User behavior data-based elasticity limit decision calculation method and system

Abstract

The invention relates to an elasticity limit decision calculation method and system based on user behavior data. The system supports real-time response to credit risk change of a user, realizes intelligent distribution and optimization of the credit line through an elastic line adjustment mechanism, gives consideration to risk control and user experience, and has the advantages that the credit line of the user is dynamically evaluated and intelligently adjusted, and the user experience is improved. And the automation and intelligence level of credit risk control and quota management is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of Internet technology application, and particularly relates to a flexible credit decision calculation method and system based on user behavior data. BACKGROUND

[0002] With the rapid development of financial technology, especially in the fields of consumer credit, Internet finance and credit payment, traditional credit limit evaluation methods have been difficult to meet the current complex and changing user behavior patterns and risk control needs. Traditional credit limit evaluation usually relies on static credit scoring systems such as credit investigation and third-party credit scoring, and makes credit decisions based on fixed rules, lacking comprehensive evaluation ability of real-time user behavior and multi-dimensional data, resulting in over-high or over-low credit limit, affecting risk control effect and user experience.

[0003] In addition, with the development of mobile Internet, user behavior data on financial platforms is increasingly rich, such as login frequency, consumption habits, repayment behavior, device information, social relationships, etc. These data contain a lot of potential value that can be used for risk identification and credit evaluation. However, existing systems often cannot effectively integrate and model these heterogeneous data, lacking dynamic perception and response mechanism for user credit risk.

[0004] In the credit scenario, how to build an intelligent decision system that can integrate multi-source behavior data, support dynamic modeling and have flexible credit adjustment capability has become a key technical challenge to improve the risk control level and operational efficiency of financial institutions.

[0005] Currently, there is no effective solution to the problems of staticity, singularity and response lag in existing credit limit evaluation methods. SUMMARY

[0006] The purpose of the present application is to solve the technical problems of staticity, singularity and response lag in existing credit limit evaluation methods by providing a flexible credit decision calculation method and system based on user behavior data.

[0007] To achieve the above-mentioned purpose, the technical solution adopted by the present application is:

[0008] The application provides a flexible quota decision calculation system based on user behavior data, comprising a data source layer, a data collection layer, a feature processing layer, an intelligent modeling layer, a decision execution layer and a feedback optimization layer, wherein the data source layer is configured to obtain multi-source user behavior data of at least one user; the data collection layer is configured to structure and clean the multi-source user behavior data to obtain a data set; the feature processing layer is configured to extract features from the multi-source user behavior data in the data set, encode the extracted feature data to obtain encoded feature data; the intelligent modeling layer is configured to process the encoded feature data through a user portrait model, a risk assessment model and a quota prediction model to obtain a user portrait corresponding to the at least one user, a score corresponding to the at least one user and a quota corresponding to the at least one user; the decision execution layer is configured to generate a corresponding quota strategy according to the user portrait corresponding to the at least one user, the score corresponding to the at least one user and the quota corresponding to the at least one user; and the feedback optimization layer is configured to obtain current operation feedback according to the quota strategy and trigger model self-learning in the intelligent modeling layer according to the current operation feedback.

[0009] Optionally, the data collection layer comprises a cleaning module, a filling module, a normalization module and an encoding module, wherein the cleaning module is configured to clean abnormal values in the multi-source user behavior data to obtain cleaned multi-source user behavior data; the filling module is configured to fill the cleaned multi-source user behavior data to obtain filled multi-source user behavior data; the normalization module is configured to normalize the filled multi-source user behavior data to obtain normalized multi-source user behavior data; and the encoding module is configured to encode the normalized multi-source user behavior data to obtain the data set.

[0010] Further, the filling module is further configured to fill missing items in the cleaned multi-source user behavior data by using mean value, mode value or default value to obtain the filled multi-source user behavior data.

[0011] Optionally, the feature processing layer comprises a feature extraction module and a feature engineering module, wherein the feature extraction module is configured to extract features from the multi-source user behavior data in the data set according to at least one of basic features, time series features, interaction features, behavior labels and social features to obtain feature data; and the feature engineering module is configured to process the feature data by using time dimension, cross-domain interaction, graph structure modeling, time series modeling and dynamic binning to obtain the encoded feature data.

[0012] Optionally, the intelligent modeling layer comprises a user portrait model, a risk assessment model, a credit limit prediction model, and a model updating module. The user portrait model is configured to classify at least one user based on a clustering algorithm, a topic model, or a deep learning model, determine a group of the at least one user, and generate a user portrait according to the group and the encoded feature data. The risk assessment model is configured to determine a score according to the encoded feature data and display the score through a specific model. The credit limit prediction model is configured to predict a credit limit through a regression model according to the encoded feature data. The model updating module is configured to trigger self-learning of the user portrait model, the risk assessment model, and the credit limit prediction model according to a specific period and / or online triggering, and manage the user portrait model, the risk assessment model, and the credit limit prediction model through a version management and A / B testing mechanism.

[0013] Further, the risk assessment model comprises a default probability model. The default probability model is trained according to the encoded feature data through at least one of a logistic regression, an XGBoost, a LightGBM, and a DeepFM, and determines the score through the trained and converged default probability model.

[0014] Optionally, the risk assessment model further comprises a specific model. The specific model is a SHAP value explanation model configured to display the score through a visualization tool.

[0015] Optionally, the decision execution layer comprises a rule engine, an approval module, and an interface service. The rule engine is configured to filter the user portrait corresponding to at least one user, the score corresponding to the at least one user, and the credit limit corresponding to the at least one user according to a preset condition, perform a weighted calculation on the filtered user portrait corresponding to the at least one user, the score corresponding to the at least one user, and the credit limit corresponding to the at least one user, obtain a calculation result, and assign a corresponding credit limit strategy according to the calculation result in combination with an application scenario. The approval module is configured to audit the credit limit strategy through automatic approval and manual review, and obtain an audited credit limit strategy. The interface service is configured to send the audited credit limit strategy to a lending system, a risk control system, and a CRM system through a specific interface.

[0016] Optionally, the feedback optimization layer comprises a behavior feedback collection module and a model optimization module. The behavior collection module is configured to obtain a feedback result of at least one user under a current operation according to the credit limit strategy. The model optimization module is configured to trigger self-learning of the user portrait model, the risk assessment model, and the credit limit prediction model according to the feedback result.

