A method and device for intelligent completion, verification and fraud analysis of user addresses

Abstract

This invention discloses a method and device for intelligent address completion, verification, and fraud analysis. The method includes semantic parsing and structural reconstruction of unstructured address information input by the user, outputting standardized address data with a unified format; using a large language model to perform compliance verification and semantic cleansing on the standardized address data; collecting the user's dynamic operation trajectory and filling logs during the address input process, and combining user attributes, machine learning algorithms, and the large language model to mine abnormal behavior patterns. This invention combines multi-pattern string matching, large language model semantic understanding, and machine learning behavior modeling techniques, combined with the data and business characteristics of the consumer finance field, to construct a complete implementation solution for standardizing customer address information and identifying fraud in the consumer finance field, addressing challenges such as incomplete and non-standardized address information, difficulty in verifying authenticity, and strong concealment of fraudulent behavior in high-concurrency scenarios.

Description

Technical Field

[0001] This invention relates to the field of bank credit risk management technology, specifically to a method and device for intelligent user address completion, verification, and fraud analysis. Background Technology

[0002] In consumer credit, the highly diversified customer acquisition channels lead to significant differences in address input formats, verification standards, and filling specifications, resulting in inconsistent address information quality—a major pain point in the industry. Missing province / city / district boundaries, insufficient address accuracy, and inconsistent formats frequently cause customers to be unreachable. With a massive daily order volume in consumer credit, manual verification of massive amounts of customer data is impractical. Inaccurate address information directly leads to misjudgments in loan approval, failure to reach customers after loan disbursement, and obstacles in legal proceedings, increasing the risk of fraud by customers deliberately obscuring address information to evade responsibility. Currently, existing technologies are not comprehensive or refined enough in address integrity, authenticity, and behavioral correlation issues. They lack sufficient insight into the behavioral intent behind customer address entry and lack customer profiling analysis based on address information. In summary, existing technologies for address information processing suffer from weak structural completion capabilities, lack of semantic understanding, and insufficient utilization of behavioral features. Effectively addressing these issues can further improve credit risk control in high-concurrency, high-risk scenarios.

[0003] In current consumer credit systems, the automated processing of address information mainly relies on traditional text matching and rule validation techniques, but these methods have certain limitations in practical applications: 1) Common methods for extracting administrative division information such as provinces, cities, and districts include fuzzy matching, word segmentation matching, and regular expression matching. These methods typically rely on explicit keywords such as "province," "city," and "district / county" for identification and structured segmentation. However, in scenarios where users input information freely, these keywords are often omitted or abbreviated, leading to matching failures or incomplete identification, making it difficult to meet the robustness requirements of address structured completion.

[0004] 2) While existing multi-pattern string matching algorithms have improved matching capabilities to some extent, they rely on statically maintained administrative division dictionaries and require templates or manual intervention to synchronize with the latest administrative division changes, resulting in poor flexibility. Furthermore, these algorithms lack compatibility with place name abbreviations and colloquialisms, easily leading to false or missed matches.

[0005] 3) For granular verification of detailed address portions, traditional solutions often employ rule engines based on regular expressions. However, due to the highly diverse forms of address representation in China and the lack of a unified format standard, fixed rules are insufficient to cover all compliance patterns. This often results in valid addresses being misclassified as invalid or obviously abnormal addresses being missed as compliant, thus limiting the accuracy of verification.

[0006] 4) Existing technologies generally focus on the static verification of the address text itself. However, the behavioral characteristics of users during the filling process can, to some extent, help to judge the authenticity of the address and the credibility of the user. This is because customers with fraudulent tendencies often exhibit specific behavioral patterns in the address filling process, such as deliberately blurring the address granularity, avoiding filling in precise information, or frequently modifying or re-filling inconsistent content after address verification prompts. However, existing technologies lack full utilization of such behavioral characteristics. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and device for intelligent user address completion, verification, and fraud analysis.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a method for intelligent completion, verification, and fraud analysis of user addresses, comprising: Step 1: Perform semantic parsing and structure reconstruction on the unstructured address information input by the user, and output standardized address data with a unified format; Step 2: Use a large language model to perform compliance verification and semantic cleanup on the standardized address data; Step 3: Collect the user's dynamic operation trajectory and filling log during the address input process, and combine user attributes, machine learning algorithms and large language models to mine abnormal behavior patterns.

