Execution thread recommendation method and system for whole-cycle clearing management

Abstract

The application discloses a kind of execution clue recommendation method and system for whole cycle clean collection management, the method includes: obtaining the multi-source data of the person executed, pre-processing obtains candidate clue set;Based on the preset prediction model, predict the feasibility score corresponding to the structured feature in the candidate clue set;Based on the preset inference model, determine the execution condition of the unstructured feature in the candidate clue set;Based on the preset case retrieval mechanism, according to the feasibility score and execution condition, determine the target recommended clue in the candidate clue set.It can be seen that the present application can realize the execution clue recommendation of high utilization rate in whole cycle clean collection management.

Description

Technical Field

[0001] This invention relates to the field of internet finance technology, and in particular to a method and system for recommending execution leads for full-cycle debt collection management. Background Technology

[0002] In the full-cycle debt collection scenarios of finance, judiciary, and enterprises, cases involving judgment debtors who fail to fulfill their obligations to use non-performing assets for repayment will go through the entire chain cycle of litigation, legal judgment, and enforcement of the judgment. Therefore, the continuity of information, process coherence, and availability of enforcement leads are extremely important throughout the entire debt collection management cycle. Among these, how to accurately discover enforcement leads with high availability has become a key technology.

[0003] Current methods for uncovering enforcement leads have the following shortcomings: First, traditional methods obtain raw data through data procurement or by acquiring publicly available data online. Then, the data value is mined and assessed based on the personal experience of appraisers. This is not only inefficient but also prone to overlooking hidden assets or connections. Second, existing mining solutions suffer from several problems: First, data from the litigation stage is scattered across different platforms, making it difficult to create unified modeling and resulting in severe data silos. Second, lead mining relies on keyword matching or simple rule engines, which cannot identify hidden connections. Third, there is a lack of dynamic feedback, and a closed-loop mechanism from prediction and execution to learning has not yet been formed. Finally, existing lead mining systems often focus on a single stage and lack the ability to structurally integrate and intelligently reason about the entire lifecycle of litigation, legal judgments, and enforcement judgments, leading to low enforcement rates and extended recovery cycles.

[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 method and system for recommending execution leads for full-cycle debt collection management, which can achieve high utilization rate of execution lead recommendation in full-cycle debt collection management.

[0006] To address the aforementioned technical problems, the first aspect of this invention discloses a method for recommending execution leads for full-cycle debt collection management, the method comprising: Obtain multi-source data on the judgment debtor and preprocess it to obtain a candidate clue set; Based on a preset prediction model, the feasibility score corresponding to the structured features in the candidate clue set is predicted; Based on a pre-defined reasoning model, the execution conditions for unstructured features in the candidate clue set are determined; Based on a pre-defined case retrieval mechanism, target recommended leads are determined from the candidate lead set according to the feasibility score and execution conditions.

[0007] As an optional implementation, in the first aspect of the present invention, the multi-source data includes structured data and unstructured data. The structured data includes basic case information, de-identified identity information of the person subject to enforcement, credit records, business registration and change information, and the unstructured data includes judgment documents, court transcripts, enforcement notices, and scanned PDF protocols.

[0008] As an optional implementation, in a first aspect of the invention, the candidate clue set includes structured features for case execution and unstructured features for determining the recommendation degree of the structured features, wherein the unstructured features include legal semantic features for matching similar cases and association graph features for determining implicit assets.

[0009] As an optional implementation, in the first aspect of the present invention, acquiring multi-source data of the person subject to enforcement and preprocessing it to obtain a candidate clue set includes: OCR is used to identify and extract related fields from scanned documents and image text in multi-source data to generate structured text; The structured text from multi-source data and the identified structured text are used to extract fields based on a predefined rule engine to generate structured features; Dense vectors corresponding to legal semantic features in multi-source data are generated based on an embedding model to capture legal terms, behavioral patterns, and points of contention in legal texts. A correlation graph is constructed for identifying implicit assets from information entities associated with the judgment debtor in multi-source data, where nodes in the correlation graph are asset entities and edges are asset relationships. The dense vectors and correlation graphs are labeled as unstructured features and added to the candidate clue set containing the structured features.

[0010] As an optional implementation, in the first aspect of the present invention, based on a preset prediction model, predicting the feasibility score corresponding to the structured features in the candidate clue set includes: The candidate clue set is input into the prediction model, and the candidate clue corresponding to any structured feature is output as the feasibility score of the potential execution clue. The prediction model is an ensemble learning model, including a first gradient boosting tree model and a second gradient boosting tree model. The first gradient boosting tree model is trained on the corresponding numerical feature subset in the clue set, and the second gradient boosting tree model is trained on the corresponding classification feature subset in the clue set. The feasibility scores output by the two models are weighted and averaged or stacked to output a unified feasibility score.

[0011] As an optional implementation, in the first aspect of the present invention, determining the execution conditions of unstructured features in the candidate clue set based on a preset inference model includes: The unstructured features are input into the reasoning model, and the corresponding execution conditions are output. The reasoning model is the base architecture of the language model and is trained by multiple clue samples containing manually labeled legal elements. The execution conditions correspond to the model training objectives, including legal element extraction, dispute focus identification, and implicit asset behavior reasoning. The legal element extraction is used to determine legal compliance.

[0012] As an optional implementation, in the first aspect of the present invention, based on a preset case retrieval mechanism, target recommended clues are determined from the candidate clue set according to the feasibility score and execution conditions, including: A historical case database is constructed by generating semantic vectors based on the case summary, enforcement measures, and enforcement results of completed historical cases. The current case corresponding to the person subject to enforcement is determined from the input candidate clue set. Based on the case semantics of the current case as the retrieval condition, the top-K historical cases are recalled from the historical case database by semantic similarity. Using the semantic vectors of the historical cases as contextual constraints, potential clues that meet the feasibility score threshold and execution conditions are judged based on the contextual constraints, and target recommendation clues are obtained by filtering.