[0017] The application provides a flexible credit decision calculation method based on user behavior data, comprising: acquiring multi-source user behavior data of at least one user; structurally cleaning the multi-source user behavior data to obtain a data set; extracting features from the multi-source user behavior data in the data set, and encoding the extracted feature data to obtain encoded feature data; processing the encoded feature data through a user portrait model, a risk assessment model and a credit prediction model respectively to obtain a user portrait corresponding to the at least one user, a score corresponding to the at least one user and a credit corresponding to the at least one user; generating a corresponding credit strategy according to the user portrait corresponding to the at least one user, the score corresponding to the at least one user and the credit corresponding to the at least one user; and acquiring current operation feedback according to the credit strategy, and triggering model self-learning in an intelligent modeling layer according to the current operation feedback.

[0018] The application adopts the above technical scheme, acquires multi-source user behavior data of at least one user, structurally cleans the multi-source user behavior data to obtain a data set, extracts features from the multi-source user behavior data in the data set, encodes the extracted feature data to obtain encoded feature data, processes the encoded feature data through a user portrait model, a risk assessment model and a credit prediction model respectively to obtain a user portrait corresponding to the at least one user, a score corresponding to the at least one user and a credit corresponding to the at least one user, generates a corresponding credit strategy according to the user portrait corresponding to the at least one user, the score corresponding to the at least one user and the credit corresponding to the at least one user, acquires current operation feedback according to the credit strategy, and triggers model self-learning in an intelligent modeling layer according to the current operation feedback, compared with the prior art, has the following technical effects: dynamic evaluation and intelligent adjustment of user credit limit, and improvement of the automation and intelligent level of credit risk control and credit management. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a schematic diagram of a flexible credit decision calculation system based on user behavior data according to embodiment one of the application;

[0020] Figure 2 is a schematic diagram of another flexible credit decision calculation system based on user behavior data according to embodiment one of the application

[0021] Figure 3 is a schematic diagram of AB testing of a model update module in a flexible credit decision calculation system based on user behavior data according to embodiment one of the application;

[0022] Figure 4 is a schematic diagram of a flexible credit decision calculation system based on user behavior data according to embodiment one of the application

[0023] Figure 5 Figure 1 is a schematic diagram of an interface service in a flexible credit decision computing system based on user behavior data according to an embodiment of the present application;

[0024] Figure 6 Figure 2 is a flowchart of a flexible credit decision computing method based on user behavior data according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the objects, technical solutions and advantages of the present application clearer, the present application is described and explained below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of the present application.

[0026] Obviously, the accompanying drawings in the following description are only some examples or embodiments of the present application, and for those of ordinary skill in the art, the present application can be applied to other similar scenarios without creative efforts based on these drawings. In addition, it can be understood that although the efforts made in this development process can be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in the present application, some designs, manufacturing or production changes based on the technical content disclosed in the present application are only routine technical means, and should not be understood as insufficient disclosure of the present application.

[0027] In the present application, "embodiments" means that the specific features, structures or characteristics described in conjunction with the embodiments can be included in at least one embodiment of the present application. The phrase appears at various places in the specification does not necessarily refer to the same embodiment, nor is it mutually exclusive or alternative embodiments to other embodiments. Those of ordinary skill in the art explicitly and implicitly understand that the embodiments described in the present application can be combined with other embodiments without conflict.

[0028] Unless otherwise defined, technical terms and scientific terms used in the present application shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. Unless otherwise defined, the terms "one" and "a" or "an" used in the present application shall not be limited to singular aspects but can include both the singular and plural aspects. The terms "including", "containing", "having" and any variations thereof in the present application are intended to cover a non-exclusive inclusion; for example, a process, method, system, product or device including a list of steps or units (units) is not limited to the listed steps or units, but can also include steps or units not listed or can also include other steps or units inherent to these processes, methods, products or devices. The terms "connected", "connected", "coupled" and the like in the present application are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The term "multiple" in the present application refers to two or more. The "and / or" describes the relationship between the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases: A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the associated objects before and after are a "or" relationship. The terms "first", "second", "third" and the like in the present application are only to distinguish similar objects, and do not represent a specific order for the objects.

[0029] Embodiment 1

[0030] An exemplary embodiment of the present application is shown in Figure 1 Figure 1 is a schematic diagram of a flexible quota decision calculation system based on user behavior data according to an embodiment of the present application. The flexible quota decision calculation system based on user behavior data provided in the present application includes:

[0031] ​The data source layer 1002, the data collection layer 1004, the feature processing layer 1006, the intelligent modeling layer 1008, the decision execution layer 1010, and the feedback optimization layer 1012, wherein the data source layer 1002 is configured to obtain multi-source user behavior data of at least one user; the data collection layer 1004 is configured to structure and clean the multi-source user behavior data to obtain a data set; the feature processing layer 1006 is configured to extract features from the multi-source user behavior data in the data set, encode the extracted feature data, and obtain encoded feature data; the intelligent modeling layer 1008 is configured to process the encoded feature data through a user portrait model, a risk assessment model, and a quota prediction model to obtain a user portrait corresponding to the at least one user, a score corresponding to the at least one user, and a quota corresponding to the at least one user; the decision execution layer 1010 is configured to generate a corresponding quota strategy according to the user portrait corresponding to the at least one user, the score corresponding to the at least one user, and the quota corresponding to the at least one user; and the feedback optimization layer 1012 is configured to obtain current operation feedback according to the quota strategy, and trigger model self-learning in the intelligent modeling layer 1008 according to the current operation feedback.

[0032] Specifically, as shown in FIG. 1, Figure 2 Figure 2 is a schematic diagram of another user behavior data-based flexible quota decision calculation system according to an embodiment of the present application. The data source layer 1002 in the user behavior data-based flexible quota decision calculation system provided in the present application is mainly configured to obtain multi-source user behavior data of at least one user, wherein the multi-source user behavior data includes online transaction data, credit data, device data, behavior data, and social data.

[0033] In the present application, the online transaction data, the credit data, the device data, the behavior data, and the social data are as follows:

[0034] The online transaction data includes user shopping records, payment frequency, amount distribution, and refund behavior.

[0035] The credit data is user credit history from a credit system or other third-party credit platform.

[0036] The device data includes device model, operating system, IP address, login location, and whether jailbroken / rooted.

