Repayment request decision method and system based on multi-source data fusion prediction

Abstract

The application discloses a repayment request decision method and system based on multi-source data fusion prediction, and the method comprises the following steps: acquiring multi-source data of a case to perform preprocessing and generating a fusion data table; constructing a feature library corresponding to the case according to the fusion data table; outputting a prediction result corresponding to the current case according to the feature library and the current case based on a two-stage prediction model; generating a business strategy corresponding to the current case based on the prediction result, and updating the feature library and the two-stage prediction model according to the feedback result of the business strategy. It can be seen that the repayment request decision of the application can realize the accurate quantitative prediction of litigation risks through the multi-source data fusion and the dynamic closed-loop two-stage prediction mechanism.

Description

Technical Field

[0001] This invention relates to the field of internet finance technology, and in particular to a repayment request decision-making method and system based on multi-source data fusion prediction. Background Technology

[0002] In the business scenarios of internet finance, the only way to make repayment requests to defaulting debtors is through litigation. This process requires quantifying the case analysis results of the entire process of filing a lawsuit, reviewing the case, and enforcing the judgment, and generating predictable judicial enforcement results to provide legal personnel with feasible litigation decisions. Among these, how to quantify and predict the litigation risks of repayment requests has become a key technical issue.

[0003] Currently, existing technologies first rely on legal professionals' subjective assessment of case materials to make experience-based judgments on litigation risks. However, these judgments are susceptible to individual cognitive biases and are difficult to standardize. Therefore, existing technologies introduce rule engines to filter case analysis results. However, rule engines ignore the complexity of legal elements, resulting in poor quantitative predictions. Furthermore, existing technologies suffer from a single data source and fragmented data at different stages, making it impossible to form an effective closed-loop feedback mechanism for case analysis results. On the other hand, they do not consider the impact of dynamic factors such as regional judicial differences, policy changes, and business adaptability on actual litigation risks. Consequently, they cannot effectively predict the success rate before litigation and the feasibility of enforcement after litigation, and therefore cannot provide a highly accurate and interpretable risk assessment.

[0004] It is evident that existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a repayment request decision-making method and system based on multi-source data fusion prediction, which can realize accurate quantitative prediction of litigation risks in repayment request decisions.

[0006] To address the aforementioned technical problems, the first aspect of this invention discloses a repayment request decision-making method based on multi-source data fusion prediction, the method comprising: Acquire multi-source data of the case, preprocess it, and generate a fused data table; A feature database corresponding to the case is constructed based on the fused data table; Based on the two-stage prediction model, the prediction result corresponding to the current case is output according to the feature library and the current case. Based on the prediction results, a business strategy corresponding to the current case is generated, and the feature library and the two-stage prediction model are updated according to the feedback results of the business strategy.

[0007] As an optional implementation, in the first aspect of the present invention, the multi-source data includes structured data, unstructured data, and judicial data, and the data sources of the multi-source data include financial business systems, judicial disclosure platforms, third-party credit reporting platforms, and regional judicial indicator databases.

[0008] As an optional implementation, in the first aspect of the present invention, acquiring multi-source data of a case and preprocessing it to generate a fused data table includes: Multi-source data collection is performed by connecting to the data source through a data pipeline. For unstructured data from the collected multi-source data, its image pages are input into the visual language model, and the structured text and target fields corresponding to the unstructured data are output. For image pages with image quality below the resolution threshold, text recognition is performed using OCR, and then the table structure and paragraph structure of the image page are restored through document analysis. The structured text and target fields corresponding to the unstructured data are output, and the output results of the unstructured data are stored in the database after being verified by structured rules. All multi-source data are aligned by case ID to form a fused data table with debtor-case granularity. The structured data in the multi-source data is written into the fused data table after being validated by structured rules, and the judicial data is written into the fused data table based on the granularity determined by case similarity.

[0009] As an optional implementation, in the first aspect of the present invention, constructing a feature library corresponding to a case based on the fused data table includes: Extract the structured text and business features corresponding to the target fields of any case in the fused data table, and write the business features as static features into the feature library; Periodically extract the judicial features corresponding to the judicial data associated with any case in the fused data table, and update the judicial features incrementally as dynamic features and write them into the feature library; For any historical case at any historical moment, a feature snapshot is generated by timestamp based on the static and dynamic features. The feature snapshot is used for consistency training of the two-stage prediction model and attribution analysis of the prediction results. A feature library is constructed based on the static features, dynamic features, and feature snapshots, wherein the feature library satisfies that the effective feature coverage is greater than the fusion threshold, the effective feature coverage is the ratio of the total number of dimensions of features obtainable from the data source to the total number of dimensions of features written into the feature library, and the fusion threshold is a preset requirement value for the complete utilization rate of the multi-source data.

[0010] As an optional implementation, in the first aspect of the present invention, the two-stage prediction model is a composite model consisting of a first-stage identification model and a second-stage prediction model. The identification model is trained using a training dataset of multiple historical cases and a corresponding training dataset of identification results containing case element labels, legal element labels, and dispute focus labels. The prediction model is trained using a training dataset of multiple historical case features and a corresponding training dataset of prediction results containing real judgment labels of historical cases. Furthermore, for the prediction model within the same training period, semantic vectors are generated based on the cases input to the recognition model. The judgment results of the top-K similar cases are retrieved from the feature library according to the semantic vectors and injected into the prediction model as auxiliary features during the second stage of training.

[0011] As an optional implementation, in the first aspect of the present invention, based on a two-stage prediction model, according to the feature library and the current case, the prediction result corresponding to the current case is output, including: The current case is input into the two-stage prediction model, and the case elements, legal requirements, and points of contention of the current case are extracted by the first-stage identification model. Based on the case elements, legal requirements, and disputed issues, the semantic features of the current case are generated. Based on the semantic features, semantic similarity is retrieved in the feature database to obtain the corresponding business features and judicial features. The semantic features of the current case, along with the retrieved business and judicial features, are concatenated to output the input vector for the prediction model in the second stage. The input vector is used as the input to the prediction model in the second stage, and the feature snapshot of the winning probability and prediction basis corresponding to the current case is output. The expected value of the current case is calculated based on the winning probability. Based on the probability of winning, feature snapshot, and expected value, the prediction result corresponding to the current case is output.