[0013] As an optional implementation, in the first aspect of the invention, after determining the target recommendation clue, the method further includes: The actual execution results corresponding to the target recommendation clues are used as feedback tags; The feedback labels are aligned with the prediction results of the recommendation clues and then stored in the feature library. When the cumulative number of valid samples of newly added execution results reaches a preset threshold, or the performance index of the preset model drops to the tolerance threshold, the preset model will be retrained based on the newly added valid samples, and updated to a new preset model after testing and verification.

[0014] A second aspect of this invention discloses an execution lead recommendation system for full-cycle debt collection management, the system comprising: The acquisition module is configured to acquire multi-source data of the person subject to enforcement and preprocess it to obtain a candidate clue set. The prediction module is configured to predict the feasibility score corresponding to the structured features in the candidate clue set based on a preset prediction model. The reasoning module is configured to determine the execution conditions of unstructured features in the candidate clue set based on a preset reasoning model. The recommendation module is configured to determine target recommended leads in the candidate lead set based on a preset case retrieval mechanism, according to the feasibility score and execution conditions.

[0015] As an optional implementation, in a second aspect of the present invention, the multi-source data includes structured data and unstructured data. The structured data includes basic case information, de-identified identity information of the person subject to enforcement, credit records, business registration and change information, and the unstructured data includes judgments, court transcripts, enforcement notices, and scanned PDF protocols.

[0016] As an optional implementation, in a second aspect of the invention, the candidate clue set includes structured features for case execution and unstructured features for determining the recommendation degree of the structured features, wherein the unstructured features include legal semantic features for matching similar cases and association graph features for determining implicit assets.

[0017] As an optional implementation, in a second aspect of the invention, acquiring multi-source data of the person subject to enforcement and preprocessing it to obtain a candidate clue set includes: OCR is used to identify and extract related fields from scanned documents and image text in multi-source data to generate structured text; The structured text from multi-source data and the identified structured text are used to extract fields based on a predefined rule engine to generate structured features; Dense vectors corresponding to legal semantic features in multi-source data are generated based on an embedding model to capture legal terms, behavioral patterns, and points of contention in legal texts. A correlation graph is constructed for identifying implicit assets from information entities associated with the judgment debtor in multi-source data, where nodes in the correlation graph are asset entities and edges are asset relationships. The dense vectors and correlation graphs are labeled as unstructured features and added to the candidate clue set containing the structured features.

[0018] As an optional implementation, in a second aspect of the invention, based on a preset prediction model, the feasibility score corresponding to the structured features in the candidate clue set is predicted, including: The candidate clue set is input into the prediction model, and the candidate clue corresponding to any structured feature is output as the feasibility score of the potential execution clue. The prediction model is an ensemble learning model, including a first gradient boosting tree model and a second gradient boosting tree model. The first gradient boosting tree model is trained on the corresponding numerical feature subset in the clue set, and the second gradient boosting tree model is trained on the corresponding classification feature subset in the clue set. The feasibility scores output by the two models are weighted and averaged or stacked to output a unified feasibility score.

[0019] As an optional implementation, in a second aspect of the invention, determining the execution conditions for unstructured features in the candidate clue set based on a preset inference model includes: The unstructured features are input into the reasoning model, and the corresponding execution conditions are output. The reasoning model is the base architecture of the language model and is trained by multiple clue samples containing manually labeled legal elements. The execution conditions correspond to the model training objectives, including legal element extraction, dispute focus identification, and implicit asset behavior reasoning. The legal element extraction is used to determine legal compliance.

[0020] As an optional implementation, in a second aspect of the invention, based on a preset case retrieval mechanism, target recommended leads are determined from the candidate lead set according to the feasibility score and execution conditions, including: A historical case database is constructed by generating semantic vectors based on the case summary, enforcement measures, and enforcement results of completed historical cases. The current case corresponding to the person subject to enforcement is determined from the input candidate clue set. Based on the case semantics of the current case as the retrieval condition, the top-K historical cases are recalled from the historical case database by semantic similarity. Using the semantic vectors of the historical cases as contextual constraints, potential clues that meet the feasibility score threshold and execution conditions are judged based on the contextual constraints, and target recommendation clues are obtained by filtering.

[0021] As an optional implementation, in a second aspect of the invention, after determining the target recommendation clue, the method further includes: The actual execution results corresponding to the target recommendation clues are used as feedback tags; The feedback labels are aligned with the prediction results of the recommendation clues and then stored in the feature library. When the cumulative number of valid samples of newly added execution results reaches a preset threshold, or the performance index of the preset model drops to the tolerance threshold, the preset model will be retrained based on the newly added valid samples, and updated to a new preset model after testing and verification.

[0022] A third aspect of this invention discloses another execution lead recommendation system for full-cycle debt collection management, 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 execution clue recommendation method for full-cycle collection management 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 execution clue recommendation method for full-cycle collection management 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 overcomes the bottleneck of data silos by integrating data from the entire litigation, judgment, and enforcement cycle to construct a unified candidate clue set, improving data coverage while reducing clue omission rate. It calculates feasibility scores for structured features using a predictive model and extracts enforcement conditions from unstructured features using a reasoning model, thus overcoming the problem of implicit association identification and improving the efficiency of clue value assessment and the accuracy of implicit association identification. A case retrieval mechanism integrates feasibility scores and enforcement conditions to generate target recommendation clues, selecting high-potential clues from candidate clues with high feasibility scores. Simultaneously, it ensures that the clue recommendation process is more aligned with actual enforcement experience, avoiding misleading outputs and significantly improving the coverage and accuracy of usable clues. Through full-cycle multi-source data fusion and a multi-level intelligent reasoning judgment mechanism, it solves the problems of data silos, missing implicit association identification, and reliance on subjective experience in debt collection scenarios, achieving an upgrade from human experience-driven to data intelligence-driven approaches. 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 an execution clue recommendation method for full-cycle debt collection management disclosed in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of an execution clue recommendation system for full-cycle debt collection management disclosed in an embodiment of the present invention.

