Personal consumption loan potential customer screening method based on multi-source big data fusion

Abstract

The application discloses a personal consumption loan potential customer screening method based on multi-source big data fusion, and the method comprises the following steps: collecting five source data of bank internal, government affairs, compliance online lending, online behavior and outbound history, splicing after desensitization and cleaning with customer number as the primary key to build a panoramic data wide table; deriving credit risk, consumption tendency and loan willingness three-dimensional feature set based on the wide table; after screening by variance, mutual information and regularization combination, it is divided into three subsets, respectively input parallel heterogeneous base model, output each special score and gather, output loan conversion probability by meta-model comprehensive research and judgment; accordingly, generate potential score and output potential customer list according to threshold layering, monitor effective feature set distribution drift, trigger incremental training and support version rollback when the drift exceeds the threshold, and complete the closed-loop iteration. The application is based on multi-source data system and integrated model architecture, realizes multi-dimensional comprehensive evaluation of customers, solves the problems of portrait distortion, screening deviation and long-term adaptation deficiency of the model.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and big data analysis and mining technology, and in particular to a method and apparatus for screening potential customers for personal consumer loans based on multi-source big data fusion. Background Technology

[0002] As a core revenue stream for the retail banking sector of major domestic commercial banks, personal consumer lending has long relied on a linear process where customers proactively initiate loan applications, and banks passively respond with approvals. This model has revealed numerous drawbacks in practice: banks lack effective means to proactively identify potential customers, marketing resources are not precisely targeted, and high-potential customers with genuine consumption needs and good repayment ability are easily lost due to a lack of timely outreach. Furthermore, banks struggle to effectively predict and manage customer risk at the marketing front end. With the continued penetration and deep application of big data, artificial intelligence, and other next-generation information technologies in the financial sector, commercial banks urgently need to build a proactive customer acquisition capability based on the fusion of multi-dimensional, multi-source, heterogeneous data and driven by machine learning intelligent algorithms, thereby achieving the dual business objectives of precise marketing and risk control.

[0003] In the existing technological system, the methods for screening potential customers for personal consumer loans mainly present the following technological paths.

[0004] The first category is the traditional manual rule-based screening scheme. This type of method relies solely on single-dimensional structured data such as bank internal assets and deposits and payroll disbursements. It uses fixed threshold rules to perform coarse-grained screening of customers, without introducing rich data sources such as external government data, third-party compliance data, and user behavior data, and it also lacks intelligent modeling capabilities based on machine learning.

[0005] Although the second type of solution introduces internal credit data to build a scoring card model, which expands the data dimensions compared to the former, the data sources are still limited to internal bank information and credit reports. It fails to access government qualification data such as housing provident fund and social security to assess the stability of customer income and the authenticity of employment, and does not include external online lending data to identify the customer's overall debt level and the risk of multiple borrowing, resulting in significant blind spots in the information foundation for customer credit assessment.

[0006] The third type of solution attempts to integrate multi-source data into the loan credit system for customer credit assessment, which broadens the data dimensions. However, this type of technology still does not collect real-time behavioral log data from mobile banking and lifestyle applications, nor does it integrate historical interaction records from the bank's outbound call center. As a result, the system cannot effectively perceive the customer's real-time consumption needs and loan intentions. Its assessment system only focuses on the credit risk side, and it is seriously lacking in the ability to quantitatively determine the customer's needs.

[0007] Although the fourth type of solution introduces mobile banking browsing behavior data to assist in customer screening, the data source only covers a single channel and fails to deeply integrate with multi-dimensional data such as assets, credit, housing provident fund, online lending, and outbound calls. In addition, the model architecture still uses a single lightweight algorithm, resulting in limited prediction accuracy, and it lacks a model adaptive update mechanism for production environments.

[0008] A comprehensive analysis of the existing technical solutions reveals that they all have varying degrees of deficiencies in terms of data fusion completeness, feature engineering refinement, multidimensionality of the evaluation system, model prediction accuracy, comprehensiveness of risk identification, and automation level of system operation and maintenance. These deficiencies make it difficult to effectively support the urgent business needs of large commercial banks in retail lending to achieve low-cost, high-efficiency, and high-precision potential customer group mining and refined marketing operations for a massive number of customers.

[0009] Specifically, the shortcomings of existing technologies are mainly reflected in the following aspects: First, the data source coverage is incomplete and the integration of multi-source heterogeneous data is incomplete. Core business data within the bank, government qualification data, external online lending data, online behavior data, and outbound call interaction data are all isolated, lacking a unified data alignment and integration mechanism. The problem of data silos is serious, making it impossible to build a complete customer profile. Second, feature engineering is simple and crude, lacking the ability to derive cross-field features and construct time-series features, and lacking a systematic and standardized preprocessing process for heterogeneous data, making it difficult to fully release the intrinsic value of the data. Third, the assessment dimensions are singular, unable to simultaneously complete the three-dimensional comprehensive quantitative judgment of customer credit risk, consumption tendency, and loan willingness. Risk identification and demand mining are disconnected, easily leading to high risk. Fourth, the model architecture often uses a single machine learning algorithm and lacks integrated design. When faced with massive amounts of real business data with high dimensions and high noise, the model's prediction accuracy and generalization ability have significant bottlenecks. Fifth, existing technologies generally do not establish performance monitoring and adaptive evolution mechanisms after the model goes live. They lack automated operation and maintenance capabilities such as real-time warning of feature drift, incremental learning updates, and version rollback. The model performance continues to decline with changes in customer structure and market environment, resulting in high long-term operation and maintenance costs. Sixth, the screening rules are rigid and the level of automation is low, making it impossible to achieve fine-grained segmentation of customer levels and differentiated matching of marketing strategies. This leads to serious waste of marketing resources and makes it difficult to adapt to the business requirements of large-scale and precise operations of large banks.

[0010] In summary, there is an urgent need for an intelligent screening method for potential personal consumer loan customers that can fully integrate multi-source heterogeneous data, deeply mine multi-dimensional customer attributes, achieve comprehensive quantitative assessment of risk and demand, and possess adaptive model evolution capabilities, in order to overcome the many shortcomings of existing technologies. Summary of the Invention

[0011] The main objective of this invention is to provide a method for screening potential customers for personal consumer loans based on multi-source big data fusion.

[0012] Another objective of this invention is to propose a device for screening potential customers for personal consumer loans based on the fusion of multi-source big data.

[0013] The third objective of this invention is to provide an electronic device.

