Financial customer precision marketing system and method based on big data and artificial intelligence

Abstract

The invention relates to a financial customer precision marketing system and method based on big data and artificial intelligence in the field of financial science and technology and intelligent marketing cross technology, and provides a financial customer precision marketing system and method based on big data and artificial intelligence by integrating multi-dimensional customer data, constructing a dynamic customer portrait, predicting customer demands by using a deep learning model and combining the real-time reaching capability of enterprise WeChat ecology. According to the method, automatic execution and closed-loop optimization of a personalized marketing strategy are realized, and the customer conversion rate and loyalty of a financial institution are improved; the method is innovatively embodied in deep combination of an emotion recognition technology and a financial scene, reinforcement learning optimization of a dynamic rule engine and a privacy compliance data desensitization mechanism, and solves the problems of lagging customer demand recognition and low strategy matching efficiency in traditional marketing.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of financial technology and intelligent marketing, and specifically relates to a financial customer precision marketing system and method combining big data analysis, deep learning model and enterprise WeChat ecology, which is suitable for efficient customer operation of banks, securities, insurance and other institutions. BACKGROUND

[0002] With the increasingly fierce competition in the financial market, traditional extensive marketing strategies have been difficult to meet the needs of financial institutions. Financial institutions need to more accurately identify potential customers and improve the conversion rate of marketing activities and customer satisfaction. However, existing financial customer marketing algorithms are often based on simple statistical analysis, which is difficult to deeply mine the potential needs and behavior patterns of customers.

[0003] There are three bottlenecks in traditional financial marketing: Data silos: customer transaction data, social behavior data, and external credit data are scattered in different systems and are difficult to integrate and analyze; Static strategy: relying on manual experience to develop rules, unable to respond to changes in customer behavior in real time (such as not reaching out in time after browsing a certain fund product); Compliance risk: marketing content lacks reasonable audit mechanism, easy to produce misleading propaganda. SUMMARY

[0004] In view of the deficiencies in the prior art, the present application provides a financial customer precision marketing system and method based on big data and artificial intelligence, which solves the following problems through technical innovation: 1. Break down data silos: Achieve millisecond-level real-time data synchronization (transaction flow delay ≤1 second) and multi-protocol automatic analysis (JSON / XML / CSV conversion error rate ≤0.1%); 2. Build dynamic customer portrait: Integrate transaction behavior time series analysis and unstructured text semantic mining (LSTM model sentiment analysis), portrait accuracy rate increased to 90%; 3. Improve model prediction accuracy: Use XGBoost ensemble learning + SHAP explainability framework, input features cover transaction, credit, and social behavior three-dimensional data, response rate prediction error ≤5%; 4. Real-time optimization of strategy: Adjust marketing strategy according to real-time customer behavior, realize "data-model-execution" closed loop.

[0005] 5. Ensure full-link compliance: The design sensitive word real-time filtering engine (miss detection rate ≤0.5%) and the hierarchical homomorphic encryption architecture meet the ISO27001 and PCIDSS financial data security standards.

[0006] In view of the above technical problems, the present application is realized by the following technical solutions: A financial customer precision marketing system based on big data and artificial intelligence, comprising data integration module, customer portrait module, prediction model module, strategy matching module and execution feedback module connected in sequence; The data integration module collects customer transaction data, enterprise micro interaction data and external credit investigation data; The customer portrait module divides customer groups by using K-means clustering algorithm, analyzes semantic features of enterprise micro chat records based on Transformer model, and identifies customer emotional tendency; The prediction model module uses LSTM network to model customer behavior sequence, predicts purchase intention probability in the next 30 days, combines XGBoost algorithm to evaluate customer risk tolerance, and generates product adaptation score; The strategy matching module matches marketing strategy library according to customer portrait and prediction results through dynamic rule engine, and optimizes push timing and content template by using reinforcement learning; The execution feedback module realizes message batch push and customer behavior data recovery through enterprise micro API, adjusts model weight based on A / B test results, and updates strategy library rules.

[0007] Further, when the customer portrait module performs sentiment analysis on enterprise micro chat records, a pre-trained BERT model is used to extract keywords.