[0012] As an optional implementation, in the first aspect of the invention, calculating the expected value of the current case based on the probability of winning includes: The expected net recovery is obtained by multiplying the probability of winning the case, the recovery rate, and the amount in dispute of the current case. The expected value is obtained by calculating the difference between the expected net recovery and the litigation costs of the current case, wherein the enforcement recovery rate is estimated based on historical enforcement cases.

[0013] As an optional implementation, in the first aspect of the invention, updating the feature library and the two-stage prediction model based on the feedback results of the business strategy includes: Collect the actual judgment results corresponding to the implemented business strategies within a unit period; For any given case, calculate the prediction error between the predicted outcome and the actual judgment outcome. The average absolute error of adjacent unit periods is calculated based on the number of cases within a unit week and the prediction error of a single case. Error case samples whose average absolute error shows a monotonically increasing trend within adjacent unit periods are added to the training dataset of the two-stage prediction model, and feature snapshots of the error case samples associated with the prediction error attribution analysis are written into the feature library. The two-stage prediction model is fine-tuned based on the updated training dataset and feature library.

[0014] A second aspect of this invention discloses a repayment request decision system based on multi-source data fusion prediction, the system comprising: The acquisition module is used to acquire multi-source data of cases, preprocess it, and generate a fused data table. The fusion module is used to construct a feature database corresponding to the case based on the fusion data table; The prediction module is used to output the prediction result corresponding to the current case based on the feature library and the current case, according to the two-stage prediction model. The decision-making module is used to generate a business strategy corresponding to the current case based on the prediction results, and update the feature library and the two-stage prediction model according to the feedback results of the business strategy.

[0015] As an optional implementation, in the second aspect of the present invention, the multi-source data includes structured data, unstructured data, and judicial data, and the data sources of the multi-source data include financial business systems, judicial disclosure platforms, third-party credit reporting platforms, and regional judicial indicator databases.

[0016] As an optional implementation, in a second aspect of the invention, acquiring multi-source case data and preprocessing it to generate a fused data table includes: Multi-source data collection is performed by connecting to the data source through a data pipeline. For unstructured data from the collected multi-source data, its image pages are input into the visual language model, and the structured text and target fields corresponding to the unstructured data are output. For image pages with image quality below the resolution threshold, text recognition is performed using OCR, and then the table structure and paragraph structure of the image page are restored through document analysis. The structured text and target fields corresponding to the unstructured data are output, and the output results of the unstructured data are stored in the database after being verified by structured rules. All multi-source data are aligned by case ID to form a fused data table with debtor-case granularity. The structured data in the multi-source data is written into the fused data table after being validated by structured rules, and the judicial data is written into the fused data table based on the granularity determined by case similarity.

[0017] As an optional implementation, in a second aspect of the invention, constructing a feature library corresponding to a case based on the fused data table includes: Extract the structured text and business features corresponding to the target fields of any case in the fused data table, and write the business features as static features into the feature library; Periodically extract the judicial features corresponding to the judicial data associated with any case in the fused data table, and update the judicial features incrementally as dynamic features and write them into the feature library; For any historical case at any historical moment, a feature snapshot is generated by timestamp based on the static and dynamic features. The feature snapshot is used for consistency training of the two-stage prediction model and attribution analysis of the prediction results. A feature library is constructed based on the static features, dynamic features, and feature snapshots, wherein the feature library satisfies that the effective feature coverage is greater than the fusion threshold, the effective feature coverage is the ratio of the total number of dimensions of features obtainable from the data source to the total number of dimensions of features written into the feature library, and the fusion threshold is a preset requirement value for the complete utilization rate of the multi-source data.

[0018] As an optional implementation, in the second aspect of the present invention, the two-stage prediction model is a composite model consisting of a first-stage identification model and a second-stage prediction model. The identification model is trained using a training dataset of multiple historical cases and a corresponding training dataset of identification results containing case element labels, legal element labels, and dispute focus labels. The prediction model is trained using a training dataset of multiple historical case features and a corresponding training dataset of prediction results containing real judgment labels of historical cases. Furthermore, for the prediction model within the same training period, semantic vectors are generated based on the cases input to the recognition model. The judgment results of the top-K similar cases are retrieved from the feature library according to the semantic vectors and injected into the prediction model as auxiliary features during the second stage of training.

[0019] As an optional implementation, in a second aspect of the invention, based on a two-stage prediction model, a prediction result corresponding to the current case is output according to the feature library and the current case, including: The current case is input into the two-stage prediction model, and the case elements, legal requirements, and points of contention of the current case are extracted by the first-stage identification model. Based on the case elements, legal requirements, and disputed issues, the semantic features of the current case are generated. Based on the semantic features, semantic similarity is retrieved in the feature database to obtain the corresponding business features and judicial features. The semantic features of the current case, along with the retrieved business and judicial features, are concatenated to output the input vector for the prediction model in the second stage. The input vector is used as the input to the prediction model in the second stage, and the feature snapshot of the winning probability and prediction basis corresponding to the current case is output. The expected value of the current case is calculated based on the winning probability. Based on the probability of winning, feature snapshot, and expected value, the prediction result corresponding to the current case is output.

[0020] As an optional implementation, in a second aspect of the invention, calculating the expected value of the current case based on the probability of winning includes: The expected net recovery is obtained by multiplying the probability of winning the case, the recovery rate, and the amount in dispute of the current case. The expected value is obtained by calculating the difference between the expected net recovery and the litigation costs of the current case, wherein the enforcement recovery rate is estimated based on historical enforcement cases.

