Credit marketing intelligent recommendation system

Abstract

The invention discloses a credit marketing intelligent recommendation system. According to the system, for the existing credit marketing problem, a customer portrait is constructed through multi-channel data acquisition, cleaning and preprocessing, and personalized credit product recommendation is carried out based on content and collaborative filtering. And training a recommendation model by using a neural network model and a reinforcement learning algorithm, updating recommendation in real time or at regular intervals according to real-time conditions and feedback of customers, establishing a system evaluation effect including recommendation accuracy, customer satisfaction and business indexes, and optimizing model parameters, improving the algorithm and updating data according to evaluation. The method can improve the customer satisfaction and the business conversion rate, reduces the risk, adapts to the market change, and effectively solves the defects of the existing credit marketing channel and strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of credit marketing, and particularly to an intelligent recommendation system for credit marketing. Background Art

[0002] Credit marketing is an important link in the financial field. In credit marketing, the main goal is to attract potential customers to apply for loan products and ensure that customers choose appropriate credit solutions.

[0003] Currently, the following problems exist in credit marketing:

[0004] In terms of marketing channels:

[0005] 1. Data quality and privacy issues in online channels: Although online marketing channels can quickly obtain a large amount of customer information, the data quality is uneven, with problems such as false information and inaccurate information, affecting the accuracy of marketing effectiveness. In addition, customer privacy protection is also an important issue. Once customer information is leaked, it will not only damage the interests of customers but also affect the reputation of credit institutions. For example, there have been incidents where a large amount of customer information was leaked on some online lending platforms, leading to public doubts about the data security of credit institutions.

[0006] 2. High cost and low efficiency in offline channels: Offline marketing channels such as bank branches, outdoor advertisements, leaflets, etc. require a large amount of manpower, material resources, and financial resources, resulting in high costs. Moreover, the coverage of offline channels is limited, and the marketing efficiency is relatively low, making it difficult to meet the needs of credit institutions to rapidly expand their business. Taking bank branches as an example, it is necessary to lease venues, allocate staff, etc., with high operating costs. At the same time, customers also need to spend a lot of time and energy to handle business at the branches.

[0007] In terms of marketing strategies:

[0008] 1. Insufficient product innovation: The innovation ability of credit products is insufficient. Most products are similar in terms of functions, interest rates, terms, etc., and cannot meet the diverse needs of customers. For example, for some emerging industries or fields, such as technology innovation enterprises and green industries, there are a lack of targeted credit products, making it difficult to support the development of these industries.

[0009] 2. Single marketing means: The marketing means of some credit institutions are relatively single, mainly relying on traditional advertising, telemarketing, etc., and lacking the effective use of emerging channels such as new media and social platforms. Moreover, in the marketing process, there is a lack of in-depth understanding and analysis of customer needs, and personalized marketing solutions cannot be provided, resulting in poor customer experience, unnecessary disturbance to customers, low marketing success rate, and low business conversion rate. Summary of the Invention

[0010] I. Overview of the Technical Solution

[0011] Data Collection and Integration

[0012] Multi-channel Data Acquisition

[0013] Internal data requirements: Collect internal data such as customer basic information (name, gender, age, etc.), account information (deposit, loan balance, etc.), transaction records (consumption, transfer, etc.) and credit ratings, through user login and registration, extraction from the bank's core business system, transaction data analysis and internal credit rating system respectively.

[0014] External data requirements: Obtain customer activity information from social media platforms, collect network behavior data using network analysis tools, and cooperate with third-party credit agencies to obtain external credit reports to enrich the sources of customer data.

[0015] Internal Data Collection

[0016] Database integration: Integrate data from various business systems within the bank to establish a unified data warehouse for easy data extraction, analysis and processing.

[0017] Data interfaces: Establish interfaces with the core business system, CRM system, etc. to achieve real-time data transmission and update, ensuring data timeliness and accuracy.

[0018] Data mining: Use data mining techniques such as association rule mining and clustering analysis to extract valuable information from a large amount of internal data to support personalized marketing.

[0019] External Data Collection

[0020] API interfaces: Establish API interfaces with social media, network analysis tools, credit agencies, etc. to automatically collect and update external data and improve collection efficiency.

[0021] Web crawlers: Use web crawler technology to scrape relevant financial information on the Internet for reference in credit market analysis.

[0022] Data cooperation: Cooperate with enterprises or institutions such as e-commerce and telecommunications operators to share data resources, expand data sources and richness.

[0023] Data Cleaning and Preprocessing

[0024] Data cleaning: Remove duplicate data, handle missing values (such as mean filling, deletion, etc.), correct incorrect data (through checksums and logical judgments) to ensure data accuracy and uniqueness.

[0025] Data preprocessing: Standardize data from different sources (unify format and unit), encode (convert categorical data to numerical type) and normalize (map to a specific interval) for easy data analysis.

