A car finance loan risk prediction method and device based on a multi-role intelligent agent and a special knowledge base, equipment and a storage medium

Abstract

The application discloses a car financial loan risk prediction method and device based on multi-role agents and special knowledge base, equipment and storage medium, relates to the field of financial technology, and comprises the following steps: planning an agent to analyze a transaction flow electronic document provided by a user authorization to obtain a data set containing a plurality of transaction records, and generating a transaction classification task; a transaction classification agent extracts a transaction counterpart field, calls a large model based on a current special knowledge base and the transaction counterpart field, determines a comprehensive classification confidence, and when a preset dynamic threshold is reached, classifies the records to form a classification result set; a data statistical agent performs statistical analysis on the set and extracts statistical features; an insight reasoning agent generates an atomic evidence set based on the transaction records and the statistical features to determine an evidence score, aggregates atomic evidences according to the score, generates an intermediate conclusion set, generates a target conclusion based on the intermediate conclusion, and outputs a loan risk prediction result, so that the efficiency of predicting car financial loan risks is improved.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and in particular to a method, apparatus, device, and storage medium for predicting auto finance loan risks based on multi-role intelligent agents and proprietary knowledge bases. Background Technology

[0002] Currently, competition in the financial industry is becoming increasingly fierce, with financial institutions vying for customers. It is unsustainable to simply reject customers with poor credit. Financial institutions need to operate with refinement while adhering to the bottom line of risk and try to increase the approval rate as much as possible. Therefore, financial institutions need to select high-quality customers from this subprime customer group for loan disbursement.

[0003] The prerequisite for refined customer screening is a thorough understanding of customer qualifications. However, with increasingly stringent personal information security protections, financial institutions have limited access to external information, often relying solely on credit reports. Credit reports, however, primarily showcase basic customer information and past business records with financial institutions, lacking sufficient depth and breadth of understanding. This is especially true for customers with poor credit histories, posing a significant risk of loan approval. To grant loans to such customers, auditors require transaction records demonstrating income levels that meet approval requirements. Currently, different financial institutions provide varying formats for transaction records, with WeChat transaction records being the most detailed. Fields such as transaction type and counterparty have significant analytical value, allowing for the creation of detailed customer profiles and providing auditors with valuable information beyond credit reports to gain a comprehensive understanding of their clients.

[0004] However, the WeChat transaction records submitted by customers are usually hundreds of pages long, covering daily consumption records such as clothing, food, housing and transportation. Under the requirement of approval timeliness, it is impossible for the reviewers to quickly analyze the transaction flow. It is necessary to use technical means to automatically analyze and summarize the data so that they can quickly obtain the basis for increasing or decreasing the trust of customers.

[0005] The current mainstream technical solution uses OCR technology to identify and standardize transaction records, and then, based on a rule engine, performs additions, subtractions, summaries, and classifications according to deterministic rules as needed. This allows for simple tasks such as calculating total transaction amounts, identifying frequently transacting accounts, and recognizing salary deposits. The drawback is that it can only perform statistics based on preset rules and cannot conduct in-depth analysis based on fields such as "transaction time," "transaction counterparty," and "transaction amount" in each customer's transaction records. Furthermore, it cannot deduce complex, hidden information such as lifestyle patterns, consumption habits, social relationships, or abnormal transactions.

[0006] As can be seen from the above, how to improve the efficiency of predicting the risks of auto finance loans is an urgent problem to be solved. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method, apparatus, device, and storage medium for predicting auto finance loan risks based on multi-role intelligent agents and a proprietary knowledge base, which can improve the efficiency of predicting auto finance loan risks in the process of predicting auto finance loan risks based on multi-role intelligent agents and a proprietary knowledge base. The specific solution is as follows: Firstly, this application provides a method for predicting auto finance loan risk based on multi-role intelligent agents and a proprietary knowledge base, including: A planning agent is used to parse the target transaction flow electronic document to obtain a transaction dataset including several transaction records. A transaction classification task is generated based on the transaction dataset. A transaction classification agent is then used to extract the counterparty field of the transaction records in the transaction dataset based on the transaction classification task. The target transaction flow electronic document is a transaction flow electronic document authorized by the user and obtained when processing an auto finance loan application. The large model is invoked and the comprehensive classification confidence level corresponding to each transaction record is determined based on the current proprietary knowledge base and the transaction counterparty field. When the comprehensive classification confidence level is not less than a preset dynamic threshold, the transaction record is classified into categories to obtain a classification result set. The data statistical agent is used to perform statistical analysis on the classification result set to determine a statistical analysis dataset including several statistical features. The insight reasoning agent is used to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset. The evidence score is determined based on the atomic evidence set. The atomic evidence in the atomic evidence set is aggregated based on the evidence score to obtain an intermediate conclusion set. Based on the intermediate conclusion set, corresponding target conclusions are generated, and based on the target conclusions, loan risk prediction results for auto finance are generated; the loan risk prediction results include user consumption characteristics, behavioral patterns, and risk assessment information.

[0008] Optionally, the step of using a planning agent to parse the target transaction flow electronic document to obtain a transaction dataset including several transaction records, generating a transaction classification task based on the transaction dataset, and using a transaction classification agent to extract the counterparty field of the transaction records in the transaction dataset based on the transaction classification task, includes: The planning agent uses a document extraction tool to perform structured parsing on the target transaction flow electronic document, obtains the parsing results, and extracts field information including transaction number, transaction time, transaction type, income and expenditure direction, transaction method, transaction amount, transaction counterparty and merchant order number from the parsing results; Standardized transaction records are constructed based on preset field mapping relationships and the field information, so as to generate a transaction dataset including several transaction records based on the transaction records; Based on the transaction dataset, a transaction classification task corresponding to each of the transaction records is generated, so as to utilize a transaction classification agent and extract the counterparty field corresponding to each transaction record in the transaction dataset based on the transaction classification task.

[0009] Optionally, the step of calling the large model and determining the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the counterparty field includes: The matching weight is determined based on the matching relationship between the counterparty field and the current proprietary knowledge base, the semantic weight is determined based on the business type keywords in the counterparty field, and the frequency weight is determined based on the frequency of the corresponding counterparty in the transaction dataset. The large model is invoked to determine the predicted probability that the transaction record belongs to each candidate category, and the comprehensive classification confidence of the corresponding transaction record is determined based on the predicted probability, the matching weight, the semantic weight, and the frequency weight.

[0010] Optionally, determining the matching weight based on the matching relationship between the counterparty field and the current proprietary knowledge base includes: If the name of the counterparty in the counterparty field is consistent with an entry in the current proprietary knowledge base, and the counterparty name in the counterparty field is completely consistent with an entry in the current proprietary knowledge base, then the matching weight is set to the first weight value; the current proprietary knowledge base is a tree-like hierarchical structure; the current proprietary knowledge base includes a first-level transaction category node and several second-level transaction category nodes corresponding to the first-level transaction category node; When the similarity between the name of the counterparty and the entry in the current proprietary knowledge base meets the preset similarity condition, the matching weight is set to the second weight value; When the name of the counterparty to the transaction does not match an entry in the current proprietary knowledge base, the matching weight is set to a third weight value; wherein the first weight value is greater than the second weight value, and the second weight value is greater than the third weight value.