[0014] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0015] To achieve the above objectives, a first aspect of the present invention proposes a method for screening potential customers for personal consumer loans based on multi-source big data fusion, comprising: Collect internal bank business data, government industry data, external compliant online lending data, online application behavior data, and historical bank outbound call data. De-identify, clean, and standardize the collected data and merge them horizontally using the bank's internal customer ID as the global primary key to construct a wide table of customer panoramic data. Based on the aforementioned customer panoramic data wide table, credit risk feature sets, consumption tendency feature sets, and loan willingness feature sets are derived respectively. The derived features are then filtered through a combination of variance filtering, mutual information filtering, and regularization filtering to obtain an effective modeling feature set. The effective modeling feature set is split into three feature subsets, which are then input into three heterogeneous base models running in parallel. Each base model outputs a corresponding special evaluation score. The special evaluation scores are then aggregated into a fusion feature set. The meta-model then performs a comprehensive evaluation on the fusion feature set and outputs the customer loan conversion probability. A comprehensive customer potential score is generated based on the customer loan conversion probability. Customers are intelligently segmented according to a preset scoring threshold, and a list of potential customers is output. The data distribution drift of the effective modeling feature set is monitored. When the drift exceeds the threshold, incremental training is triggered based on the new samples to update the model parameters, and historical version rollback is supported to complete the closed-loop iteration of the model.

[0016] Optionally, data may be collected from internal bank operations, government sector data, external compliant online lending platforms, online application behavior data, and historical bank outbound call data, including: By directly connecting to the bank's internal big data warehouse, internal business data is collected in batches and synchronously to obtain customer unit information, loan history, asset deposit and wealth management data, credit report records, card spending details, payroll flow and credit card repayment records, which serve as the basic data source for customer repayment ability assessment and qualification verification. Based on the compliant shared data channel, we connect with the housing provident fund management system and the social security management system to collect government industry data, obtain data on contribution base and continuous contribution duration, and use this as a basis for assessing the social qualifications and income stability of our clients. Connect to the encrypted data interface of licensed financial service institutions to collect external compliant online lending data, extracting online lending information, repayment plan execution information, loan application frequency, total outstanding debt, historical overdue records and the number of multiple lending institutions, as the basis for assessing the customer's comprehensive debt risk and credit behavior; Real-time user behavior data on mobile banking and lifestyle applications is captured using streaming data collection. This includes browsing history, click history, search keyword history, page dwell time, and loan pre-application behavior log data, which serve as a data source for real-time perception of customer consumption needs and loan intentions. Synchronize historical data from the bank's outbound call center system, collect historical outbound call data, and extract outbound call connection rate, customer intention level, and historical response record text as a basis for evaluating customer marketing interaction preferences and cooperation.

[0017] Optionally, the collected data can be anonymized, cleaned, and standardized for fusion. Using the bank's internal customer ID as the global primary key, a horizontally concatenated wide table of customer panoramic data can be constructed, including: The five types of heterogeneous raw data collected were subjected to hierarchical desensitization processing. For personal privacy information fields containing names, ID card numbers, and mobile phone numbers, an irreversible hash encryption algorithm combined with field partial hiding technology was used for desensitization and conversion. Cleaning is performed on the five types of data after desensitization. The entire record is traversed by a duplicate value comparison algorithm to detect and remove duplicate and redundant records. For numerical continuous fields such as income and liabilities, statistical identification algorithms are used to detect and correct outliers. For missing values ​​in regular continuous fields, a median imputation strategy is used, and for missing values ​​in high-dimensional behavioral feature fields, a model prediction imputation strategy is used to maximize the retention of effective information. Using the bank's internal customer ID as the global primary key, the five types of cleaned data are horizontally concatenated according to the customer dimension. The field naming format, unit of measurement, and time caliber of each data source are unified to eliminate format differences and semantic conflicts between multi-source heterogeneous data and construct a wide table of customer panoramic data.

[0018] Optionally, based on the aforementioned customer panoramic data wide table, credit risk feature sets, consumption propensity feature sets, and loan willingness feature sets are derived, including: Based on the credit history, asset and liability data, housing provident fund contribution duration, online loan application frequency, and online loan amount range data in the aforementioned customer panoramic data wide table, a credit risk feature set is derived and constructed. This set generates indicators such as the number of credit overdue payments and overdue days, comprehensive debt ratio, housing provident fund continuous contribution duration, online loan application frequency in the past year, and online loan amount range distribution in the past year. These indicators are used to quantitatively assess the customer's ability to fulfill obligations and credit risk level. Based on the payroll records, income level classification, average monthly consumption expenditure, frequency of large-amount consumption, distribution of consumption scenario categories, and asset-to-consumption ratio data in the customer panoramic data wide table, a consumption tendency feature set is derived and constructed to generate income level classification indicators, average monthly consumption amount indicators, frequency of large-amount consumption indicators, consumption scenario preference distribution indicators, and asset-to-consumption ratio indicators, which are used to quantitatively assess customers' consumption capacity and consumption demand intensity. Based on the online application behavior duration records, loan-related search keyword records, loan product click frequency, outbound call connection rate statistics, and historical response intention records in the aforementioned customer panoramic data wide table, a loan intention feature set is derived and constructed. This generates online behavior activity duration indicators, loan keyword search frequency indicators, loan product click frequency indicators, outbound call connection rate indicators, and historical response intention rating indicators, which are used to quantitatively assess the strength of customers' loan intentions.

[0019] Optionally, derived features can be filtered using a combination of variance filtering, mutual information filtering, and regularization filtering to obtain an effective modeling feature set, including: Variance screening is performed on all derived features. The variance value of each feature over all samples is calculated. Features with variance below a preset threshold are identified as features without discrimination and are removed. Features with sufficient sample discrimination ability are retained to enter the next round of screening. Mutual information filtering is performed on the features retained after variance filtering. The mutual information between each feature and the target variable is calculated. Features with mutual information higher than the preset correlation threshold are retained, and irrelevant features with low correlation with the target variable are removed to ensure that the retained features make an effective contribution to the prediction target. After variance screening and mutual information screening, the features retained are subjected to L1 and L2 combined regularization screening. L1 regularization drives the sparsity of redundant feature weights, and L2 regularization suppresses the influence of feature collinearity. Redundant features with weights approaching zero and duplicate contribution features in strongly correlated feature groups are removed to obtain a simplified effective modeling feature set.