[0009] Furthermore, step 1 specifically includes: Based on the latest national administrative division standard data, a localized geographic information knowledge base was established, and commonly used abbreviations, colloquial names, aliases and dialect expressions of various regions were collected to establish a mapping relationship table between standard names and non-standard expressions. An enhanced multi-pattern string matching algorithm is used to segment and recognize the address information input by the user, extract provincial, municipal and district-level administrative division information and detailed address content, and then parse and complete the extracted data to generate structurally complete standard administrative division information, which is then concatenated with the original detailed address to form standardized address data in a unified format. Based on standardized address data, the corresponding administrative division code is automatically matched and generated.

[0010] Furthermore, the specific method for parsing and completing the extracted data is as follows: For addresses with incomplete third-level regional information, a hierarchical deduction mechanism from smallest to largest is used to complete them; for addresses that do not contain explicit keywords, the administrative division level is determined by contextual semantics and location priority; for cases where place names are abbreviated or commonly used, association parsing is performed in conjunction with the mapping relationship table.

[0011] Furthermore, it also includes: designing a scheduled task mechanism to periodically connect to the internet to obtain official updated data and third-party geographic service interface information, so as to realize the automated updating and dynamic maintenance of the geographic information knowledge base.

[0012] Furthermore, step 2 specifically includes: To address the different business uses of different address types, a multi-dimensional verification strategy is implemented, as follows: For residential addresses, details are required down to the unit / room level; For company or work addresses, details such as street number and specific company / organization name are required. Based on the address type label, the corresponding verification prompt word template is dynamically loaded and the prompt word is filled in. This drives the large language model to determine whether the current address meets the required granularity and returns specific missing item prompts for addresses that do not meet the conditions. A special address naming rule knowledge base is constructed, which includes typical non-standard but effective address expression patterns. Combined with retrieval enhancement generation technology, before performing semantic analysis on the large model, the naming rules and historical cases related to the current address's region are retrieved from the special address naming rule knowledge base, and contextual information is injected into prompt words. By using a large language model, semantic-level cleanup and completion of addresses are performed to identify and eliminate redundant information, clean up illegal characters, and standardize the representation of common names.

[0013] Furthermore, step 3 specifically includes: Record the user's multiple modifications and resubmissions after address verification failure, and compare whether the administrative division has changed across regions or changed frequently based on a multi-pattern string matching algorithm; at the same time, call the large language model to perform semantic similarity analysis on the detailed address to determine whether there is a significant logical deviation or detailed address drift before and after the modification. If so, construct semantic drift features and address drift features respectively. Collect input behavior data, construct a multi-dimensional feature system including time, interaction and feedback, and introduce correlation analysis between user age information and operation rate and change frequency to identify features that do not match age; A machine learning model is constructed, in which the machine learning module outputs an address to fill in a behavioral risk score based on the semantic drift feature, address drift feature, and age-inconsistent feature, for credit decision support.

[0014] Furthermore, by archiving high-risk samples and combining them with actual post-loan performance for tag feedback, the machine learning model parameters are updated regularly to achieve continuous evolution and adaptive optimization of behavior recognition capabilities.

[0015] In a second aspect, the present invention provides a device for intelligent completion, verification and fraud analysis of user addresses, including a storage medium and a processor, wherein the storage medium stores a computer program, which, when executed by the processor, is used to implement the above-described method.

[0016] Beneficial Effects: This invention combines multi-pattern string matching, large language model semantic understanding, and machine learning behavior modeling techniques, along with data and business characteristics specific to the consumer finance sector, to construct a comprehensive solution for standardizing customer address information and identifying fraud in the consumer finance field. This solution addresses challenges such as incomplete, non-standardized, and difficult-to-verify address information, as well as the high concealment of fraudulent activities, under high-concurrency scenarios. Specific advantages include: (1) An intelligent completion mechanism based on a dynamic and scalable regional address knowledge base is proposed. By integrating national standard administrative division data with non-standard expressions such as colloquialisms, aliases, and abbreviations collected from the Internet, and combining them with a timed update mechanism, a high coverage recognition rate of information at the provincial, municipal, and district levels is achieved. To address the common problems of default keywords and place name abbreviations in user-input addresses, a hierarchical derivation and missing character matching strategy is adopted to significantly improve the structured conversion capability of unstructured addresses and effectively solve the matching failure problem caused by the reliance on explicit keywords in traditional methods.