[0021] As an optional implementation, in a second aspect of the invention, updating the feature library and the two-stage prediction model based on the feedback results of the business strategy includes: Collect the actual judgment results corresponding to the implemented business strategies within a unit period; For any given case, calculate the prediction error between the predicted outcome and the actual judgment outcome. The average absolute error of adjacent unit periods is calculated based on the number of cases within a unit week and the prediction error of a single case. Error case samples whose average absolute error shows a monotonically increasing trend within adjacent unit periods are added to the training dataset of the two-stage prediction model, and feature snapshots of the error case samples associated with the prediction error attribution analysis are written into the feature library. The two-stage prediction model is fine-tuned based on the updated training dataset and feature library.

[0022] A third aspect of this invention discloses another repayment request decision system based on multi-source data fusion prediction, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the repayment request decision method based on multi-source data fusion prediction disclosed in the first aspect of the present invention.

[0023] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the repayment request decision-making method based on multi-source data fusion prediction disclosed in the first aspect of the present invention.

[0024] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention addresses the issues of single data sources and fragmented stages by generating structured fusion data tables through multi-source data fusion. It introduces static case features and dynamic features of judicial changes into a feature library, constructing a multi-dimensional feature engineering system across stages. A two-stage prediction model analyzes cases in the first stage and combines case analysis and feature engineering to predict and quantify litigation risks in the second stage, overcoming the bottleneck of identifying complex legal elements and providing precise quantitative support for subsequent business decisions. A dynamic feedback loop feeds the results of business strategy execution back to the feature library, triggering retraining of the two-stage model, resolving prediction inaccuracies and missing loops, improving long-term prediction accuracy, and providing interpretability for business strategies to eliminate decision-making blind spots. This forms a dynamic closed loop of data fusion, two-stage prediction, and dynamic feedback, enabling intelligent decision-making regarding litigation risks. It solves the problems of subjective experience dependence, superficial legal elements, data fragmentation, and lack of dynamic factors in internet finance litigation scenarios, promoting high-quality development of litigation management for defaulting debtors' repayment requests from experience-driven to intelligent decision-making. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating a repayment request decision-making method based on multi-source data fusion prediction disclosed in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of a repayment request decision system based on multi-source data fusion prediction disclosed in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of another repayment request decision system based on multi-source data fusion prediction disclosed in an embodiment of the present invention. Detailed Implementation

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0030] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0031] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0032] This invention discloses a repayment request decision-making method and system based on multi-source data fusion prediction. It generates a structured fusion data table by fusing multi-source data, solving the problems of single data and stage-specific fragmentation. A feature library is used to introduce static features of cases and dynamic features of judicial changes, constructing a multi-dimensional feature engineering system across stages. A two-stage prediction model analyzes cases in the first stage, and combines case analysis and feature engineering prediction to quantify litigation risks in the second stage, overcoming the bottleneck of identifying complex legal elements and providing accurate quantitative support for subsequent business decisions. A dynamic feedback loop feeds the results of business strategy execution back to the feature library, triggering retraining of the two-stage model, solving the problems of prediction inaccuracies and missing loops, improving the accuracy of long-term predictions, and providing interpretability for business strategies to eliminate decision-making blind spots. This forms a dynamic closed loop of data fusion-two-stage prediction-dynamic feedback, thereby achieving intelligent decision-making on litigation risks. It addresses the problems of subjective experience dependence, superficial legal elements, data fragmentation, and lack of dynamic factors in internet finance litigation scenarios, promoting the high-quality development of litigation management of defaulting debtors' repayment requests from experience-driven to intelligent decision-making. Detailed explanations follow.

[0033] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a repayment request decision-making method based on multi-source data fusion prediction disclosed in an embodiment of the present invention. Figure 1 The described repayment request decision-making method based on multi-source data fusion prediction can be applied to data processing systems / data processing equipment / data processing servers (wherein, the server includes local processing servers or cloud processing servers). Figure 1 As shown, the repayment request decision-making method based on multi-source data fusion prediction may include the following operations: 101. Obtain multi-source data of the case, perform preprocessing, and generate a fused data table.

[0034] Optionally, multi-source data includes structured data, unstructured data, and judicial data. Structured data includes, but is not limited to, contract terms, repayment records, and credit reports. Unstructured data includes, but is not limited to, complaints, answers, and court transcripts. Judicial data includes, but is not limited to, historical precedents, enforcement termination rates, leads on the assets of judgment debtors, case closure cycles within court jurisdictions, regional judicial efficiency indices, trends in winning similar cases, and defendant performance ratings. It is understood that unstructured data consists of scanned image pages, which are not specific to any particular page, paragraph, or table, and require preprocessing to extract the text data. Judicial data is dynamic and varies geographically and temporally. Judicial data from different times and regions has a dynamic impact on the litigation outcomes of actual cases, therefore, the impact of this data needs to be quantified.

[0035] Optionally, the data sources for multi-source data include financial business systems, judicial disclosure platforms, third-party credit reporting platforms, and regional judicial indicator databases. Financial business systems can obtain, but are not limited to, basic debtor information, contract texts, repayment records, and repayment reminders. Judicial disclosure platforms, such as the China Judgments Online or the Enforcement Information Disclosure Website, can obtain, but are not limited to, historical litigation records, enforcement status, and information on cases terminated due to lack of assets. Third-party credit reporting platforms can obtain, but are not limited to, personal credit scores, corporate credit scores, and related risk events. Regional judicial indicator databases can obtain, but are not limited to, dynamic statistical indicators such as the recent case completion rate, average trial period, and success rate of similar cases in various intermediate people's courts. It is understood that, except for financial business systems, all of the above data sources are publicly available. The multi-source data obtained from financial business systems consists of case data of defaulting debtors, and the personal information portions of the obtained multi-source data have been anonymized.

[0036] 102. Construct a feature database corresponding to the case based on the fused data table.

[0037] 103. Based on the two-stage prediction model, output the prediction result corresponding to the current case according to the feature library and the current case.