[0026] Customer Portrait Construction

[0027] Machine Learning-Based Customer Portrait

[0028] Classification Algorithms: Classification algorithms such as decision trees, random forests, and support vector machines are adopted to classify customers according to characteristics such as basic customer information, credit ratings, and consumption behaviors. For example, they can be classified into high-net-worth customers, small and medium-sized enterprise owners, young white-collar workers, etc. Taking the CART decision tree algorithm as an example, its Gini index is used to measure the impurity of the dataset. The optimal feature and splitting point are selected by calculating the Gini index of each feature to construct a classification tree, and the construction is recursively carried out until the stopping conditions are met (the samples in the node are of the same category, the maximum depth is reached, the number of samples is lower than the threshold, or there are no available features).

[0029] Portrait Data Tagging and Application: Tag the portrait data, screen valuable data and customers in combination with the business scenario, and locate the target customers. The customer portrait design follows the principles of mainly based on demographic attributes and credit information, mainly based on strongly relevant information, and mainly based on qualitative data. The user portrait information is divided into five categories: demographic attributes, credit attributes, consumption characteristics, hobbies, and social attributes, which are used to screen target customers and recommend credit products.

[0030] Dynamic Customer Portrait Update

[0031] Real-Time Data Monitoring: Real-time monitor customer behaviors and transaction data, and update the customer portrait in a timely manner. For example, when there is a large consumption, update the consumption behavior characteristics to adjust the recommended products.

[0032] Feedback Mechanism: Establish a customer feedback mechanism, and adjust the customer portrait and recommendation strategies according to the feedback. For example, when customers are not satisfied with the recommended products, collect opinions and analyze the demand preferences.

[0033] Personalized Credit Product Recommendation

[0034] Content-Based Recommendation

[0035] Product Feature Extraction: Extract the features of credit products, such as loan amount, interest rate, term, repayment method, etc., and classify them (for example, the amount is divided into high, medium, and low).

[0036] Customer Demand Matching: Match the customer portrait features with the credit product features to recommend products that meet the customer's needs. For example, recommend housing loans to customers who are buying houses.

[0037] Collaborative Filtering Recommendation

[0038] User-Based Collaborative Filtering: Find other customers who are similar to the target customer in terms of interests and needs, and recommend similar products to the target customer according to their credit product selections, which is achieved by calculating the customer similarity.

[0039] Project-based collaborative filtering: Find other products similar to the credit products that the target customer is interested in and recommend them to the target customer by calculating the product similarity.

[0040] Recommendation model construction and training

[0041] Neural network model

[0042] Determine the problem and data: Clearly define the credit product recommendation as a classification problem, and collect customer portrait features, credit product features, and corresponding label data (customer application status).

[0043] Data preprocessing: Perform data cleaning (handling missing and outlier values), feature scaling (normalizing or standardizing numerical features), and encoding categorical features (such as one-hot encoding or label encoding of occupation information).

[0044] Determine the model structure

[0045] Input layer: The number of nodes is equal to the total number of customer portrait and credit product features.

[0046] Hidden layer: Determine the number of hidden layers and nodes, select the ReLU activation function, and it can be adjusted according to the model performance.

[0047] Output layer: The number of nodes depends on the problem type. In a classification problem, if there are multiple credit products, the number of nodes is equal to the number of products. For binary classification, the sigmoid activation function is used to output the application probability, and for multi-classification, the softmax activation function is used to output the probability distribution.

[0048] Model training

[0049] Select the loss function: For classification problems, use the cross-entropy loss function (binary cross-entropy or categorical cross-entropy), and for regression problems, use the mean squared error, etc.

[0050] Select the optimization algorithm: Such as stochastic gradient descent, Adam, RMSprop, etc., and minimize the loss function by adjusting the model parameters.

[0051] Divide the training set and validation set: Divide them according to a certain ratio (such as 8:2 or 7:3). Use the training set to train the model and the validation set to monitor the performance to prevent overfitting.

[0052] Train the model: Use the training set data to train, calculate the loss function, calculate the gradient by backpropagation, and update the parameters using the optimization algorithm until the training rounds are reached or the stopping condition is met.

[0053] Reinforcement learning: In the interaction between the recommendation system and the customer, continuously learn and optimize the recommendation strategy through the reinforcement learning algorithm, and adjust the strategy according to the customer feedback (click, application, rejection, etc.) to improve the accuracy and satisfaction.

[0054] Implementation of personalized credit product recommendations

[0055] Real-time recommendation: When a customer initiates a credit application, the system recommends the most suitable products based on the real-time profile and product situation.

[0056] Regular update of recommendations: Regularly update the recommendation results according to the changes in customer situations and products to ensure accuracy and timeliness.

[0057] Effect evaluation and optimization

[0058] Establishment of an index system

[0059] Recommendation accuracy indicators: Include hit rate, accuracy rate, recall rate, etc. Calculate the proportions of true positives, false positives, true negatives, and false negatives after sample classification to evaluate the accuracy of the recommendation system. For example, the hit rate is the proportion of applications among the recommended products, the accuracy rate is the proportion of recommended products that meet the requirements, and the recall rate is the proportion of recommended products among the applied products.