[0011] Optionally, when the comprehensive classification confidence level is not less than a preset dynamic threshold, classifying the transaction records into categories to obtain a classification result set includes: If the current transaction is the first round of classification, the preset basic threshold is determined as the dynamic threshold. If the current transaction is not the first round of classification, the classification accuracy corresponding to the previous round of classification is obtained, and the initial threshold is determined based on the classification accuracy. The initial threshold is then determined as the dynamic threshold corresponding to the current round of transaction classification. The comprehensive classification confidence level is compared with the dynamic threshold. When the comprehensive classification confidence level is not less than the dynamic threshold, the transaction record is classified into the corresponding category to obtain a classification result set. When the overall classification confidence level is less than the dynamic threshold, the transaction record is classified into other categories, resulting in an undetermined classification set.

[0012] Optionally, the step involves using a data statistical agent to perform statistical analysis on the classification result set, determining a statistical analysis dataset including several statistical features, and using an insight reasoning agent to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset. Evidence scores are then determined based on the atomic evidence set, and each atomic piece of evidence in the atomic evidence set is aggregated based on the evidence scores to obtain an intermediate conclusion set, including: The transaction records in the classification result set are loaded using a data statistical agent, and statistical analysis is performed on the transaction records according to the dimensions of transaction classification, counterparty, direction, month, and amount to obtain a statistical analysis dataset including several statistical features. An explicit evidence chain model is constructed using an insight-based reasoning agent and based on the statistical analysis dataset and corresponding transaction records; the explicit evidence chain model is a directed acyclic graph; the directed acyclic graph includes atomic evidence nodes, intermediate conclusion nodes, and final conclusion nodes; A set of atomic evidence is generated based on the transaction records and the corresponding statistical features. The explicit chain of evidence model is used to determine the evidence score based on the set of atomic evidence, evidence strength, evidence novelty, and evidence timeliness. The evidence strength is used to characterize the degree to which the current atomic evidence supports the corresponding reasoning conclusion. The evidence novelty is used to characterize the degree of deviation of the current atomic evidence from the historical transaction pattern. The evidence timeliness is used to characterize the degree of time decay of the current atomic evidence from the current moment. Based on the evidence scores, determine the first set of atomic evidence supporting the target intermediate conclusion and the second set of atomic evidence opposing the target intermediate conclusion among the atomic evidence in the atomic evidence set; The intermediate conclusion score is determined based on the evidence score corresponding to the first atomic evidence set and the evidence score corresponding to the second atomic evidence set, and the corresponding intermediate conclusion is generated based on the intermediate conclusion score. Based on the intermediate conclusions, a first set of intermediate conclusions supporting the final target conclusion and a second set of intermediate conclusions opposing the final target conclusion are determined, and an intermediate conclusion set is generated based on the first set of intermediate conclusions and the second set of intermediate conclusions.

[0013] Optionally, the step of generating a corresponding target conclusion based on the intermediate conclusion set, and generating a loan risk prediction result for auto finance based on the target conclusion, includes: Based on the first set of intermediate conclusions and the second set of intermediate conclusions, the conclusion weights corresponding to each intermediate conclusion are determined, and the target conclusion score is determined based on each conclusion weight and the intermediate conclusion score. Based on the target conclusion score, a conclusion to be processed is generated, and a customer profile analysis result is generated based on the target conclusion. Then, the counterparty field of each transaction record in the undetermined classification set is deduplicated to obtain a candidate knowledge set. The data access agent obtains the counterparty information corresponding to the candidate knowledge set, and the knowledge base maintenance agent updates the current proprietary knowledge base based on the counterparty information to obtain a new current proprietary knowledge base. The jump stops when a preset iteration termination condition is met; wherein, the preset iteration termination condition includes the undetermined classification set being empty or the number of iterations reaching a preset threshold. If the preset iteration termination condition is not met, the process jumps back to the large model and determines the comprehensive classification confidence level of each transaction record based on the current proprietary knowledge base and the counterparty field, thus obtaining the target conclusion. Based on the aforementioned target conclusions, loan risk prediction results are generated, including user consumption characteristics, behavioral patterns, and risk assessment information.

[0014] Secondly, this application provides an auto finance loan risk prediction device based on multi-role intelligent agents and a proprietary knowledge base, comprising: The counterparty field extraction module is used to parse the target transaction flow electronic document using a planning agent to obtain a transaction dataset including several transaction records, generate a transaction classification task based on the transaction dataset, and extract the counterparty field of the transaction records in the transaction dataset using a transaction classification agent based on the transaction classification task; the target transaction flow electronic document is a transaction flow electronic document authorized by the user and obtained when processing an auto finance loan application; The classification result set generation module is used to call the large model and determine the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the transaction counterparty field. When the comprehensive classification confidence level is not less than a preset dynamic threshold, the module classifies the transaction record into the corresponding category to obtain the classification result set. The intermediate conclusion set acquisition module is used to perform statistical analysis on the classification result set using a data statistical agent, determine a statistical analysis dataset including several statistical features, and use an insight reasoning agent to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset, determine the evidence score based on the atomic evidence set, and aggregate each atomic evidence in the atomic evidence set based on the evidence score to obtain the intermediate conclusion set. The loan risk prediction result generation module is used to generate corresponding target conclusions based on the intermediate conclusion set, and to generate loan risk prediction results for auto finance based on the target conclusions; the loan risk prediction results include user consumption characteristics, behavioral patterns and risk assessment information.

[0015] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned method for predicting auto finance loan risks based on multi-role intelligent agents and proprietary knowledge bases.

[0016] Fourthly, this application provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the aforementioned method for predicting auto finance loan risk based on multi-role intelligent agents and a proprietary knowledge base.

[0017] As can be seen from the above, before performing auto finance loan risk prediction based on multi-role intelligent agents and proprietary knowledge bases, this application needs to use a planning intelligent agent to parse the electronic documents of the target transaction flow, obtain a transaction dataset including several transaction records, generate a transaction classification task based on the transaction dataset, and use a transaction classification intelligent agent to extract the counterparty field of the transaction records in the transaction dataset based on the transaction classification task; call the large model and determine the comprehensive classification confidence of each transaction record based on the current proprietary knowledge base and the counterparty field, and classify the transaction records into categories when the comprehensive classification confidence is not less than a preset dynamic threshold, to obtain a classification result set; use a data statistics intelligent agent to perform statistical analysis on the classification result set, determine a statistical analysis dataset including several statistical features, and use an insight reasoning intelligent agent to generate an atomic evidence set based on the statistical features in the transaction records and statistical analysis dataset, determine the evidence score based on the atomic evidence set, and aggregate the atomic evidence in the atomic evidence set based on the evidence score to obtain an intermediate conclusion set; generate the corresponding target conclusion based on the intermediate conclusion set, and generate the auto finance loan risk prediction result based on the target conclusion.

[0018] Therefore, this application first utilizes a planning agent to parse user-authorized auto finance transaction records and extract the counterparty field; it then calls a large model combined with a proprietary knowledge base to classify each transaction based on comprehensive confidence levels, obtaining classification results after reaching the required thresholds; next, a data statistics agent generates statistical features, which the insight reasoning agent uses to generate atomic evidence and score, aggregating them into intermediate conclusions; finally, a loan risk prediction result is generated. This improves the efficiency of predicting auto finance loan risk, thereby enhancing the user experience. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This application discloses a flowchart of a method for predicting the risk of auto finance loans based on multi-role intelligent agents and proprietary knowledge bases. Figure 2 This application discloses a specific flowchart of a method for predicting auto finance loan risks based on multi-role intelligent agents and proprietary knowledge bases. Figure 3 This is a schematic diagram illustrating a specific knowledge base update process disclosed in this application; Figure 4 This is a schematic diagram of the structure of an auto finance loan risk prediction device based on a multi-role intelligent agent and a proprietary knowledge base disclosed in this application. Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Currently, competition in the financial industry is increasingly fierce, with financial institutions vying for customers. A blanket rejection of customers with poor credit is unsustainable. Financial institutions need to operate with meticulous care while adhering to risk control standards, striving to maximize approval rates. Therefore, financial institutions need to select high-quality customers from this subprime customer group for loan disbursement. To address this, this application provides a method for predicting auto finance loan risk based on multi-role intelligent agents and a proprietary knowledge base, which improves the efficiency of predicting auto finance loan risk.