[0038] 104. Generate a business strategy corresponding to the current case based on the prediction results, and update the feature library and the two-stage prediction model according to the feedback results of the business strategy.

[0039] As can be seen, the above-mentioned embodiments of the invention generate structured fused data tables through multi-source data fusion, solving the problems of single data and stage fragmentation. By introducing static features of cases and dynamic features of judicial changes through a feature library, a multi-dimensional feature engineering system across stages is constructed. A two-stage prediction model analyzes cases in the first stage and combines case analysis and feature engineering to predict and quantify litigation risks in the second stage, breaking through the bottleneck of identifying the complexity of legal elements and providing strong support for accurate quantification for subsequent business decisions. Through a dynamic feedback loop, the results of business strategy execution are fed back to the feature library, triggering the retraining of the two-stage model, solving the problems of prediction inaccuracy and lack of closed loop, improving the accuracy of long-term prediction, and providing interpretability for business strategies to eliminate decision blind spots. This forms a dynamic closed loop of data fusion-two-stage prediction-dynamic feedback, thereby realizing intelligent decision-making on litigation risks. It solves the problems of subjective experience dependence, superficial legal elements, data fragmentation, and lack of dynamic factors in Internet finance litigation scenarios, and promotes the high-quality development of litigation management of defaulting debtors' repayment requests from experience-driven to intelligent decision-making.

[0040] As an optional embodiment, the step of obtaining multi-source data of the case and preprocessing it to generate a fused data table in the above steps includes: Multi-source data collection is performed by connecting to the data source through a data pipeline. For unstructured data from the collected multi-source data, its image pages are input into the visual language model, and the structured text and target fields corresponding to the unstructured data are output. For image pages with image quality below the resolution threshold, text recognition is performed using OCR, and then the table structure and paragraph structure of the image page are restored through document analysis. The structured text and target fields corresponding to the unstructured data are output, and the output results of the unstructured data are stored in the database after being verified by structured rules. All multi-source data are aligned by case ID to form a fused data table with debtor-case granularity. The structured data in the multi-source data is written into the fused data table after being validated by structured rules, and the judicial data is written into the fused data table based on the granularity determined by case similarity.

[0041] Optionally, the data pipeline can be a FastAPI data pipeline or a data pipeline built with the Apache NiFi orchestration engine. The data pipeline obtains multi-source data by uniformly accessing the above-mentioned data sources, and this application does not impose any restrictions.

[0042] Optionally, unstructured data refers to data extracted from scanned copies of unstructured documents such as paper complaints, judgments, and repayment agreements, which therefore require visual recognition processing.

[0043] Optionally, structured rules are business rules that constrain structured data, such as amount format and ID card verification code, which can ensure the reliability of features in the subsequent feature database.

[0044] Optionally, judicial data can be written to different fusion data table locations at multiple granularities for multiple suitable cases.

[0045] As can be seen, through the above optional embodiments, data coverage and access efficiency are improved by pipelined access to multi-source data, fundamentally addressing the bottlenecks of data silos and fragmentation. By converting unstructured data of different image qualities into structured text, the efficiency and accuracy of text processing are improved. By aligning multi-source data by case ID to construct a fused data table, the accuracy of case granularity alignment and the completeness of the fused data table are improved, reducing the impact of single data sources and fragmented data at different stages on risk assessment. Through the intelligent fusion and structured conversion mechanism of multi-source data, the problems of single data sources, inefficient unstructured processing, and lack of case granularity alignment in Internet finance litigation scenarios are solved, achieving a standardized upgrade of automated and precise fusion of multi-source data.

[0046] As an optional embodiment, the step of constructing a feature library corresponding to the case based on the fused data table in the above steps includes: Extract the structured text and business features corresponding to the target fields of any case in the fused data table, and write the business features as static features into the feature library; Periodically extract the judicial features corresponding to the judicial data associated with any case in the fused data table, and update the judicial features incrementally as dynamic features and write them into the feature library; For any historical case at any historical moment, a feature snapshot is generated by timestamp based on the static and dynamic features. The feature snapshot is used for consistency training of the two-stage prediction model and attribution analysis of the prediction results. A feature library is constructed based on the static features, dynamic features, and feature snapshots, wherein the feature library satisfies that the effective feature coverage is greater than the fusion threshold, the effective feature coverage is the ratio of the total number of dimensions of features obtainable from the data source to the total number of dimensions of features written into the feature library, and the fusion threshold is a preset requirement value for the complete utilization rate of the multi-source data.

[0047] Optional, static features include, but are not limited to, business features such as the contract signing date or the initial principal amount of the case.

[0048] Optional, dynamic features include, but are not limited to, dynamically changing external features of the case, such as the number of times the defendant is currently subject to enforcement or the case closure rate of the court in the previous month.

[0049] As can be seen, through the above optional embodiments, by extracting structured text from the fused data table to generate static features, which are then written into the feature library after rule verification, the impact of shallow legal elements on risk assessment is reduced, and the coverage of legal elements and the completeness of the feature library are improved. By periodically extracting dynamic feature increments from judicial data and updating the feature library, the impact of the quantification of missing dynamic factors on the bias of risk assessment prediction is reduced, and the dynamic adaptability and timeliness of the prediction results are improved. By constructing feature snapshots and applying them to model training and attribution analysis, a closed-loop association of historical data is achieved, improving model consistency and the interpretability of decisions. By defining an effective feature coverage rate, the data integrity of the feature library is ensured, data redundancy is reduced, and data utilization is improved, thereby providing accurate and usable data support for the decision-making system.

[0050] As an optional embodiment, in the above steps, the two-stage prediction model is a composite model consisting of a first-stage identification model and a second-stage prediction model. The identification model is trained using a training dataset of multiple historical cases and a corresponding training dataset of identification results containing case element labels, legal element labels, and dispute focus labels. The prediction model is trained using a training dataset of multiple historical case features and a corresponding training dataset of prediction results containing real judgment labels of historical cases. Furthermore, for the prediction model within the same training period, semantic vectors are generated based on the cases input to the recognition model. The judgment results of the top-K similar cases are retrieved from the feature library according to the semantic vectors and injected into the prediction model as auxiliary features during the second stage of training.