[0017] (2) A large language model is introduced to perform semantic-level compliance verification of detailed addresses, breaking through the limitations of traditional regular expression rules in terms of address granularity, regional differences and expression diversity. By setting up multi-scenario differentiated verification strategies and combining RAG technology to call a special address naming logic knowledge base, accurate identification of complex naming patterns in regions and urban and rural areas can be achieved, improving the adaptability and accuracy of address verification.

[0018] (3) Design a closed-loop correction mechanism of "semantic purification + intelligent feedback". Use the preliminary standardized address output by the large language model in the previous steps to clean up redundancy, unify the format and fix typos, eliminate the problems of duplicate provinces, cities and districts and detailed addresses, and illegal characters mixed in, and ensure that the address text is concise and consistent. For seriously abnormal or unparseable input, generate suggested address and feed it back to the front-end interface to guide the user to confirm or modify, forming an interactive closed loop of "system verification - intelligent suggestion - user confirmation", which takes into account both the efficiency of automated processing and user experience.

[0019] (4) Comprehensively collect operation trajectory data such as input duration, field modification frequency, and number of re-entry times, and combine this with user age-related operation familiarity analysis to establish a hierarchical behavior judgment benchmark. At the same time, establish an address behavior risk identification model based on machine learning, integrate multi-dimensional features such as semantic drift, address drift, and operation behavior matching degree to model and output "address entry behavior risk score" to assist credit decision-making. The system supports tag feedback and model iteration through post-loan actual performance data, realizing the improvement from static verification to dynamic intent recognition, enhancing the forward-looking identification capability of hidden fraud behaviors such as deliberate ambiguity and evasion of responsibility, and improving the level of risk control intelligence of financial institutions in high-concurrency credit scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method for intelligent user address completion, verification, and fraud analysis according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0022] like Figure 1 As shown, this embodiment of the invention provides a method for intelligent user address completion, verification, and fraud analysis, including: Step 1: Perform semantic parsing and structured reconstruction on the unstructured address information input by the user to achieve intelligent identification and completeness of provincial, municipal, and district-level administrative divisions, outputting standardized address data with a unified format and clear hierarchy. Specifically, this includes: 1. Construct a dynamically scalable regional address knowledge base Based on the latest national administrative division standards, a localized geographic information knowledge base will be established, covering multiple levels of administrative units, including provinces, cities, districts (counties), and townships, along with their historical change records. Furthermore, using web crawling technology, commonly used abbreviations, colloquialisms, aliases, and dialect expressions from various regions will be collected to establish a mapping table between standard names and non-standard expressions. Simultaneously, a scheduled task mechanism can be designed to periodically access officially updated data and third-party geographic service interface information, enabling automated updates and dynamic maintenance of the address knowledge base, ensuring comprehensive and timely address coverage.

[0023] 2. Multi-pattern string matching and address structured parsing Based on the aforementioned dynamically expandable regional address knowledge base, an enhanced multi-pattern string matching algorithm is employed to segment and recognize patterns in the original address text, accurately extracting the provincial, municipal, and district-level administrative division information and detailed address content. Specific processing logic is designed to address common issues: for addresses with incomplete third-level regional information, a hierarchical derivation mechanism is used to back-complete the higher-level administrative affiliation (municipal, provincial) based on the identified smaller regions (e.g., districts); for addresses lacking explicit keywords such as "province," "municipal," and "district," default keyword matching is implemented, determining the administrative division level through contextual semantics and location priority; for cases involving abbreviations or colloquial names, association parsing is performed using an extended mapping table. After parsing and completion, a complete and hierarchically clear standard administrative division information is generated and concatenated with the original detailed address to form standardized address data in a unified format.

[0024] 3. Region code generation and reverse lookup Based on the standardized administrative division results, the corresponding administrative division code (based on the national standard GB / T 2260 code) is automatically matched and generated as the basic data for subsequent address verification and external system integration, supporting functions such as regional risk label matching.