[0008] Further, the K-means clustering algorithm is used to divide data samples into pre-defined KK clusters, the goal being to make the similarity of samples in the same cluster high and the difference between samples in different clusters large; The steps of K-means algorithm are as follows: (1) Initialize the center point: Randomly select KK samples as initial cluster centers; (2) Assign samples to clusters: For each sample xixi, calculate the distance between xixi and all cluster centers μjμj, and assign xixi to the nearest cluster CjCj; (3) Update cluster center: Recalculate the center point μjμj of each cluster; (4) Iterative optimization: Repeat steps (2)-(3) until the cluster center no longer changes; Minimize the within-cluster sum of squares: WCSS = j = 1 ∑ K xi ∈ Cj ∑ || xi - μj || 2.

[0009] Further, the dynamic rule engine comprises New customer activation: automatically send a new customer package to the customer who first adds WeChat; Silent customer wake-up: generate customized activity content according to historical behavior; Push WeChat one-on-one service of exclusive financial consultant for high net worth customers; Push WeChat community batch fission activities for long tail customers; Adjust the reach frequency and content according to the customer life cycle.

[0010] Further, the execution feedback module comprises a data compliance processing mechanism: Customer sensitive information is stored and transmitted using AES-256 encryption; Before pushing marketing content, check the language through the compliance audit engine to avoid illegal language.

[0011] The application also provides a marketing method of the above-mentioned financial customer precision marketing system based on big data and artificial intelligence, comprising the following steps: (1) Data collection and integration Collect internal and external data of financial customers, and perform data cleaning and standardization processing; (2) Customer feature extraction and portrait construction Through the collected financial customer data, perform feature engineering on the data to form a layered architecture of the label system; (3) Training of deep learning prediction model According to the constructed customer portrait, perform model training and reasoning on the customer portrait to form a model architecture; (4) Match and execute marketing strategies according to the prediction model Group customers to form a priority matrix, and match and execute different marketing strategies for customers with different priorities; (5) Evaluate marketing effect and optimize model Collect customer feedback and interaction data, and adjust the strategy and model parameters based on the feedback data.

[0012] In step (1), The internal data includes credit system data and CRM system data, and the external data includes enterprise WeChat platform data; Data cleaning and standardization processing includes protocol conversion and format unification of data, cleaning of unstructured data, data exception processing, and hierarchical storage and encryption of data.

[0013] In step (2), the step (2) is performed by using the following formula: The feature engineering includes time sequence feature extraction and unstructured feature generation. The label system hierarchical architecture includes a basic attribute layer, a behavior feature layer, a financial feature layer, and a customer value layer.

[0014] In step (3), the step (3) is performed by using the following formula: The input features of the model include transaction features, credit features, and behavior features. The model architecture includes a base model, an output structure, and a dynamic updating mechanism.

[0015] The beneficial effects of the present application are as follows: 1. Real-time heterogeneous data fusion technology: The first flow batch data pipeline: Kafka real-time flow (processing delay ≤1 second) and HDFS batch data (T+1) unified access, data synchronization efficiency improved by 80%; Propose multi-protocol adaptive parsing engine: automatically identify JSON / XML / CSV format differences, field mapping accuracy ≥99.9%; 2. Dynamic portrait modeling method: Design time sequence-text joint feature extractor: combine transaction cycle detection (Fourier transform) and text sentiment analysis (Bi-LSTM), portrait accuracy improved to 90%; Build KOL influence power model: calculate customer influence index based on social interaction data, used to identify high transmission value customers; 3. Compliance embedded algorithm: Develop sensitive word variant recognition algorithm: combine AC automaton and edit distance (Levenshtein Distance ≤2), intercept "guaranteed" variant words (such as "guaranteed" "guaranteed"); Implement group calculation under homomorphic encryption: customer labels are clustered in ciphertext state, meeting the requirements of GDPR cross-border data transmission. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The structure diagram of the financial customer precision marketing system based on big data and artificial intelligence of the present application; Figure 2 The flowchart of the financial customer precision marketing method based on big data and artificial intelligence of the present application. DETAILED DESCRIPTION

[0017] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the protection scope of the present application can be more clearly defined.

[0018] As Figure 1 shown in a financial customer precision marketing system based on big data and artificial intelligence, comprising data integration module, customer portrait module, prediction model module, strategy matching module, execution feedback module connected in turn.

