agricultural fine-tuning calendar stored locally on the terminal that records regional time deviation information.

4. The cross-institutional rural finance modeling method based on federated learning according to claim 1, characterized in that, It also includes a step: the participating terminal caches and accumulates the remaining gradients that were not identified as high-value gradient subsets in step c as gradient residuals locally, and adds the accumulated gradient residuals to the newly generated initial gradients during the next local model training.

5. The cross-institutional rural finance modeling method based on federated learning according to claim 1, characterized in that, The agricultural financial event history records one or more of the following cyclical agricultural events: the spring planting preparation period, the crop growth period, the harvest and settlement period, the period of fluctuation in agricultural product market prices, and the period of natural disaster warning.

6. The cross-institutional rural finance modeling method based on federated learning according to claim 4, characterized in that, The process also includes performing a feature confidence update step after step d. This update step involves the participating terminal calculating and updating a confidence score C for each feature in a feature dictionary based on the gradient residuals over multiple consecutive periods. i,t The update rule for the confidence score is: C i,t =(1-α)C i,t-1 +αδ i,t , where C i,t C is the confidence score of feature i in the current period. i,t-1 Let α be the confidence score of feature i in the previous period, α be a learning rate factor between 0 and 1, and δ be the learning rate factor. i,t It is a state indicator. Its value is 1 when the gradient of feature i is determined to be a high-value gradient subset in the current period, and its value is 0 when the gradient of feature i is classified into the gradient residual in the current period.

7. A cross-institutional rural finance modeling method based on federated learning according to claim 6, characterized in that, It also includes a feature drift alert step, which includes: when the confidence score C of a certain feature... i,t When the number of drift alerts falls below a drift threshold, a drift alert containing the feature index is generated, and the drift alert is sent to the central server along with the next high-value gradient subset uploaded. When the number of drift alerts received by the central server for the same feature index reaches a consensus threshold, the weight of the feature is reduced when performing a global model aggregation update.

8. The cross-institutional rural finance modeling method based on federated learning according to claim 1, characterized in that, The method further includes: before performing local model training in step b, performing an event spatiotemporal fingerprint generation step, which includes: calling the satellite navigation module on the terminal to obtain the satellite timing timestamp and geographic grid code; and generating an event spatiotemporal fingerprint by hashing the event type code parsed from the synchronization beacon, the obtained satellite timing timestamp, and the geographic grid code; and uploading the generated event spatiotemporal fingerprint together with a high-value gradient subset in step d.

9. A cross-institutional rural finance modeling method based on federated learning according to claim 8, characterized in that, The method also includes: before the central server performs global model aggregation update, performing an event consensus arbitration step, which includes: performing consistency comparison on the event spatiotemporal fingerprints received from multiple participating terminals that are geographically close; when it is identified that the event spatiotemporal fingerprint of a certain terminal has a deviation greater than a deviation threshold from the spatiotemporal reference value calculated from the event spatiotemporal fingerprints of more than one neighboring terminal, the high-value gradient subset uploaded by that terminal this time is judged as abnormal data and isolated.

10. A cross-institutional rural finance modeling method based on federated learning according to claim 1, characterized in that, The method also includes: after the global model is aggregated and updated, a post-decision arbitration step is performed. The arbitration step includes: after a preliminary credit decision is generated using the updated global model, the preliminary credit decision is verified according to one or more financial risk control rules. The financial risk control rules include: verifying whether there is any abnormal rejection of the preliminary credit decision based on a whitelist of high-quality customers, and verifying whether the credit limit of the preliminary credit decision exceeds the upper limit constraint based on the applicant's historical financial and tax data. When the preliminary credit decision conflicts with any financial risk control rule, the decision is marked and transferred to the manual review process.

Images