[0020] Optionally, the effective modeling feature set is split into three feature subsets, which are then input into three heterogeneous base models running in parallel. Each base model outputs a corresponding specific evaluation score. These specific evaluation scores are aggregated into a fusion feature set, which is then comprehensively evaluated by a meta-model to output the customer loan conversion probability, including: The effective modeling feature set is divided into a credit risk feature subset, a consumption propensity feature subset, and a loan intention feature subset according to their source. Each feature subset corresponds to three assessment dimensions: customer credit risk, consumption demand, and loan intention. A subset of credit risk features is input into the first heterogeneous base model, which performs nonlinear relationship mining and feature interaction learning on the credit risk features based on the gradient boosting tree algorithm, and outputs a credit risk score to quantify the customer's default risk level. A subset of consumption propensity features is input into a second heterogeneous base model. This base model performs complex pattern recognition and nonlinear mapping learning on the consumption propensity features based on the gradient boosting tree algorithm, and outputs a consumption propensity score to quantify the customer's consumption needs and consumption capacity. A subset of loan intention features is input into a third heterogeneous base model. This base model performs linear weighted discrimination and probability mapping on the loan intention features based on a linear classification algorithm, and outputs a loan intention score to quantify the strength of the customer's loan intention, while maintaining the interpretability of the evaluation results in this dimension. By aggregating the credit risk score, consumption propensity score, and loan willingness score, a three-dimensional score fusion feature set is constructed; The three-dimensional score fusion feature set is input into the meta-model, which uses a logistic regression algorithm to learn the weights of the three special assessment scores and make a comprehensive judgment, outputting the customer loan conversion probability.

[0021] Optionally, a comprehensive customer potential score is generated based on the customer loan conversion probability. Customers are intelligently segmented according to a preset scoring threshold, and a potential customer list is output. Data distribution drift of the effective modeling feature set is monitored. When the drift exceeds the threshold, incremental training is triggered based on new samples to update the model parameters, and historical version rollback is supported, including: The customer loan conversion probability is converted into a percentage score according to a linear mapping relationship, which is used as the customer's comprehensive potential score, realizing the transformation of the abstract probability value output by the model into an intuitive business score; Based on the preset high-potential and low-potential scoring thresholds, customers are divided into three levels: high-potential customers, medium-potential customers, and low-potential customers, and marketing priorities and service strategies are differentiated for different levels of customers. From the segmented customers, high-potential and medium-potential customers are selected, and customers on the system's preset gray list and those exempted from compliance and safety are automatically matched and removed. A standard formatted list of potential customers is generated and pushed to the customer marketing system for the business side to carry out differentiated marketing. The system continuously monitors the data distribution trends of each feature in the effective modeling feature set, calculates the mean, variance, and distribution morphology statistics of the features, and automatically triggers a drift warning signal when the drift measure of any feature exceeds the preset warning threshold. In response to the drift warning signal, new business sample data is extracted monthly, and incremental training is performed on the current model in combination with the drift analysis results to automatically update the model parameters without the need for manual full retraining. Each time the model is iterated and updated, all model parameters and configuration information of the current version are retained to form a multi-version history archive. When an abnormal performance of the model is detected after the update, it supports one-click rollback to the previous stable version, realizing automated closed-loop operation and maintenance of the model throughout its entire life cycle.

[0022] To achieve the above objectives, a second aspect of the present invention provides a device for screening potential customers for personal consumer loans based on multi-source big data fusion, comprising: The multi-source data fusion unit is used to collect internal bank business data, government industry data, external compliant online lending data, online application behavior data, and bank outbound call history data. The collected data is de-identified, cleaned, and standardized for fusion. The data is then horizontally assembled using the bank's internal customer ID as the global primary key to construct a wide table of customer panoramic data. The multi-dimensional feature filtering unit is used to derive credit risk feature sets, consumption tendency feature sets, and loan willingness feature sets based on the wide table of customer panoramic data, and to filter the derived features through a combination of variance filtering, mutual information filtering, and regularization filtering to obtain an effective modeling feature set; The dual-layer integrated inference unit is used to split the effective modeling feature set into three feature subsets, which are respectively input into three heterogeneous base models running in parallel. Each base model outputs a corresponding special evaluation score, and the special evaluation scores are aggregated into a fusion feature set. Then, the meta-model performs a comprehensive evaluation on the fusion feature set and outputs the customer loan conversion probability. The intelligent stratified iteration unit is used to generate a comprehensive customer potential score based on the customer loan conversion probability, intelligently stratify customers according to a preset scoring threshold and output a list of potential customers, monitor the data distribution drift of the effective modeling feature set, and when the drift exceeds the threshold, trigger incremental training to update the model parameters based on the new samples, and support the rollback of historical versions to complete the closed-loop iteration of the model.

[0023] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0024] To achieve the above objectives, a third aspect of this application provides an electronic device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a method for screening potential customers for personal consumer loans based on multi-source big data fusion as described in the first aspect embodiment.

[0025] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for screening potential customers for personal consumer loans based on multi-source big data fusion as described in the first aspect embodiment.

[0026] The embodiments of the present invention have the following beneficial effects: (1) A five-source heterogeneous data fusion system covering internal bank business data, government industry data, external compliant online lending data, online application behavior data and bank outbound call history data was constructed. The system restores the true status of customers from multiple dimensions such as repayment ability, debt risk, consumption demand and loan intention, effectively eliminating the problems of customer profile distortion and one-sided assessment caused by the lack of data dimensions in traditional solutions. (2) A comprehensive quantitative assessment framework integrating credit risk, consumption tendency and loan willingness has been established, realizing the coordinated unity of credit risk control, customer consumption demand identification and marketing acceptance judgment, overcoming the technical defects of existing technologies that separate risk identification and demand mining and are prone to mis-screening of "capable but not demanding, demanding but not willing". (3) A two-layer integrated machine learning architecture of parallel evaluation of heterogeneous base models and fusion decision of meta-models is adopted to give full play to the differentiated advantages of different algorithms in their respective evaluation dimensions, significantly improve the prediction accuracy and generalization ability of potential customer screening, and achieve screening accuracy and marketing conversion efficiency higher than traditional manual rule schemes and single model schemes. (4) By introducing external compliant online lending data to build a comprehensive debt and multiple borrowing risk identification mechanism, high-risk customers were effectively intercepted, which made up for the shortcomings of the traditional solution that relied solely on internal data and could not identify the risk of customers’ external implicit debt. This improved the accuracy of marketing while strengthening the front-end control capability of credit risk. (5) A feature drift real-time monitoring and incremental learning automatic iteration mechanism for production environment has been established, which supports the periodic automatic update of model parameters and multi-version rollback management, so that the model can continuously adapt to changes in customer behavior patterns and market environment, maintain stable prediction performance in the long term, and effectively reduce the operation and maintenance costs caused by manual full retraining. Attached Figure Description

[0027] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for screening potential customers for personal consumer loans based on multi-source big data fusion, provided as an embodiment of the present invention; Figure 2 This is a diagram illustrating a five-layer progressive feature engineering processing architecture for multi-source compliant data provided in this embodiment of the invention. Figure 3 This is a schematic diagram of the multi-base model parallel fusion meta-logistic regression stacking modeling hierarchy provided in an embodiment of the present invention; Figure 4 This is a structural diagram of a potential customer screening device for personal consumer loans based on multi-source big data fusion, provided in an embodiment of the present invention. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0030] The following description, with reference to the accompanying drawings, describes a method and apparatus for screening potential customers for personal consumer loans based on multi-source big data fusion, according to an embodiment of the present invention.