[0060] Customer satisfaction indicators: Collect feedback through customer satisfaction surveys, complaint rates, etc. to evaluate satisfaction, and conduct surveys regularly to obtain improvement suggestions.

[0061] Business indicators: Analyze business indicators such as loan application volume, disbursement volume, and business income to evaluate the contribution of the recommendation system to the business, and compare the changes before and after use to evaluate the effect.

[0062] Model optimization and adjustment

[0063] Parameter adjustment: Adjust the parameters of the recommendation model according to the evaluation results, such as the learning rate, number of layers, and number of nodes of the neural network, to optimize performance.

[0064] Algorithm improvement: Try new recommendation algorithms and technologies, such as introducing new deep learning or reinforcement learning algorithms, to improve the accuracy and personalization of recommendations.

[0065] Data update: Regularly update data, such as updating transaction and behavior data monthly, and retraining the model to improve accuracy.

[0066] Beneficial effects

[0067] Improve customer satisfaction: Meet the specific needs of customers through personalized recommendations and enhance customer satisfaction with credit product recommendations.

[0068] Improve business conversion rate: Precise recommendations increase customers' interest and probability of choosing credit products, thereby improving the business conversion rate, enhancing the marketing precision rate, and reducing marketing costs.

[0069] Reduce risks: Recommend suitable products based on customers' risk preferences and repayment abilities to reduce the credit risks of financial institutions.

[0070] Adapting to Market Changes: Leveraging AI technology to promptly capture changes in market and customer demands, adjust recommendation strategies, and adapt to dynamic market changes. Description of the Drawings

[0071] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only for the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0072] Figure 1 Schematic diagram of data collection and integration for the embodiments of the present invention;

[0073] Figure 2 Schematic diagram of data cleaning and preprocessing for the embodiments of the present invention;

[0074] Figure 3 Schematic diagram of customer portraits based on machine learning for the embodiments of the present invention;

[0075] Figure 4 Schematic diagram of personalized credit product recommendations for the embodiments of the present invention;

[0076] Figure 5 Schematic diagram of model training for the embodiments of the present invention. Detailed Embodiment

[0077] The following will describe the present invention in detail in combination with the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; and the drawings are only for more specifically describing the embodiments and are not intended to specifically limit the present invention.

[0078] It should be pointed out that in the specification, when referring to "one embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc., it indicates that the described embodiments may include specific features, structures or characteristics, but not necessarily every embodiment includes such specific features, structures or characteristics. Additionally, when combining embodiments to describe specific features, structures or characteristics, implementing such features, structures or characteristics in combination with other embodiments (whether explicitly described or not) should be within the knowledge scope of those skilled in the relevant art.

[0079] Generally, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or property in the singular sense, or can be used to describe a combination of features, structures, or properties in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but rather, depending at least in part on the context, can allow for the existence of other factors that are not necessarily explicitly described.

[0080] Example 1

[0081] Refer to Figures 1 to 5

[0082] I. Data Collection and Integration

[0083] 1. Multi-channel Data Collection

[0084] 1.1 Internal Data Requirements

[0085] Basic customer information: including name, gender, age, contact information, occupation, income, etc. It can be collected when the user logs in and registers.

[0086] Account information: such as deposit account balance, loan account balance, credit card limit, repayment records, etc. These data can be extracted from the bank's core business system.

[0087] Transaction records: including consumption records, transfer records, payment records, etc. By analyzing the customer's transaction records, the bank can understand the customer's consumption habits and fund flow, providing a reference for credit marketing.

[0088] Credit rating: The bank's internal credit rating system will evaluate the customer's credit status, including credit scores, default records, etc. This information is crucial for determining the customer's credit risk.

[0089] 1.2 External Data Requirements

[0090] Social media data: By collaborating with social media platforms, the bank can obtain information about the customer's activities on social media, such as followed topics, posted content, social relationships, etc. These data can help the bank understand the customer's interests and lifestyle, thus better conducting personalized marketing.

[0091] Web behavior data: Using web analysis tools, the bank can collect the customer's browsing records, search keywords, shopping behavior, etc. on the Internet. These data can reflect the customer's needs and preferences, providing a basis for the recommendation of credit products.

[0092] Third-party credit data: Collaborate with professional credit agencies to obtain external credit reports of customers, including loan records, credit card usage, overdue records, etc. from other financial institutions. These data can supplement the bank's internal credit ratings and improve the accuracy of credit risk assessment.

[0093] 2. Internal data collection

[0094] Database integration: Integrate the data of various business systems within the bank to establish a unified data warehouse. Through the data warehouse, the required data can be conveniently extracted and subjected to data analysis and processing.

[0095] Data interface: Establish data interfaces with the bank's core business system, CRM system, etc. to achieve real-time data transmission and update. This can ensure the timeliness and accuracy of the data.