[0051] Optionally, the recognition model can be a language model, such as BERT, RoBERTa, or Qwen, etc., and this application does not impose any restrictions.

[0052] Specifically, the prediction model can be a neural network model or an ensemble model, which can be a time series variant of LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), or Transformer, or an ensemble model of XGBoost and LightGBM. This invention does not limit the model.

[0053] As can be seen, through the above optional embodiments, the accuracy of legal element identification and risk quantification prediction is improved through precise collaboration of the two-stage model, the adaptability to the dynamic judicial environment is improved through knowledge transfer from similar cases, the fine-tuning training loop of knowledge accumulation-transfer-optimization is realized through the dynamic fusion of auxiliary features, and the problems of incomplete legal element identification, weak generalization ability of prediction model and slow adaptation to dynamic judicial environment in Internet finance litigation scenarios are solved through the collaboration of the two-stage model and the knowledge transfer mechanism of similar cases. It realizes an intelligent closed loop from single model prediction to identification-reasoning-transfer.

[0054] As an optional embodiment, the step above, based on the two-stage prediction model, outputs the prediction result corresponding to the current case according to the feature library and the current case, including: The current case is input into the two-stage prediction model, and the case elements, legal requirements, and points of contention of the current case are extracted by the first-stage identification model. Based on the case elements, legal requirements, and disputed issues, the semantic features of the current case are generated. Based on the semantic features, semantic similarity is retrieved in the feature database to obtain the corresponding business features and judicial features. The semantic features of the current case, along with the retrieved business and judicial features, are concatenated to output the input vector for the prediction model in the second stage. The input vector is used as the input to the prediction model in the second stage, and the feature snapshot of the winning probability and prediction basis corresponding to the current case is output. The expected value of the current case is calculated based on the winning probability. Based on the probability of winning, feature snapshot, and expected value, the prediction result corresponding to the current case is output.

[0055] Optionally, case elements include, but are not limited to, the validity of the contract in the current case, the status of the statute of limitations, a description of the breach of contract, and grounds for defense, which are not limited in this application.

[0056] As an example, the legal requirements may include whether the provisions of Article 675 of the Civil Code regarding the repayment of loans are met; this application does not specify such requirements.

[0057] Optionally, the focus of the dispute may be whether a valid contract has been signed or whether there is a repayment ability, etc., which are not limited in this application.

[0058] As can be seen, through the above optional embodiments, the first stage automatically extracts case elements, legal requirements, and points of contention, generating structured semantic features to reduce the superficial impact of element identification and improve identification accuracy and coverage. By retrieving TOP-K similar cases from the feature library using semantic features, and injecting business features and judicial features as auxiliary inputs into the prediction model, the adaptability and prediction robustness of the dynamic judicial environment are improved. By concatenating the current case semantic features with the search results to generate an input vector, the second-stage prediction model outputs the probability of victory and feature snapshots. Multi-feature fusion achieves a dual improvement in prediction accuracy and interpretability. By dynamically quantifying expected value, it drives the precise formulation of business strategies. Through the prediction mechanism of two-stage model collaboration and semantic knowledge transfer, the problems of incomplete legal element identification, uninterpretable prediction results, and slow adaptation to the dynamic judicial environment in Internet finance litigation scenarios are solved, realizing intelligent decision support from fuzzy experience judgment to accurate and interpretable prediction.

[0059] As an optional embodiment, the step described above, calculating the expected value of the current case based on the probability of winning, includes: The expected net recovery is obtained by multiplying the probability of winning the case, the recovery rate, and the amount in dispute of the current case. The expected value is obtained by calculating the difference between the expected net recovery and the litigation costs of the current case, wherein the enforcement recovery rate is estimated based on historical enforcement cases.

[0060] As an optional embodiment, the step above, updating the feature library and the two-stage prediction model based on the feedback results of the business strategy, includes: Collect the actual judgment results corresponding to the implemented business strategies within a unit period; For any given case, calculate the prediction error between the predicted outcome and the actual judgment outcome. The average absolute error of adjacent unit periods is calculated based on the number of cases within a unit week and the prediction error of a single case. Error case samples whose average absolute error shows a monotonically increasing trend within adjacent unit periods are added to the training dataset of the two-stage prediction model, and feature snapshots of the error case samples associated with the prediction error attribution analysis are written into the feature library. The two-stage prediction model is fine-tuned based on the updated training dataset and feature library.

[0061] As can be seen, through the above optional embodiments, by calculating the prediction error within a unit period and screening erroneous cases according to the trend of average absolute error, a dynamic error monitoring mechanism is formed to reduce the impact of model performance degradation and improve the accuracy of long-term prediction. Only monotonically increasing erroneous cases are screened to avoid noise data interfering with training, thereby improving the effectiveness of training data and thus improving training efficiency. By writing the attribution feature snapshots of erroneous cases into the feature library, a decision interpretability closed loop is achieved, reducing the decision misjudgment rate and improving attribution interpretability. By fine-tuning the two-stage model based on the updated training dataset, historical knowledge is retained to build a prediction-execution-learning closed loop. Through a model self-evolution mechanism driven by dynamic feedback, the problems of model performance degradation, slow response to judicial changes, and low data utilization efficiency in Internet finance litigation scenarios are solved, realizing the dynamic evolution from static model application to intelligent self-iterative system.

[0062] In summary, the technical solutions disclosed in the embodiments of the present invention have the following advantages: 1. Reduce costs and increase efficiency: Reduce invalid litigation and lower sunk costs such as attorney fees and litigation costs.

[0063] 2. Precise decision-making: Transform fuzzy experience into quantitative indicators to support the optimal allocation of business resources for repayment requests.