[0023] See Figure 1 As shown, this embodiment of the invention discloses a method for predicting auto finance loan risk based on multi-role intelligent agents and a proprietary knowledge base, including: Step S11: Use a planning agent to parse the target transaction flow electronic document to obtain a transaction dataset including several transaction records. Generate a transaction classification task based on the transaction dataset, and use a transaction classification agent to extract the counterparty field of the transaction records in the transaction dataset based on the transaction classification task. The target transaction flow electronic document is a transaction flow electronic document authorized by the user and obtained when processing an auto finance loan application.

[0024] In this embodiment, the multi-role intelligent agent collaboration system of this application can uniformly predefine roles, tasks, and available tools for each intelligent agent through standardized prompt word templates. In one specific implementation, this application embodiment can uniformly set the temperature of the large model below 0.3 to ensure the consistency and stability of the large model's output.

[0025] Taking the prompt words of the large model in the transaction classification intelligent agent as an example, the role and task are as follows: You are a transaction flow classification expert of a financial institution. Based on the "counterpartie" field information in the transaction record, you combine a proprietary knowledge base to classify each transaction record and mark whether the current "counterpartie" appears in the knowledge base. For those that cannot be clearly classified, they are classified into the "other" category. The results are output in JSON format.

[0026] The following is a flowchart illustrating the intelligent agent roles involved in the system. These roles communicate via APIs and connect tasks through a workflow. The flowchart illustrates the process of risk prediction for auto finance loans based on multiple intelligent agents and a proprietary knowledge base. Figure 2 As shown: 1. Planning Agent: Responsible for planning transaction flow profile analysis tasks, updating proprietary knowledge base tasks, and scheduling other agents; 2. Data access agent: responsible for calling database read / write tools to interact with the database; 3. Transaction Classification Intelligent Agent: Combining a proprietary knowledge base, it classifies each transaction record and marks whether the "counterpartie" in the current transaction record has appeared in the proprietary knowledge base; 5. Data Statistics Intelligent Agent: Calls statistical tools to perform statistics on the categorized transaction records from multiple preset dimensions and outputs reports; 6. Insight Reasoning Agent: Based on the classified transaction records and multi-dimensional statistical reports, perform positive and negative evidence chain reasoning and output customer profile reports. The explicit reasoning process is required: "hypothesis generation -> evidence search -> positive and negative weighing -> conclusion confirmation". 7. Knowledge Base Maintenance Agent: Responsible for updating the transaction category directory of the proprietary knowledge base.

[0027] Secondly, construct and maintain a proprietary knowledge base: 1. The knowledge base is constructed in Markdown format, and knowledge is entered according to the heading hierarchy. In order to adapt to large model parsing and paragraph semantic segmentation, knowledge is stored in vector form through word embedding model.

[0028] 2. The knowledge base is divided into two levels. The classification rules for the two levels are defined manually, and the initial sample of the second level is provided to the transaction classification agent for reference. The knowledge base maintenance agent handles the knowledge base update task.

[0029] 3. For transaction records whose classification result by the transaction classification agent is "other", after manual classification, the knowledge base maintenance agent handles the knowledge base update task.

[0030] The auditors transmit the electronic transaction records provided by the client to the system via network communication with the server. After collaboration among multiple agents within the system, a client transaction record analysis report is generated.

[0031] In this embodiment, the large model can be any general-purpose large model, and the knowledge base can be any vector database. The specific implementation steps are as follows: I. Multi-role intelligent agents collaborate using a proprietary knowledge base to perform transaction flow insight and analysis tasks: During the data processing phase: Transaction classification is performed on the electronic documents of transaction records uploaded by users. The task execution workflow is as follows: 1) The planning agent first calls the document extraction tool to parse the document. Based on the large model context capacity, it creates n transaction classification tasks for transaction records and notifies the transaction classification agent to perform transaction classification; 2) The transaction classification agent, in conjunction with a proprietary knowledge base, classifies each transaction record based on the "counterpartie" field information in the transaction record and marks whether the current "counterpartie" appears in the knowledge base.

[0032] Specifically, a planning agent is used to parse the target transaction flow electronic document to obtain a transaction dataset containing several transaction records. A transaction classification task is then generated based on this dataset. Finally, a transaction classification agent is used to extract the counterparty fields from the transaction records in the transaction dataset based on the transaction classification task. This process can include: using the planning agent to call a document extraction tool to perform structured parsing of the target transaction flow electronic document, obtaining the parsing results, and extracting field information from the parsing results, including transaction number, transaction time, transaction type, payment / receipt direction, transaction method, transaction amount, counterparty, and merchant order number; constructing standardized transaction records based on preset field mapping relationships and field information, and generating a transaction dataset containing several transaction records; and generating a transaction classification task corresponding to each transaction record based on the transaction dataset, and using a transaction classification agent to extract the counterparty fields corresponding to each transaction record in the transaction dataset based on the transaction classification task.

[0033] Step S12: Call the large model and determine the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the transaction counterparty field. When the comprehensive classification confidence level is not less than the preset dynamic threshold, classify the transaction record into the corresponding category to obtain a classification result set.

[0034] In this embodiment, items that cannot be clearly categorized are classified into the "Other" category. The specific scheme for the classification task is as follows: The large model generates the probability of predicting the classification, and the final classification confidence score is obtained using the comprehensive confidence score formula: .

[0035] Specifically, calling the large model and determining the comprehensive classification confidence of each transaction record based on the current proprietary knowledge base and the counterparty field can include: determining the matching weight based on the matching relationship between the counterparty field and the current proprietary knowledge base, determining the semantic weight based on the business type keywords in the counterparty field, and then determining the frequency weight based on the frequency of the corresponding counterparty in the transaction dataset; calling the large model to determine the predicted probability of the transaction record belonging to each candidate category, and then determining the comprehensive classification confidence of the corresponding transaction record based on the predicted probability, matching weight, semantic weight, and frequency weight.

[0036] In this embodiment, the weighting factors need to be set from multiple dimensions. In one specific implementation, dimension 1 is: the matching type of the counterparty in the knowledge base: exact match: the counterparty name is completely consistent with the existing entries in the knowledge base, then the weighting factor Wmatch=0.95; fuzzy match: the similarity between the counterparty name and the knowledge base entries is >0.8 (e.g., “A milk tea shop” matches “A milk tea shop (a certain shop)”), then the weighting factor Wmatch=0.7; no match: the counterparty does not appear in the knowledge base, Wmatch=0.5.