[0037] The behavior data includes APP usage time, page stay, click path, search keywords, and login frequency.

[0038] The social data includes user behavior track, friend circle, and interaction frequency on a social platform (authorization required).

[0039] ​Example: Obtain user consumption behavior through API docking Xbao, Xxin payment, Xdong mall and other platforms; Obtain user credit report through Xma credit, Xhang credit and other institutions.

[0040] Optionally, the data collection layer 1004 comprises a cleaning module, a filling module, a normalization module and an encoding module, wherein the cleaning module is configured to clean abnormal values in the multi-source user behavior data to obtain cleaned multi-source user behavior data; the filling module is configured to fill the cleaned multi-source user behavior data to obtain filled multi-source user behavior data; the normalization module is configured to normalize the filled multi-source user behavior data to obtain normalized multi-source user behavior data; and the encoding module is configured to encode the normalized multi-source user behavior data to obtain a data set.

[0041] Further, the filling module is further configured to fill missing items in the cleaned multi-source user behavior data by using mean value, mode value or default value to obtain the filled multi-source user behavior data.

[0042] Specifically, in the embodiment of the application, the data collection layer 1004 uses an ETL tool (such as Apache Nifi, DataX) for data extraction, conversion and loading, uses Kafka or Flume to realize high-concurrency data stream collection, uses HDFS or object storage (such as OSS) to save raw data, realizes structured cleaning, removes abnormal values, fills missing values, normalizes fields, encodes mapping, etc., and outputs a data table in a unified format (such as Parquet, ORC) for use by the feature processing layer 1006.

[0043] For example, suppose that online transaction data, credit data, device data, behavior data and social data are obtained.

[0044] The cleaning module deletes or corrects unreasonable extreme data, for example, a user's page click count is as high as “100,000 times” in one day, which obviously does not conform to the normal behavior pattern; the processing method can be to identify as a crawler behavior and filter out the data (i.e., the cleaning module in the embodiment of the application is configured to clean abnormal values in the multi-source user behavior data to obtain cleaned multi-source user behavior data).

[0045] The filling module uses mean value, mode value, default value to fill missing items, for example, missing credit score, the processing method can be to estimate a default value according to occupation, age, region and other information; or set to “not provided” (i.e., the filling module in the embodiment of the application is configured to fill the cleaned multi-source user behavior data to obtain filled multi-source user behavior data; and the filling module is further configured to fill missing items in the cleaned multi-source user behavior data by using mean value, mode value or default value to obtain the filled multi-source user behavior data).

[0046] The different dimensional data is unified to the same scale by the normalization module, facilitating model processing, for example, behavior click times: from 0~1000 to standardization (i.e., the normalization module in the embodiment of the application, used for normalizing the filled multi-source user behavior data to obtain normalized multi-source user behavior data).

[0047] The category type field is converted into a numerical form that can be understood by the machine learning model by the encoding module, for example, social activity level: high activity is defined as 2; medium activity is defined as 1; and low activity is defined as 0 (i.e., the encoding module in the embodiment of the application, used for encoding the normalized multi-source user behavior data to obtain the data set).

[0048] The final data collection layer 1004 can output a structured, clean, and unified format data table (i.e., the data set in the embodiment of the application), which can be used for feature engineering of the feature processing layer 1006.

[0049] Optionally, the feature processing layer 1006 includes a feature extraction module and a feature engineering module, wherein the feature extraction module is configured to perform feature extraction on the multi-source user behavior data in the data set according to at least one of the basic features, the time series features, the interaction features, the behavior labels, and the social features, to obtain feature data; and the feature engineering module is configured to process the feature data through time dimension, cross-domain interaction, graph structure modeling, time series modeling, and dynamic binning, to obtain encoded feature data.

[0050] Specifically, the feature extraction module in the feature processing layer 1006 in the embodiment of the application performs feature extraction on the multi-source user behavior data in the data set according to at least one of the basic features, the time series features, the interaction features, the behavior labels, and the social features, to obtain feature data;

[0051] The basic features can include gender, age, occupation, income level, marital status, etc.

[0052] The time series features can include average consumption amount in the last 30 days, consumption volatility, activity trend, etc.

[0053] The interaction features can include the number of communications with customer service, APP opening frequency, advertisement click conversion rate, etc.

[0054] The behavior labels can include high-frequency consumption users, night active users, cross-regional login users, etc.

[0055] The social features can include social network density, whether there is a delinquent user among friends, social activity level, etc.

[0056] In a preferred example, feature extraction is performed on multi-source user behavior data in the data set according to at least one of the basic features, the time series features, the interaction features, the behavior labels, and the social features, to obtain feature data, as shown in Table 1:

[0057] Table 1

[0058]

[0059] That is, in the embodiments of the present application, corresponding feature extraction types are adopted according to the data types in the data set to perform feature extraction, to obtain feature data.

[0060] For example, analyze whether a certain user is suitable for increasing the amount

[0061] Input: User A has traded 15 times in the last 30 days, with an average single amount of 300 yuan; the credit report shows no overdue records, and the credit score is 720; the login device is iPhone 14, which has never been changed in the past week; the average time spent on browsing product pages is 2 minutes; there are no blacklisted users among friends, and the social activity is moderate.

[0062] Based on the input, the features are extracted to obtain the results shown in Table 2:

[0063] Table 2

[0064]

[0065] After merging, a unified feature table (structured form) is obtained, as shown in Table 3, that is, the feature data in the embodiments of the present application;

[0066] Table 3

[0067]

[0068] The feature engineering module in the feature processing layer 1006 in the embodiments of the present application processes the feature data through time dimension, cross-domain interaction, graph structure modeling, time series modeling, and dynamic binning, to obtain encoded feature data;

[0069] Among them, the processing of the feature data mainly includes standardization, encoding, aggregation function, discretization, and feature cross processing; in the embodiments of the present application, the standardization, encoding, aggregation function, discretization, and feature cross are as follows:

[0070] Standardization: Z-score, Min-Max, etc.

[0071] Encoding: One-Hot, Label Encoding, Embedding, etc.

[0072] Aggregation functions: Sum, Mean, Max, Min, Count, Ratio, etc.

[0073] Discretization: Binning, e.g. age bracket, income interval, etc.

[0074] Feature cross: Combine different dimension features to enhance predictive power.