[0028] Figure 3 This is a schematic diagram of another execution clue recommendation system for full-cycle collection management 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 is applicable to the full-cycle collection management in the financial or quasi-financial fields. Collection management involves financial processes, judicial processes, and corporate collection processes. For cases involving judgment debtors who fail to fulfill their obligations to use non-performing assets for repayment, the entire chain cycle will sequentially go through litigation (including initiating litigation), legal judgment, and execution of the judgment. Therefore, the continuity of information, process coherence, and utilization rate of execution clues are extremely important throughout the entire collection management cycle. Among these, how to accurately discover execution clues with high utilization rates has become a key technology.

[0033] Current methods for uncovering enforcement leads have the following shortcomings: On the one hand, traditional methods obtain raw data through data procurement or by acquiring publicly available online data. Then, the value of this data is mined and assessed based on the personal experience of the appraisers, such as checking the property, vehicles, bank accounts, and social security payment records of the person subject to enforcement. This approach is not only inefficient but also prone to overlooking hidden assets or related relationships. On the other hand, existing data mining schemes suffer from several issues. Firstly, case documents, judgments, and party information from the litigation stage are scattered across different platforms compared to data from the enforcement stage, including property searches, blacklists, and restrictions on high-end consumption. Firstly, existing clue mining systems suffer from several shortcomings. Firstly, they rely on keyword matching or simple rule engines, failing to identify implicit relationships. Secondly, they lack a dynamic feedback mechanism, failing to optimize the model based on newly acquired execution results, thus hindering the achievement of a closed loop of prediction-execution-learning. Finally, existing clue mining systems often focus on a single stage (e.g., only supporting case filing or only providing execution document generation), lacking the ability to structurally integrate and intelligently reason about the entire lifecycle of litigation, legal judgments, and execution judgments. This results in low execution rates and extended payment cycles. Therefore, existing technologies have deficiencies that urgently need to be addressed.

[0034] This invention discloses a method and system for recommending enforcement leads for full-cycle debt collection management. By integrating data from the entire litigation, judgment, and enforcement cycle to construct a unified candidate lead set, it overcomes the bottleneck of data silos, improves data coverage while reducing lead omission rates. It calculates feasibility scores for structured features using a predictive model and extracts enforcement conditions by analyzing unstructured features using a reasoning model, overcoming the problem of implicit association identification and improving the efficiency of lead value assessment and the accuracy of implicit association identification. Through a case retrieval mechanism, it integrates feasibility scores and enforcement conditions to generate target recommended leads, selecting high-potential leads from candidate leads with high feasibility scores. Simultaneously, it ensures that the lead recommendation process is more aligned with actual enforcement experience, avoiding misleading results and significantly improving the coverage and accuracy of usable leads. Through full-cycle multi-source data fusion and a multi-level intelligent reasoning judgment mechanism, it solves the problems of data silos, missing implicit association identification, and reliance on subjective experience in debt collection scenarios, achieving an upgrade from human experience-driven to data intelligence-driven approaches. These are described in detail below.

[0035] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating an execution lead recommendation method for full-cycle debt collection management disclosed in an embodiment of the present invention. Figure 1 The described execution lead recommendation method for full-cycle collection management can be applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). For example... Figure 1As shown, this execution lead recommendation method for full-cycle collection management may include the following operations: 101. Obtain multi-source data of the judgment debtor and preprocess it to obtain a candidate clue set.

[0036] Optionally, multi-source data can be either structured or unstructured data. Multi-source data can also be heterogeneous multi-source data, including heterogeneous combinations of structured and unstructured data. The source of multi-source data is multi-source data related to the judgment debtor at various stages of litigation, judgment, and enforcement. This invention does not limit the source of multi-source data.

[0037] Optionally, the structured data may include basic case information, de-identified identity information of the person subject to enforcement, credit records, business registration and change information, etc. The structured data may be stored in a relational database (such as MySQL or PostgreSQL) to facilitate efficient querying and correlation analysis of the structured data. The basic case information refers to case information related to the person subject to enforcement that has been judged but has not been executed in accordance with the judgment. This invention does not limit the scope of the case information.

[0038] Optionally, unstructured data can be full text of judgments, court transcripts, enforcement notices, or scanned PDF documents. Semi-structured and unstructured data can be archived using the MinIO object storage system, and full-text indexing and keyword retrieval can be achieved using Elasticsearch to ensure high availability and traceability of multi-source data. This invention does not impose any limitations.

[0039] Optionally, the candidate clue set can be structured and unstructured features of multi-source data, including high-dimensional judicial indicators, credit data of the judgment debtor, relationship graph of the judgment debtor, and semantic vectors of legal documents, etc., which are not limited in this application.

[0040] 102. Based on a preset prediction model, predict the feasibility score corresponding to the structured features in the candidate clue set.