[0096] Data mining: Utilize data mining techniques to extract valuable information from the bank's large amount of data. For example, through association rule mining, the relationship between customers' consumption behaviors and credit demands can be discovered; through clustering analysis, customers can be divided into different groups to provide support for personalized marketing.

[0097] 3. External data collection

[0098] API interface: Establish API interfaces with social media platforms, web analytics tools, credit agencies, etc. to obtain external data. Through the API interface, automated data collection and update can be achieved, improving the efficiency of data collection.

[0099] Web crawler: Use web crawler technology to scrape relevant data from the Internet. For example, financial information on news websites, financial forums, etc. can be scraped to provide a reference for the analysis of the credit market.

[0100] Data cooperation: Conduct data cooperation with other enterprises or institutions to share data resources. For example, cooperate with e-commerce platforms to obtain customers' shopping records; cooperate with telecommunications operators to obtain customers' communication records, etc. Through data cooperation, the data sources can be expanded and the richness of the data can be improved.

[0101] 4. Data cleaning and preprocessing

[0102] 4.1 Data cleaning

[0103] Remove duplicate data: De-duplicate the collected data to ensure data uniqueness. Technologies such as hash algorithms can be used to compare and de-duplicate the data.

[0104] Handling missing values: For missing values in the data, methods such as filling and deletion can be used for processing. For example, methods such as mean filling and median filling can be used to fill the missing values of numerical data; for missing values that cannot be filled, the corresponding data entry can be considered for deletion.

[0105] Correcting incorrect data: Correct the errors in the data to ensure data accuracy. Methods such as data verification and logical judgment can be used to detect and correct errors in the data. For example, for data with incorrect formats such as ID numbers and mobile phone numbers, format verification and correction can be performed.

[0106] 4.2 Data preprocessing

[0107] Data standardization: Standardize data from different sources to unify data formats and units. For example, unify customers' income data into monthly income or annual income, and unify date data into a specific format, etc.

[0108] Data encoding: Encode categorical data and convert it into numerical data for data analysis and processing. For example, encode categorical data such as customers' genders and occupations into digital forms.

[0109] Data normalization: Normalize numerical data and map it to a specific interval to eliminate the influence of data dimensions. For example, normalize data such as customers' income and assets so that they take values in the interval [0, 1].

[0110] II. Customer portrait construction

[0111] 1. Customer portrait based on machine learning

[0112] Classification algorithms: Use classification algorithms to classify customers, such as decision trees, random forests, support vector machines, etc. According to customers' basic information, credit ratings, consumption behaviors and other characteristics, classify customers into different categories, such as high-net-worth customers, small and medium-sized enterprise owners, young white-collar workers, etc. Extract characteristics such as consumption amount, consumption frequency, and consumption location from customers' transaction records, and extract characteristics such as hobbies and social relationships from social media data

[0113] For example, through the CART decision tree algorithm, according to customers' age, income, occupation and other characteristics, classify customers into different risk levels, and recommend different credit products for customers with different risk levels.

[0114] The Gini index of the CART decision tree is an evaluation index used in the CART (Classification And Regression Tree) algorithm for classification tasks, mainly used to measure the impurity or uncertainty of a data set

[0115] The Gini Index represents the probability that two randomly selected samples from a dataset have different class labels. For a dataset D with K classes, the formula for calculating its Gini Index is as follows:

[0116]

[0117]

[0118] Where pk represents the proportion of class k in dataset D. The value range of the Gini Index is between [0, 1]. The smaller the value, the higher the purity of the dataset, that is, the larger the proportion of samples belonging to the same class.

[0119] When constructing a CART classification tree, the algorithm selects the optimal feature for dataset splitting based on the Gini Index. The specific steps are as follows:

[0120] Calculate the Gini Index: For each feature, the algorithm tries all possible split points and calculates the Gini Index of the left and right subsets after splitting.

[0121] Select the best split: Select the feature and split point that minimize the weighted sum of the Gini Index after partitioning. The weighted sum is calculated based on the subset size (number of samples).

[0122] Recursively construct the tree: Split the dataset with the selected feature and threshold, and then repeat the above process for each subset until the stopping condition is met (such as all samples in the node belong to the same class, reaching the preset maximum depth, the number of samples in the node is lower than a certain threshold, etc.).

[0123] The construction process of the CART decision tree is a recursive process. By continuously selecting the optimal feature and split point to split the dataset until the stopping condition is met. The Gini Index plays a key role in this process, helping the algorithm select the feature and split point that can most effectively reduce the impurity of the dataset. Here is an example.

[0124] Example:

[0125] Loan Application Sample Data Table

[0126]

[0127] First, calculate the Gini Index of each feature, select the optimal feature and its optimal split point. Denote the four features of age, having a job, mortgage, and credit status as A1, A2, A3, and A4 respectively. Denote the values of age as young, middle-aged, and old as 1, 2, 3, the values of having a job and having a mortgage as yes and no as 1, 2, and the values of credit status as excellent, good, and average as 1, 2, 3.

