Financial product intelligent recommendation method and device
By using XGBOOST and random forest classification models to predict the user's financial product preference types and weighted summary, solve the problem of low bank financial product recommendation and high customer loss rate, and realize more efficient customer service and financial product recommendation recommendation Essence
Patent Information
- Application Number
- CN202310926225.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-25
- Publication Date
- 2025-01-28
AI Technical Summary
The existing banks recommend financial products with low recommendation success rate and high customer churn rate.
By obtaining the data of multiple users, setting labels to indicate the user's financial product preference type, use XGBOOST and random forest classification models to predict the user preference type, and weigh the prediction results to determine the degree of preferences of various types of financial products to various types of financial products. , Make intelligent recommendation.
It has improved the success rate of financial products, reduced labor costs, improved the work efficiency of customer managers, enhanced customers' interest in bank financial products, and reduced customer loss rates.
Smart Images
Figure CN119377794A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer technology, and specifically relates to a method and device for intelligently recommending financial products. Background Art
[0002] With the accelerated development of Internet technology and the implementation of digital transformation strategies, our society is also accelerating into the digital age, and online consumption and online financial management have gradually become the mainstream of society. While people hope to increase the value of their property, they also find it difficult to choose from the mixed wealth management products in society. The recommendation of financial products by banks generally adopts manual recommendation and simple waterfall machine recommendation, without exploring the inherent relationship between data. As a result, when recommending financial products to customers, account managers cannot systematically grasp the relationship between customer information and product purchase information. In order to improve the success rate, account managers often set up personal small accounts to record some characteristics and important information of customers, collect some product information through major websites, and record some information about financial products they care about. This way of working is relatively inefficient. Account managers only have the energy to provide services to familiar customers, and give up providing services to other customers, which causes customers to lose interest in the bank's financial products, and ultimately causes the phenomenon of continuous customer loss.
[0003] In general, the existing way banks recommend financial products has problems such as low recommendation success rate and high customer churn rate. Summary of the invention
[0004] The purpose of the present invention is to provide a method and device for intelligently recommending financial products, so as to solve the problems of low recommendation success rate and high customer churn rate in the existing way of recommending financial products by banks.
[0005] In order to solve the above technical problems, the present invention provides a method for intelligently recommending financial products, comprising the following steps:
[0006] 1) Acquire data of multiple users, including the types of long-term and / or short-term financial products purchased by the users and the information of the users, and set labels for the data of each user, the labels are used to indicate the user's preference type for financial products, and serve as training data;
[0007] 2) Preprocessing the training data, and using the preprocessed training data to respectively train two classification models constructed for classifying the user's financial product preference types;
[0008] 3) Use the two trained classification models to predict the financial product preference types of the users to be predicted, and obtain the preference probabilities of the users for each type of financial product predicted by the two classification models. Perform a weighted summation on the two prediction results to obtain the user's preference degree for each type of financial product, and recommend financial products to the user based on the user's preference degree.
[0009] The beneficial effects are as follows: two classification models are trained according to existing user data to obtain a classification model that can accurately predict the user's preference probability for each financial product, and then the preference probabilities predicted by the two classification models are weighted and summed to obtain the user's preference degree for each financial product; the present invention adopts the method of weighted summation of the preference probabilities predicted by the two classification models to ensure the credibility and robustness of the prediction results, so that account managers can better serve customers, improve service quality, reduce labor costs, thereby improving the work efficiency of account managers and increasing the success rate of financial product recommendations.
[0010] Furthermore, the preprocessing in step 2) includes: removing useless fields of the training data and removing redundant and repeated rows.
[0011] The beneficial effects are: deleting data that is irrelevant to predicting user preferences for various types of financial products, reducing the amount of computational work, and improving the training efficiency of model classification types.
[0012] Furthermore, the preprocessing also includes: processing outliers in the training data, specifically: if the missing data ratio is greater than a set threshold, it is considered a useless feature and the entire column is deleted; if the missing data ratio is less than or equal to the set threshold, it is processed according to the field information, specifically filling the missing data with numeric 0 for the amount field information, and using Newton interpolation to fill the missing data for other field information.