[0019] Specifically, the data integration module collects customer transaction data (amount, frequency, product type), enterprise-micro interaction data (chat keywords, click behavior), and external credit data (credit score, risk level).

[0020] Specifically, the customer portrait module uses K-means clustering algorithm to divide customer groups (high net worth, long tail, loss warning), and analyzes the semantic features of enterprise-micro chat records based on Transformer model to identify customer emotional tendencies (such as anxiety, interest, hesitation); When the customer portrait module performs sentiment analysis on enterprise-micro chat records, a pre-trained BERT model is used to extract keywords; K-means clustering algorithm is used to divide customer groups. K-means is a classic unsupervised learning algorithm that is used to divide data samples into pre-defined KK clusters. The goal is to make the samples in the same cluster similar and the samples in different clusters different; The steps of K-means algorithm are as follows: (1) Initialize the center point: Randomly select KK samples as the initial cluster center; (2) Assign samples to clusters: For each sample xixi, calculate the distance (commonly used Euclidean distance) between xixi and all cluster centers μjμj, and assign xixi to the nearest cluster CjCj; (3) Update cluster center: Recalculate the center point μjμj of each cluster (take the mean of the samples in the cluster); (4) Iterative optimization: Repeat steps (2)-(3) until the cluster center no longer changes or the change is less than a threshold; Minimize the Within-Cluster Sum of Squares (WCSS): WCSS = j = 1 ∑ K xi ∈ Cj ∑ ∥ xi - μj ∥2.

[0021] Specifically, the prediction model module uses LSTM network to model customer behavior sequence and predict the purchase intention probability in the next 30 days, and combines XGBoost algorithm to evaluate customer risk tolerance and generate product adaptation score.

[0022] Specifically, the strategy matching module matches the customer portrait with the prediction results according to the dynamic rule engine to match the marketing strategy library (such as triggering one-to-one WeChat service for high-intention customers and pushing a split activity for low-intention customers), and uses reinforcement learning (Q-Learning) to optimize the pushing time and content template (text, applet, and live link). The dynamic rule engine includes New customer activation: automatically send a new customer package to the customer who first adds WeChat; Silent customer awakening: generate customized activity content according to historical behavior; Pushing WeChat one-to-one service of exclusive financial consultants for high net worth customers; Pushing a split activity (such as inviting friends to win rewards) through a WeChat community for long-tail customers; Adjusting the reach frequency and content according to the customer life cycle (new customers, silent customers, and early warning customers).

[0023] Specifically, the execution feedback module realizes batch pushing of messages and collection of customer behavior data (click rate, forwarding volume, and conversion path) through WeChat API, adjusts the model weight based on the A / B test results, and updates the strategy library rules; The execution feedback module includes the following data compliance processing mechanism: Customer sensitive information (ID number and bank card number) is stored and transmitted using AES-256 encryption; Before pushing the marketing content, the compliance audit engine checks the language to avoid illegal language.

[0024] In combination with Figure 2 As shown in the above marketing method of the financial customer precision marketing system based on big data and artificial intelligence, the method specifically includes the following steps: (1) Data collection and integration Collecting multi-source data (internal data and external data) of financial customers, and performing data cleaning and standardization processing.

[0025] The internal data includes credit system data and CRM system data, Credit system: capturing transaction stream (fields include: transaction timestamp, amount, counterparty account type, IP geographic location) in real time through the Kafka stream processing engine, with a data delay of ≤1 second; CRM system: extracting customer risk assessment results (such as risk tolerance level R1-R5) and customer tag data in batches according to T+1 period; The external data includes enterprise WeChat platform data, Enterprise WeChat platform: Real-time monitoring of customer conversation keywords (such as "redemption" "interest coupon") through Webhook, and recording the frequency of community speech (times / hour).

[0026] Data cleaning and standardization includes data protocol conversion and format unification, unstructured data cleaning, data anomaly handling, and hierarchical data storage and encryption.

[0027] Protocol conversion and format unification: For JSON / XML / CSV data from bank credit systems, enterprise WeChat and other interfaces, use adaptive parsing engine to automatically map fields (such as "txn_amt" mapped to "transaction amount"), format error rate ≤0.1%; Timestamp calibration: unify multi-time zone data to UTC+8 format (such as New York transaction data UTC-5 converted to UTC+8), error compensation ≤1 millisecond.