[0064] 3. Closed-loop feedback: The actual judgment and execution results are automatically fed back to the model for continuous iterative optimization.

[0065] 4. Compliance and controllability: All forecasts are traceable and explainable, meeting the requirements of financial regulatory audits.

[0066] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of a repayment request decision-making system based on multi-source data fusion prediction, as disclosed in an embodiment of the present invention. Figure 2 The described repayment request decision-making system based on multi-source data fusion prediction can be applied to data processing systems / data processing equipment / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 2 As shown, the repayment request decision system based on multi-source data fusion prediction may include: The acquisition module 201 is used to acquire multi-source data of the case, perform preprocessing, and generate a fused data table. The fusion module 202 is used to construct a feature database corresponding to the case based on the fusion data table; Prediction module 203 is used to output the prediction result corresponding to the current case based on the feature library and the current case, according to the two-stage prediction model; The decision module 204 is used to generate a business strategy corresponding to the current case based on the prediction results, and update the feature library and the two-stage prediction model according to the feedback results of the business strategy.

[0067] As can be seen, the above-mentioned embodiments of the invention generate structured fused data tables through multi-source data fusion, solving the problems of single data and stage fragmentation. By introducing static features of cases and dynamic features of judicial changes through a feature library, a multi-dimensional feature engineering system across stages is constructed. A two-stage prediction model analyzes cases in the first stage and combines case analysis and feature engineering to predict and quantify litigation risks in the second stage, breaking through the bottleneck of identifying the complexity of legal elements and providing strong support for accurate quantification for subsequent business decisions. Through a dynamic feedback loop, the results of business strategy execution are fed back to the feature library, triggering the retraining of the two-stage model, solving the problems of prediction inaccuracy and lack of closed loop, improving the accuracy of long-term prediction, and providing interpretability for business strategies to eliminate decision blind spots. This forms a dynamic closed loop of data fusion-two-stage prediction-dynamic feedback, thereby realizing intelligent decision-making on litigation risks. It solves the problems of subjective experience dependence, superficial legal elements, data fragmentation, and lack of dynamic factors in Internet finance litigation scenarios, and promotes the high-quality development of litigation management of defaulting debtors' repayment requests from experience-driven to intelligent decision-making.

[0068] As an optional embodiment, the multi-source data includes structured data, unstructured data, and judicial data, and the data sources of the multi-source data include financial business systems, judicial disclosure platforms, third-party credit reporting platforms, and regional judicial indicator databases.

[0069] As an optional embodiment, the process of acquiring multi-source case data, preprocessing it, and generating a fused data table includes: Multi-source data collection is performed by connecting to the data source through a data pipeline. For unstructured data from the collected multi-source data, its image pages are input into the visual language model, and the structured text and target fields corresponding to the unstructured data are output. For image pages with image quality below the resolution threshold, text recognition is performed using OCR, and then the table structure and paragraph structure of the image page are restored through document analysis. The structured text and target fields corresponding to the unstructured data are output, and the output results of the unstructured data are stored in the database after being verified by structured rules. All multi-source data are aligned by case ID to form a fused data table with debtor-case granularity. The structured data in the multi-source data is written into the fused data table after being validated by structured rules, and the judicial data is written into the fused data table based on the granularity determined by case similarity.

[0070] As can be seen, through the above optional embodiments, data coverage and access efficiency are improved by pipelined access to multi-source data, fundamentally addressing the bottlenecks of data silos and fragmentation. By converting unstructured data of different image qualities into structured text, the efficiency and accuracy of text processing are improved. By aligning multi-source data by case ID to construct a fused data table, the accuracy of case granularity alignment and the completeness of the fused data table are improved, reducing the impact of single data sources and fragmented data at different stages on risk assessment. Through the intelligent fusion and structured conversion mechanism of multi-source data, the problems of single data sources, inefficient unstructured processing, and lack of case granularity alignment in Internet finance litigation scenarios are solved, achieving a standardized upgrade of automated and precise fusion of multi-source data.

[0071] As an optional embodiment, a feature library corresponding to the case is constructed based on the fused data table, including: Extract the structured text and business features corresponding to the target fields of any case in the fused data table, and write the business features as static features into the feature library; Periodically extract the judicial features corresponding to the judicial data associated with any case in the fused data table, and update the judicial features incrementally as dynamic features and write them into the feature library; For any historical case at any historical moment, a feature snapshot is generated by timestamp based on the static and dynamic features. The feature snapshot is used for consistency training of the two-stage prediction model and attribution analysis of the prediction results. A feature library is constructed based on the static features, dynamic features, and feature snapshots, wherein the feature library satisfies that the effective feature coverage is greater than the fusion threshold, the effective feature coverage is the ratio of the total number of dimensions of features obtainable from the data source to the total number of dimensions of features written into the feature library, and the fusion threshold is a preset requirement value for the complete utilization rate of the multi-source data.

[0072] As can be seen, through the above optional embodiments, by extracting structured text from the fused data table to generate static features, which are then written into the feature library after rule verification, the impact of shallow legal elements on risk assessment is reduced, and the coverage of legal elements and the completeness of the feature library are improved. By periodically extracting dynamic feature increments from judicial data and updating the feature library, the impact of the quantification of missing dynamic factors on the bias of risk assessment prediction is reduced, and the dynamic adaptability and timeliness of the prediction results are improved. By constructing feature snapshots and applying them to model training and attribution analysis, a closed-loop association of historical data is achieved, improving model consistency and the interpretability of decisions. By defining an effective feature coverage rate, the data integrity of the feature library is ensured, data redundancy is reduced, and data utilization is improved, thereby providing accurate and usable data support for the decision-making system.