[0013] The beneficial effect is: some data with too many missing data is deleted, and data with less missing data is filled, so as to obtain better training effect and make the prediction results of the trained classification model more accurate.
[0014] Furthermore, the preprocessing also includes: first mapping the discrete data in the training data, converting it into continuous data and then normalizing it; directly normalizing the continuous data; converting the string data in the training data into discrete variables.
[0015] The beneficial effects are: normalizing data with a large number of words, reducing the memory occupied by these data, and effectively improving the efficiency of training.
[0016] Furthermore, the two classification models constructed are the XGBoost classification model and the random forest classification model.
[0017] The beneficial effects are: both the XGBoost classification model and the random forest classification model can perform fast and effective classification based on the existing data and predict the required classification results. They are simple and easy to use and can ensure the robustness of the prediction results.
[0018] Furthermore, when performing weighted summation in step 3), the weight proportion of the XGBoost classification model is higher than the weight proportion of the random forest classification model.
[0019] The beneficial effects are: it can further ensure the credibility and robustness of the prediction results, thereby obtaining the user's preference for various types of financial products and improving the success rate of financial product recommendations.
[0020] Furthermore, the types of financial products in step 1) include: conservative, stable, aggressive and risky.
[0021] Furthermore, the useless fields to be eliminated include any one of the following fields: customer number, institution code and mobile phone number.
[0022] To solve the above technical problems, the present invention also provides a device for intelligent recommendation of financial products, including a memory and a processor, wherein the processor is used to execute computer program instructions stored in the memory to implement the above-mentioned method for intelligent recommendation of financial products.
[0023] The beneficial effect is that the device ensures the effective and reliable execution of a method for intelligent recommendation of financial products. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0025] The basic concept of the present invention is to find the association information between the isolated data of existing banks through two classification models, to mine the implicit relationship in the association information between various data, and to use the weighted sum of the prediction probabilities of the two classification models to ensure the credibility and robustness of the prediction results, thereby improving the work efficiency of the account manager and the success rate of financial product recommendations. Based on this concept, a method and device for intelligent recommendation of financial products of the present invention can be realized.
[0026] The present invention will be described in detail below with reference to the accompanying drawings and method embodiments.
[0027] Method Example:
[0028] A method for intelligently recommending financial products according to the present invention is shown in the flowchart as follows: Figure 1 As shown, the specific implementation steps include:
[0029] Step 1: Obtain static and dynamic historical data authorized by the customer, classify the existing financial products, and set data labels based on the classification results. The labels are used to indicate the preference types of each user for financial products. The preference types include conservative (for example, preference for bank regular deposits), stable, aggressive, and risky (for example, preference for stocks); integrate the obtained static data and dynamic data as a complete data set; among them, static historical data includes the types of long-term financial products purchased by users and user information, and dynamic historical data includes the types of short-term financial products purchased by users and user information.
[0030] Step 2: Perform a series of preprocessing such as data cleaning and data normalization on the data set, and divide the data set into training sets and test sets, which specifically includes the following steps:
[0031] 1) Manually screen fields based on expert experience, remove useless fields, such as customer number, institution code, mobile phone number, etc., and remove redundant and repeated rows;
[0032] 2) Process the outliers in the data set. If the missing data ratio is greater than 70%, it is considered as a useless feature and the entire column is deleted. For abnormal field information that does not meet the deletion conditions, it is processed according to the specific meaning of the field. If the field information is an amount type, it is filled with a numeric value of 0. Other fields are filled using Newton interpolation.