[0031] Example 1 This invention provides a method for screening potential customers for personal consumer loans based on multi-source big data fusion. Figure 1 This is a flowchart illustrating a method for screening potential customers for personal consumer loans based on multi-source big data fusion, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step S1: Collect internal bank business data, government industry data, external compliant online lending data, online application behavior data, and historical bank outbound call data. De-identify, clean, and standardize the collected data and merge them horizontally using the bank's internal customer ID as the global primary key to construct a wide table of customer panoramic data.

[0032] It should be understood that this step corresponds to the entire data source collection and data preprocessing chain of the entire screening system, fully covering... Figure 2 The diagram shown in the Five-Source Big Data Feature Engineering Layered Processing Framework illustrates two main layers: the raw data layer and the data preprocessing layer. Figure 2 As shown, the entire feature engineering system is arranged from top to bottom as follows: raw data layer, data preprocessing layer, feature derivation layer, feature selection layer, and modeling feature output layer. This step completes the top-level raw data acquisition and the second-level preprocessing and management work, which is the data foundation for all subsequent feature derivation and modeling operations.

[0033] It should be further explained that the original data layer is divided into five independent data sources, namely internal bank data, government data, external online lending data, online behavior data, and outbound call history data. The five types of data respectively build a customer panoramic data foundation from five dimensions: customer repayment ability, social qualifications, comprehensive debt, real-time behavior, and marketing interaction. The embodiment of this application adopts a multi-interface hierarchical collection architecture to complete the compliant collection of the five types of data.

[0034] For example, for internal bank business data, structured data such as customer unit information, loan history, asset deposits and wealth management, credit information, card spending, payroll, and credit card repayments are synchronized in batches through direct connection interfaces to the bank's big data warehouse; for government sector data, qualified data such as contribution base and continuous contribution duration are collected by connecting to the housing provident fund and social security systems through compliant shared channels; for external compliant online lending data, encrypted interfaces of licensed service providers are connected to obtain online lending and repayment information, application frequency, outstanding debt, overdue and multiple borrowing records; for online real-time behavioral data, behavioral log data such as browsing, clicking, searching, dwell time, and pre-application of mobile banking and lifestyle apps are captured in a streaming manner; for outbound call historical data, data from the outbound call center system is synchronized to extract outbound call connection rate, customer intent, and historical response records. The system is configured with a timed incremental update mechanism to complete the full data synchronization on a daily basis, continuously ensuring the real-time performance and integrity of multi-source heterogeneous data. The entire data collection process implements the financial industry's full-process compliance and pre-control requirements, simultaneously taking into account both static basic stock data and dynamic real-time behavioral incremental data. It fills in the gaps in traditional solutions by providing four key data types: government data, external online lending, online behavior, and outbound call records, thus avoiding the defects of one-sided customer profiles and distorted assessments caused by a single data source.

[0035] To further clarify, after all five types of heterogeneous raw data have been collected, they are uniformly transferred to [a platform / system / etc.]. Figure 2 The second layer, the data preprocessing layer, performs standardized governance, relying on three functional modules—de-identification, cleaning, and fusion—to eliminate problems such as format barriers, privacy leakage risks, and difficulties in matching customer associations among multiple data sources.

[0036] First, a tiered data anonymization process was implemented, strictly adhering to relevant financial data security regulations. For core privacy fields such as names, ID cards, and mobile phone numbers, a composite processing scheme combining irreversible hash encryption and partial field hiding was adopted to fundamentally mitigate the risk of leakage of sensitive customer information. Second, a full-scale data cleaning operation was conducted, using a duplicate value comparison algorithm to traverse all customer records, automatically identifying and eliminating duplicate and redundant data. For continuous numerical fields such as income and liabilities, the 3σ criterion combined with a box plot algorithm was used to identify and correct extreme outlier data. Differentiated missing value filling strategies were configured based on field type, with median filling used for regular continuous fields and model prediction filling used for high-dimensional online behavioral feature fields, preserving the effective business information within the original data to the greatest extent possible. Finally, using the bank's internal customer unique ID as the global primary key, the five types of cleaned and anonymized data were horizontally spliced ​​and integrated according to the customer dimension. The naming conventions, numerical units of measurement, and statistical time calibers of all data sources were unified, resolving semantic conflicts and format differences inherent in heterogeneous data, and ultimately constructing a standardized wide table of customer panoramic data.

[0037] It is worth mentioning that the core technical features of this step are reflected in the collaborative collection of multi-source heterogeneous data, targeted filtering of financial business fields, hierarchical privacy desensitization control, repair of differentiated missing values, and single primary key full-domain data fusion. The overall working principle is to complete three core transformations of the original multi-source data: noise reduction, compliance, and normalization. This breaks down the long-standing data silos between internal bank departments, government departments, and third-party service providers, integrating scattered and independent multi-channel data into a unified, standardized, and usable dataset. On the one hand, this meets the regulatory requirements for the security and compliance of financial customer data; on the other hand, it removes data noise and fills in data defects, providing a basis for... Figure 2 Subsequent feature mining work in the feature derivation layer lays a solid foundation for high-quality data.

[0038] Step S2: Based on the customer panoramic data wide table, credit risk feature set, consumption tendency feature set and loan willingness feature set are derived respectively. The derived features are then filtered through a combination of variance screening, mutual information screening and regularization screening to obtain an effective modeling feature set.

[0039] It should be understood that this step corresponds completely. Figure 2The five-source big data feature engineering layered processing flowchart consists of a three-layer progressive processing flow: feature derivation layer, feature selection layer, and modeling feature output layer. It receives the wide table of customer panoramic data output from S1 and performs full-dimensional feature engineering operations. Feature engineering serves as the core transformation bridge between raw business data and machine learning models, and is also a key innovative module that distinguishes this application's embodiment from traditional manual rule modeling schemes. Figure 2 As shown, after the data flows into the feature derivation layer through the preprocessing layer, it will be split into three independent and parallel feature sets. The three feature sets are simultaneously sent to the feature selection layer for simplification. Finally, the modeling feature output layer outputs fixed-dimensional standardized modeling features.

[0040] It should be noted that in the feature derivation layer, relying on the multi-dimensional business fields built into the customer panoramic data wide table, three non-overlapping feature sets are derived according to the three core objectives of customer assessment. The three features correspond to the three core assessment dimensions of customer credit risk, consumption needs, and loan intention.