[0073] As an optional embodiment, the two-stage prediction model is a composite model consisting of a first-stage identification model and a second-stage prediction model. The identification model is trained using a training dataset of multiple historical cases and a corresponding training dataset of identification results containing case element labels, legal element labels, and dispute focus labels. The prediction model is trained using a training dataset of multiple historical case features and a corresponding training dataset of prediction results containing real judgment labels of historical cases. Furthermore, for the prediction model within the same training period, semantic vectors are generated based on the cases input to the recognition model. The judgment results of the top-K similar cases are retrieved from the feature library according to the semantic vectors and injected into the prediction model as auxiliary features during the second stage of training.

[0074] As can be seen, through the above optional embodiments, the accuracy of legal element identification and risk quantification prediction is improved through precise collaboration of the two-stage model, the adaptability to the dynamic judicial environment is improved through knowledge transfer from similar cases, the fine-tuning training loop of knowledge accumulation-transfer-optimization is realized through the dynamic fusion of auxiliary features, and the problems of incomplete legal element identification, weak generalization ability of prediction model and slow adaptation to dynamic judicial environment in Internet finance litigation scenarios are solved through the collaboration of the two-stage model and the knowledge transfer mechanism of similar cases. It realizes an intelligent closed loop from single model prediction to identification-reasoning-transfer.

[0075] As an optional embodiment, based on the two-stage prediction model, according to the feature library and the current case, the prediction result corresponding to the current case is output, including: The current case is input into the two-stage prediction model, and the case elements, legal requirements, and points of contention of the current case are extracted by the first-stage identification model. Based on the case elements, legal requirements, and disputed issues, the semantic features of the current case are generated. Based on the semantic features, semantic similarity is retrieved in the feature database to obtain the corresponding business features and judicial features. The semantic features of the current case, along with the retrieved business and judicial features, are concatenated to output the input vector for the prediction model in the second stage. The input vector is used as the input to the prediction model in the second stage, and the feature snapshot of the winning probability and prediction basis corresponding to the current case is output. The expected value of the current case is calculated based on the winning probability. Based on the probability of winning, feature snapshot, and expected value, the prediction result corresponding to the current case is output.

[0076] As can be seen, through the above optional embodiments, the first stage automatically extracts case elements, legal requirements, and points of contention, generating structured semantic features to reduce the superficial impact of element identification and improve identification accuracy and coverage. By retrieving TOP-K similar cases from the feature library using semantic features, and injecting business features and judicial features as auxiliary inputs into the prediction model, the adaptability and prediction robustness of the dynamic judicial environment are improved. By concatenating the current case semantic features with the search results to generate an input vector, the second-stage prediction model outputs the probability of victory and feature snapshots. Multi-feature fusion achieves a dual improvement in prediction accuracy and interpretability. By dynamically quantifying expected value, it drives the precise formulation of business strategies. Through the prediction mechanism of two-stage model collaboration and semantic knowledge transfer, the problems of incomplete legal element identification, uninterpretable prediction results, and slow adaptation to the dynamic judicial environment in Internet finance litigation scenarios are solved, realizing intelligent decision support from fuzzy experience judgment to accurate and interpretable prediction.

[0077] As an optional embodiment, calculating the expected value of the current case based on the probability of winning includes: The expected net recovery is obtained by multiplying the probability of winning the case, the recovery rate, and the amount in dispute of the current case. The expected value is obtained by calculating the difference between the expected net recovery and the litigation costs of the current case, wherein the enforcement recovery rate is estimated based on historical enforcement cases.

[0078] As an optional embodiment, updating the feature library and the two-stage prediction model based on the feedback results of the business strategy includes: Collect the actual judgment results corresponding to the implemented business strategies within a unit period; For any given case, calculate the prediction error between the predicted outcome and the actual judgment outcome. The average absolute error of adjacent unit periods is calculated based on the number of cases within a unit week and the prediction error of a single case. Error case samples whose average absolute error shows a monotonically increasing trend within adjacent unit periods are added to the training dataset of the two-stage prediction model, and feature snapshots of the error case samples associated with the prediction error attribution analysis are written into the feature library. The two-stage prediction model is fine-tuned based on the updated training dataset and feature library.

[0079] As can be seen, through the above optional embodiments, by calculating the prediction error within a unit period and screening erroneous cases according to the trend of average absolute error, a dynamic error monitoring mechanism is formed to reduce the impact of model performance degradation and improve the accuracy of long-term prediction. Only monotonically increasing erroneous cases are screened to avoid noise data interfering with training, thereby improving the effectiveness of training data and thus improving training efficiency. By writing the attribution feature snapshots of erroneous cases into the feature library, a decision interpretability closed loop is achieved, reducing the decision misjudgment rate and improving attribution interpretability. By fine-tuning the two-stage model based on the updated training dataset, historical knowledge is retained to build a prediction-execution-learning closed loop. Through a model self-evolution mechanism driven by dynamic feedback, the problems of model performance degradation, slow response to judicial changes, and low data utilization efficiency in Internet finance litigation scenarios are solved, realizing the dynamic evolution from static model application to intelligent self-iterative system.

[0080] Example 3 Please see Figure 3 , Figure 3 This is another repayment request decision system based on multi-source data fusion prediction disclosed in the embodiments of the present invention. Figure 3 The described repayment request decision-making system based on multi-source data fusion prediction is applied in a data processing system / data processing equipment / data processing server (wherein, the server includes a local processing server or a cloud processing server). Figure 3 As shown, the repayment request decision system based on multi-source data fusion prediction may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the repayment request decision method based on multi-source data fusion prediction described in Embodiment 1.

[0081] Example 4 This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps of the repayment request decision method based on multi-source data fusion prediction described in Embodiment 1.

[0082] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the repayment request decision method based on multi-source data fusion prediction described in Embodiment 1.