[0037] Dimension 2 is: Semantic clarity of the counterparty: Merchant names containing explicit business type keywords (such as "hot pot," "supermarket," "hospital"), Wsematch=0.9; Merchant names containing some vague business type hints (such as generic terms like "food" and "service"): W sematch =0.7; Merchant names are personal nicknames or meaningless characters, Wsematch=0.5. Dimension 3 is: Transaction frequency and regularity: If the number of times the same counterparty appears (freq) ≥ 12, then Wfreq=0.9; if 5≤freq<12, then W... freq =0.7; if 1≤freq<5, then W freq =0.5, if freq=0, then Wfreq=0.

[0038] Where α is the balance coefficient (e.g., set to 0.7), controlling the weights of the model's prediction probability and feature matching quality, and P(y|x) is the highest probability that the large model predicts transaction record x belongs to category y. The lower the final confidence level, the more uncertain the model is. Example: If the final calculated result is " The confidence level of "Ice City" belonging to "Daily Consumption - Catering" is 0.9, which indicates high certainty. It does not need to be classified into the "Other" category, and therefore does not require manual labeling.

[0039] Specifically, determining the matching weight based on the matching relationship between the counterparty field and the current proprietary knowledge base can include: determining whether the counterparty name in the counterparty field is consistent with an entry in the current proprietary knowledge base; if the counterparty name in the counterparty field is completely consistent with an entry in the current proprietary knowledge base, then the matching weight is set as the first weight value; the current proprietary knowledge base has a tree-like hierarchical structure; the current proprietary knowledge base includes a first-level transaction category node and several second-level transaction category nodes corresponding to the first-level transaction category node; when the similarity between the counterparty name and an entry in the current proprietary knowledge base meets a preset similarity condition, the matching weight is set as the second weight value; when the counterparty name does not match an entry in the current proprietary knowledge base, the matching weight is set as the third weight value; wherein, the first weight value is greater than the second weight value, and the second weight value is greater than the third weight value.

[0040] In one specific implementation, when At that time, transaction records are divided into "Other" category, and the dynamic threshold function is as follows: ; Among them, Threshold_base (e.g., set to 0.8) is the base threshold to ensure the minimum quality requirements, β (0.2-0.5) is the accuracy influence coefficient, used to control the influence of historical accuracy on the threshold, and Accuracy(t-1) is the manually calculated classification accuracy of the previous round.

[0041] Example: If Threshold_base=0.8, the accuracy of the previous round=0.95, and β=0.3, then threshold_t=0.515, and more samples can be automatically classified without much manual intervention.

[0042] Secondly, after the transaction classification agent completes its task, the embodiment of this application sends the result to the planning agent. At the same time, the transaction classification agent sends the transaction records that do not appear in the knowledge base and the "other" category to the data access agent for database entry after deduplication based on the "transaction counterparty".

[0043] Table 1. Information on WeChat transaction history headers and database stored field names.

[0044] The data statistics phase involves using the planning agent to call the data statistics agent to obtain multi-dimensional statistical analysis reports. The task execution workflow is as follows: 1) The data statistics agent loads the categorized transaction records, calls the statistical tool, performs statistical analysis from multiple preset dimensions, and outputs a report with headers including: transaction category, counterparty, transaction direction, transaction month, and transaction amount; 2) After completing the task, the data statistics agent sends the results to the planning agent.

[0045] Specifically, when the overall classification confidence level is not less than a preset dynamic threshold, the transaction records are classified into the corresponding categories to obtain a classification result set. This may include: if the current transaction is in its first round of classification, the preset basic threshold is determined as the dynamic threshold; if the current transaction is not in its first round of classification, the classification accuracy corresponding to the previous round of classification is obtained, and an initial threshold is determined based on the classification accuracy, which is then set as the dynamic threshold corresponding to the current round of transaction classification; the overall classification confidence level is compared with the dynamic threshold; when the overall classification confidence level is not less than the dynamic threshold, the transaction records are classified into the corresponding categories to obtain a classification result set; when the overall classification confidence level is less than the dynamic threshold, the transaction records are classified into other categories to obtain an undetermined classification set.

[0046] Step S13: Utilize a data statistical agent to perform statistical analysis on the classification result set, determine a statistical analysis dataset including several statistical features, and utilize an insight reasoning agent to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset. Determine the evidence score based on the atomic evidence set, and aggregate each atomic evidence in the atomic evidence set based on the evidence score to obtain an intermediate conclusion set.

[0047] In this embodiment, during the insight reasoning stage, this application embodiment selects to utilize the planning agent to call the insight reasoning agent and obtain the final analysis report. The task execution workflow is as follows: 1) The insight reasoning agent loads the categorized transaction records and multidimensional statistical reports, executes the preset reasoning task, and outputs a profile analysis report; 2) After completing the task, the insight reasoning agent sends the results to the planning agent; 3) The planning agent displays the customer profile analysis report to the reviewers through the UI.

[0048] The insightful reasoning agent needs to perform multi-dimensional profiling analysis tasks: Life trajectory and consumption habit analysis: analyzing customers' daily consumption based on statistical reports; Special timeline analysis: including consumption records outside of residential areas, determining whether someone is an office worker or single / living alone based on consumption records; Social relationship analysis: analyzing possible relationship networks based on statistical reports; Specific potential risk analysis: analyzing risky transactions such as gambling, pawnbroking, POS machine cash-out, revolving accounts, and large-amount periodic repayments based on statistical reports and transaction details; Summarizing the profile: summarizing and analyzing the above to refine customer profiles and assess risk and income levels.

[0049] Furthermore, insightful reasoning agents need to perform analytical tasks through explicit chains of evidence: Explicit chains of evidence: Each chain of evidence is a directed acyclic graph (DAG) containing three core elements: atomic evidence, intermediate conclusions, and final conclusions. Atomic evidence is a single transaction or statistical feature in the flow of data; intermediate conclusions are stage-by-stage judgments derived from multiple atomic evidences; and the final conclusion is derived from the intermediate conclusions.

[0050] Subsequently, the evidence chain reasoning process is carried out: hypothesis generation -> evidence search -> evidence scoring -> positive and negative weighing -> conclusion confirmation, and the results are visualized: the evidence chain graph structure is presented in the form of a directed graph, where the node types include atomic evidence nodes, intermediate conclusion nodes and final conclusion nodes, and the edge types include support relationships and opposition relationships.

[0051] Evidence annotation: Hyperlinks are created for atomic evidence, allowing reviewers to interactively view the original transaction details in the generated report.

[0052] It is worth mentioning that the evidence score is calculated as follows: Atomic evidence score, which integrates three dimensions: Dimension 1: Strength of evidence, with a base score based on the type of evidence: Direct evidence (e.g., notes containing the word "gambling"): base score 0.9; Indirect evidence (e.g., high-frequency two-way trading): base score 0.6; Statistical characteristic evidence: base score 0.7; Contextual evidence (e.g., trading in the early morning): base score 0.4.

[0053] Dimension 2: Novelty of Evidence, the degree to which the evidence deviates from common patterns: whether the transaction amount deviates from the average level of this customer; whether the transaction time deviates from normal working hours, and the corresponding calculation formula is as follows: ; Where IQR is the interquartile range, e is a preset decimal to prevent calculation errors, and the higher the deviation, the higher the novelty score.

[0054] Dimension 3: Timeliness of Evidence: Recent transactions have higher weighting. ; Where t is the half-life parameter (which can be set to 90 days, meaning that the weight of evidence from 90 days ago decays to 37%).

[0055] Overall score of atomic evidence: ; The weights are w1=0.5, w2=0.3, and w3=0.2.