[0028] ‌Unstructured data cleaning‌: Sensitive word filtering, using AC automatic machine algorithm to match financial advertising compliance guidelines prohibited word library (such as "guaranteed return" "stable profit and no loss"), and identifying variant words (such as "guaranteed" replaced by "guaranteed"); Semantic extraction, use BERT model to extract customer intent from customer service chat records (such as "consult fund purchase"); Real-time desensitization of customer conversation text (such as bank card number replaced by "**** 1234").

[0029] Data anomaly handling: For missing customer age field, based on ID number to calculate and fill (such as "32010219900101****" calculated as 35 years old); Use linear interpolation method to repair the timestamp breakpoint in transaction flow (such as missing 10:00:00 data filled with average value before and after); Label conflict resolution, for "high net worth customers" but "no transaction in the past 30 days" contradictory labels, corrected to "potential loss customers" and others through cluster analysis; Based on LSTM model to analyze customer behavior sequence, identify abnormal labels. Hierarchical data storage and encryption: ‌Hot data layer: Storage medium uses Redis Cluster cluster, sharding strategy is consistent hash, supports ≥150,000 concurrent read / write per second; Data range uses near 30-day transaction flow, real-time conversation record, and customer activity label.

[0030] ‌Cold data layer‌: The storage medium uses HDFS 3.0 + Parquet columnar storage and uses the Snappy compression algorithm (compression rate ≥ 75%); The encryption specification uses PII data such as ID numbers and bank card numbers to use AES-256-GCM encryption.

[0031] (2) Customer feature extraction and portrait construction Through the collected financial customer data, feature engineering is performed on the data to form a layered architecture of the label system.

[0032] Among them, the feature engineering includes time series feature extraction and unstructured feature generation.

[0033] ‌Time series feature extraction‌: Transaction periodicity analysis, detect customer fund transfer-in / out cycle (such as salary payment day, quarterly financial maturity day) through Fourier transform; Risk volatility calculation, based on 30-day transaction amount standard deviation σ, define risk level (σ < 10,000: low risk; 10,000 ≤ σ < 100,000: medium risk; σ ≥ 100,000: high risk).

[0034] ‌Unstructured feature generation‌: Text sentiment analysis, use Bi-LSTM model to score customer service chat records (interval -1 to +1), identify customer emotional tendency (such as "anxious" "satisfied"); Social influence calculation, generate KOL index (0-100 points) according to the number of likes and forwards of customer's speech in the enterprise WeChat community.

[0035] Among them, the layered architecture of the label system includes multi-dimensional label layering (basic attribute layer, behavior feature layer, financial feature layer, and customer value layer).

[0036] ‌Basic attribute layer‌: Contains customer age, gender, occupation, etc. Static information, automatically generated through ID number, WeChat data; geographic label is refined to city commercial district level.

[0037] Behavior feature layer: Transaction behavior label, including high-frequency transfer, large-financial preference, etc. Channel usage preference, including mobile banking activity, offline branch visit frequency, etc.

[0038] Financial feature layer: Risk credit label, including credit limit level (such as "Q3 + 1.5 IQR dynamic credit"), overdue record identification; Asset holding label, including deposit size binning (such as "AUM 50-100 million"), product cross-holding matrix.

[0039] Customer value stratification: Based on the RFM model dynamic grouping, including high-value customers, loss warning customers, and long-tail customers.

[0040] (3) Train deep learning prediction model According to the customer portrait constructed, train and infer the model for the customer portrait, and form the model architecture.

[0041] Among them, the input features of the model include transaction features, credit features and behavior features.

[0042] Transaction features: transaction frequency in the past 30 days, amount volatility, net difference between transfer-in and transfer-out; Credit features: central bank credit scoring, historical overdue times; Behavioral features: social KOL index, customer service consultation sentiment score.

[0043] Among them, the model architecture includes base model, output result and dynamic updating mechanism.