[0128] For subset A1, we calculate the Gini index Gini(A1). Similarly, for subsets A2, A3, and A4, we calculate Gini(A2, A3, A4) and substitute them into the formula

[0129] Gini coefficient for the youth group (5 people, 2 with loans):

[0130] Gini coefficient (D1) = 5 / 2 * (1 - 5 / 2) + 5 / 3 * (1 - 5 / 3) = 0.48

[0131] Gini coefficient for the non - youth group (10 people, 7 with loans):

[0132] Gini coefficient (D2) = 2 * 7 / 10 * (1 - 7 / 10) = 0.42

[0133] Gini index of D under the condition A1 = 1 (youth):

[0134] Gini index (D, A1 = 1) = 5 / 15 * 0.48 + 10 / 15 * 0.42 = 0.44

[0135] The general formula is:[[]]

[0136] Gini index (D, A1 = 1) = 5 / 15 * [2 * 2 / 5 * (1 - 2 / 5)] + 10 / 15 * [2 * 7 / 10 * (1 - 7 / 10)] = 0.44

[0137] Gini index of D under the condition A1 = 2 (middle - aged):

[0138] Gini index (D, A1 = 2) = 5 / 15 * [2 * 3 / 5 * (1 - 3 / 5)] + 10 / 15 * [2 * 6 / 10 * (1 - 6 / 10)] = 0.48

[0139] Gini index of D under the condition A1 = 3 (elderly):

[0140] Gini index (D, A1 = 3) = 5 / 15 * [2 * 4 / 5 * (1 - 4 / 5)] + 10 / 15 * [2 * 5 / 10 * (1 - 5 / 10)] = 0.44

[0141] Step 1: Select the optimal feature

[0142] Since Gini index (D, A1 = 1) and Gini index (D, A1 = 3) are equal and the smallest, both A1 = 1 and A1 = 3 can be selected as the optimal cut - off points of A1

[0143] The same principle applies to the other Gini indices

[0144] Calculate the Gini index of features A2 and A3

[0145] Gini index(D, A2 = 1) = 0.32

[0146] Gini index(D, A3 = 1) = 0.27

[0147] Calculate the Gini index of feature A4

[0148] Gini index(D, A4 = 1) = 0.36

[0149] Gini index(D, A4 = 2) = 0.47

[0150] Gini index(D, A4 = 3) = 0.32

[0151] The Gini index of (D, A4 = 3) is the smallest, so A4 = 3 is the optimal splitting point of A

[0152] II. Recursive construction

[0153] 1. For each sub - dataset, go back to the step of selecting the optimal splitting feature and splitting point, and repeat this process to construct the sub - tree.

[0154] Calculate the initial value of the Gini index of the current sub - dataset.

[0155] Traverse the possible features and splitting points, calculate the Gini index, and select the feature and splitting point that minimize the Gini index as the optimal splitting of the sub - dataset.

[0156] Continue to split the sub - dataset and recurse like this.

[0157] III. Stopping condition judgment

[0158] In each recursive process, it is necessary to check whether the stopping condition is met. Common stopping conditions are:

[0159] 1. The number of samples in the node is less than a certain threshold:

[0160] If the number of samples in the sub - dataset is very small, continuing to split may lead to overfitting because a small number of samples may not represent the overall distribution, and very small sub - nodes may be obtained after splitting, lacking statistical significance.

[0161] For example, set the threshold to 10. If the number of samples in the sub - dataset is less than 10, stop recursively constructing the sub - tree.

[0162] 2. The Gini index of the node is less than a certain threshold:

[0163] When the Gini index is very small, it means that the purity of the data is already very high, and continuing to split may not bring significant performance improvement.

[0164] For example, set the Gini index threshold to 0.1. If the Gini index of the current node is less than 0.1, stop recursively constructing the sub-tree.

[0165] 3. The depth of the tree reaches the preset maximum value:

[0166] Limiting the depth of the tree can prevent the decision tree from being too complex and avoid overfitting.

[0167] For example, set the maximum depth of the tree to 5. When the depth of recursively constructing the sub-tree reaches 5, stop further construction.

[0168] 4. There are no more features available for splitting:

[0169] If all features have been used, or in the current sub-dataset, no feature can effectively perform splitting, then the recursive construction of the sub-tree should also be stopped.

[0170] Label the portrait data, use machine learning algorithms to find similar populations, deeply integrate with the business scenario, screen out valuable data and customers, locate target customers, reach customers, and record and feedback on the marketing effect.

[0171] The design of the user portrait should adhere to three principles, namely: mainly based on demographic attributes and credit information, mainly based on strongly relevant information, and mainly based on qualitative data.

[0172] Mainly based on credit information and demographic attributes. There is a lot of information to describe a user. Credit information is an important information in the user portrait. Credit information is the information describing a person's consumption ability in society. Credit information can directly prove the customer's consumption ability and is the most important and basic information in the user portrait. It includes information such as the consumer's job, income, education level, property, etc.