[0025] Step 2: Utilize the semantic understanding capabilities of large language models to perform compliance verification and semantic cleansing on the standardized address data, further improving the accuracy and business applicability of address information. This focuses on resolving issues that traditional rule engines struggle to handle, such as semantic ambiguity, regional differences, and redundant expressions, ensuring that addresses meet downstream business requirements such as post-loan outreach and protocol filling. Specifically, this includes: 1. Multi-scenario differentiated granularity verification To address the varying business purposes of different address types, a multi-dimensional verification strategy is implemented: for residential addresses, verification must be detailed down to the unit / room level; for work or business addresses, verification must be detailed down to the street number and the specific company / organization name. Based on the address type tag, the corresponding verification prompt template is dynamically loaded and filled in, driving the large language model to determine whether the current address meets the required granularity. For addresses that do not meet the conditions, specific missing item prompts are returned, achieving precise guidance.

[0026] 2. Regional Adaptability Verification Based on RAG Enhancement Considering the significant differences in address naming habits between urban and rural areas and between regions in China, this step aims to avoid misjudgments due to regional specificities. It constructs a special address naming rule knowledge base, including typical non-standard but effective address expression patterns. Combining retrieval enhancement generation techniques, before the large model performs semantic analysis, it first searches the special address naming rule knowledge base for naming rules and historical cases related to the current address's location, injecting contextual information into prompt words. This assists the model in more accurately judging the rationality and accessibility of addresses, improving its compatibility with special scenarios.

[0027] 3. Address semantic cleanup and intelligent feedback correction For non-standard user addresses that cannot be identified in step 1 (such as typos, homophones, or duplicate province / city / district addresses and garbled characters), this step utilizes the contextual understanding capabilities of a large language model to perform semantic-level cleanup and completion of the addresses. It identifies and eliminates redundant information, removes illegal characters, and standardizes generic terminology (such as "lane," "alley," "number," and "room") to ensure concise, accurate, and consistent addresses. The processing results are fed back to the front end, prompting users to confirm or modify, forming a closed-loop mechanism of "system verification—intelligent suggestion—user confirmation," improving the accuracy of address collection and the user experience.

[0028] Step 3: Collect the user's dynamic operation trajectory and filling log during the address input process, and combine this with user attributes, machine learning algorithms, and large language models to mine abnormal behavior patterns, assisting in the identification of potential fraudulent intentions. Details are as follows: 1. Address refilling behavior and semantic consistency analysis Record the user's multiple modifications and resubmissions after address verification failure. Use a multi-pattern string matching algorithm to compare whether the administrative division (province, city, district) has changed across regions or changed frequently. At the same time, call a large language model to perform semantic similarity analysis on the detailed address to determine whether there is a significant logical deviation or detailed address drift before and after the modification. If so, construct semantic drift features and address drift features respectively.

[0029] 2. Multidimensional behavioral feature extraction and population attribute adaptation modeling Input behavior data such as input duration, field modification frequency, page navigation, and paste operation are collected to construct a multi-dimensional feature system that includes time, interaction, and feedback categories. Furthermore, the correlation analysis between user age information and operation speed and change frequency is introduced to identify characteristics that are inconsistent with age, such as elderly customers "completing the form in a very short time" and "repeatedly verifying".

[0030] 3. Fraudulent Intent Recognition Model and Closed-Loop Optimization Mechanism A lightweight machine learning model is constructed. The machine learning module outputs a behavioral risk score based on the aforementioned semantic drift features, address drift features, and age-inconsistent features, using the address for credit decision support. Furthermore, high-risk samples can be archived and their post-loan performance (such as loss of contact, delinquency, fraud, etc.) can be used for label feedback. The machine learning model parameters are then updated periodically to achieve continuous evolution and adaptive optimization of behavioral recognition capabilities.

[0031] Based on the above embodiments, those skilled in the art can easily understand that the present invention also provides a device for intelligent completion, verification and fraud analysis of user addresses, including a storage medium and a processor. The storage medium stores a computer program, which is executed by the processor to implement the above-described method.

[0032] The following will use the consumer credit sector as an example for a detailed explanation: 1. User front-end operation The user front-end provides a page for filling in user information and address, supporting users to complete forward path operations such as address entry, submission, and operation record collection. It also has feedback path functions such as message prompts and intelligent suggestions to achieve two-way interaction.

[0033] 1.1 User Address Input: Users fill in their address information through the consumer credit APP, H5 page, or third-party partner platform. Different channels use adapted address information input controls. Users can select administrative division information such as province, city, and district through map selection or drop-down menu, and manually enter detailed address content such as street and house number. The front-end system automatically concatenates and generates complete address text, such as "Room 303, Building 2, No. 1, XX Road, Jianye District, Nanjing City, Jiangsu Province".