[0033] 3) The data in the data set processed in step 2) are then normalized, which is specifically divided into discrete normalization and continuous normalization; the discrete data are first mapped and converted into continuous data, and then normalized; the continuous data are normalized to 0-1;
[0034] 4) One-hot encode the string data in the data set to convert it into discrete variables;
[0035] 5) Since the amount of data after preliminary processing is still large and there are many fields, PCA dimensionality reduction is used to extract the main features. The definition of PCA is: XX T W = λW, where X represents the matrix of input data, X T represents the transpose of X, W represents the matrix representation of parameter data, and λ represents the weight;
[0036] 6) Divide 80% of the dataset (i.e. training data) into a training set and the rest as a test set.
[0037] Step three, input 80% of the training set into the initially constructed XGBoost classification model for training, and obtain the initially trained xgboost_model classification model, which specifically includes the following steps: first call XGBClassifier, and initialize the XGBoost classification model xgb and related parameters; then use the fit function to train xgb, enter the epoch loop, and use f1_score to calculate the score of each type of financial product in the user; at the end of the loop, find the classification model with the highest score and save it to xgboost_model.
[0038] Step 4: Input 20% of the training data into the initially constructed Random Forest classification model for training to obtain a preliminarily trained random_forest classification model, which specifically includes the following steps: first call RandomForestClassifier, and initialize the RandomForest classification model forest and related parameters; then use the fit function to train forest, enter the epoch loop, and use f1_score to calculate the score value of each type of financial product; at the end of the loop, find the model with the highest score and save it to random_forest.
[0039] Step 5: Input the test set into the trained xgboost_model classification model and random_forest classification model respectively, and obtain the customer's preference probability and F1 value for various financial products respectively; then optimize the xgboost_model classification model and random_forest classification model respectively to improve the F1 value. The specific optimization includes the following contents:
[0040] 1) Adjust the learning_rate, gamma, max_depth, subsample and random_state of the xgboost_model classification model, where learning_rate is the learning rate, and [0.01, 0.025, 0.05, 0.1] are taken for training respectively; gamma is the minimum loss reduction required for specifying leaf node branches, and the larger the value, the more conservative it is, and [0, 0.1, 0.3] are taken respectively; max_depth is the maximum depth of the specified tree, and a reasonable setting can prevent overfitting, and [3, 5, 8, 12] are taken here; subsample is used to train weak learners, and [0.6, 0.7, 0.75, 0.8] are taken here; random_state is the random number seed, and [13, 27, 35, 42] are taken here.
[0041] 2) First, perform a grid search on the n_estimators of the random_forest classification model to obtain the optimal number of iterations of the weak learner, then perform a grid search on the maximum depth of the decision tree max_depth and the minimum number of samples min_samples_split required for internal node re-division, and finally adjust the maximum number of features max_features.
[0042] Step 6: Record the model structures of the optimized xgboost_model classification model and the random_forest classification model at the maximum F1 value, and obtain the predicted preference probabilities under the F1 value, as follows:
[0043] TP: True positive, predicted as 1, actual as 1; FP: False positive, predicted as 1, actual as 0;
[0044] FN: false negative, predicted as 0, actual as 1; TN: true negative, predicted as 0, actual as 0;
[0045]
[0046] Among them, P is the accuracy;
[0047]
[0048] Among them, R is the recall rate;
[0049]
[0050] Then, the predicted preference probabilities of the two classification models are weighted averaged, referring to the ratio of the F1 value and the training data. The specific formula is:
[0051]
[0052] Among them, P d Predict probabilities for preferences; is the maximum F1 value of the XGBoost_model model; is the maximum F1 value of the Random_Forest model; XGB pre is the predicted probability value of the XGBoost_model model; RF pre is the predicted probability value of the Random_Forest model.
[0053] Step 7: Input the data information of the user to be predicted into the two trained classification models respectively, obtain the preference probability of the user for each type of financial product predicted by the two classification models respectively, and then perform weighted summation on the two prediction results to obtain the user's preference for each type of financial product. According to the obtained preference of each user for each type of financial product, the user's preference for each type of financial product is imported into the recommendation device of the business department. The account manager recommends financial products with high preference to the user, conducts actual business recommendation tests, and periodically feeds back the recommendation results to update the user's data set, such as monthly information; perform step 2 on the newly updated data set for retraining to continuously update the classification model, thereby improving the recommendation success rate of financial products. Among them, the actual business recommendation test can be: the account manager pushes financial products to the customer according to the customer's preference probability for the product predicted by the model through outbound calls, text messages or mobile banking envelopes, and timely feeds back the customer's purchase status of the product on a monthly basis.