[0075] In the embodiments of the present application, the feature data is processed through the time dimension, which can process the time dimension for the time decay aggregation function based on user behavior;

[0076] For example, scenario: User behavior data (such as page clicks, transaction frequency) has timeliness with time change.

[0077] Use time decay weighted aggregation to enhance the importance of recent behavior.

[0078] Example formula: WeightedSum = Σ (value_i × e^(-λ × t_i))

[0079] Where:

[0080] value_i is the value of the i-th behavior;

[0081] t_i is the time difference (unit: day) from the behavior to now;

[0082] λ is the decay coefficient, which controls the speed of time weight decline.

[0083] Application example:

[0084] Example 1: Weighted sum of click behavior in the past 7 days > original click count;

[0085] Example 2: Assign different weights to recent small high-frequency transactions vs. long-term large transactions;

[0086] Example 3: Simulate the "heat value" of user activity.

[0087] By using time decay weighted aggregation, the sensitivity of the model to the dynamic changes of user behavior is improved.

[0088] In the embodiments of the present application, the feature data is processed through cross-domain interaction, which can process the feature data through cross-dimension feature interaction mapping;

[0089] For example, scenario: Cross information between multiple data sources often reveals deeper risk or credit signals.

[0090] By constructing cross-data-source feature interaction items, as shown in Table 4, for example:

[0091] Table 4

[0092]

[0093] By cross-dimension feature interaction mapping, hidden association patterns are mined to improve model discrimination ability.

[0094] In the embodiments of the present application, the feature data is processed by graph structure modeling, which can be processed by social graph embedding;

[0095] For example, scenario: social data not only includes the number of friends, but also includes the relationship network between friends.

[0096] By using graph neural network (GNN) or Node2Vec to embed the social graph, a social feature vector (Embedding Vector) of each user is generated as part of the model input.

[0097] Example steps:

[0098] Step 1, build a user social graph;

[0099] Step 2, train node embedding using GraphSAGE, Node2Vec algorithm;

[0100] Step 3, extract the embedding vector and splice it into the structured feature table.

[0101] By processing the feature data through social graph embedding, the implicit features such as "social influence" and "circle risk" of the user can be captured.

[0102] In the embodiments of the present application, the feature data is processed by time series modeling, which can be processed by behavior sequence temporal pattern modeling:

[0103] For example, scenario: user behavior has time series characteristics (such as daily activity, weekly consumption peak).

[0104] By introducing time series modeling methods (such as LSTM, Transformer, CNN) to extract periodic and trend patterns in behavior sequences.

[0105] Example application:

[0106] Input: daily click count in the past 30 days;

[0107] Output: a compressed "behavior time series feature vector" for judging whether the user is abnormally active or has a loss tendency.

[0108] The feature data is processed through the timing modeling of the behavior sequence to improve the prediction ability of the model on the user behavior trajectory.

[0109] In the embodiments of the present application, the feature data is processed through dynamic binning, which can be processed through an adaptive binning strategy.

[0110] For example, scenario: fixed binning strategy cannot adapt to the difference in behavior distribution of different groups of people.

[0111] By using a clustering-based dynamic binning method (such as K-means, Quantile Clustering), the intervals are adaptively divided according to the user groups.

[0112] Example application:

[0113] The income level binning is no longer fixed as [0-5k, 5k-1w, 1w+]; but the binning boundary is automatically adjusted according to the income distribution of different city / professional groups.

[0114] The feature data is processed through the dynamic binning strategy to avoid the deviation caused by one-size-fits-all and improve the generalization ability of the model.

[0115] Optionally, the intelligent modeling layer 1008 includes: a user portrait model, a risk assessment model, a limit prediction model, and a model updating module, wherein the user portrait model is used to classify at least one user based on a clustering algorithm, a topic model or a deep learning model, determine the group of at least one user, and generate a user portrait according to the group and the encoded feature data; the risk assessment model is used to determine a score according to the encoded feature data, and display the score through a specific model; the limit prediction model is used to predict through a regression model according to the encoded feature data to obtain a limit; and the model updating module is used to trigger the user portrait model, the risk assessment model and the limit prediction model to learn according to a specific period and / or online, and manage the user portrait model, the risk assessment model and the limit prediction model through a version management and A / B test mechanism.

[0116] Further, optionally, the risk assessment model includes a default probability model, wherein the default probability model is obtained by training at least one of a logistic regression, an XGBoost, a LightGBM and a DeepFM according to the encoded feature data, and the score is determined by the trained default probability model.

[0117] Optionally, the risk assessment model further includes a specific model, wherein the specific model is a SHAP value explanation model, and is used to display the score through a visualization tool.

[0118] Specifically, the user portrait model in the intelligent modeling layer 1008 in the embodiment of the present application is constructed based on a clustering algorithm (such as K-means), a topic model (such as LDA), or a deep learning model (such as AutoEncoder), wherein at least one user is classified to determine a group of the at least one user, and a user portrait is generated according to the group and the encoded feature data.

[0119] In the embodiment of the present application, based on a clustering algorithm (such as K-means), a topic model (such as LDA), or a deep learning model (such as AutoEncoder), the user can be grouped from different angles, and a specific strategy can be formulated for each group.

[0120] Generating a user portrait based on a clustering algorithm (such as K-means):

[0121] Applicable scenario: when the user needs to be grouped according to certain numerical features (such as transaction amount, login frequency, etc.), K-means is used as a preferred example to illustrate:

[0122] Step 1, data preprocessing: standardize or normalize numerical features to ensure that data of different scales does not affect the clustering results.

[0123] Step 2, select K value: elbow method (Elbow Method) or silhouette coefficient (Silhouette Score) can be used to determine the optimal K value.

[0124] Step 3, model training: use K-means algorithm to cluster users.

[0125] Step 4, result analysis: check the center point of each cluster to understand the main features of each group.

[0126] Step 5, construct portrait: based on the features of the cluster center point, combine domain knowledge to construct a portrait for each group.

[0127] Generating a user portrait based on a topic model (such as LDA):

[0128] Applicable scenario: when the data contains a large amount of text information (such as user comments, social media posts, etc.), and it is desired to discover potential topics or interest points through text information, LDA is used as a preferred example to illustrate:

[0129] Step 1, data preparation: convert text data into a bag of words (Bag of Words) or TF-IDF representation.