[0083] The foregoing has described specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily have to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0084] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0085] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0086] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0091] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0092] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0093] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0094] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0095] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0096] Finally, it should be noted that the repayment request decision-making method and system based on multi-source data fusion prediction disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A repayment request decision-making method based on multi-source data fusion prediction, characterized in that, The method includes: Acquire multi-source data of the case, preprocess it, and generate a fused data table; A feature database corresponding to the case is constructed based on the fused data table; Based on the two-stage prediction model, the prediction result corresponding to the current case is output according to the feature library and the current case. Based on the prediction results, a business strategy corresponding to the current case is generated, and the feature library and the two-stage prediction model are updated according to the feedback results of the business strategy.

2. The repayment request decision-making method based on multi-source data fusion prediction according to claim 1, characterized in that, The multi-source data includes structured data, unstructured data, and judicial data. The data sources of the multi-source data include financial business systems, judicial disclosure platforms, third-party credit reporting platforms, and regional judicial indicator databases.

3. The repayment request decision-making method based on multi-source data fusion prediction according to claim 2, characterized in that, The process involves acquiring multi-source case data, preprocessing it to generate a fused data table, including: Multi-source data collection is performed by connecting to the data source through a data pipeline. For unstructured data from the collected multi-source data, its image pages are input into the visual language model, and the structured text and target fields corresponding to the unstructured data are output. For image pages with image quality below the resolution threshold, text recognition is performed using OCR, and then the table structure and paragraph structure of the image page are restored through document analysis. The structured text and target fields corresponding to the unstructured data are output, and the output results of the unstructured data are stored in the database after being verified by structured rules. All multi-source data are aligned by case ID to form a fused data table with debtor-case granularity. The structured data in the multi-source data is written into the fused data table after being validated by structured rules, and the judicial data is written into the fused data table based on the granularity determined by case similarity.

4. The repayment request decision-making method based on multi-source data fusion prediction according to claim 1, characterized in that, Based on the fused data table, a feature database corresponding to the case is constructed, including: Extract the structured text and business features corresponding to the target fields of any case in the fused data table, and write the business features as static features into the feature library; Periodically extract the judicial features corresponding to the judicial data associated with any case in the fused data table, and update the judicial features incrementally as dynamic features and write them into the feature library; For any historical case at any historical moment, a feature snapshot is generated by timestamp based on the static and dynamic features. The feature snapshot is used for consistency training of the two-stage prediction model and attribution analysis of the prediction results. A feature library is constructed based on the static features, dynamic features, and feature snapshots, wherein the feature library satisfies that the effective feature coverage is greater than the fusion threshold, the effective feature coverage is the ratio of the total number of dimensions of features obtainable from the data source to the total number of dimensions of features written into the feature library, and the fusion threshold is a preset requirement value for the complete utilization rate of the multi-source data.

5. The repayment request decision-making method based on multi-source data fusion prediction according to claim 1, characterized in that, The dual-stage prediction model is a composite model consisting of a first-stage identification model and a second-stage prediction model. The identification model is trained using training datasets of multiple historical cases and corresponding training datasets of identification results containing case element labels, legal element labels, and dispute focus labels. The prediction model is trained using training datasets of multiple historical case features and corresponding training datasets of prediction results containing actual judgment labels of historical cases. Furthermore, for the prediction model within the same training period, semantic vectors are generated based on the cases input to the recognition model. The judgment results of the top-K similar cases are retrieved from the feature library according to the semantic vectors and injected into the prediction model as auxiliary features during the second stage of training.

6. The repayment request decision-making method based on multi-source data fusion prediction according to claim 5, characterized in that, Based on the two-stage prediction model, and according to the feature library and the current case, the prediction result corresponding to the current case is output, including: The current case is input into the two-stage prediction model, and the case elements, legal requirements, and points of contention of the current case are extracted by the first-stage identification model. Based on the case elements, legal requirements, and disputed issues, the semantic features of the current case are generated. Based on the semantic features, semantic similarity is retrieved in the feature database to obtain the corresponding business features and judicial features. The semantic features of the current case, along with the retrieved business and judicial features, are concatenated to output the input vector for the prediction model in the second stage. The input vector is used as the input to the prediction model in the second stage, and the feature snapshot of the winning probability and prediction basis corresponding to the current case is output. The expected value of the current case is calculated based on the winning probability. Based on the probability of winning, feature snapshot, and expected value, the prediction result corresponding to the current case is output.

7. The repayment request decision-making method based on multi-source data fusion prediction according to claim 6, characterized in that, The calculation of the expected value of the current case based on the probability of winning includes: The expected net recovery is obtained by multiplying the probability of winning the case, the recovery rate, and the amount in dispute of the current case. The expected value is obtained by calculating the difference between the expected net recovery and the litigation costs of the current case, wherein the enforcement recovery rate is estimated based on historical enforcement cases.

8. The repayment request decision-making method based on multi-source data fusion prediction according to claim 1, characterized in that, The feature library and the two-stage prediction model are updated based on the feedback results of the business strategy, including: Collect the actual judgment results corresponding to the implemented business strategies within a unit period; For any given case, calculate the prediction error between the predicted outcome and the actual judgment outcome. The average absolute error of adjacent unit periods is calculated based on the number of cases within a unit week and the prediction error of a single case. Error case samples whose average absolute error shows a monotonically increasing trend within adjacent unit periods are added to the training dataset of the two-stage prediction model, and feature snapshots of the error case samples associated with the prediction error attribution analysis are written into the feature library. The two-stage prediction model is fine-tuned based on the updated training dataset and feature library.

9. A repayment request decision system based on multi-source data fusion prediction, characterized in that, The system includes: The acquisition module is used to acquire multi-source data of cases, preprocess it, and generate a fused data table. The fusion module is used to construct a feature database corresponding to the case based on the fusion data table; The prediction module is used to output the prediction result corresponding to the current case based on the feature library and the current case, according to the two-stage prediction model. The decision-making module is used to generate a business strategy corresponding to the current case based on the prediction results, and update the feature library and the two-stage prediction model according to the feedback results of the business strategy.

10. A repayment request decision system based on multi-source data fusion prediction, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the repayment request decision method based on multi-source data fusion prediction as described in any one of claims 1-8.

Images