[0056] Intermediate conclusion scores are derived from the aggregate of supporting atomic evidence: ; Where S is the set of atomic evidence supporting the intermediate conclusion, O is the set of atomic evidence opposing the intermediate conclusion, |S| and |O| represent the number of supporting and opposing evidence, respectively, λ represents the total number of evidences participating in the vote, and is a normalization factor; λ is the penalty coefficient for opposing evidence (recommended 0.7, indicating that the weight of opposing evidence is lower than that of supporting evidence).

[0057] Specifically, a statistical analysis dataset is determined based on the classification result set, and an atomic evidence set is generated based on transaction records and corresponding statistical features. Evidence scores are then determined based on these atomic evidence sets. Finally, the atomic evidence sets are aggregated based on these scores to obtain an intermediate conclusion set. This process may include: using a data statistical agent to load transaction records from the classification result set and performing statistical analysis on these records according to transaction classification, counterparty, direction, month, and amount dimensions to obtain a statistical analysis dataset containing several statistical features; using an insight reasoning agent to construct an explicit evidence chain model based on the statistical analysis dataset and corresponding transaction records; the explicit evidence chain model is a directed acyclic graph (DAG); the DAG includes atomic evidence nodes, intermediate conclusion nodes, and final conclusion nodes; generating an atomic evidence set based on transaction records and corresponding statistical features, and then using the explicit evidence chain model to aggregate the atomic evidence... Evidence scores are determined based on set, strength of evidence, novelty of evidence, and timeliness of evidence. Strength of evidence characterizes the degree to which current atomic evidence supports the corresponding reasoning conclusion; novelty of evidence characterizes the degree of deviation of current atomic evidence from historical transaction patterns; and timeliness of evidence characterizes the degree of time decay of current atomic evidence from the present moment. Based on the evidence scores, a first set of atomic evidence supporting the target intermediate conclusion and a second set of atomic evidence opposing the target intermediate conclusion are determined from each atomic evidence in the atomic evidence set. Intermediate conclusion scores are determined based on the evidence scores corresponding to the first and second atomic evidence sets, and corresponding intermediate conclusions are generated based on these scores. A first set of intermediate conclusions supporting the target final conclusion and a second set of intermediate conclusions opposing the target final conclusion are determined based on the intermediate conclusions, and an intermediate conclusion set is generated based on the first and second intermediate conclusion sets.

[0058] Step S14: Generate corresponding target conclusions based on the intermediate conclusion set, and generate loan risk prediction results for auto finance based on the target conclusions; the loan risk prediction results include user consumption characteristics, behavior patterns and risk assessment information.

[0059] In this embodiment, the final conclusion score corresponding to this application embodiment is as follows: ; Where Cs is the set of intermediate conclusions that support the final conclusion, Co is the set of intermediate conclusions that oppose the final conclusion, Weight(c) is the weight of the intermediate conclusion (preset according to the conclusion type, such as "there is a gambling transaction counterparty" weight 0.4, "there is fund transfer" weight 0.3), and μ is the evidence coefficient (such as set to 0.5).

[0060] Specifically, the process of generating a target conclusion based on an intermediate conclusion set and then generating a loan risk prediction result for auto finance based on the target conclusion can include: determining the conclusion weights corresponding to each intermediate conclusion based on a first and second intermediate conclusion set, and determining the target conclusion score based on the conclusion weights and intermediate conclusion scores; generating a conclusion to be processed based on the target conclusion score, and generating a customer profile analysis result based on the target conclusion; then deduplicating the counterparty field of each transaction record in the undetermined classification set to obtain a candidate knowledge set; using a data access agent to obtain counterparty information corresponding to the candidate knowledge set, and using a knowledge base maintenance agent to update the current proprietary knowledge base based on the counterparty information to obtain a new current proprietary knowledge base; stopping the jump when a preset iteration termination condition is met; wherein, the preset iteration termination condition includes the undetermined classification set being empty or the number of iterations reaching a preset threshold; if the preset iteration termination condition is not met, then jumping back to call the large model and determining the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the counterparty field to obtain the target conclusion; and generating a loan risk prediction result including user consumption characteristics, behavioral patterns, and risk assessment information based on the target conclusion.

[0061] In this embodiment, the present application requires the construction and updating of a proprietary knowledge base, and the corresponding knowledge base update process is illustrated in the diagram below. Figure 3 As shown: First, the knowledge base is constructed: it is written in Markdown format, and knowledge is entered according to heading hierarchy. This facilitates parsing and paragraph semantic segmentation, and finally, it is vectorized and stored using a word embedding model.

[0062] It is worth mentioning that the knowledge base has a two-level tree structure. The first level is defined by human classification rules; the second level is defined by human initial examples, and the large model updates itself through reasoning. Alternatively, the model can manually mark the examples that cannot be classified, and the knowledge base can then maintain the intelligent agent for updates.

[0063] In this embodiment, the present application requires the construction of an initial primary transaction type classification based on human expert experience, including 7 categories: Table 2 Example Information Table of Transaction Types

[0064] In this embodiment, the knowledge base needs to be updated: The knowledge generation stage is as follows: When the transaction classification agent analyzes the "counterpartie" field of each transaction record, it classifies the transaction record by combining the knowledge at the current time. Those that cannot be clearly classified are classified into the "other" category and the "counterpartie" of the current transaction record is marked as being in the knowledge base. After deduplication based on the "counterpartie" field, the data storage knowledge base is called and the data is stored in the database.

[0065] The knowledge update phase is as follows: the planning agent creates knowledge base update tasks according to the set period, calls the data access agent to obtain transaction records that are not "other" categories, sends them to the knowledge base maintenance agent, which updates the corresponding secondary categories. After the task is completed, the planning agent calls the data access agent to mark the read transaction records as "updated".

[0066] Manual Update: The user requests transaction records in the "Other" category from the system through the UI interface. The planning agent calls the data access agent to retrieve the corresponding transaction records from the database and display them to the user. After the user categorizes the transactions, the planning agent sends all manually marked transaction records to the knowledge base maintenance agent, which updates the corresponding secondary categories. After completing the task, the planning agent calls the data access agent again to mark the retrieved transaction records as "updated".

[0067] In one specific implementation, single young people in small towns are: I. Lifestyle and Consumption Habits: 1) The core activity area is in City B of Province A. Basis: Local businesses frequently appearing in phrases such as "a hospital in City B", "District C of City B", and "District D of City B".

[0068] 2) He frequently eats out and orders takeout, spending mostly between 10 and 30 yuan a day. He also goes to internet cafes often, which is a typical lifestyle for young men.

[0069] II. Special Timeline Analysis: 1) Before December 2025, frequent consumption records in "City E" (such as "City E Rail Transit") indicate that their life trajectory may have migrated from City B to City E.

[0070] 2) Their dining and entertainment consumption times are not fixed, and they are not typical office workers.

[0071] III. Social Relations Analysis

[0072] 1) "Interpreting a way of life" is its core source of funding. It receives transfers of tens to hundreds of yuan almost daily. This is highly likely from its parents or guardians, providing for its daily living expenses.

[0073] 2) "Zuo": The person with the most frequent two-way fund flows. There are records of both receiving payments from and transferring large sums of money to the other party. The amounts range from a few yuan to several thousand yuan. The relationship could be that of close friends, business partners, or financial intermediaries.

[0074] 3) "Suiyuan Ruomeng", "Houmou", "Nanfeng", etc.: Important supplementary sources of funds. Transfers of tens to hundreds of yuan are made irregularly, which may be payments from other relatives and friends, or online / offline customers.