[0041] Optionally, the prediction model is a lightweight model, which can be a CNN model with a 5-layer 1D convolutional structure: Layer 1: 64 filters, kernel=5, ReLU; Layer 2: 128 filters, kernel=3, ReLU; Layer 3: 256 filters, kernel=3, ReLU; Layer 4: global average pooling; Layer 5: fully connected to output the predicted score. Alternatively, it can be an ensemble model of XGBoost and LightGBM. Both models are trained 100 times on a subset of 100,000 structured features with potential cue labels. The ensemble model can fuse the output scores through weighted averaging or stacking. This invention does not impose any limitations.

[0042] Optionally, the structured features may include the regional judicial enforcement intensity index, historical performance rate, social security payment status of the judgment debtor, and the number of related enterprises; this invention does not impose any limitations on these features.

[0043] 103. Based on a preset reasoning model, determine the execution conditions for the unstructured features in the candidate clue set.

[0044] Optionally, unstructured features can be case texts obtained from multi-source data processing, such as indictments, court transcripts, mediation records, etc., which are not limited in this invention.

[0045] Optionally, the inference model is a language model, which can be a QWEN large language model or an LLA large language model. It is fine-tuned by low-rank adaptation LoRA. The fine-tuning data includes manually labeled legal elements (such as "there is a guarantee relationship" and "the amount of debt has been confirmed") and compliant language samples. The training task is legal element extraction (identifying whether the statutory enforcement conditions are met from the text). This invention does not limit the scope of the model.

[0046] Optionally, the reasoning model can also identify the points of contention in legal judgments, thereby pinpointing the core points of disagreement between the parties, such as whether to acknowledge the debt or whether there is a dispute over the amount; this invention does not limit this.

[0047] Optionally, the inference model can also infer from unstructured features, based on semantics, whether there are asset transfers or nominee shareholdings that evade enforcement. This invention does not limit this.

[0048] Optionally, the fine-tuning process of the inference model supports supervised fine-tuning (SFT) and preference optimization (DPO) to ensure that the output meets judicial compliance requirements.

[0049] 104. Based on the preset case retrieval mechanism, and according to the feasibility score and execution conditions, determine the target recommended clues in the candidate clue set.

[0050] Optionally, the case retrieval mechanism first pre-constructs a historical execution case database, which contains cases that have been executed. By encoding the case summary, execution measures, and execution results of the cases, semantic vectors are generated and stored in a vector database (such as FAISS or Milvus). When dealing with pending cases of the judgment debtor, the case retrieval mechanism identifies the current case semantics of the judgment debtor and retrieves the top-K most similar historical cases from the historical execution case database through semantic similarity. The semantic vectors of the historical cases are then used as contextual constraints in the judgment process of candidate clues to guide the generation of judgments that are more in line with actual execution experience and avoid illusionary output.

[0051] As can be seen, the above-mentioned embodiments of the invention, by integrating data from the entire litigation, judgment, and enforcement cycle to construct a unified candidate clue set, overcome the bottleneck of data silos, improve data coverage while reducing clue omission rate, calculate feasibility scores for structured features through predictive models, and extract execution conditions by analyzing unstructured features through reasoning models, thus overcoming the problem of implicit association identification, improving the efficiency of clue value assessment and the accuracy of implicit association identification, and generating target recommended clues by integrating feasibility scores and execution conditions through a case retrieval mechanism, selecting high-potential clues from candidate clues with high feasibility scores, while ensuring that the clue recommendation process is more in line with the judgment of actual execution experience, avoiding illusory output, and significantly improving the coverage and accuracy of usable clues. Through the fusion of multi-source data throughout the entire cycle and a multi-level intelligent reasoning judgment mechanism, the invention solves the problems of data silos, lack of implicit association identification, and reliance on subjective experience in the collection scenario, realizing an upgrade from human experience-driven to data intelligence-driven.

[0052] As an optional embodiment, the steps described above, including obtaining multi-source data of the judgment debtor and preprocessing it to obtain a candidate clue set, include: OCR is used to identify and extract related fields from scanned documents and image text in multi-source data to generate structured text; The structured text from multi-source data and the identified structured text are used to extract fields based on a predefined rule engine to generate structured features; Dense vectors corresponding to legal semantic features in multi-source data are generated based on an embedding model to capture legal terms, behavioral patterns, and points of contention in legal texts. A correlation graph is constructed for identifying implicit assets from information entities associated with the judgment debtor in multi-source data, where nodes in the correlation graph are asset entities and edges are asset relationships. The dense vectors and correlation graphs are labeled as unstructured features and added to the candidate clue set containing the structured features.

[0053] As an example, scanned documents and image text can be documents such as repayment slips or debt transfer agreements. The recognition process first classifies the pages, such as recognizing agreement pages, transaction pages, ID card pages, etc. Then, OCR recognition is performed and character errors are corrected by combining contextual semantics. For table-type documents, the column boundaries and table header positions are dynamically recognized through visual layout analysis, and key fields (such as contract number, amount, and date) can be accurately extracted without the need for preset templates.

[0054] Optionally, the rule engine supports dynamic matching based on regular expressions, keyword anchors, and relative position logic. For example, after locating the contract number in a text message or PDF, it can automatically search for strings that conform to the encoding rules to the right or down to achieve highly robust field extraction. This invention does not impose any limitations on this.

[0055] Optionally, the embedding model can be a Law-BERT model, a BCE model, or a BGE model. This invention does not impose any limitations. The dense vectors generated by the embedding model are used as input features for subsequent similar case vector retrieval or model learning.

[0056] Optionally, entities in the association graph can be based on the name, ID number, or unified social credit code of the judgment debtor. By constructing a network of relationships between the judgment debtor and their relatives, related companies, and historical counterparties, nodes represent asset entity information (such as asset type, amount, and rights holder), while edges represent relationship types such as equity, guarantees, shared address, and fund transfers. The association graph can be used for path queries to discover the judgment debtor's hidden asset transfer paths, such as judgment debtor → spouse → newly registered vehicle owner.