[0044] Base model: XGBoost, parameter setting: learning rate η=0.05, tree depth =6, regularization λ=1.0; Output result: customer response probability P to target product (0≤P≤1), life cycle value LTV (predicted contribution income in the next year); ‌Dynamic updating mechanism‌: Incremental training, loading new data every day at 0 o'clock, model iteration period ≤4 hours; Feature importance analysis, identify Top 10 influencing factors (such as "transfer-in amount in the past 7 days" contribution ratio ≥35%) through SHAP value.

[0045] (4) Match and execute marketing strategy according to prediction model Group customers and form priority matrix, match and execute different marketing strategies for customers with different priorities.

[0046] Among them, the priority matrix is shown in the following table:

[0047] High-priority customers (high LTV): send personalized product links (including customer name, preferred product name) through enterprise WeChat API; Long-tail customers (low LTV): generate WeChat templates (such as "Dear customer, you may be interested in money fund annual rate 2.8%"), and send them in batches at regular intervals.

[0048] (5) Evaluate marketing effect and optimize model Collect customer feedback and interaction data (click rate, conversion rate, chat record, etc.), adjust strategy and model parameters based on feedback data.

[0049] The following examples are described in detail.

[0050] Example 1: High net worth customer wealth management product matching system Core algorithm and data flow: ‌Customer portrait construction ‌Data integration: Integrate bank core system (AUM / product holdings), tax data (annual income ≥ 1 million), social behavior (enterprise WeChat chat keyword extraction) to build customer 360 portrait; Identify customer associated enterprise equity structure through knowledge graph (hold ≥ 10% as "enterprise owner customer").

[0051] ‌Life cycle prediction: Use LSTM model to analyze asset variation trend in the past 24 months to predict the probability of fund inflow in the next 3 months (output value ≥ 0.6 triggers product recommendation).

[0052] ‌Product matching engine ‌Feature vectorization: Product dimensions: Sharpe ratio, maximum drawdown, lock-in period, purchase amount (normalized to 0-1 interval value); Customer dimensions: risk assessment level, historical product holding period, expected return value (updated dynamically through questionnaire).

[0053] ‌Similarity calculation: Use improved cosine similarity algorithm, add time decay factor (behavior weight of the past 3 months accounts for 60%); TOP3 matching products are pushed through enterprise WeChat exclusive customer managers, with product comparison analysis report.

[0054] Example 2: Retail customer loss warning and recovery system Technical implementation details: ‌Warning signal identification ‌Multi-dimensional monitoring indicators: Key behaviors: decrease in login frequency by ≥60% in the past 30 days, product redemption amount accounts for ≥40%; Fund movement: trigger secondary warning when single-day transfer amount ≥ 50% of account balance.

[0055] ‌Classification intervention strategy ‌Automatic recovery mechanism: First-level warning (probability of loss ≥75%): "Exclusive financial income + 1.2%" coupons are given, valid for 3 days; Second-level warning (probability 50-75%): Targeted placement of WeChat circle of friends advertisements through the DMP advertising platform (material click rate ≥8%); For non-responding customers, AI voice call is triggered on the 5th day (success rate of recovery increased by 18%).

[0056] Privacy protection design Customer behavior data is processed using differential privacy technology (ε=0.5) to prevent individual data from being reverse deduced; Marketing effect analysis only shows aggregated data (such as group conversion rate), and does not expose individual behavior tracks.

[0057] Example 3: Retail credit scenario marketing Core algorithm and data flow: Multi-dimensional behavior data collection Analyze mobile banking APP operation tracks (loan calculator usage times, limit query frequency), and construct a credit demand intensity index (0-100 points); Integrate e-commerce platform consumption data (installment payment records, return rate) to calculate customer compliance ability score.

[0058] Real-time triggered marketing Scenario recognition model: When a customer browses school district house information for >5 minutes, automatically push education renovation loan (up to 500,000, with 30BP interest rate discount); Detect that the customer's credit card bill installment times ≥3 times / month, recommend low-interest debt replacement plan (annual interest rate 4.35%).

[0059] Risk hedging mechanism: Start manual review process for high-risk customers to intercept potential multi-head lending risks.