[0034] 1.2 User input behavior log collection: Deploy data tracking points in the address input area of ​​the application front end to collect user operation behavior records within the system page within the scope of compliance (in accordance with privacy protection and data security specifications), including: the time taken for the first input, the number of times the field was modified, whether pasting was used, the duration of page stay and other key behavioral indicators.

[0035] 1.3 Address Information Submission: After the user clicks the submit button, the front end packages the original address text and the collected behavior logs, and transmits them to the backend processing engine in real time through the API interface, triggering the subsequent address standardization parsing and risk analysis process.

[0036] 1.4 Address Verification Feedback: When address verification fails, the backend service will push a verification failure message to the client's frontend and return the reason for the verification failure, guiding the user to re-enter the address; when the user's address is ambiguous, after standardization processing, the result will be pushed to the frontend to guide the user to confirm or re-enter the address.

[0037] 2. Administrative Division Analysis and Standardization Completion This stage involves constructing a standard and extended address knowledge base, combined with an efficient string matching algorithm, to perform hierarchical identification, regional mapping, and structural reconstruction of unstructured addresses input by users, thereby achieving accurate parsing of administrative divisions and automatic completion of national standard codes.

[0038] 2.1 Acquisition of Address Information and Operation Log Information: The backend engine of the credit system receives customer information submitted by users, and parses, transforms and standardizes the address information and operation log records for use in subsequent processes.

[0039] 2.2 Loading of Standard Administrative Division Knowledge Base: Based on the standard coding of administrative divisions in GB / T 2260, maintain the standard administrative division knowledge base, load the knowledge base data into shared memory, and provide matching calls for subsequent algorithms.

[0040] 2.3. Loading of Extended Address Knowledge Base: Based on network data sources and through web crawling technology, commonly used abbreviations, colloquialisms, aliases and dialect expressions of various regions are collected, updated and maintained regularly to establish a mapping relationship between standard names and non-standard expressions, forming an extended address knowledge base to provide matching and calling for subsequent algorithms.

[0041] 2.4 Administrative Division Resolution: Based on the AC algorithm principle, using the aforementioned extended backend standard address, a Trie tree is constructed, and a "finite state automaton" and failure pointer mechanism are built to perform word segmentation and pattern recognition on the original address text, accurately dividing it into the provincial, municipal, and district-level administrative division parts and the detailed address part, as shown in the following example: Example 1: The user inputs the address "Room 303, Building 2, No. 1, XX Road, Jianye District". The algorithm, based on the loaded standard address knowledge base, matches "Jianye District" as a third-level administrative division, and then automatically reverses the process to deduce its parent city "Nanjing City" and "Jiangsu Province", thus completing the three-level structure.

[0042] Example 2: “No. 1, XX Village, Shuangjiang County”. The algorithm is based on the loaded extended address knowledge base. It identifies “Shuangjiang County” as the abbreviation of “Shuangjiang Lahu, Wa, Blang and Dai Autonomous County” through the mapping table, thereby completing the standardization and completion, and deducing its corresponding “Lincang City” and “Yunnan Province”.

[0043] 2.5 User Address Standardization Reconstruction and Regional Code Reverse Lookup: After parsing, a standard structured address is generated, such as "Room 303, Building 2, No. 1, XX Road, Jianye District, Nanjing City, Jiangsu Province", and a national standard code (such as 320105) is automatically added for temporary storage for use in subsequent processes.

[0044] 3. Address compliance verification and semantic cleanup Based on Retrieval Enhanced Generation (RAG) technology and the semantic understanding capabilities of large language models, the standardized addresses undergo multi-dimensional compliance verification and semantic-level cleaning, improving the system's tolerance for fuzzy expressions, typos, and region-specific formats, thus ensuring the accuracy and accessibility of address information.

[0045] 3.1 Address Naming Knowledge Base Retrieval: An address naming knowledge base is constructed, including common non-standard but effective address expression patterns (such as colloquialisms, aliases, and dialect expressions) from urban and rural areas. This knowledge base combines the semantic similarity retrieval capabilities of vector databases (such as ChromaDB) with the full-text matching capabilities of unstructured databases (such as ElasticSearch) to achieve efficient retrieval of multimodal address information. This knowledge base serves as the foundation for Retrieval Enhancement Generation (RAG) technology. Before address verification in the large model, context-related historical cases and regional naming rules are dynamically injected to improve the understanding and accuracy of complex and ambiguous addresses.