[0130] Step 2, select the number of topics: determine the number of topics according to experience or through cross-validation.

[0131] Step3, Model Training: Use LDA algorithm to model the topics from the text data.

[0132] Step4, Result Analysis: Look at the keywords under each topic to understand what each topic represents.

[0133] Step5, Build Profile: Based on the user's interest distribution on different topics, build a profile for the user.

[0134] Generate user profile based on deep learning model (e.g. AutoEncoder):

[0135] Applicable scenario: When you want to automatically discover the internal structure of user data in high-dimensional space and may need to reduce dimensionality to reduce computational complexity, AutoEncoder is a preferred example:

[0136] Step1, Data Preprocessing: Standardize or normalize the input data.

[0137] Step2, Model Design: Design an AutoEncoder network, including encoder and decoder parts.

[0138] Step3, Model Training: Train the AutoEncoder using user data to minimize reconstruction error.

[0139] Step4, Hidden layer representation: Use the intermediate hidden layer output of the AutoEncoder as the low-dimensional representation of the user.

[0140] Step5, Further clustering or classification: The hidden layer representation can be input into other algorithms (such as K-means) for further analysis.

[0141] Step6, Build Profile: Based on the hidden layer representation or the results of further analysis, build a profile for the user.

[0142] The risk assessment model in the intelligent modeling layer 1008 in the embodiment of the application adopts algorithms such as logistic regression, XGBoost, LightGBM, and DeepFM to train a default probability model.

[0143] The input variables include the feature data encoded by the feature processing layer 1006, i.e., online transaction data, credit data, device data, behavior data, and social data.

[0144] The output is the user credit score (Score) or default probability (PD) (i.e., the score in the embodiment of the application);

[0145] Specifically, the default probability model is trained in a cascaded or multi-model integrated manner, the advantages of different models are complementary, and the overall performance is improved.

[0146] The integration strategy includes:

[0147] Strategy one: simple average method: take the average of the prediction results of the logistic regression, XGBoost, LightGBM, and DeepFM models.

[0148] Strategy two: weighted average method: assign different weights to the logistic regression, XGBoost, LightGBM, and DeepFM models based on their performance on the validation set.

[0149] Strategy three: stacking: use a layer of meta-models (such as logistic regression) to combine the outputs of XGBoost, LightGBM, and DeepFM models.

[0150] The specific process of training the default probability model in the embodiments of the present application is as follows:

[0151] Step 1, data preparation: load and preprocess the encoded feature data.

[0152] Step 2, multi-model training: train the logistic regression, XGBoost, LightGBM, and DeepFM models respectively. Each model is trained independently and its performance is evaluated on the validation set.

[0153] Step 3, model integration: if it is a simple average / weighted average: directly merge the prediction results of each model. If it is stacking: use a meta-model (such as logistic regression) to retrain the prediction results of the base model.

[0154] Step 4, model evaluation: evaluate the performance of the integrated model on the test set to ensure better performance than individual models.

[0155] Step 5, model deployment: save the integrated model and its components and deploy them to the production environment.

[0156] The SHAP value interpretation model in the embodiments of the present application can be:

[0157] 1. Model training and saving

[0158] First, one or more risk assessment models (such as XGBoost, LightGBM, etc.) need to be trained and saved for subsequent interpretation.

[0159] 2. Introduce SHAP library

[0160] The SHAP library needs to be installed and imported, which is a Python library for explaining machine learning models. Then, import SHAP in the code.

[0161] 3. Create an explainer

[0162] Create a SHAP explainer object to calculate the contribution of each feature to the model output.

[0163] 4. Calculate SHAP values

[0164] Select a portion of the test data (usually a random portion) to calculate SHAP values.

[0165] 5. Visualize SHAP values

[0166] SHAP provides multiple visualization tools to help understand the model output, as follows:

[0167] SHAP Summary Plot: Display the importance of all features and their impact direction on the model output (positive or negative).

[0168] SHAP Dependence Plot: View the relationship between a single feature and its corresponding SHAP value, which can identify nonlinear relationships between features and prediction results.

[0169] SHAP Force Plot: Generate a force plot for a single instance, showing the specific impact of each feature on the prediction result of that instance.

[0170] 6. Explain user credit score or default probability: Assuming the default probability (PD) of a certain user has been calculated, now you can use SHAP values to explain the prediction result.

[0171] For example:

[0172] # Assume we have the feature data of a user

[0173] user_features = X_sample.iloc[0:1] # Get the data of the first user

[0174] # Calculate the SHAP value of this user

[0175] user_shap_values = explainer(user_features)

[0176] # Print the predicted default probability of this user

[0177] Predicted_pd = model.predict_proba(user_features)[:, 1]

[0178] Print(f"Predicted pd: {predicted_pd[0]:4f}")

[0179] # Use Force Plot to explain the prediction result for this user

[0180] Shap.force_plot(explainer.expected_value, user_shap_values.values[0], user_features)

[0181] Predicted PD: The output is the probability of the user being predicted to default.

[0182] Force Plot: Shows the specific contribution of each feature to the prediction result. Positive values indicate an increase in the probability of default, while negative values indicate a decrease in the probability of default. In this way, it can be clearly seen which features have the greatest impact on the default probability prediction for this user.

[0183] Through the above, SHAP values can be used to explain the output of the model, whether it is the user's credit score or the probability of default. Not only does it improve the transparency of the model, but it also enhances the trust of business personnel in the model, especially in financial risk management, which has important application value.

[0184] In the present application, the credit limit prediction model in the intelligent modeling layer 1008 uses a regression model (such as linear regression, random forest regression) to predict a reasonable credit limit; considers factors such as user debt capacity, repayment willingness, and historical usage rate; sets upper and lower limits for the credit limit (such as a minimum of 500 yuan and a maximum of 100,000 yuan); dynamic adjustment mechanism: automatically adjust the credit limit according to recent changes in user behavior (such as increased consumption, good repayment).

[0185] In the present application, the use of a regression model to predict a reasonable credit limit is as follows:

[0186] Step 1, data preparation: First, relevant data needs to be collected and processed. Data typically includes but is not limited to:

[0187] User's basic information: age, gender, occupation, etc.

[0188] Financial status: income level, debt situation, asset valuation, etc.

[0189] Behavioral data: historical consumption patterns, repayment records, transaction frequency, etc.