[0173] Again, financial enterprises need to reach customers. Demographic attribute information plays the role of reaching customers. Demographic attribute information includes: name, gender, phone number, email address, home address, etc. These information can help financial enterprises contact customers and promote products and services to customers.

[0174] Adopt strongly relevant information and ignore weakly relevant information

[0175] Strongly relevant information: is the information directly related to the scenario requirements, which can be causal information or information with a very high degree of correlation.

[0176] For example: Analysis reveals that, on the premise that other conditions are the same, the average salary of people around 35 years old is higher than that of people with an average age of 30 years old. The average salary of students graduating from computer science majors is higher than that of students majoring in philosophy. The average salary of those working in the financial industry is higher than that of those working in the textile industry. The average salary in Shanghai exceeds that in Hainan Province. These pieces of information tell us that a person's age, education level, occupation, and location have a significant impact on income, and there is a strong correlation with income level. It also means that information with a greater impact on credit attributes is strongly correlated information, and vice versa for weakly correlated information.

[0177] Other information of users, such as their height, weight, name, constellation, etc., is difficult to analyze probabilistically for its impact on consumption ability and is weakly correlated information.

[0178] These pieces of information are not included in the user portrait for analysis, as they have little impact on the user's credit consumption ability and no commercial value.

[0179] Convert quantitative information into qualitative information

[0180] Quantitative information is not conducive to screening customers. It is necessary to convert quantitative information into qualitative information to screen people by information categories.

[0181] For example: Convert customers divided by age groups into qualitative information. Those aged 18 - 25 are defined as young people, those aged 25 - 35 are defined as young and middle-aged people, those aged 36 - 45 are defined as middle-aged people, etc.

[0182] Referring to personal income information, people can be defined as high-income groups, middle-income groups, and low-income groups.

[0183] Referring to asset information, customers can also be defined as high, medium, and low levels.

[0184] Classify various quantitative information of financial enterprises into qualitative information categories. Qualification is beneficial for screening users and quickly positioning target customers. Financial enterprises create user portraits in combination with business needs. From a practical perspective, user portrait information can be divided into five categories of information. They are: demographic attributes, credit attributes, consumption characteristics, hobbies, and social attributes.

[0185] They basically cover the strongly correlated information required by business needs, and combining with external scenario data will generate huge commercial value. Let's first understand the roles of the five categories of information in the user portrait and the strongly correlated information involved.

[0186] 2. Dynamic Customer Portrait Update

[0187] Real-time data monitoring: By monitoring customers' behavioral data and transaction data in real time, the customer profile is updated in a timely manner. For example, when a customer makes a large consumption, the system can promptly update the customer's consumption behavior characteristics and adjust the recommended credit products for the customer.

[0188] Feedback mechanism: Establish a customer feedback mechanism and adjust the customer profile according to the customer's feedback information. For example, when a customer is not satisfied with the recommended credit product, the system can collect the customer's feedback, analyze the customer's needs and preferences, and adjust the customer profile and recommendation strategy.

[0189] III. Personalized credit product recommendation

[0190] 1. Content-based recommendation

[0191] Product feature extraction: Extract features of credit products, such as loan amount, interest rate, term, repayment method, etc. For example, divide the loan amount into three levels: high, medium, and low; divide the interest rate into three levels: low, medium, and high; divide the term into short-term, medium-term, and long-term; and divide the repayment method into equal principal and interest, equal principal, interest first and principal later, etc.

[0192] Customer demand matching: According to the features in the customer profile, match them with the features of credit products to recommend credit products that meet the customer's needs. For example, for customers with housing purchase needs, recommend housing loan products; for customers with entrepreneurship needs, recommend entrepreneurship loan products.

[0193] 2. Collaborative filtering recommendation

[0194] User-based collaborative filtering: Find other customers who have similar interests and needs as the target customer, and recommend similar products to the target customer based on the credit product selections of these customers. For example, by calculating the similarity between customers, find other customers with a high similarity to the target customer, and recommend the credit products that these customers have applied for to the target customer.

[0195] Item-based collaborative filtering: Find other products that are similar to the credit products that the target customer is interested in and recommend them to the target customer. For example, by calculating the similarity between credit products, find other products that are similar to the credit products that the target customer has previously paid attention to and recommend them to the target customer.

[0196] IV. Recommendation model construction and training

[0197] 1. Neural network model: Utilize neural network models in deep learning, such as multi-layer perceptrons, convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), etc., to learn customer portraits and credit product features, and predict the interest level and application probability of customers for different credit products. For example, by constructing a multi-layer perceptron model, taking the features in the customer portrait and credit product features as inputs, and outputting the application probability of the customer for each credit product, and recommending credit products with a higher application probability to the customer.

[0198] Steps for constructing a multi-layer perceptron model:

[0199] I. Determine the problem and data

[0200] 1. Clearly define the problem type, such as whether it is a classification problem or a regression problem. In credit product recommendation, it is usually a classification problem, predicting whether a customer will apply for a certain credit product.