[0041] The first category is the credit risk feature set, which is derived from the existing credit history, assets and liabilities, housing provident fund contributions, and online lending related data in the broad table. Specifically, it generates indicators such as the number of credit overdue payments and overdue days, the customer's comprehensive debt ratio, the continuous contribution period of the housing provident fund, the frequency of online lending applications in the past year, and the distribution of online lending amount ranges in the past year. All indicators are used to quantitatively assess the customer's long-term repayment ability and overall credit risk level. The second category is the consumption tendency feature set, which is derived from data such as payroll records, income levels, consumption details, and asset size. Specifically, it generates indicators such as income level grading, average monthly consumption amount, frequency of large-amount consumption, consumption scenario preference distribution, and asset-to-consumption ratio, which are used to quantitatively assess customers' actual consumption capacity and the intensity of potential consumption demand. The third category is the loan intention feature set, which is derived from online APP behavior tracking, loan search click logs, and outbound call interaction records. Specifically, it generates online product activity duration indicators, loan keyword search frequency indicators, credit product click frequency indicators, outbound call connection rate indicators, and historical outbound call response intention rating indicators, which are used to quantify the strength of customers' intention to actively apply for loans in real time.

[0042] Optionally, the three complete feature sets are generated and then simultaneously entered into the process. Figure 2The feature selection layer shown employs a triple-filtering algorithm to eliminate invalid and redundant features layer by layer. The entire filtering process sequentially performs variance filtering, mutual information filtering, and L1+L2 combination regularization filtering to complete feature simplification step by step. First, variance screening is performed on all derived features. The variance of each feature across all customer samples is calculated. Features with variances below a preset threshold are deemed invalid due to insufficient sample discrimination and are directly removed. Only features with basic discrimination capabilities are retained and proceed to the next stage. Second, mutual information screening is performed. The mutual information between retained features and loan conversion target variables is calculated one by one. Only features with mutual information values ​​above a preset correlation threshold are retained, while irrelevant features with extremely weak correlation to the prediction target are removed, ensuring that the remaining features can contribute effectively to the prediction results for potential customers. Finally, a combination of L1 and L2 regularization is performed. L1 regularization promotes the sparsity of redundant feature weights to zero, while L2 regularization constrains prediction bias caused by feature multicollinearity. Invalid redundant features with weights approaching zero and duplicate contribution fields within highly correlated feature groups are simultaneously removed. After this three-layer algorithm simplification, a set of 80-dimensional, non-redundant, and highly discriminative effective modeling features is finally selected and output. Figure 2 The bottommost modeling feature output layer outputs a unified value.

[0043] To further clarify, the core technical features of this step are three-dimensional targeted feature decomposition and multi-algorithm cascaded feature screening. The overall technical approach strictly revolves around the three major customer assessment objectives of risk, needs, and willingness. It transforms unstructured and fragmented raw business data into standardized quantitative features that can be directly recognized by machine learning models. The triple screening algorithm filters low-value information from three perspectives: data distribution dispersion, feature target correlation, and model fitting weights, balancing the comprehensiveness of the feature system with the simplicity of the input dimensions. This step effectively solves the technical drawbacks of traditional bank potential customer screening schemes that rely on manually constructed features and lack standardized modeling input features. It achieves a quantitative and digital expression of the three core attributes of customers. The 80-dimensional refined output features not only avoid problems such as model overfitting and slow online inference caused by ultra-high-dimensional feature input, but also lock in the upper limit of the prediction accuracy of the subsequent Stacking integrated model from the data input end.

[0044] Step S3: The effective modeling feature set is split into three feature subsets, which are then input into three heterogeneous base models running in parallel. Each base model outputs a corresponding special evaluation score. The special evaluation scores are then aggregated into a fusion feature set. The meta-model then performs a comprehensive evaluation on the fusion feature set and outputs the customer loan conversion probability.

[0045] It should be understood that this step is the core algorithm execution stage for achieving high-precision potential customer identification in the embodiments of this application, and it fully corresponds to... Figure 3 The complete five-level hierarchical structure, such as Figure 3As shown, the entire model architecture is divided into an input layer, a parallel base model layer, a secondary feature layer, a meta-model fusion layer, and an output layer from top to bottom. This step relies on a layered and stacked integrated learning architecture to complete the multi-dimensional special evaluation and global comprehensive judgment. The entire data flow and model division of labor are completely consistent with the hierarchical logic shown in the diagram.

[0046] It should be added that, Figure 3 The top layer is the model input layer, which is directly connected to S2. Figure 2 The 80-dimensional effective modeling feature set output by the modeling feature output layer is split according to the source of feature derivation before entering the parallel base model operation. The 80-dimensional features are divided into three independent feature subsets: credit risk feature subset, consumption tendency feature subset, and loan intention feature subset. The three subsets are matched with the three major assessment dimensions of customers respectively and are separately sent to three types of heterogeneous parallel base models to carry out special score calculation.

[0047] For example, the three base models are in Figure 3 The middle layer operates independently in parallel, with no interference between the computational processes of the models. Different algorithms match their own exclusive feature subsets based on their adaptability: the LightGBM gradient boosting tree model on the left receives the credit risk feature subset, and relies on the powerful nonlinear relationship mining and high-order feature interaction learning capabilities of the tree model to complete the quantitative assessment of customer default risk and output an independent credit risk score; the XGBoost gradient boosting tree model in the middle receives the consumption propensity feature subset, and completes nonlinear mapping learning for complex hidden behavioral patterns in consumption data, outputting a consumption propensity score that quantifies the strength of customer consumption demand; the LR logistic regression linear model on the right receives the loan intention feature subset, and completes intention vectorization by relying on linear weighted discrimination and probabilistic linear mapping, outputting a loan intention score. The logistic regression model simultaneously retains the high interpretability inherent in linear algorithms, fully adapting to the hard requirements of financial credit risk control supervision for model traceability and interpretability.

[0048] It is worth mentioning that after the three base models output credit risk scores, consumption propensity scores, and loan willingness scores respectively, the three scores are uniformly merged into... Figure 3 The second-level feature layer, or secondary feature layer, is responsible for fusing the outputs of the three base models, aggregating the three types of one-dimensional scores to construct a three-dimensional score fusion secondary feature set, completing the secondary feature construction of the first-level model results, and providing composite input features for the upper-level meta-model.

[0049] To further explain, the 3D score fusion feature set is then passed down to... Figure 3In the meta-model fusion layer, this embodiment selects a meta-logistic regression model as the upper-layer fusion meta-model. The meta-logistic regression model automatically learns the optimal weight allocation relationship corresponding to the scores of the three special categories, neutralizing and offsetting the evaluation biases brought by the single base model, and completing a comprehensive global assessment of the customer's overall loan potential. After the meta-model completes the weighted calculation, it sends the comprehensive judgment result to... Figure 3 The bottom output layer outputs the potential conversion probability of customer loans in the form of continuous numerical values ​​in the range of 0 to 1. This probability value intuitively represents the likelihood that the customer will eventually apply for a consumer loan.