[0046] 3.2 Validation Rule Prompt Generation: Based on address type tags such as "residential address," "work address," and "mailing address," and combined with the granular requirements of credit business for different address types, differentiated validation prompt templates are preset. During the inference phase, the system automatically matches the corresponding template through the address type classification results, and integrates the search results from the address naming knowledge base and the actual input content of the user to dynamically fill in contextual information and generate appropriate prompts to drive the large language model for accurate validation.

[0047] Example: For residential addresses, the details should be refined to the level of locatable information such as house number, building, and unit. The prompt template is set as follows: "...If the detailed address includes locatable information such as building number, unit number, and house number, it is likely to be considered a compliant address. Note: Some addresses may omit explicit keywords such as 'room', 'number', and 'building'. Inferences should be made based on the context and residential address naming conventions. Please refer to the area rule information {CONTEXT_RAG_RESULT} retrieved from the address naming knowledge base to determine whether the address has actual reachability and reasonableness...".

[0048] 3.3 Large Language Model Address Compliance Verification: Based on business verification requirements, the compliance judgment standard knowledge, output format specifications and examples are embedded into the prompt word template. The general large language model interface (such as Qwen3) in private domain deployment or cloud service mode is called to perform multi-dimensional compliance judgment on the detailed address. The model output is required to follow the predefined JSON format, including the verification status and explanatory information, so as to facilitate automatic system parsing and subsequent processing.

[0049] Example 1: The input residential address is the city address: "No. 10, XX Road, Qinhuai District, Nanjing City, Jiangsu Province". The model output is: {"status":"N";"msg":"Specific building and house number are missing";}; Example 2: The input address is a rural address: "Group 3, XX Village, Jiangning District, Nanjing City, Jiangsu Province". Based on the RAG search results "rural areas commonly use the 'village + group' structure as an effective location unit", the model identifies that the address conforms to rural naming habits and is reachable, and determines it to be compliant. The model output is: {"status":"Y";"msg":"Passed";}.

[0050] 3.4 Address Semantic Cleaning: Addresses containing typos, homophones, duplicate provinces / cities / districts, special symbols, or garbled characters in non-standard user addresses are cleaned semantically using the contextual understanding capabilities of a large language model and knowledge base retrieval to construct prompt words. This process identifies and eliminates redundant information and removes illegal characters. If conflicting or ambiguous information is found in the address, feedback prompts are generated to guide the user to confirm or re-enter the address.

[0051] Example 1: When the address contains non-conflicting information and special characters, such as: "No. 1, Building 1, Unit 1-202, First Avenue, Qinhuai District, Nanjing City, Jiangsu Province", the model output is: {"status":"Y";"msg":"Standardization successful";stdAddr:"No. 1, Building 1, Unit 1, Room 202, First Avenue, Qinhuai District, Nanjing City, Jiangsu Province"}.

[0052] Example 2: When there is conflicting address information, such as "No. 1, Building 1, No. 1, First Avenue, Jiangning District, Nanjing City, Jiangsu Province", the model output is: {"status":"N";"msg":"Standardization failed. There is conflicting area information 'Qinhuai District' and 'Jiangning District' in the address. We guess your address is 'Jiangning District'. Please confirm the address or re-enter it!";stdAddr:"No. 1, Building 1, No. 1, First Avenue, Jiangning District, Nanjing City, Jiangsu Province"}.

[0053] 3.5 Address Information and Operation Log Storage: Associate the original address, address type, standardized address, verification result, intermediate results (decomposed province, city, district and area codes, etc.) with the customer number and save them to a structured database.

[0054] 4. User behavior modeling and fraud risk identification stage By integrating semantic analysis and quantitative behavioral features, a dynamic risk identification model based on machine learning is constructed. By modeling information from user operations and log entries, an interpretable behavioral risk score is output, and the model can be continuously iterated and optimized to achieve intelligent inference of fraudulent intent.

[0055] 4.1 Extraction of semantic features of user address operation behavior: Based on the large language model, semantic-level comparative analysis is performed on the multiple modification records generated by the user during the address filling process. By calculating the semantic similarity of the address before and after the modification, the regional jump characteristics are reflected, and the feature values ​​of "type migration" and "semantic drift" are calculated.