[0060] The above-described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A financial customer precision marketing system based on big data and artificial intelligence, characterized in that, It includes a data integration module, a customer profiling module, a predictive model module, a strategy matching module, and an execution feedback module, which are connected in sequence. The data integration module collects customer transaction data, enterprise WeChat interaction data, and external credit data. The customer profiling module uses the K-means clustering algorithm to divide customer groups and analyzes the semantic features of WeChat chat records based on the Transformer model to identify customer sentiment. The prediction model module uses an LSTM network to model customer behavior sequences, predicts the probability of purchase intention in the next 30 days, and combines the XGBoost algorithm to assess the customer's risk tolerance and generate a product suitability score. The strategy matching module uses a dynamic rule engine to match a marketing strategy library based on customer profiles and prediction results, and uses reinforcement learning to optimize push timing and content templates. The execution feedback module uses the Enterprise WeChat API to push messages in batches and collect customer behavior data. It adjusts model weights and updates strategy library rules based on A / B test results.

2. The financial customer precision marketing system based on big data and artificial intelligence according to claim 1, characterized in that, When performing sentiment analysis on WeChat Work chat logs, the customer profiling module uses a pre-trained BERT model to extract keywords.

3. The financial customer precision marketing system based on big data and artificial intelligence according to claim 2, characterized in that, The K-means clustering algorithm is used to divide data samples into predefined K clusters, with the goal of making samples within the same cluster highly similar and samples between different clusters highly different. The steps of the K-means algorithm are as follows: (1) Initialize the center point: Randomly select K samples as the initial cluster centers; (2) Assigning samples to clusters: For each sample xi, calculate its distance to all cluster centers μj, and assign xi to the nearest cluster Cj. (3) Update cluster center: Recalculate the center point μj for each cluster; (4) Iterative optimization: Repeat steps (2)-(3) until the cluster center no longer changes or the change is less than the threshold; Minimize the in-cluster squared error: WCSS=j=1∑Kxi∈Cj∑∥xi-μj∥2.

4. The financial customer precision marketing system based on big data and artificial intelligence according to claim 1, characterized in that, The dynamic rule engine includes New Customer Activation: Automatically send a new customer gift pack to customers who add WeChat for the first time; Reactivating dormant customers: Generating customized campaign content based on historical behavior; Offer one-on-one WeChat-based financial advisor services to high-net-worth clients. Targeting long-tail customers by sending mass promotional campaigns through WeChat official accounts and social media groups; Adjust the frequency and content of outreach based on the customer lifecycle.

5. The financial customer precision marketing system based on big data and artificial intelligence according to claim 1, characterized in that, The execution feedback module includes a data compliance processing mechanism: Sensitive customer information is stored and transmitted using AES-256 encryption. Before marketing content is pushed out, the wording is verified through a compliance review engine to avoid the appearance of non-compliant wording.

6. A marketing method for a financial customer precision marketing system based on big data and artificial intelligence as described in any one of claims 1-5, characterized in that, Includes the following steps: (1) Data collection and integration Collect internal and external data from financial clients, and perform data cleaning and standardization. (2) Customer feature extraction and profile construction By collecting financial customer data, feature engineering is performed on the data to form a hierarchical tagging system architecture; (3) Training deep learning prediction models Based on the constructed customer profile, the customer profile model is trained and inferred to form a model architecture; (4) Match and execute marketing strategies based on the predictive model. Segment customers to form a priority matrix, and match and execute different marketing strategies with customers of different priorities. (5) Evaluate marketing effectiveness and optimize the model Collect customer feedback and interaction data, and adjust strategies and model parameters based on the feedback data.

7. The financial customer precision marketing method based on big data and artificial intelligence according to claim 6, characterized in that, In step (1), Internal data includes credit system data and CRM system data, while external data includes data from the WeChat Work platform. Data cleaning and standardization processes include protocol conversion and format standardization of data, cleaning of unstructured data, handling of data anomalies, and layered storage and encryption of data.

8. The financial customer precision marketing method based on big data and artificial intelligence according to claim 6, characterized in that, In step (2), Feature engineering includes temporal feature extraction and unstructured feature generation; The tagging system's hierarchical architecture includes a basic attribute layer, a behavioral feature layer, a financial feature layer, and a customer value layer.

9. The financial customer precision marketing method based on big data and artificial intelligence according to claim 6, characterized in that, In step (3), The model's input features include transaction features, credit features, and behavioral features; The model architecture includes the base model, output structure, and dynamic update mechanism.

Images