[0050] Optionally, the base model and meta-model architecture of this application embodiment have flexible replacement and expansion capabilities. The entire Stacking two-layer stacked integrated structure can be replaced with a Bagging integrated architecture or a Boosting serial boosting architecture. The three basic algorithms of the parallel layer LightGBM, XGBoost, and LR can be replaced with heterogeneous algorithms such as CatBoost gradient boosting model and DNN deep neural network according to business needs, flexibly adapting to the data distribution characteristics of different bank customers. This step relies on the core technical idea of ​​heterogeneous model division of labor for special evaluation and two-layer stacking hierarchical fusion decision-making, giving full play to the differentiated advantages of different machine learning algorithms. Gradient boosting tree algorithms are good at mining complex nonlinear relationships within financial data, and logistic regression meets the interpretability requirements of financial industry supervision. The two-layer stacking architecture first scores by dimension and then globally merges to correct biases. Compared with traditional single rule models and single algorithm statistical models, the generalization ability and prediction accuracy are greatly improved. It can accurately screen high-quality consumer loan potential customers with low default risk, high consumption demand, and strong loan intention, and is highly adaptable to the business scenario of commercial banks' credit marketing and risk control in parallel.

[0051] Step S4: Generate a comprehensive customer potential score based on the customer loan conversion probability, intelligently stratify customers according to a preset scoring threshold and output a list of potential customers, monitor the data distribution drift of the effective modeling feature set, and when the drift exceeds the threshold, trigger incremental training based on new samples to update the model parameters, and support historical version rollback to complete the closed-loop iteration of the model.

[0052] It should be understood that this step is divided into two major business logic modules. The first module is the transformation of model prediction results into banking business scenarios, completing the segmentation of potential customers and the output of marketing lists. The second module is the automated operation and maintenance module for the entire life cycle of the model, realizing a closed-loop operation mechanism of feature drift monitoring, incremental iteration, and version rollback. It takes over the customer loan conversion probability output by S3 to carry out subsequent full-process business processing and long-term model maintenance, forming a complete closed-loop operation link of the entire potential customer screening system from data collection, modeling and prediction, business implementation, and automatic updates.

[0053] It should be noted that, regarding the potential customer prediction and intelligent tiered business conversion process, the loan conversion probability of customers in the 0-1 range output by the S3 model is first converted into a comprehensive potential score of 100 points in the 0-100 range using a fixed linear mapping conversion rule. This completes the conversion of abstract probability values ​​from machine learning into an intuitive and easy-to-understand business score for bank marketing personnel. Secondly, the system retrieves preset tiering threshold rules to complete the intelligent tiering of customers. Customers with scores of 85 or higher are classified as high-potential customers, those with scores between 65 and 84 are classified as medium-potential customers, and those with scores below 65 are classified as low-potential customers. Different levels of customers correspond to differentiated marketing outreach priorities and supporting service strategies. After tiering, the system automatically filters the two core marketing customer groups of high-potential and medium-potential, and simultaneously matches them with the bank's preset gray list and compliance exemption customer list, automatically removing prohibited marketing customers, generating a standardized structured potential customer list, and automatically pushing it to the bank's front-end customer marketing business system for offline account managers and online digital marketing channels to carry out precise and differentiated marketing.

[0054] Optionally, the customer segmentation scoring threshold and segmentation rules can be adaptively adjusted according to the customer structure and credit business operation characteristics of different commercial banks to meet the differentiated marketing needs of different institutions. This module connects the underlying machine learning algorithm with the front-end banking business application, reduces ineffective marketing outreach to low-intent customers by relying on customer segmentation screening, lowers the overall customer acquisition cost of banks, and simultaneously improves the marketing conversion efficiency of consumer loan products.

[0055] To further clarify, the automated operation and maintenance module for the entire model lifecycle is a key closed-loop design that ensures the long-term stable and high-precision operation of the entire screening system. It continuously monitors the full-dimensional distribution of the 80-dimensional effective modeling feature set output by S2. The system continuously collects the mean, variance, and overall distribution statistics of all modeling features in real time, and dynamically tracks feature distribution shifts caused by changes in customer behavior and market fluctuations. When the drift metric of any single feature exceeds the system's preset warning threshold, the system automatically triggers a feature drift warning signal to remind operation and maintenance personnel that there is a significant shift in data distribution.

[0056] In response to drift warning signals, the system automatically extracts new real business customer sample data from banks monthly. Combined with the offset feature details generated by drift analysis, it performs incremental training on the currently running Stacking integrated model. The model weight parameters are fine-tuned only based on the new samples, without the need for manual intervention to retrain all historical samples, significantly reducing the computing power consumption of model iteration and the workload of manual operation and maintenance. After each round of incremental iteration, the system completely retains all model parameters, hyperconfiguration files, and training logs of this iteration, generating an independent historical version archive. When the system detects abnormal decline in core performance indicators such as model AUC, KS, and recall after the iteration update, it supports operation and maintenance personnel to perform a one-click rollback operation to quickly switch to the previous stable historical model version, avoiding the operational risk of a significant decline in the online business screening effect.

[0057] It is worth mentioning that the core technical features of the operation and maintenance module in this step are real-time monitoring of feature drift across all dimensions, periodic incremental automatic training and updates, and rollback of historical versions of multiple models. Relying on the long-term operation and maintenance approach of cyclical iteration, it continuously adapts to changes in the data distribution of market and customer behavior, dynamically corrects model prediction bias, and fundamentally solves the industry pain point of traditional bank screening schemes where business rules and prediction models are fixed and prediction accuracy continues to decline after long-term online operation. This enables the entire potential customer screening system to perform long-term, stable, and high-precision automated screening for millions of existing customers of banks, and significantly reduces the manpower cost of manual model operation and maintenance.

[0058] Integrating the complete closed-loop execution process from S1 to S4, Figure 2 Feature engineering layered processing architecture Figure 3 The Stacking dual-layer integrated model algorithm architecture in this application uses S1 multi-source compliant data fusion and S2 standardized 3D feature engineering as the underlying data processing foundation, and S3 heterogeneous dual-layer Stacking integrated machine learning model as the core high-precision prediction algorithm. It relies on S4 business layered implementation and automated model iteration closed-loop operation and maintenance to form a complete commercial technology system. The entire solution is interconnected and logically self-consistent in each step. It systematically solves many inherent technical defects of traditional consumer loan potential customer screening solutions, such as incomplete data dimensions, one-sided customer assessment logic, insufficient prediction accuracy of single models, weak credit risk management capabilities, and the inability of models to adapt and iterate for a long time. The entire process of data collection, feature derivation, modeling and prediction, and operation and maintenance iteration all comply with all the hard specifications of commercial bank big data compliance application, machine learning credit modeling, and financial customer privacy and security supervision. It has strong practicality for offline implementation and wide-ranging industry promotion value.