[0057] As can be seen, through the above optional embodiments, by using OCR recognition on scanned documents / image text and combining it with a predefined rule engine to automatically extract execution-related fields to generate structured text, the efficiency of data conversion and the accuracy of structured extraction are improved. By using an embedding model finely tuned based on the financial judicial field, legal text is encoded into dense vectors, capturing semantic patterns such as mortgage behavior and related party control, breaking through the bottleneck of implicit association recognition in legal text, improving the efficiency of dispute focus positioning and the accuracy of legal semantic recognition, constructing an implicit asset association graph to remove blind spots in implicit asset recognition, and significantly improving the implicit asset recognition rate and the accuracy of association mining. By integrating the dense legal semantic vectors and unstructured features such as association graphs into a structured feature set, a high-dimensional candidate clue set is formed, improving the quality of clues and achieving accurate quantification of clue value. Through multi-source data deep structuring and implicit association intelligent mining mechanisms, the core problems of data fragmentation, implicit asset recognition blind spots, and shallow legal semantic understanding in collection scenarios are solved, realizing the intelligent generation of high-value clues from raw data accumulation.

[0058] As an optional embodiment, the step above, predicting the feasibility score corresponding to the structured features in the candidate clue set based on a preset prediction model, includes: The candidate clue set is input into the prediction model, and the candidate clue corresponding to any structured feature is output as the feasibility score of the potential execution clue. The prediction model is an ensemble learning model, including a first gradient boosting tree model and a second gradient boosting tree model. The first gradient boosting tree model is trained on the corresponding numerical feature subset in the clue set, and the second gradient boosting tree model is trained on the corresponding classification feature subset in the clue set. The feasibility scores output by the two models are weighted and averaged or stacked to output a unified feasibility score.

[0059] Optionally, the first gradient boosting tree model can be an XGBoost model.

[0060] Optionally, the second gradient boosting tree model can be the LightGBM model.

[0061] As can be seen, through the above optional embodiments, the bottlenecks of feature heterogeneity and model bias are overcome by using dual-model professional training. At the same time, the strong correlation of numerical features and the high-dimensional sparsity of classification features are captured, which improves the accuracy of feasibility prediction and model robustness. The weights are dynamically adjusted according to the case features to avoid misjudgment caused by uniform thresholds, and the assessment inaccuracy caused by differences in case types is eliminated, thereby improving the accuracy of high-value clue screening.

[0062] As an optional embodiment, the step above, determining the execution conditions for unstructured features in the candidate clue set based on a preset inference model, includes: The unstructured features are input into the reasoning model, and the corresponding execution conditions are output. The reasoning model is the base architecture of the language model and is trained by multiple clue samples containing manually labeled legal elements. The execution conditions correspond to the model training objectives, including legal element extraction, dispute focus identification, and implicit asset behavior reasoning. The legal element extraction is used to determine legal compliance.

[0063] As can be seen, through the above optional embodiments, the inaccuracy of compliance judgment is eradicated by automatically extracting legal elements, thereby improving the accuracy of compliance judgment and the interception rate of legal risks. The bottleneck of implicit association identification is broken through by accurately identifying the focus of disputes, thereby improving the value of clues and the coverage of the focus of disputes. By combining the association graph with legal semantic reasoning of asset behavior paths, the accuracy of execution conditions and the identification rate of implicit assets are improved. Through the legal semantic-driven reasoning model, the problems of lack of legal compliance judgment, superficial positioning of the focus of disputes, and blind spots in the reasoning of implicit asset behavior in the collection scenario are solved, thereby realizing a paradigm upgrade from reliance on human experience to semantic intelligent reasoning.

[0064] As an optional embodiment, the step described above, determining the target recommended clue in the candidate clue set based on a preset case retrieval mechanism, according to the feasibility score and execution conditions, includes: A historical case database is constructed by generating semantic vectors based on the case summary, enforcement measures, and enforcement results of completed historical cases. The current case corresponding to the person subject to enforcement is determined from the input candidate clue set. Based on the case semantics of the current case as the retrieval condition, the top-K historical cases are recalled from the historical case database by semantic similarity. Using the semantic vectors of the historical cases as contextual constraints, potential clues that meet the feasibility score threshold and execution conditions are judged based on the contextual constraints, and target recommendation clues are obtained by filtering.

[0065] As can be seen, through the above optional embodiments, by encoding historical cases into semantic vectors to construct a historical case database, the bottlenecks of data silos and fragmented experience are overcome, the utilization rate of historical data and the efficiency of implicit association mining are improved, the similarity is calculated by using the semantic vector of the current case as a query condition, and the top-K most relevant historical cases are recalled as prompts, thereby improving the accuracy of recall relevance and the coverage of high-value clues. By using the semantic vector of historical cases as contextual constraints to perform secondary verification of potential clues, dynamic filtering realizes the closed-loop integration of decision-making and historical experience, reduces the probability of misjudgment and solves the illusion problem. Through the dynamic clue filtering mechanism driven by the historical case semantic database, the problems of the difficulty in reusing historical experience, the passive nature of implicit association mining, and the lack of contextual constraints in decision-making in the collection scenario are completely solved, realizing a paradigm upgrade from isolated case processing to intelligent empowerment of historical experience.

[0066] As an optional embodiment, the above steps, after determining the target recommendation clues, further include: The actual execution results corresponding to the target recommendation clues are used as feedback tags; The feedback labels are aligned with the prediction results of the recommendation clues and then stored in the feature library. When the cumulative number of valid samples of newly added execution results reaches a preset threshold, or the performance index of the preset model drops to the tolerance threshold, the preset model will be retrained based on the newly added valid samples, and updated to a new preset model after testing and verification.