[0190] Risk Assessment Result: Credit Score or Probability of Default (PD) obtained from the risk assessment model.

[0191] Step 2, Feature Engineering: Before building the model, some preprocessing of features may be needed, such as standardizing numerical features, encoding categorical features, etc. At the same time, consider adding score and pd as additional risk assessment features to the model.

[0192] Step 3, Model Training: Choose a suitable regression model for training. Take random forest regression as an example:

[0193] from sklearn.ensemble import RandomforestRegressor

[0194] from sklearn.model_selection import train_test_split

[0195] from sklearn.metrics import mean_squard_error

[0196] # Divide the dataset

[0197] X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.2, random_state=42)

[0198] # Initialize and train the model

[0199] model = RandomforestRegressor(n_estimators=100, random_state=42)

[0200] model.fit(X_train, y_train)

[0201] # Predict and evaluate the model

[0202] predictions = model.predict(X_test)

[0203] mse = mean_squrared_error(y_test, predictions)

[0204] Print(f"Mean Squrared Error: {mse:.4f}")

[0205] The "reasonable" credit limit should be determined based on the actual repayment ability and potential risks of the user. This can be achieved in the following ways:

[0206] Method 1: Upper and lower limit control: Set the minimum and maximum limit, for example, the minimum is 500 yuan, and the maximum is 100,000 yuan.

[0207] Method 2: Risk-based adjustment: Adjust the predicted limit based on score and pd value. For example, for high-risk users (low score or high pd), their limit can be appropriately reduced; for low-risk users, the limit can be appropriately increased.

[0208] def adjust_credit_limit(base_limit,score,pd);

[0209] if score>700and pd<0.05;

[0210] return base_limit*1.1#For high-quality users, increase by 10%

[0211] elif pd>0.1;

[0212] return base_limit*0.9#For high-quality users, reduce by 10%

[0213] else:

[0214] return base_limit

[0215] adjusted_limits=[adjust_credit_limit(pred,score,pd)for pred,score,pdin zip{predictions,df[‘scor

[0216] As can be seen from the above, the user credit limit is dynamically adjusted by credit score and default probability, with risk management as the goal.

[0217] The model updating module in the intelligent modeling layer 1008 in the embodiments of the present application learns the user portrait model, risk assessment model and limit prediction model according to a specific period and / or online triggering, wherein the specific period in the embodiments of the present application can be daily or weekly;

[0218] The model updating module in the embodiments of the present application supports online learning (Online Learning) to quickly respond to data changes; and manages the user portrait model, risk assessment model and limit prediction model through version management and A / B testing mechanism. For example,Figure 3 As shown, Figure 3 is a schematic diagram of AB testing of a model update module in a flexible quota decision-making system based on user behavior data according to an embodiment of the application; the details are as follows:

[0219] Step 1, receiving a user request;

[0220] Step 2, AB testing according to the user request through the shunting strategy of the traffic distributor; wherein the shunting strategy of the traffic distributor is configured by the experiment configuration center through the configuration of experiment ID, shunting ratio and enabled state;

[0221] Step 3, selecting version A or version B according to the shunting strategy, wherein version A is the current online model, supports one-key rollback, and is tested through prediction service A; version B is a new model to be verified, and if version B is selected, it is tested through prediction service B; and the results of prediction service A or prediction service B are stored in the record log; if version A or version B completes the test, the model is deployed;

[0222] Step 4, storing the results in the record log in the data lake / data warehouse to support offline analysis to finally generate an AB test report, wherein the AB test report also includes real-time indicators;

[0223] Step 5, returning the AB test report, and two results can be obtained according to the AB test report: result one, if the performance is significantly improved, the new model is put into full-scale online, and the old model is stopped running; result two, if the performance is not significantly improved, continue to observe or iterate.

[0224] Optionally, the decision execution layer 1010 includes a rule engine, an approval module and an interface service, wherein the rule engine is configured to filter at least one user portrait corresponding to at least one user, at least one score corresponding to at least one user and at least one quota corresponding to at least one user according to a preset condition, perform weighted calculation on the filtered at least one user portrait, at least one score and at least one quota, obtain a calculation result, and combine an application scenario to distribute a corresponding quota strategy according to the calculation result; the approval module is configured to audit the quota strategy through automatic approval and manual review to obtain an audited quota strategy; and the interface service is configured to send the audited quota strategy to a loan system, a risk control system and a CRM system through a specific interface.

[0225] Specifically, the preset condition in the embodiment of the application is a hard condition, such as age limit and blacklist filtering;

[0226] The weighted calculation in the embodiment of the application is as follows:

[0227] The user portrait corresponding to the at least one user, the score corresponding to the at least one user and the quota corresponding to the at least one user are weighted and calculated according to the weights of the respective models;

[0228] That is, Final_Score = w1*Score_Model1 + w2*Score_Model2 +... + wn*Score_Modeln.

[0229] Wherein w represents the weight of each model.

[0230] The application scenario in the embodiment of the present application can be holiday promotion or new user first loan discount; as shown in Figure 4 Figure 4 is a schematic diagram of allocating a corresponding quota strategy according to a calculation result in an elastic quota decision calculation system based on user behavior data according to Embodiment One of the present application; specifically as follows:

[0231] Step 1, receiving a user request;

[0232] Step 2, triggering a rule engine according to the user request, identifying the user scenario, obtaining a user portrait, calling a risk model and a quota prediction model, comprehensively processing the user portrait, risk assessment and quota prediction results, and obtaining a processing result; wherein the user portrait can be obtained by K-means / LDA / AE; the risk assessment model can be XGBoost / DeepFM; and the quota prediction model can be RandomForest / Regression;

[0233] Step 3, simultaneously, loading a current strategy in a strategy configuration center, wherein the current strategy can be a combined strategy, for example: holiday promotion, new user discount and high credit quota increase;

[0234] Step 4, calling an application engine rule according to the combined strategy and the processing result, judging whether it meets the risk control rule; if the judgment result is yes, executing Step 5; if the judgment result is no, executing Step 6;

[0235] Step 5, determining the final credit quota, synchronizing to the loan system and recording in the strategy execution log;

[0236] Step 6, freezing or manual review.

[0237] In the embodiment of the present application, the part of manual review is the data of high-risk users.