[0059] For example, the embodiments of this application are equipped with a multi-dimensional technology replacement and expansion scheme, which can flexibly adjust the module implementation method without changing the core execution logic of S1-S4: the triple concatenation screening algorithm of the feature selection layer can be replaced by the PCA principal component dimensionality reduction algorithm and the RFE recursive feature elimination algorithm to complete feature simplification; the original data source can be expanded to include online shopping consumption data and operator user behavior data to further enrich the feature derivative dimensions and improve the completeness of customer profile.

[0060] Example 2 This invention provides a device for screening potential customers for personal consumer loans based on multi-source big data fusion. Figure 4 This is a schematic diagram of a potential customer screening device for personal consumer loans based on multi-source big data fusion, provided in an embodiment of the present invention. Figure 4 As shown, the device includes: The multi-source data fusion unit 100 is used to collect internal bank business data, government industry data, external compliant online lending data, online application behavior data, and bank outbound call history data. The collected data is de-identified, cleaned, and standardized and fused. The data is horizontally spliced ​​with the bank's internal customer number as the global primary key to build a wide table of customer panoramic data. The multi-dimensional feature filtering unit 200 is used to derive credit risk feature set, consumption tendency feature set and loan willingness feature set based on the wide table of customer panoramic data, and to filter the derived features through a combination of variance filtering, mutual information filtering and regularization filtering to obtain an effective modeling feature set; The dual-layer integrated inference unit 300 is used to split the effective modeling feature set into three feature subsets, which are respectively input into three heterogeneous base models running in parallel. Each base model outputs a corresponding special evaluation score, and the special evaluation scores are aggregated into a fusion feature set. Then, the meta-model performs a comprehensive evaluation on the fusion feature set and outputs the customer loan conversion probability. The intelligent stratification iteration unit 400 is used to generate a comprehensive customer potential score based on the customer loan conversion probability, intelligently stratify customers according to a preset scoring threshold and output a list of potential customers, monitor the data distribution drift of the effective modeling feature set, and when the drift exceeds the threshold, trigger incremental training to update the model parameters based on the new samples, and support the rollback of historical versions to complete the closed-loop iteration of the model.

[0061] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0062] Example 3 To implement the methods of the above embodiments, the present invention also provides an electronic device, which includes a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the various steps of the methods described above.

[0063] Example 4 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0065] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0066] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for screening potential customers for personal consumer loans based on multi-source big data fusion, characterized in that, include: Collect internal bank business data, government industry data, external compliant online lending data, online application behavior data, and historical bank outbound call data. De-identify, clean, and standardize the collected data and merge them horizontally using the bank's internal customer ID as the global primary key to construct a wide table of customer panoramic data. Based on the aforementioned customer panoramic data wide table, credit risk feature sets, consumption tendency feature sets, and loan willingness feature sets are derived respectively. The derived features are then filtered through a combination of variance filtering, mutual information filtering, and regularization filtering to obtain an effective modeling feature set. The effective modeling feature set is split into three feature subsets, which are then input into three heterogeneous base models running in parallel. Each base model outputs a corresponding special evaluation score. The special evaluation scores are then aggregated into a fusion feature set. The meta-model then performs a comprehensive evaluation on the fusion feature set and outputs the customer loan conversion probability. A comprehensive customer potential score is generated based on the customer loan conversion probability. Customers are intelligently segmented according to a preset scoring threshold, and a list of potential customers is output. The data distribution drift of the effective modeling feature set is monitored. When the drift exceeds the threshold, incremental training is triggered based on the new samples to update the model parameters, and historical version rollback is supported to complete the closed-loop iteration of the model.

2. The method according to claim 1, characterized in that, Collect internal bank business data, government sector data, external compliant online lending data, online application behavior data, and historical bank outbound call data, including: By directly connecting to the bank's internal big data warehouse, internal business data is collected in batches and synchronously to obtain customer unit information, loan history, asset deposit and wealth management data, credit report records, card spending details, payroll flow and credit card repayment records, which serve as the basic data source for customer repayment ability assessment and qualification verification. Based on the compliant shared data channel, we connect with the housing provident fund management system and the social security management system to collect government industry data, obtain data on contribution base and continuous contribution duration, and use this as a basis for assessing the social qualifications and income stability of our clients. Connect to the encrypted data interface of licensed financial service institutions to collect external compliant online lending data, extracting online lending information, repayment plan execution information, loan application frequency, total outstanding debt, historical overdue records and the number of multiple lending institutions, as the basis for assessing the customer's comprehensive debt risk and credit behavior; Real-time user behavior data on mobile banking and lifestyle applications is captured using streaming data collection. This includes browsing history, click history, search keyword history, page dwell time, and loan pre-application behavior log data, which serve as a data source for real-time perception of customer consumption needs and loan intentions. Synchronize historical data from the bank's outbound call center system, collect historical outbound call data, and extract outbound call connection rate, customer intention level, and historical response record text as a basis for evaluating customer marketing interaction preferences and cooperation.

3. The method according to claim 2, characterized in that, The collected data was anonymized, cleaned, and standardized before being merged. Using the bank's internal customer ID as the global primary key, a wide table summarizing the customer's overall data was constructed, including: The five types of heterogeneous raw data collected were subjected to hierarchical desensitization processing. For personal privacy information fields containing names, ID card numbers, and mobile phone numbers, an irreversible hash encryption algorithm combined with field partial hiding technology was used for desensitization and conversion. Cleaning is performed on the five types of data after desensitization. The entire record is traversed by a duplicate value comparison algorithm to detect and remove duplicate and redundant records. For numerical continuous fields such as income and liabilities, statistical identification algorithms are used to detect and correct outliers. For missing values ​​in regular continuous fields, a median imputation strategy is used, and for missing values ​​in high-dimensional behavioral feature fields, a model prediction imputation strategy is used to maximize the retention of effective information. Using the bank's internal customer ID as the global primary key, the five types of cleaned data are horizontally concatenated according to the customer dimension. The field naming format, unit of measurement, and time caliber of each data source are unified to eliminate format differences and semantic conflicts between multi-source heterogeneous data and construct a wide table of customer panoramic data.