[0067] As an example, to support the stable operation and continuous model optimization of large-scale execution clue mining tasks, this embodiment constructs an MLOps system with asynchronous task scheduling and closed-loop feedback as its core.

[0068] Furthermore, Celery is used as a distributed task queue and Redis as a message broker to achieve asynchronous batch prediction of tens of thousands of cases. When the business system submits a batch of cases to be processed (e.g., 10,000 cases), the scheduler breaks them down into multiple subtasks and distributes them in parallel to multiple inference worker nodes. Each node independently calls the Triton inference service to complete feature scoring and semantic analysis, and the results are aggregated and written to the database. For processing tasks involving local files (e.g., scanning PDFs), the system can switch to a self-developed thread pool framework to bypass network I / O overhead and directly read and process files on the local disk, significantly improving the throughput efficiency of I / O-intensive tasks.

[0069] Following this, the internal MLOps platform runs through the entire lifecycle of model development, deployment, monitoring, and retraining. The system automatically collects real execution results (such as "whether assets were successfully seized" and "actual amount recovered") as feedback labels, aligns them with the original prediction results, and stores them in the feature library. When the cumulative number of newly added valid samples reaches a preset threshold, or when the model performance indicators (such as AUC and F1) drop beyond the tolerance range, the platform automatically triggers the retraining process: calling the LLaMA-Factory or LightGBM training script to generate a new version of the model, and after verifying the effect through A / B testing, it is gradually rolled out, forming a positive closed loop of "prediction → execution → learning".

[0070] As can be seen, through the above optional embodiments, by collecting the actual execution results of the target recommendation clues as feedback labels, and aligning them with the prediction results to store them in the feature library to form a prediction-execution-feedback closed loop, the data drift bottleneck is overcome based on the dynamic triggering mechanism of two thresholds, improving the model iteration efficiency and the speed of adapting to new judicial regulations. Through dynamic closed-loop feedback and adaptive model iteration mechanism, the problems of model performance degradation, inability to cope with data drift, and lack of long-term repayment stability in the debt collection scenario are solved, realizing a paradigm upgrade from static model application to intelligent self-evolving system.

[0071] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an execution clue recommendation system for full-cycle debt collection management disclosed in an embodiment of the present invention. Figure 2 The described execution lead recommendation system for full-cycle debt collection management can be applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). For example... Figure 2 As shown, this execution lead recommendation system for full-cycle debt collection management may include: The acquisition module 201 is configured to acquire multi-source data of the person subject to enforcement and preprocess it to obtain a candidate clue set. The prediction module 202 is configured to predict the feasibility score corresponding to the structured features in the candidate clue set based on a preset prediction model. The reasoning module 203 is configured to determine the execution conditions of the unstructured features in the candidate clue set based on a preset reasoning model. The recommendation module 204 is configured to determine the target recommended clues in the candidate clue set based on the feasibility score and execution conditions, according to a preset case retrieval mechanism.

[0072] As can be seen, the above-mentioned embodiments of the invention, by integrating data from the entire litigation, judgment, and enforcement cycle to construct a unified candidate clue set, overcome the bottleneck of data silos, improve data coverage while reducing clue omission rate, calculate feasibility scores for structured features through predictive models, and extract execution conditions by analyzing unstructured features through reasoning models, thus overcoming the problem of implicit association identification, improving the efficiency of clue value assessment and the accuracy of implicit association identification, and generating target recommended clues by integrating feasibility scores and execution conditions through a case retrieval mechanism, selecting high-potential clues from candidate clues with high feasibility scores, while ensuring that the clue recommendation process is more in line with the judgment of actual execution experience, avoiding illusory output, and significantly improving the coverage and accuracy of usable clues. Through the fusion of multi-source data throughout the entire cycle and a multi-level intelligent reasoning judgment mechanism, the invention solves the problems of data silos, lack of implicit association identification, and reliance on subjective experience in the collection scenario, realizing an upgrade from human experience-driven to data intelligence-driven.

[0073] As an optional implementation, the multi-source data includes structured data and unstructured data. The structured data includes basic case information, de-identified identity information of the person subject to enforcement, credit records, business registration and change information, and the unstructured data includes judgment documents, court transcripts, enforcement notices, and scanned PDF protocols.

[0074] As an optional implementation, the candidate clue set includes structured features for case execution and unstructured features for judging the recommendation degree of the structured features, wherein the unstructured features include legal semantic features for matching similar cases and association graph features for judging implicit assets.

[0075] As an optional implementation, multi-source data of the person subject to enforcement is obtained, and preprocessed to obtain a candidate clue set, including: OCR is used to identify and extract related fields from scanned documents and image text in multi-source data to generate structured text; The structured text from multi-source data and the identified structured text are used to extract fields based on a predefined rule engine to generate structured features; Dense vectors corresponding to legal semantic features in multi-source data are generated based on an embedding model to capture legal terms, behavioral patterns, and points of contention in legal texts. A correlation graph is constructed for identifying implicit assets from information entities associated with the judgment debtor in multi-source data, where nodes in the correlation graph are asset entities and edges are asset relationships. The dense vectors and correlation graphs are labeled as unstructured features and added to the candidate clue set containing the structured features.