[0201] 2. Collect and prepare data, including customer portrait features (such as age, income, occupation, etc.), credit product features (such as interest rate, amount, term, etc.), and corresponding label data (the application situation of customers for credit products).

[0202] II. Data preprocessing

[0203] 1. Data cleaning: Handle missing values, outliers, etc.

[0204] 2. Feature scaling: Normalize or standardize numerical features so that the value ranges of different features are on a similar scale, which helps to improve the stability and convergence speed of model training. Common methods include Min-Max normalization and Z-score standardization.

[0205] 3. Encode categorical features: If there are categorical features in the data, such as occupation, one-hot encoding or label encoding can be used to convert them into numerical representations.

[0206] III. Determine the model structure

[0207] 1. Input layer: The number of nodes in the input layer is equal to the total number of customer portrait features and credit product features.

[0208] 2. Hidden layer:

[0209] Determine the number of hidden layers and the number of nodes in each hidden layer. Generally, you can start with a small number of hidden layers and then gradually adjust according to the performance of the model. Increasing the number of hidden layers and nodes can increase the complexity and expressive power of the model, but it may also lead to overfitting.

[0210] Select an activation function. Common activation functions include ReLU (Rectified Linear Unit), sigmoid, tanh, etc. For the hidden layer, the ReLU activation function is usually a good choice because it is computationally simple and can effectively alleviate the vanishing gradient problem.

[0211] 3. Output layer:

[0212] The number of nodes in the output layer depends on the type of problem. For the classification problem of credit product recommendation, if there are multiple credit products, the number of nodes in the output layer is usually equal to the number of credit products, and the output value of each node represents the application probability of the customer for the corresponding credit product.

[0213] Select the activation function for the output layer. For binary classification problems, the sigmoid activation function can be used, and the output value is between 0 and 1, representing the application probability; for multi-classification problems, the softmax activation function can be used, and the output value is the probability distribution of each class.

[0214] 1. The probability of application; for multi-classification problems, the softmax activation function can be used, and the output value is the probability distribution of each class.

[0215] IV. Model Training

[0216] 1. Select a loss function:

[0217] For classification problems, the cross-entropy loss function can be used, such as binary cross-entropy for binary classification problems and categorical cross-entropy for multi-classification problems.

[0218] For regression problems, loss functions such as mean squared error can be used.

[0219] 2. Select an optimization algorithm: Common optimization algorithms include Stochastic Gradient Descent (SGD), Adam, RMSProp, etc. These optimization algorithms continuously adjust the parameters of the model to minimize the loss function.

[0220] 3. Divide the training set and the validation set: Divide the dataset into a training set and a validation set, usually in a certain ratio (such as 8:2 or 7:3). The training set is used to train the model, and the validation set is used to monitor the performance of the model during training to prevent overfitting.

[0221] 4. Training the model: Use the training set data to train the model. In each training batch, input the data into the model, calculate the loss function, then calculate the gradients through the backpropagation algorithm, and update the model's parameters using the optimization algorithm. Repeat this process until the preset number of training epochs is reached or other stopping conditions are met.

[0222] 2. Reinforcement learning: Through the reinforcement learning algorithm, enable the recommendation system to continuously learn and optimize the recommendation strategy during the interaction with customers. For example, when the recommendation system recommends a credit product to a customer, adjust the recommendation strategy based on the customer's feedback (such as click, application, rejection, etc.) to improve the accuracy of the recommendation and customer satisfaction.

[0223] V. Personalized Credit Product Recommendations

[0224] 1. Real-time Recommendation

[0225] When a customer initiates a credit application, the system recommends the most suitable credit product to the customer in real time based on the customer's profile and the current credit product situation.

[0226] 2. Regularly Update Recommendations

[0227] As the customer situation and credit products change, the system regularly updates the customer's recommendation results to ensure the accuracy and timeliness of the recommendations.

[0228] VI. Effect Evaluation and Optimization

[0229] 1. Establishment of the Index System

[0230] Recommendation accuracy metrics: Such as hit rate, accuracy rate, recall rate, etc. By calculating these metrics, evaluate the accuracy and effectiveness of the recommendation system. For example, the hit rate refers to the proportion of credit products recommended that are applied for by customers, the accuracy rate refers to the proportion of credit products recommended that meet the customer's needs, and the recall rate refers to the proportion of credit products applied for by customers that are recommended.

[0231] Customer satisfaction metrics: Such as customer satisfaction surveys, customer complaint rates, etc. By collecting customer feedback information, evaluate the customer satisfaction of the recommendation system. For example, regularly conduct customer satisfaction surveys to understand customers' satisfaction with the recommendation system and improvement suggestions.

[0232] Business metrics: Such as loan application volume, loan disbursement volume, business income, etc. By analyzing these business metrics, evaluate the contribution of the recommendation system to the business. For example, compare the changes in loan application volume and business income before and after using the recommendation system to evaluate the effect of the recommendation system.