[0238] As shown in Figure 5 Figure 5 is a schematic diagram of interface services in an elastic quota decision calculation system based on user behavior data according to Embodiment One of the present application; the interface services in the embodiment of the present application are specifically as follows:​​

[0239] Step1, the front-end / user system sends a credit application to the credit decision engine, wherein the credit application contains user information and behavior data;

[0240] Step2, the credit decision engine calls the model service to perform user profiling, scoring and credit prediction; performs strategy configuration according to the user profiling, scoring and credit prediction; and sends the strategy configuration and / or the executed strategy configuration to the lending system, the risk control system, the CRM system and the message queue in the form of a credit event through asynchronous notification;

[0241] Step3, the credit decision engine returns the credit result to the front-end / user system.

[0242] It should be noted that in the embodiments of the present application, the credit strategy includes a credit operation type, and the credit operation type can include: credit issuance, credit increase, credit freezing, credit recovery and credit recovery. Among them, the credit issuance can be triggered by the first credit; the credit increase can be triggered based on good repayment records and behavior improvement; the credit freezing can be triggered by detecting abnormal behavior or overdue; the credit recovery can be triggered by automatic unfreezing after overdue settlement or resuming after manual review; the credit recovery can be triggered by automatically reducing the credit of long-term inactive users.

[0243] In addition, as shown in Figure 5 , the synchronous credit decision result can be the final credit or approval state; the reported risk assessment data can be the credit score Score or the default probability PD; the user tag and the life cycle state are updated; and the subsequent business processes such as loan execution, manual review and marketing activities are triggered.

[0244] Optionally, the feedback optimization layer 1012 includes a behavior feedback collection module and a model optimization module, wherein the behavior collection module is configured to obtain a feedback result of at least one user under a current operation according to the credit strategy; and the model optimization module is configured to trigger self-learning of a user profiling model, a risk assessment model and a credit prediction model according to the feedback result.

[0245] Specifically, in the embodiments of the present application, the behavior collection module can include: user subsequent behavior, repayment performance, overdue label and model bias analysis; wherein,

[0246] The user subsequent behavior can be, for example, reapplying for a loan, consumption frequency, APP activity;

[0247] The repayment performance can include: on-time repayment, partial repayment, overdue days, default amount;

[0248] The overdue label can include: defining M1, M2, M3 and above overdue levels;

[0249] The model bias analysis can include: comparing the prediction with the actual result, and identifying the model bias point.

[0250] The model optimization module in the embodiments of the present application runs a model optimization mechanism, which can include: automatically triggering model retraining tasks; adding the latest sample data to improve the model generalization ability; analyzing misjudgment cases to optimize feature selection and weight distribution; introducing techniques such as adversarial sample enhancement and transfer learning to improve robustness.

[0251] The elastic quota decision-making system based on user behavior data provided by the embodiments of the present application relates to the field of financial technology and is suitable for intelligent credit quota management and risk control scenarios. The system uses multi-factor dynamic modeling to integrate historical credit data, real-time behavior characteristics, transaction frequency, repayment habits, device login patterns, social graph relationships, and other multi-dimensional information of users, constructs high-dimensional feature vectors, and uses machine learning algorithms for dynamic quota evaluation and adjustment. The system supports real-time response to changes in user credit risk, realizes intelligent allocation and optimization of credit lines through an elastic quota adjustment mechanism, and balances risk control and user experience. It can be widely used in consumer credit, credit payment, virtual credit quota setting, etc., and has the advantages of high efficiency, high accuracy and high scalability. In addition, it can improve the scientificity and individuality of credit quota decision-making, enhance the identification and control ability of risk customers, and support dynamic adjustment and life cycle management of quotas, thereby improving the efficiency of credit resource allocation and reducing potential risks for financial institutions.

[0252] The elastic quota decision-making system based on user behavior data provided by the embodiments of the present application solves the problems of staticity, singularity and response lag in existing credit quota evaluation methods, and realizes intelligent prediction, flexible adjustment and continuous optimization of user credit lines by introducing a multi-factor dynamic modeling mechanism. It has the technical effects of strong dynamicity, real-time quota adjustment according to user behavior, high individuality, precise risk control, good scalability, and suitability for various credit products such as credit cards, consumer installment, and small loans.

[0253] The application adopts the above technical scheme, obtains multi-source user behavior data of at least one user, performs structured cleaning on the multi-source user behavior data to obtain a data set, performs feature extraction on the multi-source user behavior data in the data set, encodes the extracted feature data to obtain encoded feature data, processes the encoded feature data through a user portrait model, a risk assessment model and a quota prediction model respectively to obtain a user portrait corresponding to the at least one user, a score corresponding to the at least one user and a quota corresponding to the at least one user, generates a corresponding quota strategy according to the user portrait corresponding to the at least one user, the score corresponding to the at least one user and the quota corresponding to the at least one user, and obtains current operation feedback according to the quota strategy and triggers model self-learning in the intelligent modeling layer, compared with the prior art, the application has the following technical effects: dynamic evaluation and intelligent adjustment of user credit quota, and improved automation and intelligent level of credit risk control and quota management.

[0254] Embodiment 2

[0255] An exemplary embodiment of the application is shown in Figure 6 , Figure 6 is a flowchart of a flexible quota decision calculation method based on user behavior data according to Embodiment Two of the application, applied to the flexible quota decision calculation system based on user behavior data in Embodiment One, the flexible quota decision calculation method based on user behavior data provided in the application includes the following steps:

[0256] Step S600, obtaining multi-source user behavior data of at least one user;

[0257] Step S602, performing structured cleaning on the multi-source user behavior data to obtain a data set;

[0258] Step S604, performing feature extraction on the multi-source user behavior data in the data set and encoding the extracted feature data to obtain encoded feature data;

[0259] Step S606, processing the encoded feature data through a user portrait model, a risk assessment model and a quota prediction model respectively to obtain a user portrait corresponding to the at least one user, a score corresponding to the at least one user and a quota corresponding to the at least one user;

[0260] Step S608, generating a corresponding quota strategy according to the user portrait corresponding to the at least one user, the score corresponding to the at least one user and the quota corresponding to the at least one user;

[0261] Step S610, obtaining current operation feedback according to the quota strategy and triggering model self-learning in the intelligent modeling layer.