[0076] As can be seen, through the above optional embodiments, by using OCR recognition on scanned documents / image text and combining it with a predefined rule engine to automatically extract execution-related fields to generate structured text, the efficiency of data conversion and the accuracy of structured extraction are improved. By using an embedding model finely tuned based on the financial judicial field, legal text is encoded into dense vectors, capturing semantic patterns such as mortgage behavior and related party control, breaking through the bottleneck of implicit association recognition in legal text, improving the efficiency of dispute focus positioning and the accuracy of legal semantic recognition, constructing an implicit asset association graph to remove blind spots in implicit asset recognition, and significantly improving the implicit asset recognition rate and the accuracy of association mining. By integrating the dense legal semantic vectors and unstructured features such as association graphs into a structured feature set, a high-dimensional candidate clue set is formed, improving the quality of clues and achieving accurate quantification of clue value. Through multi-source data deep structuring and implicit association intelligent mining mechanisms, the core problems of data fragmentation, implicit asset recognition blind spots, and shallow legal semantic understanding in collection scenarios are solved, realizing the intelligent generation of high-value clues from raw data accumulation.

[0077] As an optional implementation, based on a preset prediction model, the feasibility score corresponding to the structured features in the candidate clue set is predicted, including: The candidate clue set is input into the prediction model, and the candidate clue corresponding to any structured feature is output as the feasibility score of the potential execution clue. The prediction model is an ensemble learning model, including a first gradient boosting tree model and a second gradient boosting tree model. The first gradient boosting tree model is trained on the corresponding numerical feature subset in the clue set, and the second gradient boosting tree model is trained on the corresponding classification feature subset in the clue set. The feasibility scores output by the two models are weighted and averaged or stacked to output a unified feasibility score.

[0078] As can be seen, through the above optional embodiments, the bottlenecks of feature heterogeneity and model bias are overcome by using dual-model professional training. At the same time, the strong correlation of numerical features and the high-dimensional sparsity of classification features are captured, which improves the accuracy of feasibility prediction and model robustness. The weights are dynamically adjusted according to the case features to avoid misjudgment caused by uniform thresholds, and the assessment inaccuracy caused by differences in case types is eliminated, thereby improving the accuracy of high-value clue screening.

[0079] As an optional implementation, based on a preset inference model, the execution conditions for the unstructured features in the candidate clue set are determined, including: The unstructured features are input into the reasoning model, and the corresponding execution conditions are output. The reasoning model is the base architecture of the language model and is trained by multiple clue samples containing manually labeled legal elements. The execution conditions correspond to the model training objectives, including legal element extraction, dispute focus identification, and implicit asset behavior reasoning. The legal element extraction is used to determine legal compliance.

[0080] As can be seen, through the above optional embodiments, the inaccuracy of compliance judgment is eradicated by automatically extracting legal elements, thereby improving the accuracy of compliance judgment and the interception rate of legal risks. The bottleneck of implicit association identification is broken through by accurately identifying the focus of disputes, thereby improving the value of clues and the coverage of the focus of disputes. By combining the association graph with legal semantic reasoning of asset behavior paths, the accuracy of execution conditions and the identification rate of implicit assets are improved. Through the legal semantic-driven reasoning model, the problems of lack of legal compliance judgment, superficial positioning of the focus of disputes, and blind spots in the reasoning of implicit asset behavior in the collection scenario are solved, thereby realizing a paradigm upgrade from reliance on human experience to semantic intelligent reasoning.

[0081] As an optional implementation, based on a preset case retrieval mechanism, and according to the feasibility score and execution conditions, target recommended leads are determined from the candidate lead set, including: A historical case database is constructed by generating semantic vectors based on the case summary, enforcement measures, and enforcement results of completed historical cases. The current case corresponding to the person subject to enforcement is determined from the input candidate clue set. Based on the case semantics of the current case as the retrieval condition, the top-K historical cases are recalled from the historical case database by semantic similarity. Using the semantic vectors of the historical cases as contextual constraints, potential clues that meet the feasibility score threshold and execution conditions are judged based on the contextual constraints, and target recommendation clues are obtained by filtering.

[0082] As can be seen, through the above optional embodiments, by encoding historical cases into semantic vectors to construct a historical case database, the bottlenecks of data silos and fragmented experience are overcome, the utilization rate of historical data and the efficiency of implicit association mining are improved, the similarity is calculated by using the semantic vector of the current case as a query condition, and the top-K most relevant historical cases are recalled as prompts, thereby improving the accuracy of recall relevance and the coverage of high-value clues. By using the semantic vector of historical cases as contextual constraints to perform secondary verification of potential clues, dynamic filtering realizes the closed-loop integration of decision-making and historical experience, reduces the probability of misjudgment and solves the illusion problem. Through the dynamic clue filtering mechanism driven by the historical case semantic database, the problems of the difficulty in reusing historical experience, the passive nature of implicit association mining, and the lack of contextual constraints in decision-making in the collection scenario are completely solved, realizing a paradigm upgrade from isolated case processing to intelligent empowerment of historical experience.

[0083] As an optional implementation, after determining the target recommendation clues, the method further includes: The actual execution results corresponding to the target recommendation clues are used as feedback tags; The feedback labels are aligned with the prediction results of the recommendation clues and then stored in the feature library. When the cumulative number of valid samples of newly added execution results reaches a preset threshold, or the performance index of the preset model drops to the tolerance threshold, the preset model will be retrained based on the newly added valid samples, and updated to a new preset model after testing and verification.

[0084] As can be seen, through the above optional embodiments, by collecting the actual execution results of the target recommendation clues as feedback labels, and aligning them with the prediction results to store them in the feature library to form a prediction-execution-feedback closed loop, the data drift bottleneck is overcome based on the dynamic triggering mechanism of two thresholds, improving the model iteration efficiency and the speed of adapting to new judicial regulations. Through dynamic closed-loop feedback and adaptive model iteration mechanism, the problems of model performance degradation, inability to cope with data drift, and lack of long-term repayment stability in the debt collection scenario are solved, realizing a paradigm upgrade from static model application to intelligent self-evolving system.