[0075] IV. Analysis of Specific Potential Risks: Account funds move frequently, but the balance is usually not high: large sums of money (such as 20,000, 10,000, 4,600 yuan, etc.) are quickly transferred in and then withdrawn through transfers, consumption or cash withdrawal.

[0076] There is no record of a fixed salary income. The main sources of income are bank transfers from others and QR code payments (suspected to be offline transactions).

[0077] There are signs of small-scale business: the amounts received and paid are similar (e.g., 200 yuan in income followed by 200 yuan in expenditure), and the counterparties are mostly individuals, with characteristics of mobile phone repair and digital products (e.g., "mobile phone recycling", "digital transactions", "mobile phone after-sales service").

[0078] 4) Frequent small-amount (a few cents to a few yuan) top-ups and withdrawals between WeChat Wallet and bank card (debit card number ending in 6773), with very irregular amounts (such as 0.37, 0.95, 1.01 yuan), may be due to testing the account or unintentional operation.

[0079] V. Summary Profile: A young, single male currently residing in City E, likely living alone, reliant on food delivery and ride-hailing services, and with regular expenses on gaming and entertainment. His financial resources may heavily depend on family support, and he is suspected of engaging in individual mobile phone repair / digital product repair to earn extra income. Overall, he lacks a stable source of income and experiences significant cash flow fluctuations.

[0080] The chain of evidence for the inference that "income depends on family support": 1) Atomic evidence: Transaction details: Daily transfer amount of 50-200 yuan for "Interpreting a Life", with remarks including "living expenses" (direct evidence strength 0.9); Statistical characteristics: Irregular wage entries (statistical evidence strength 0.7); Transaction details: There are sporadic QR code payment records (indirect evidence strength 0.6); Statistical characteristics: There are receipts with the note "mobile phone repair", but the amounts are very small (statistical evidence strength 0.7). 2) Interim conclusion: Irregular wage entries support evidence 1 and 2, and oppose evidence 3 and 4. Assume that only the above 4 pieces of evidence exist. Interim conclusion score: 1.6 - 0.7 1.3 = 0.69, and 0.7 is the penalty coefficient for opposing evidence.

[0081] 3) Final conclusion: Income source depends on family support, confidence level 0.86 (highly confident)

[0082] Final score: 0.69 / 0.8=0.86, where 0.8 is the preset weight for "lack of stable income".

[0083] In another specific implementation, the knowledge base of this application can be updated: Table 3 Knowledge Base Information Table

[0084] In the above transaction records, if the large model does not have background knowledge of the "transaction counterparty", the transaction records cannot be classified into the daily consumption - catering category. Manual classification is required, and then the system will trigger the corresponding update task to complete the update of the proprietary knowledge base.

[0085] As can be seen from the above, the embodiments of this application first require the use of a planning intelligent agent to parse the electronic document of the target transaction flow, obtaining a transaction dataset including several transaction records. A transaction classification task is then generated based on this dataset, and a transaction classification intelligent agent is used to extract the counterparty field of the transaction records from the dataset. Secondly, a large model is invoked, and the comprehensive classification confidence level corresponding to each transaction record is determined based on the current proprietary knowledge base and the counterparty field. When the comprehensive classification confidence level is not less than a preset dynamic threshold, the transaction records are classified into their corresponding categories, resulting in a classification result set. Then, a statistical analysis dataset is determined based on the classification result set, and an atomic evidence set is generated based on the transaction records and corresponding statistical features. Evidence scores are determined based on the atomic evidence set, and the atomic evidence in the atomic evidence set is aggregated based on the evidence scores to obtain an intermediate conclusion set. Finally, the intermediate conclusion set is derived to obtain the target conclusion, and a loan risk prediction result for auto finance is generated based on the target conclusion. The loan risk prediction result includes user consumption characteristics, behavioral patterns, and risk assessment information. In this way, the efficiency of predicting auto finance loan risk is improved during the auto finance loan risk prediction process, thereby enhancing the user experience.

[0086] Accordingly, see Figure 4 As shown, this application also provides an auto finance loan risk prediction device based on multi-role intelligent agents and a proprietary knowledge base, including: The counterparty field extraction module 11 is used to parse the target transaction flow electronic document using a planning agent to obtain a transaction dataset including several transaction records, generate a transaction classification task based on the transaction dataset, and extract the counterparty field of the transaction records in the transaction dataset using a transaction classification agent based on the transaction classification task; the target transaction flow electronic document is a transaction flow electronic document authorized by the user and obtained when processing an auto finance loan application; The classification result set generation module 12 is used to call the large model and determine the comprehensive classification confidence of each transaction record based on the current proprietary knowledge base and the transaction counterparty field. When the comprehensive classification confidence is not less than a preset dynamic threshold, the transaction record is classified into the corresponding category to obtain the classification result set. The intermediate conclusion set acquisition module 13 is used to perform statistical analysis on the classification result set using a data statistical agent, determine a statistical analysis dataset including several statistical features, and use an insight reasoning agent to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset, determine the evidence score based on the atomic evidence set, and aggregate each atomic evidence in the atomic evidence set based on the evidence score to obtain an intermediate conclusion set. The loan risk prediction result generation module 14 is used to generate corresponding target conclusions based on the intermediate conclusion set, and generate loan risk prediction results for auto finance based on the target conclusions; the loan risk prediction results include user consumption characteristics, behavior patterns and risk assessment information.

[0087] In some specific embodiments, the counterparty field extraction module 11 may specifically include: The parsing result determination unit is used to use the planning agent to call the document extraction tool to perform structured parsing on the target transaction flow electronic document, obtain the parsing result, and extract field information including transaction number, transaction time, transaction type, income and expenditure direction, transaction method, transaction amount, transaction counterparty and merchant order number from the parsing result; A transaction dataset generation unit is used to construct standardized transaction records based on a preset field mapping relationship and the field information, so as to generate a transaction dataset including several transaction records based on the transaction records; The counterparty field extraction unit is used to generate a transaction classification task corresponding to each transaction record based on the transaction dataset, so as to use a transaction classification agent and extract the counterparty field corresponding to each transaction record in the transaction dataset based on the transaction classification task.

[0088] In some specific embodiments, the classification result set generation module 12 may specifically include: The matching weight determination unit is used to determine the matching weight based on the matching relationship between the counterparty field and the current proprietary knowledge base, and to determine the semantic weight based on the business type keywords in the counterparty field, and then to determine the frequency weight based on the frequency of occurrence of the corresponding counterparty in the transaction dataset. The comprehensive classification confidence determination unit is used to call the large model to determine the predicted probability that the transaction record belongs to each candidate category, so as to determine the comprehensive classification confidence of the corresponding transaction record based on the predicted probability, the matching weight, the semantic weight and the frequency weight.

[0089] In some specific embodiments, the classification result set generation module 12 may specifically include: The first weight value determination unit is used to determine whether the name of the counterparty in the counterparty field is consistent with the entry in the current proprietary knowledge base. If the name of the counterparty in the counterparty field is completely consistent with the entry in the current proprietary knowledge base, the matching weight is set as the first weight value. The current proprietary knowledge base is a tree-like hierarchical structure. The current proprietary knowledge base includes a first-level transaction category node and several second-level transaction category nodes corresponding to the first-level transaction category node. The second weight value determination unit is used to set the matching weight as the second weight value when the similarity between the name of the counterparty and the entry in the current proprietary knowledge base meets the preset similarity condition. The third weight value determination unit is used to set the matching weight as the third weight value when the name of the counterparty to the transaction does not match an entry in the current proprietary knowledge base; wherein the first weight value is greater than the second weight value, and the second weight value is greater than the third weight value.