[0056] Example 1: The user changed the address from "Building A, Zhongguancun Software Park" to "Wudaokou Internet Cafe". Although both are locations in Haidian District, Beijing, the functional attributes are significantly different (office area → public place). The big model judges from the context that the user has a tendency to avoid responsibility or hide the real place of residence, and assigns a high score to the "type migration" feature.

[0057] Example 2: The user re-entered the address from "Sunshine Community Building 5 Unit 1" to "Moonlight Community Building 8 Unit 2 No. 101". There is a drift in the community name, building number and unit number, which is given a high score by the "semantic drift" feature.

[0058] 4.2 User Address Operation Behavior (Quantitative) Feature Extraction: Fine-grained operation behavior data is collected during the user's input process through front-end event tracking. A multi-dimensional quantitative feature system covering time-based, interaction-based, and feedback-based data is constructed, including: total input time, field modification frequency, paste / fill behavior, etc. This is combined with user attributes (such as age) for correlation modeling, establishing hierarchical judgment benchmarks, and calculating features such as "regional jumps," "time anomalies," and "method anomalies." Example 1: If a user modifies or re-enters their address 5 times within 3 minutes, and the province, city, and district are different each time (e.g., from "Beijing Chaoyang" → "Shanghai Pudong" → "Guangzhou Tianhe"), then the "regional jump" feature is given a high score. Example 2: If an elderly user (age > 50 years old) completes the entire process within 5 seconds with zero modifications, which deviates significantly from the average time taken by the group (> 30 seconds), then the "time anomaly" feature is given a high score; Example 3: If a user quickly fills all address fields by pasting without manual input, and the "number of pasting operations" matches the "number of fields" perfectly, and the user's IP address is located in a high-risk area, then the "abnormality of method" feature is given a high score.

[0059] 4.3 Dynamic Optimization of the Correlation Model Between Operational Behavior and Fraudulent Intent: An iterative behavioral risk identification model training mechanism is constructed. User address operational behavior samples are periodically extracted from the production environment and aligned with actual risk scores calculated from post-loan performance labels (such as delinquency level, loss of contact status, and confirmed fraud cases) to form a high-quality training dataset. Supervised learning methods are employed, based on high-performance, interpretable machine learning models such as XGBoost, to train the mapping relationship between "behavioral features → fraudulent intent," continuously optimizing model parameters and feature weights. Monthly / quarterly model version updates are supported, and A / B testing is used to verify the effectiveness, ensuring the model possesses foresight and adaptability.

[0060] 4.4 Model Invocation and Address Operation Behavior Risk Score Calculation: In the real-time processing flow, the pre-trained behavioral risk identification model is invoked, and the current user's multi-dimensional behavioral feature vector (including the aforementioned feature dimensions such as "type migration," "semantic drift," "regional jump," "time anomaly," and "method anomaly") is input. A standardized "address entry behavior risk score" (0–100 points) is output. This score reflects the degree of anomaly exhibited by the user during address entry and the potential for fraud. The score can be flexibly integrated into the risk control rule engine, supporting differentiated strategy configurations.

[0061] Example: Table 1 shows an example of calculating the risk score for address operation behavior. The "weights" and "offsets" are continuously optimized and updated through the model training mechanism described in Section 4.3; the "offset" is generated by fusing multiple training parameters, comprehensively reflecting the interaction relationships, nonlinear correlations, and joint risk tendencies between different features, embodying the deep risk-additive effect inherent in the feature combination.

[0062] Table 1: Example Table for Calculating Risk Score of Address Operation Behavior ; Based on the final calculated address, fill in the behavioral risk score and implement a tiered response strategy: 0–30 points: Normal behavior, automatically passed; 31–75 points: Medium risk, records kept for reference in credit approval models; >75 points: High risk. The score will be pushed to the manual review queue and will be used as a strong risk signal in the credit decision engine, triggering additional verification or denial of access.

[0063] 4.5 Processing Result Archiving and Data Feedback: All data from the entire address processing process will be archived, including: original input addresses, standardized completion results, detailed address semantic verification logs, user operation behavior characteristics, and behavioral risk scores. All data is indexed by user ID and timestamp, supporting subsequent post-loan tracking, audit retrospection, model feature retrospection, and tag feedback. This also provides a high-quality data foundation for subsequent model iterations, forming a complete data loop of "collection → analysis → decision-making → feedback → optimization."