[0085] Example 3 Please see Figure 3 , Figure 3 This is another execution clue recommendation system for full-cycle debt collection management disclosed in the embodiments of the present invention. Figure 3 The described execution lead recommendation system for full-cycle debt collection management is applied in data processing systems / data processing equipment / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 3 As shown, this execution lead recommendation system for full-cycle debt collection management 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 execution clue recommendation method for full-cycle collection management described in Embodiment 1.

[0086] 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 execution clue recommendation method for full-cycle collection management described in Embodiment 1.

[0087] 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 execution clue recommendation method for full-cycle collection management described in Embodiment 1.

[0088] 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.

[0089] 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.

[0090] 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.

[0091] 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.

[0092] 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.

[0093] 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.

[0094] 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.

[0095] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0096] 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.

[0097] 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.

[0098] 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.

[0099] 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.

[0100] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. 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.

[0101] Finally, it should be noted that the execution clue recommendation method and system for full-cycle collection management 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 method for recommending execution leads for full-cycle debt collection management, characterized in that, The method includes: Obtain multi-source data on the judgment debtor and preprocess it to obtain a candidate clue set; Based on a preset prediction model, the feasibility score corresponding to the structured features in the candidate clue set is predicted; Based on a pre-defined reasoning model, the execution conditions for unstructured features in the candidate clue set are determined; Based on a pre-defined case retrieval mechanism, target recommended leads are determined from the candidate lead set according to the feasibility score and execution conditions.

2. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, The multi-source data includes structured data and unstructured data. The structured data includes basic case information, de-identified identity information of the person subject to enforcement, credit records, business registration and change information. The unstructured data includes judgment documents, court transcripts, enforcement notices and scanned PDF protocols.

3. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, The candidate clue set includes structured features for case execution and unstructured features for judging the recommendation degree of the structured features, wherein the unstructured features include legal semantic features for matching similar cases and association graph features for judging implicit assets.

4. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, Obtain multi-source data on the judgment debtor, and preprocess it to obtain a candidate clue set, including: OCR is used to identify and extract related fields from scanned documents and image text in multi-source data to generate structured text; The structured text from multi-source data and the identified structured text are used to extract fields based on a predefined rule engine to generate structured features; Dense vectors corresponding to legal semantic features in multi-source data are generated based on an embedding model to capture legal terms, behavioral patterns, and points of contention in legal texts. A correlation graph is constructed for identifying implicit assets from information entities associated with the judgment debtor in multi-source data, where nodes in the correlation graph are asset entities and edges are asset relationships. The dense vectors and correlation graphs are labeled as unstructured features and added to the candidate clue set containing the structured features.

5. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, Based on a pre-defined prediction model, the feasibility score corresponding to the structured features in the candidate clue set is predicted, including: The candidate clue set is input into the prediction model, and the candidate clue corresponding to any structured feature is output as the feasibility score of the potential execution clue. The prediction model is an ensemble learning model, including a first gradient boosting tree model and a second gradient boosting tree model. The first gradient boosting tree model is trained on the corresponding numerical feature subset in the clue set, and the second gradient boosting tree model is trained on the corresponding classification feature subset in the clue set. The feasibility scores output by the two models are weighted and averaged or stacked to output a unified feasibility score.

6. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, Based on a pre-defined reasoning model, the execution conditions for unstructured features in the candidate clue set are determined, including: The unstructured features are input into the reasoning model, and the corresponding execution conditions are output. The reasoning model is the base architecture of the language model and is trained by multiple clue samples containing manually labeled legal elements. The execution conditions correspond to the model training objectives, including legal element extraction, dispute focus identification, and implicit asset behavior reasoning. The legal element extraction is used to determine legal compliance.

7. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, Based on a pre-defined case retrieval mechanism, and according to the feasibility score and execution conditions, target recommended leads are determined from the candidate lead set, including: A historical case database is constructed by generating semantic vectors based on the case summary, enforcement measures, and enforcement results of completed historical cases. The current case corresponding to the person subject to enforcement is determined from the input candidate clue set. Based on the case semantics of the current case as the retrieval condition, the top-K historical cases are recalled from the historical case database by semantic similarity. Using the semantic vectors of the historical cases as contextual constraints, potential clues that meet the feasibility score threshold and execution conditions are judged based on the contextual constraints, and target recommendation clues are obtained by filtering.

8. The execution lead recommendation method for full-cycle debt collection management according to claim 1, characterized in that, After identifying the target recommendation clues, the process also includes: The actual execution results corresponding to the target recommendation clues are used as feedback tags; The feedback labels are aligned with the prediction results of the recommendation clues and then stored in the feature library. When the cumulative number of valid samples of newly added execution results reaches a preset threshold, or the performance index of the preset model drops to the tolerance threshold, the preset model will be retrained based on the newly added valid samples, and updated to a new preset model after testing and verification.

9. An execution lead recommendation system for full-cycle debt collection management, characterized in that, The system includes: The acquisition module is configured to acquire multi-source data of the person subject to enforcement and preprocess it to obtain a candidate clue set. The prediction module is configured to predict the feasibility score corresponding to the structured features in the candidate clue set based on a preset prediction model. The reasoning module is configured to determine the execution conditions of unstructured features in the candidate clue set based on a preset reasoning model. The recommendation module is configured to determine target recommended leads in the candidate lead set based on a preset case retrieval mechanism, according to the feasibility score and execution conditions.

10. An execution lead recommendation system for full-cycle debt collection management, 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 execution clue recommendation method for full-cycle collection management as described in any one of claims 1-8.

Images