[0262] The present application adopts the above technical scheme, acquires multi-source user behavior data of at least one user, carries out structured cleaning on the multi-source user behavior data to obtain a data set, carries out feature extraction on the multi-source user behavior data in the data set, encodes the extracted feature data to obtain encoded feature data, processes the encoded feature data through a user portrait model, a risk assessment model and a quota prediction model respectively to obtain a user portrait corresponding to at least one user, a score corresponding to at least one user and a quota corresponding to at least one user, generates a corresponding quota strategy according to the user portrait corresponding to at least one user, the score corresponding to at least one user and the quota corresponding to at least one user, and acquires current operation feedback according to the quota strategy, and triggers model self-learning in the intelligent modeling layer according to the current operation feedback. Compared with the prior art, the present application has the following technical effects: dynamic evaluation and intelligent adjustment of user credit quota, and improvement of the automation and intelligent level of credit risk control and quota management.

[0263] The above description is only the preferred embodiment of the present application, and does not limit the implementation and protection scope of the present application. It should be realized by those skilled in the art that any equivalent replacement and obvious change made by applying the content of the present application should be included in the protection scope of the present application.

Claims

1. A flexible credit limit decision calculation system based on user behavior data, characterized in that, include: The data source layer, data acquisition layer, feature processing layer, intelligent modeling layer, decision execution layer, and feedback optimization layer are as follows: The data source layer is used to acquire multi-source user behavior data of at least one user; The data acquisition layer is used to perform structured cleaning on the multi-source user behavior data to obtain a dataset; The feature processing layer is used to extract features from the multi-source user behavior data in the dataset, and encode the extracted feature data to obtain the encoded feature data. The intelligent modeling layer is used to process the encoded feature data through a user profile model, a risk assessment model, and a credit limit prediction model to obtain a user profile, a score, and a credit limit corresponding to the at least one user. The decision execution layer is used to generate a corresponding credit limit strategy based on the user profile of the at least one user, the score of the at least one user, and the credit limit of the at least one user. The feedback optimization layer is used to obtain current operation feedback based on the quota strategy, and to trigger model self-learning in the intelligent modeling layer based on the current operation feedback.

2. The elastic credit limit decision calculation system based on user behavior data according to claim 1, characterized in that, The data acquisition layer includes: a cleaning module, a filling module, a normalization module, and an encoding module, wherein, The cleaning module is used to clean outliers in the multi-source user behavior data to obtain the cleaned multi-source user behavior data. The filling module is used to fill the cleaned multi-source user behavior data to obtain the filled multi-source user behavior data. The normalization module is used to normalize the filled multi-source user behavior data to obtain normalized multi-source user behavior data. The encoding module is used to encode the normalized multi-source user behavior data to obtain the data set.

3. The flexible credit limit decision calculation system based on user behavior data according to claim 2, characterized in that, The filling module is further configured to fill in missing items in the cleaned multi-source user behavior data using the mean, mode, or default value to obtain the filled multi-source user behavior data.

4. The flexible credit limit decision calculation system based on user behavior data according to claim 1 or 2, characterized in that, The feature processing layer includes a feature extraction module and a feature engineering module, wherein... The feature extraction module is used to extract features from the multi-source user behavior data in the dataset based on at least one of basic features, time series features, interaction features, behavioral tags, and social features, to obtain the feature data. The feature engineering module is used to process the feature data through time dimension, cross-domain interaction, graph structure modeling, temporal modeling and dynamic binning to obtain the encoded feature data.

5. The elastic credit limit decision calculation system based on user behavior data according to claim 4, characterized in that, The intelligent modeling layer includes: the user profile model, the risk assessment model, the credit limit prediction model, and a model update module, wherein, The user profile model is used to classify at least one user based on a clustering algorithm, a topic model, or a deep learning model, determine a group of at least one user, and generate the user profile based on the group and the encoded feature data. The risk assessment model is used to determine the score based on the coded feature data and to display the score through a specific model. The credit limit prediction model is used to predict the credit limit based on the encoded feature data through a regression model. The model update module is used to trigger the user profile model, the risk assessment model, and the credit limit prediction model to learn on a specific period and / or online, and to manage the user profile model, the risk assessment model, and the credit limit prediction model through version management and A / B testing mechanisms.

6. The elastic credit limit decision calculation system based on user behavior data according to claim 5, characterized in that, The risk assessment model includes a default probability model, wherein the default probability model is trained using at least one of logistic regression, XGBoost, LightGBM, and DeepFM based on the encoded feature data, and the score is determined by training the convergent default probability model.

7. The elastic credit limit decision calculation system based on user behavior data according to claim 6, characterized in that, The risk assessment model also includes: the specific model, wherein the specific model is a SHAP value interpretation model, used to display the score through visualization tools.

8. The elastic credit limit decision calculation system based on user behavior data according to claim 5, characterized in that, The decision execution layer includes: a rule engine, an approval module, and interface services, wherein... The rule engine is used to filter the user profile, rating and credit limit of the at least one user according to preset conditions, perform weighted calculation on the filtered user profile, rating and credit limit of the at least one user to obtain the calculation result, and allocate the corresponding credit limit strategy according to the calculation result in combination with the application scenario. The approval module is used to review the credit limit strategy through automatic approval and manual review to obtain the approved credit limit strategy. The interface service is used to send the approved credit limit policy to the loan disbursement system, risk control system, and CRM system through a specific interface.

9. The elastic credit limit decision calculation system based on user behavior data according to claim 8, characterized in that, The feedback optimization layer includes a behavior feedback acquisition module and a model optimization module, wherein... The behavior collection module is used to obtain the feedback result of the at least one user under the current operation based on the quota strategy; The model optimization module is used to trigger the self-learning of the user profile model, the risk assessment model, and the credit limit prediction model based on the feedback results.

10. A method for calculating flexible credit limits based on user behavior data, characterized in that, include: Obtain multi-source user behavior data for at least one user; The multi-source user behavior data is structured and cleaned to obtain a dataset; Feature extraction is performed on the multi-source user behavior data in the dataset, and the extracted feature data is encoded to obtain the encoded feature data; The encoded feature data is processed through a user profile model, a risk assessment model, and a credit limit prediction model to obtain a user profile, a score, and a credit limit for the at least one user. Generate a corresponding credit limit strategy based on the user profile, rating, and credit limit of the at least one user; The current operation feedback is obtained based on the quota strategy, and the model self-learning in the intelligent modeling layer is triggered based on the current operation feedback.

Images