[0090] In some specific embodiments, the classification result set generation module 12 may specifically include: The dynamic threshold determination unit is used to determine the preset basic threshold as the dynamic threshold if the current transaction is the first round of transaction classification, and to obtain the classification accuracy corresponding to the previous round of transaction classification if the current transaction is not the first round of transaction classification, and to determine the initial threshold based on the classification accuracy, and to determine the initial threshold as the dynamic threshold corresponding to the current round of transaction classification. A sub-unit for generating the classification result set is used to compare the comprehensive classification confidence with the dynamic threshold. When the comprehensive classification confidence is not less than the dynamic threshold, the transaction record is classified into the corresponding category to obtain the classification result set. An undetermined classification set generation unit is used to classify the transaction records into other categories when the comprehensive classification confidence is less than the dynamic threshold, thereby obtaining an undetermined classification set.

[0091] In some specific embodiments, the intermediate conclusion set acquisition module 13 may specifically include: The transaction record loading unit is used to load the transaction records in the classification result set using the data statistical intelligence agent, and to perform statistical analysis on the transaction records according to the transaction classification dimension, transaction counterparty dimension, transaction direction dimension, transaction month dimension and transaction amount dimension to obtain a statistical analysis dataset including several statistical features; An explicit evidence chain model construction unit is used to construct an explicit evidence chain model using an insight reasoning agent and based on the statistical analysis dataset and corresponding transaction records; the explicit evidence chain model is a directed acyclic graph; the directed acyclic graph includes atomic evidence nodes, intermediate conclusion nodes, and final conclusion nodes; The evidence score determination unit is used to generate an atomic evidence set based on the transaction records and the corresponding statistical features, and to determine the evidence score based on the atomic evidence set, evidence strength, evidence novelty, and evidence timeliness using the explicit evidence chain model; wherein, the evidence strength is used to characterize the degree to which the current atomic evidence supports the corresponding reasoning conclusion; the evidence novelty is used to characterize the degree of deviation of the current atomic evidence from the historical transaction pattern; and the evidence timeliness is used to characterize the degree of time decay of the current atomic evidence from the current moment. An atomic evidence set generation unit is used to determine, based on the evidence score, a first atomic evidence set that supports the target intermediate conclusion and a second atomic evidence set that opposes the target intermediate conclusion among the atomic evidence in the atomic evidence set. The intermediate conclusion determination unit is used to determine the intermediate conclusion score based on the evidence score corresponding to the first atomic evidence set and the evidence score corresponding to the second atomic evidence set, and to generate the corresponding intermediate conclusion based on the intermediate conclusion score. The intermediate conclusion set generation unit is used to determine a first intermediate conclusion set that supports the target final conclusion and a second intermediate conclusion set that opposes the target final conclusion based on the intermediate conclusions, so as to generate an intermediate conclusion set based on the first intermediate conclusion set and the second intermediate conclusion set.

[0092] In some specific embodiments, the loan risk prediction result generation module 14 may specifically include: The conclusion weight determination unit is used to determine the conclusion weight corresponding to each intermediate conclusion based on the first intermediate conclusion set and the second intermediate conclusion set, so as to determine the target conclusion score based on each conclusion weight and the intermediate conclusion score. The candidate knowledge set generation unit is used to generate a conclusion to be processed based on the target conclusion score, and to generate a customer profile analysis result based on the target conclusion. Then, the counterparty field of each transaction record in the undetermined classification set is deduplicated to obtain the candidate knowledge set. The proprietary knowledge base update unit is used to obtain the counterparty information corresponding to the candidate knowledge set by the data access agent, and to update the current proprietary knowledge base based on the counterparty information by the knowledge base maintenance agent to obtain a new current proprietary knowledge base. The jump stops when a preset iteration termination condition is met; wherein, the preset iteration termination condition includes the undetermined classification set being empty or the number of iterations reaching a preset threshold. The target conclusion determination unit is used to, if the preset iteration termination condition is not met, jump back to the calling large model and determine the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the transaction counterparty field, so as to obtain the target conclusion; The loan risk prediction result generation subunit is used to generate loan risk prediction results, including user consumption characteristics, behavioral patterns and risk assessment information, based on the target conclusion.

[0093] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the auto finance loan risk prediction method based on multi-role intelligent agents and proprietary knowledge base disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0094] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0095] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0096] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the auto finance loan risk prediction method based on multi-role intelligent agents and proprietary knowledge base disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0097] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed method for predicting auto finance loan risk based on multi-role intelligent agents and a proprietary knowledge base. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0098] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0099] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0100] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0101] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 limitations, 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.

[0102] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting auto finance loan risk based on multi-role intelligent agents and a proprietary knowledge base, characterized in that, include: A planning agent is used to parse the target transaction flow electronic document to obtain a transaction dataset including several transaction records. A transaction classification task is generated based on the transaction dataset. A transaction classification agent is then used to extract the counterparty field of the transaction records in the transaction dataset based on the transaction classification task. The target transaction flow electronic document is a transaction flow electronic document authorized by the user and obtained when processing an auto finance loan application. The large model is invoked and the comprehensive classification confidence level corresponding to each transaction record is determined based on the current proprietary knowledge base and the transaction counterparty field. When the comprehensive classification confidence level is not less than a preset dynamic threshold, the transaction record is classified into categories to obtain a classification result set. The data statistical agent is used to perform statistical analysis on the classification result set to determine a statistical analysis dataset including several statistical features. The insight reasoning agent is used to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset. The evidence score is determined based on the atomic evidence set. The atomic evidence in the atomic evidence set is aggregated based on the evidence score to obtain an intermediate conclusion set. Based on the intermediate conclusion set, corresponding target conclusions are generated, and based on the target conclusions, loan risk prediction results for auto finance are generated; the loan risk prediction results include user consumption characteristics, behavioral patterns, and risk assessment information.

2. The method for predicting auto finance loan risk based on multi-role intelligent agents and proprietary knowledge base according to claim 1, characterized in that, The process involves using a planning agent to parse the electronic document containing the target transaction flow, obtaining a transaction dataset including several transaction records, generating a transaction classification task based on the transaction dataset, and using a transaction classification agent to extract the counterparty fields of the transaction records in the transaction dataset based on the transaction classification task, including: The planning agent uses a document extraction tool to perform structured parsing on the target transaction flow electronic document, obtains the parsing results, and extracts field information including transaction number, transaction time, transaction type, income and expenditure direction, transaction method, transaction amount, transaction counterparty and merchant order number from the parsing results; Standardized transaction records are constructed based on preset field mapping relationships and the field information, so as to generate a transaction dataset including several transaction records based on the transaction records; Based on the transaction dataset, a transaction classification task corresponding to each of the transaction records is generated, so as to utilize a transaction classification agent and extract the counterparty field corresponding to each transaction record in the transaction dataset based on the transaction classification task.