[0233] Index calculation methods:

[0234] The samples are divided into positive and negative classes. After classifying a set of samples, the following four situations will occur:

[0235] True Positive (TP): Actually positive class, predicted as positive class.

[0236] False Positive (FP): Actually negative class, predicted as positive class.

[0237] True Negative (TN): Actually negative class, predicted as negative class.

[0238] False Negative (FN): Actually positive class, predicted as negative class.

[0239] 1. Accuracy:

[0240] Definition: It represents the proportion of the number of correctly classified samples to the total number of samples.

[0241] Calculation formula: Accuracy = (TP + TN) / (TP + TN + FP + FN).

[0242] 2. Precision:

[0243] Definition: Also known as the precision rate, it represents the proportion of truly positive samples among all samples predicted as positive.

[0244] Calculation formula: Precision = TP / (TP + FP).

[0245] 3. Recall:

[0246] Definition: Also known as the recall rate, it represents the proportion of samples correctly predicted as positive among all actually positive samples.

[0247] Calculation formula: Recall = TP / (TP + FN).

[0248] For example, suppose there are 100 samples, among which 30 are positive-class samples and 70 are negative-class samples. After prediction by the classification model, there are 25 true positives, 5 false positives, 65 true negatives, and 10 false negatives.

[0249] Then the accuracy is: (25 + 65) / 100 = 0.9.

[0250] The precision is: 25 / (25 + 5) = 0.833.

[0251] The recall is: 25 / (25 + 10) = 0.714

[0252] VII. Model Optimization and Adjustment

[0253] Parameter Adjustment: According to the results of the effect evaluation, adjust the parameters of the recommendation model to improve the accuracy and effectiveness of the recommendation. For example, adjust parameters such as the learning rate, number of layers, and number of nodes of the neural network model to optimize the performance of the model.

[0254] Algorithm Improvement: Try new recommendation algorithms and technologies to continuously improve the performance of the recommendation system. For example, introduce new deep learning algorithms, reinforcement learning algorithms, etc. to improve the accuracy and personalization of the recommendation.

[0255] Data Update: Regularly update the data to ensure the timeliness and accuracy of the data of the recommendation system. For example, update the transaction data and behavior data of customers monthly and retrain the recommendation model to improve the accuracy of the recommendation.

[0256] This invention covers any alternatives, modifications, equivalent methods, and solutions made to the essence and scope of this invention. For the public to have a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments of this invention. However, those skilled in the art can fully understand this invention even without the description of these details. Additionally, well-known methods, processes, procedures, components, and circuits, etc. are not described in detail to avoid unnecessary confusion to the essence of this invention.

[0257] The above are only the preferred embodiments of this invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of this invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of this invention.

Claims

1. An intelligent credit marketing recommendation system, characterized in that: include: Data collection module, used to collect internal and external data of users, including but not limited to basic information of users, social media data, network behavior data, and third-party credit data; Data cleaning and preprocessing module, used to remove duplicates, process missing values, correct erroneous data, and perform standardization and normalization on the collected data; The customer profile building module builds customer profiles based on the collected data through classification algorithms, including the customer's income level, consumption characteristics, and credit rating; Credit product feature extraction module, used to extract the features of credit products, including loan amount, interest rate, term and repayment method; The personalized recommendation module uses a neural network model to learn user portraits and credit product features to provide users with the most suitable credit product recommendations.

2. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The data collection module establishes a data interface with the bank's core business system and CRM system to achieve real-time transmission and updating of data.

3. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The customer portrait building module uses classification algorithms such as decision trees, random forests and support vector machines to divide customers into different categories such as high net worth customers, small and medium-sized business owners, young white-collar workers, etc. based on their basic information, credit ratings, consumption behaviors, etc.

4. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The credit product feature extraction module divides the interest rates of credit products into three levels: low, medium and high, and the loan terms into three levels: short-term, medium-term and long-term.

5. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The personalized recommendation module constructs a multi-layer perceptron neural network model, conducts multiple training on customer portrait features and credit product features, and predicts the customer's interest in credit products and the probability of application.

6. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The system further includes a feedback mechanism module for adjusting the customer profile and recommendation strategy based on the customer's feedback on the recommended credit products to optimize the recommendation results.

7. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The personalized recommendation module uses a reinforcement learning algorithm to continuously optimize the recommendation strategy based on customer interaction feedback to improve the accuracy of recommendations and customer satisfaction.

8. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The system monitors customers’ behavioral and transaction data in real time, dynamically updates customer profiles, and ensures that recommended credit products match customers’ latest needs.

9. The intelligent credit marketing recommendation system according to claim 1, characterized in that: Based on the customer preference information in the user profile, the system uses collaborative filtering methods to find other customers with similar needs as the target customer, and recommends credit products that these customers have previously selected to the target customer.

10. The intelligent credit marketing recommendation system according to claim 1, characterized in that: The data cleaning and preprocessing module normalizes the numerical data such as the user's income and assets, and maps them into a specific interval, so as to eliminate the dimensional influence of the data and improve the accuracy of the model prediction.

Images