[0054] In this embodiment, the XGBoost classification model and the random forest classification model are used to classify users. In addition, the decision tree, GBDT, Adaboost or LightGBM classification model can also be used for classification. In this embodiment, when the predicted preference probabilities of the two classification models are weighted averaged, the XGBoost classification model accounts for 0.8, and the random forest classification model accounts for 0.2, which can also be changed according to actual conditions.
[0055] The present invention is based on the intelligent recommendation of bank financial products based on XGBoost and random forest algorithms. Financial products are recommended to customers through XGBoost and random forest algorithms, so that account managers can better serve customers, improve recommendation quality, reduce labor costs, and improve recommendation efficiency and recommendation success rate.
[0056] Device Example:
[0057] A device for intelligently recommending financial products of the present invention includes a memory and a processor, wherein the processor is used to execute computer program instructions stored in the memory to implement a method for intelligently recommending financial products described in the method embodiment of the present invention. The processor may be a processing device such as a programmable logic device FPGA. The memory may be a variety of memories that use electrical energy to store information, such as RAM, ROM, etc., or other types of memories.
Claims
1. A method for intelligently recommending financial products, characterized in that: The steps include: 1) Acquire data of multiple users, including the types of long-term and / or short-term financial products purchased by the users and the information of the users, and set labels for the data of each user, the labels are used to indicate the user's preference type for financial products, and serve as training data; 2) Preprocessing the training data, and using the preprocessed training data to respectively train two classification models constructed for classifying the user's financial product preference types; 3) Use the two trained classification models to predict the financial product preference types of the users to be predicted, and obtain the preference probabilities of the users for each type of financial product predicted by the two classification models. Perform a weighted summation on the two prediction results to obtain the user's preference degree for each type of financial product, and recommend financial products to the user based on the user's preference degree.
2. The method for intelligent recommendation of financial products according to claim 1, characterized in that: The preprocessing in step 2) includes: removing useless fields of the training data and removing redundant and repeated rows.
3. The method for intelligent recommendation of financial products according to claim 2, characterized in that: Preprocessing also includes: processing outliers in the training data, specifically: if the missing data ratio is greater than the set threshold, it is considered a useless feature and the entire column is deleted; if the missing data ratio is less than or equal to the set threshold, it is processed according to the field information, specifically filling the missing data with numeric 0 for the amount field information, and using Newton interpolation to fill the missing data for other field information.
4. The method for intelligent recommendation of financial products according to claim 2 or 3, characterized in that: Preprocessing also includes: mapping the discrete data in the training data first, converting it into continuous data and then normalizing it; directly normalizing the continuous data; and converting the string data in the training data into discrete variables.
5. The method for intelligent recommendation of financial products according to claim 1, characterized in that: The two classification models constructed are the XGBoost classification model and the random forest classification model.
6. The method for intelligent recommendation of financial products according to claim 5, characterized in that: In step 3), when performing weighted summation, the weight ratio of the XGBoost classification model is higher than that of the random forest classification model.
7. The method for intelligent recommendation of financial products according to claim 1, characterized in that: The types of financial products in step 1) include: conservative, stable, aggressive and risky.
8. The method for intelligent recommendation of financial products according to claim 2, characterized in that: Eliminate useless fields including any of the following fields: customer number, institution code, and mobile phone number.
9. A device for intelligently recommending financial products, characterized in that: It comprises a memory and a processor, wherein the processor is used to execute computer program instructions stored in the memory to implement the method for intelligent recommendation of financial products as described in any one of claims 1 to 8.