3. The method for predicting auto finance loan risk based on multi-role intelligent agents and proprietary knowledge base according to claim 1, characterized in that, The process of calling the large model and determining the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the counterparty field includes: The matching weight is determined based on the matching relationship between the counterparty field and the current proprietary knowledge base, the semantic weight is determined based on the business type keywords in the counterparty field, and the frequency weight is determined based on the frequency of the corresponding counterparty in the transaction dataset. The large model is invoked to determine the predicted probability that the transaction record belongs to each candidate category, and the comprehensive classification confidence of the corresponding transaction record is determined based on the predicted probability, the matching weight, the semantic weight, and the frequency weight.

4. The method for predicting auto finance loan risk based on multi-role intelligent agents and proprietary knowledge base according to claim 3, characterized in that, The step of determining the matching weight based on the matching relationship between the counterparty field and the current proprietary knowledge base includes: If the name of the counterparty in the counterparty field is consistent with an entry in the current proprietary knowledge base, and the counterparty name in the counterparty field is completely consistent with an entry in the current proprietary knowledge base, then the matching weight is set to the first weight value; the current proprietary knowledge base is a tree-like hierarchical structure; the current proprietary knowledge base includes a first-level transaction category node and several second-level transaction category nodes corresponding to the first-level transaction category node; When the similarity between the name of the counterparty and the entry in the current proprietary knowledge base meets the preset similarity condition, the matching weight is set to the second weight value; When the name of the counterparty to the transaction does not match an entry in the current proprietary knowledge base, the matching weight is set to a third weight value; wherein the first weight value is greater than the second weight value, and the second weight value is greater than the third weight value.

5. The method for predicting auto finance loan risk based on multi-role intelligent agents and proprietary knowledge base according to any one of claims 1 to 4, characterized in that, When the comprehensive classification confidence level is not less than a preset dynamic threshold, the transaction records are classified into categories to obtain a classification result set, including: If the current transaction is the first round of classification, the preset basic threshold is determined as the dynamic threshold. If the current transaction is not the first round of classification, the classification accuracy corresponding to the previous round of classification is obtained, and the initial threshold is determined based on the classification accuracy. The initial threshold is then determined as the dynamic threshold corresponding to the current round of transaction classification. The comprehensive classification confidence level is compared with the dynamic threshold. When the comprehensive classification confidence level is not less than the dynamic threshold, the transaction record is classified into the corresponding category to obtain a classification result set. When the overall classification confidence level is less than the dynamic threshold, the transaction record is classified into other categories, resulting in an undetermined classification set.

6. The method for predicting auto finance loan risk based on multi-role intelligent agents and proprietary knowledge base according to claim 5, characterized in that, The process involves using a data statistical agent to perform statistical analysis on the classification result set, determining a statistical analysis dataset including several statistical features, and using an insight reasoning agent to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset. Evidence scores are then determined based on these atomic evidence sets, and the atomic evidence in the atomic evidence set is aggregated based on the evidence scores to obtain an intermediate conclusion set, including: The transaction records in the classification result set are loaded using a data statistical agent, and statistical analysis is performed on the transaction records according to the dimensions of transaction classification, counterparty, direction, month, and amount to obtain a statistical analysis dataset including several statistical features. An explicit evidence chain model is constructed using an insight-based reasoning agent and based on the statistical analysis dataset and corresponding transaction records; the explicit evidence chain model is a directed acyclic graph; the directed acyclic graph includes atomic evidence nodes, intermediate conclusion nodes, and final conclusion nodes; A set of atomic evidence is generated based on the transaction records and the corresponding statistical features. The explicit chain of evidence model is used to determine the evidence score based on the set of atomic evidence, evidence strength, evidence novelty, and evidence timeliness. The evidence strength is used to characterize the degree to which the current atomic evidence supports the corresponding reasoning conclusion. The evidence novelty is used to characterize the degree of deviation of the current atomic evidence from the historical transaction pattern. The evidence timeliness is used to characterize the degree of time decay of the current atomic evidence from the current moment. Based on the evidence scores, determine the first set of atomic evidence supporting the target intermediate conclusion and the second set of atomic evidence opposing the target intermediate conclusion among the atomic evidence in the atomic evidence set; The intermediate conclusion score is determined based on the evidence score corresponding to the first atomic evidence set and the evidence score corresponding to the second atomic evidence set, and the corresponding intermediate conclusion is generated based on the intermediate conclusion score. Based on the intermediate conclusions, a first set of intermediate conclusions supporting the final target conclusion and a second set of intermediate conclusions opposing the final target conclusion are determined, and an intermediate conclusion set is generated based on the first set of intermediate conclusions and the second set of intermediate conclusions.

7. The method for predicting auto finance loan risk based on multi-role intelligent agents and proprietary knowledge base according to claim 6, characterized in that, The process of generating a corresponding target conclusion based on the intermediate conclusion set, and generating a loan risk prediction result for auto finance based on the target conclusion, includes: Based on the first set of intermediate conclusions and the second set of intermediate conclusions, the conclusion weights corresponding to each intermediate conclusion are determined, and the target conclusion score is determined based on each conclusion weight and the intermediate conclusion score. Based on the target conclusion score, a conclusion to be processed is generated, and a customer profile analysis result is generated based on the target conclusion. Then, the counterparty field of each transaction record in the undetermined classification set is deduplicated to obtain a candidate knowledge set. The data access agent obtains the counterparty information corresponding to the candidate knowledge set, and the knowledge base maintenance agent updates the current proprietary knowledge base based on the counterparty information to obtain a new current proprietary knowledge base. The jump stops when a preset iteration termination condition is met; wherein, the preset iteration termination condition includes the undetermined classification set being empty or the number of iterations reaching a preset threshold. If the preset iteration termination condition is not met, the process jumps back to the large model and determines the comprehensive classification confidence level of each transaction record based on the current proprietary knowledge base and the counterparty field, thus obtaining the target conclusion. Based on the aforementioned target conclusions, loan risk prediction results are generated, including user consumption characteristics, behavioral patterns, and risk assessment information.

8. A risk prediction device for auto finance loans based on multi-role intelligent agents and a proprietary knowledge base, characterized in that, include: The counterparty field extraction module is used to parse the target transaction flow electronic document using a planning agent to obtain a transaction dataset including several transaction records, generate a transaction classification task based on the transaction dataset, and extract the counterparty field of the transaction records in the transaction dataset using a transaction classification agent based on the transaction classification task; the target transaction flow electronic document is a transaction flow electronic document authorized by the user and obtained when processing an auto finance loan application; The classification result set generation module is used to call the large model and determine the comprehensive classification confidence level corresponding to each transaction record based on the current proprietary knowledge base and the transaction counterparty field. When the comprehensive classification confidence level is not less than a preset dynamic threshold, the module classifies the transaction record into the corresponding category to obtain the classification result set. The intermediate conclusion set acquisition module is used to perform statistical analysis on the classification result set using a data statistical agent, determine a statistical analysis dataset including several statistical features, and use an insight reasoning agent to generate an atomic evidence set based on the transaction records and the statistical features in the statistical analysis dataset, determine the evidence score based on the atomic evidence set, and aggregate each atomic evidence in the atomic evidence set based on the evidence score to obtain the intermediate conclusion set. The loan risk prediction result generation module is used to generate corresponding target conclusions based on the intermediate conclusion set, and to generate loan risk prediction results for auto finance based on the target conclusions; the loan risk prediction results include user consumption characteristics, behavioral patterns and risk assessment information.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the auto finance loan risk prediction method based on multi-role intelligent agents and proprietary knowledge base as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the auto finance loan risk prediction method based on a multi-role intelligent agent and a proprietary knowledge base as described in any one of claims 1 to 7.

Images