4. The method according to claim 3, characterized in that, Based on the aforementioned customer panoramic data wide table, credit risk feature sets, consumption propensity feature sets, and loan willingness feature sets are derived, including: Based on the credit history, asset and liability data, housing provident fund contribution duration, online loan application frequency, and online loan amount range data in the aforementioned customer panoramic data wide table, a credit risk feature set is derived and constructed. This set generates indicators such as the number of credit overdue payments and overdue days, comprehensive debt ratio, housing provident fund continuous contribution duration, online loan application frequency in the past year, and online loan amount range distribution in the past year. These indicators are used to quantitatively assess the customer's ability to fulfill obligations and credit risk level. Based on the payroll records, income level classification, average monthly consumption expenditure, frequency of large-amount consumption, distribution of consumption scenario categories, and asset-to-consumption ratio data in the customer panoramic data wide table, a consumption tendency feature set is derived and constructed to generate income level classification indicators, average monthly consumption amount indicators, frequency of large-amount consumption indicators, consumption scenario preference distribution indicators, and asset-to-consumption ratio indicators, which are used to quantitatively assess customers' consumption capacity and consumption demand intensity. Based on the online application behavior duration records, loan-related search keyword records, loan product click frequency, outbound call connection rate statistics, and historical response intention records in the aforementioned customer panoramic data wide table, a loan intention feature set is derived and constructed. This generates online behavior activity duration indicators, loan keyword search frequency indicators, loan product click frequency indicators, outbound call connection rate indicators, and historical response intention rating indicators, which are used to quantitatively assess the strength of customers' loan intentions.

5. The method according to claim 4, characterized in that, The derived features are filtered using a combination of variance filtering, mutual information filtering, and regularization filtering to obtain an effective modeling feature set, including: Variance screening is performed on all derived features. The variance value of each feature over all samples is calculated. Features with variance below a preset threshold are identified as features without discrimination and are removed. Features with sufficient sample discrimination ability are retained to enter the next round of screening. Mutual information filtering is performed on the features retained after variance filtering. The mutual information between each feature and the target variable is calculated. Features with mutual information higher than the preset correlation threshold are retained, and irrelevant features with low correlation with the target variable are removed to ensure that the retained features make an effective contribution to the prediction target. After variance screening and mutual information screening, the features retained are subjected to L1 and L2 combined regularization screening. L1 regularization drives the sparsity of redundant feature weights, and L2 regularization suppresses the influence of feature collinearity. Redundant features with weights approaching zero and duplicate contribution features in strongly correlated feature groups are removed to obtain a simplified effective modeling feature set.

6. The method according to claim 5, characterized in that, The effective modeling feature set is split into three feature subsets, which are then input into three heterogeneous base models running in parallel. Each base model outputs a corresponding specific evaluation score. These specific evaluation scores are aggregated into a fused feature set, which is then comprehensively evaluated by a meta-model to output the customer loan conversion probability, including: The effective modeling feature set is divided into a credit risk feature subset, a consumption propensity feature subset, and a loan intention feature subset according to their source. Each feature subset corresponds to three assessment dimensions: customer credit risk, consumption demand, and loan intention. A subset of credit risk features is input into the first heterogeneous base model, which performs nonlinear relationship mining and feature interaction learning on the credit risk features based on the gradient boosting tree algorithm, and outputs a credit risk score to quantify the customer's default risk level. A subset of consumption propensity features is input into a second heterogeneous base model. This base model performs complex pattern recognition and nonlinear mapping learning on the consumption propensity features based on the gradient boosting tree algorithm, and outputs a consumption propensity score to quantify the customer's consumption needs and consumption capacity. A subset of loan intention features is input into a third heterogeneous base model. This base model performs linear weighted discrimination and probability mapping on the loan intention features based on a linear classification algorithm, and outputs a loan intention score to quantify the strength of the customer's loan intention, while maintaining the interpretability of the evaluation results in this dimension. By aggregating the credit risk score, consumption propensity score, and loan willingness score, a three-dimensional score fusion feature set is constructed; The three-dimensional score fusion feature set is input into the meta-model, which uses a logistic regression algorithm to learn the weights of the three special assessment scores and make a comprehensive judgment, outputting the customer loan conversion probability.

7. The method according to claim 6, characterized in that, A comprehensive customer potential score is generated based on the customer loan conversion probability. Customers are intelligently segmented according to a preset scoring threshold, and a potential customer list is output. Data distribution drift of the effective modeling feature set is monitored. When the drift exceeds the threshold, incremental training is triggered based on new samples to update the model parameters. Historical version rollback is supported, including: The customer loan conversion probability is converted into a percentage score according to a linear mapping relationship, which is used as the customer's comprehensive potential score, realizing the transformation of the abstract probability value output by the model into an intuitive business score; Based on the preset high-potential and low-potential scoring thresholds, customers are divided into three levels: high-potential customers, medium-potential customers, and low-potential customers, and marketing priorities and service strategies are differentiated for different levels of customers. From the segmented customers, high-potential and medium-potential customers are selected, and customers on the system's preset gray list and those exempted from compliance and safety are automatically matched and removed. A standard formatted list of potential customers is generated and pushed to the customer marketing system for the business side to carry out differentiated marketing. The system continuously monitors the data distribution trends of each feature in the effective modeling feature set, calculates the mean, variance, and distribution morphology statistics of the features, and automatically triggers a drift warning signal when the drift measure of any feature exceeds the preset warning threshold. In response to the drift warning signal, new business sample data is extracted monthly, and incremental training is performed on the current model in combination with the drift analysis results to automatically update the model parameters without the need for manual full retraining. Each time the model is iterated and updated, all model parameters and configuration information of the current version are retained to form a multi-version history archive. When an abnormal performance of the model is detected after the update, it supports one-click rollback to the previous stable version, realizing automated closed-loop operation and maintenance of the model throughout its entire life cycle.

8. A device for screening potential customers for personal consumer loans based on multi-source big data fusion, characterized in that, include: The multi-source data fusion unit is used to collect internal bank business data, government industry data, external compliant online lending data, online application behavior data, and bank outbound call history data. The collected data is de-identified, cleaned, and standardized for fusion. The data is then horizontally assembled using the bank's internal customer ID as the global primary key to construct a wide table of customer panoramic data. The multi-dimensional feature filtering unit is used to derive credit risk feature sets, consumption tendency feature sets, and loan willingness feature sets based on the wide table of customer panoramic data, and to filter the derived features through a combination of variance filtering, mutual information filtering, and regularization filtering to obtain an effective modeling feature set; The dual-layer integrated inference unit is used to split the effective modeling feature set into three feature subsets, which are respectively input into three heterogeneous base models running in parallel. Each base model outputs a corresponding special evaluation score, and the special evaluation scores are aggregated into a fusion feature set. Then, the meta-model performs a comprehensive evaluation on the fusion feature set and outputs the customer loan conversion probability. The intelligent stratified iteration unit is used to generate a comprehensive customer potential score based on the customer loan conversion probability, intelligently stratify customers according to a preset scoring threshold and output a list of potential customers, monitor the data distribution drift of the effective modeling feature set, and when the drift exceeds the threshold, trigger incremental training to update the model parameters based on the new samples, and support the rollback of historical versions to complete the closed-loop iteration of the model.

9. An electronic device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images