[0064] The above description is merely a preferred embodiment of the present invention. It should be noted that for those skilled in the art, other parts not specifically described are existing technology or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent completion, verification, and fraud analysis of user addresses, characterized in that, include: Step 1: Perform semantic parsing and structure reconstruction on the unstructured address information input by the user, and output standardized address data with a unified format; Step 2: Use a large language model to perform compliance verification and semantic cleanup on the standardized address data; Step 3: Collect the user's dynamic operation trajectory and filling log during the address input process, and combine user attributes, machine learning algorithms and large language models to mine abnormal behavior patterns.

2. The method for intelligent completion, verification, and fraud analysis of user addresses according to claim 1, characterized in that, Step 1 specifically includes: Based on the latest national administrative division standard data, a localized geographic information knowledge base was established, and commonly used abbreviations, colloquial names, aliases and dialect expressions of various regions were collected to establish a mapping relationship table between standard names and non-standard expressions. An enhanced multi-pattern string matching algorithm is used to segment and recognize the address information input by the user, extract provincial, municipal and district-level administrative division information and detailed address content, and then parse and complete the extracted data to generate structurally complete standard administrative division information, which is then concatenated with the original detailed address to form standardized address data in a unified format. Based on standardized address data, the corresponding administrative division code is automatically matched and generated.

3. The method for intelligent completion, verification, and fraud analysis of user addresses according to claim 2, characterized in that, The specific method for parsing and completing the extracted data is as follows: For addresses with incomplete third-level regional information, a hierarchical deduction mechanism from smallest to largest is used to complete them; for addresses that do not contain explicit keywords, the administrative division level is determined by contextual semantics and location priority; for cases where place names are abbreviated or commonly used, association parsing is performed in conjunction with the mapping relationship table.

4. The method for intelligent completion, verification, and fraud analysis of user addresses according to claim 2, characterized in that, Also includes: Design a scheduled task mechanism to periodically connect to the internet to obtain official updated data and third-party geographic service interface information, thereby realizing the automated updating and dynamic maintenance of the geographic information knowledge base.

5. The method for intelligent completion, verification, and fraud analysis of user addresses according to claim 1, characterized in that, Step 2 specifically includes: To address the different business uses of different address types, a multi-dimensional verification strategy is implemented, as follows: For residential addresses, details are required down to the unit / room level; For company or work addresses, details such as street number and specific company / organization name are required. Based on the address type label, the corresponding verification prompt word template is dynamically loaded and the prompt word is filled in. This drives the large language model to determine whether the current address meets the required granularity and returns specific missing item prompts for addresses that do not meet the conditions. A special address naming rule knowledge base is constructed, which includes typical non-standard but effective address expression patterns. Combined with retrieval enhancement generation technology, before performing semantic analysis on the large model, the naming rules and historical cases related to the current address's region are retrieved from the special address naming rule knowledge base, and contextual information is injected into prompt words. By using a large language model, semantic-level cleanup and completion of addresses are performed to identify and eliminate redundant information, clean up illegal characters, and standardize the representation of common names.

6. The method for intelligent completion, verification, and fraud analysis of user addresses according to claim 1, characterized in that, Step 3 specifically includes: Record the user's multiple modifications and resubmissions after address verification failure, and compare whether the administrative division has changed across regions or changed frequently based on a multi-pattern string matching algorithm; at the same time, call the large language model to perform semantic similarity analysis on the detailed address to determine whether there is a significant logical deviation or detailed address drift before and after the modification. If so, construct semantic drift features and address drift features respectively. Collect input behavior data, construct a multi-dimensional feature system including time, interaction and feedback, and introduce correlation analysis between user age information and operation rate and change frequency to identify features that do not match age; A machine learning model is constructed, in which the machine learning module outputs an address to fill in a behavioral risk score based on the semantic drift feature, address drift feature, and age-inconsistent feature, for credit decision support.

7. The method for intelligent completion, verification, and fraud analysis of user addresses according to claim 6, characterized in that, By archiving high-risk samples and combining them with post-loan performance for tag feedback, the machine learning model parameters are updated regularly to achieve continuous evolution and adaptive optimization of behavior recognition capabilities.

8. A device for intelligent completion, verification, and fraud analysis of user addresses, comprising a storage medium and a processor, wherein the storage medium stores a computer program, characterized in that, When the computer program is executed by a processor, it is used to implement the method described in any one of claims 1-7.

Images