Financial customer product personalized recommendation based on multi-classifier fusion model

Abstract

The invention discloses a customer product personalized recommendation method and device based on a multi-classifier fusion model and computer equipment, and the method comprises the steps: extracting effective features from the original feature data of multiple dimensions in a preset customer range in a set historical time period through a self-encoding algorithm based on deep learning; a model training sample of the target customer group is formed, and the model training efficiency is improved. According to the method, a stack generalization integrated learning training strategy is utilized, and a plurality of machine learning models are fused, so that the prediction accuracy of the recommendation model obtained by training is greatly improved, and then business personnel can accurately judge the product preference of a customer, adopt a personalized marketing strategy to cope with the customer demand and improve the sales success rate.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence, and particularly relates to a financial customer product individual recommendation method, device and system based on a multi-classifier fusion model and a storage medium. BACKGROUND

[0002] In the field of banks, customer product individual recommendation is a core technology for realizing accurate marketing and improving customer satisfaction. However, the existing technology faces the following challenges when processing complex financial customer data: first, a traditional recommendation system relies on a single classification model, and it is difficult to fully mine the nonlinear correlation of customer multi-dimensional features, and the feature expression of a small number of class samples is ignored, which affects the recommendation effect of long-tail customers; second, the traditional method based on rule-based feature extraction cannot capture the dynamic changes of customer behavior over time, and relies on manual customer portrait feature screening, which also has problems such as low efficiency and introduction of subjective bias. In view of the above problems, the application provides an individual recommendation method based on multi-weak classifier fusion, which realizes high-precision financial product recommendation through a multi-model stacking hybrid strategy, and improves the efficiency of bank financial services. SUMMARY

[0003] The application is mainly realized through the following technical solutions.

[0004] The financial customer product individual recommendation based on the multi-weak classifier fusion model comprises the following steps.

[0005] S1: Source data acquisition, including customer portrait data, behavior time series data, product feature data, market environment data and other data in four fields, and storing the data in a database.

[0006] S2: Data cleaning, including missing value filling, data standardization and normalization processing and the like.

[0007] S3: Feature extraction, adopting an autoencoder AutoEncoder to perform low-dimensional coding on data features.

[0008] S4: Model construction, using an xgboost model based on a decision tree and a lightGBM model as basic models of a weak classifier cluster, using a stacking generalization strategy (Stacking) to fuse the models, establishing multiple weak classifiers and fusing the models based on the stacking method, and combining cross-validation and grid search technologies to find the best hybrid model parameters.

[0009] S5: Data set division and model training, dividing the data set into a training set and a test set.

[0010] S6: Model evaluation, using AUC value and recall rate indexes to quantitatively evaluate the prediction results of the model.

[0011] S7: Model application, applying the evaluated and validated data to customer product recommendations.

[0012] In step S3, this application employs a deep learning-based autoencoder for user behavior feature extraction. By learning, the input data is encoded into a low-dimensional representation, reducing the size of the source data space during training. The autoencoder model used in this application comprises two parts: an encoder and a decoder. The encoder maps the input data to a low-dimensional space, and the decoder reconstructs the features. The encoder calculation formula is as follows:

[0013] Where x represents the input data, z represents the low-dimensional space after the encoder, and W e Let b be the encoder weight matrix. e σ is the encoder bias term, and σ is the activation function. This application uses the ReLU function as the activation term.

[0014] The decoder calculation formula is as follows:

[0015] in, x̂ W represents the high-dimensional representation returned by the mapping. d Let b be the weight matrix of the decoder. d σ is the encoder bias term, and σ is the activation function.

[0016] In step S4, this application employs a stacking ensemble learning method to train the combined model. Multi-layered model results are constructed in a queue manner, using the prediction results of preceding models as the training input for subsequent models. The prediction results of multiple machine learning models are then fused to improve overall prediction accuracy and performance. The specific steps are as follows:

[0017] S4.1 Dataset partitioning: The training dataset is divided into multiple subsets, and the subsets are partitioned using cross-validation.

[0018] S4.2 Base Learner Training and Prediction: For each subset, different machine learning models are selected for training and prediction. This patent uses the XGBoost model and the LightGBM model as base learners. Each base learner is trained independently on the subset data and uses different algorithms or parameters to generate prediction results for the data.

[0019] S4.3 Feature Reorganization: The prediction results of each base model are used as new features, supplementing the original features, and combined to form a new training set. For the training set, the output of each model is regarded as a newly added feature column. For the test set, the prediction values ​​obtained from the base models are also used as features.

[0020] The S4.4 meta-model training uses a new model as a secondary learner, whose task is to learn how to optimally combine the predictions of the base model. This meta-model is trained using the features generated by the base model as input and the true value of the target variable as output.

[0021] S4.5 Cross-validation: To avoid overfitting, cross-validation is typically used when constructing these new features. This means that the original training set is divided into multiple subsets, and each base model is trained on a different fold of the cross-validation and generates predictions for other folds. This ensures that the features used by the meta-model during training are not directly generated from the data it will be predicting.

[0022] S4.6 Final Prediction: The trained model is used to predict the data. These predictions are combined by the meta-model to generate the final prediction result.

[0023] In step S4, this application uses the XGBoost model and the LightGBM model as the base learners for Stacking. The core of the XGBoost model is to construct multiple decision trees to form a strong classifier. Training is performed based on the residuals of each decision tree, and the objective function is minimized through iterative iteration. During the iteration process, a gradient boosting-based algorithm is used to solve for the objective function, and the formula for calculating the objective function is as follows:

[0024] in, It is the model's predicted value, g i It is the gradient with respect to the predicted value, h i It is the Hessian matrix of the predicted values, where i represents the i-th sample.

[0025] In the XGBoost model, the objective functions of each leaf node are independent. During training, each leaf node reaches its optimal value, and the overall objective function also reaches its optimal solution. Therefore, the formulas for calculating the weight and objective value of each leaf node are as follows:

[0026] To reduce training computation time and model complexity, a quadratic regularization term is added to the objective function to reduce the number and depth of child nodes in the model. The calculation formula is as follows:

[0027] Where w is the score of the leaf nodes, T is the number of leaf nodes, and γ and λ are regularization parameters.

[0028] In step S4, LightGBM is selected as one of the base learners. The LightGBM model is a decision tree-based Boosting model. This patent uses an additive model and forward distribution algorithm for training. The objective function is optimized by iteratively calculating the negative gradient of each tree. Information gain is used as the evaluation metric for the classification objective function. Therefore, the objective function calculation formula is as follows:

[0029] Where Ω(f) is the regularization term, l(y) i ,f(x i )) represents the error between the predicted value and the actual value.

[0030] This application uses the Histogram algorithm to preprocess the input data. The specific processing steps are as follows: First, the continuous numerical values ​​are discretized into N integers, and a histogram with a width of k is constructed. The data is traversed, and the cumulative values ​​are counted and recorded in the histogram based on the discretized values ​​as indices. After the traversal is completed, the histogram stores all the statistics. Subsequently, based on the statistical values ​​of the histogram, the optimal split point of the tree model is found through traversal.

[0031] This application uses GOSS one-sided gradient sampling when calculating the split of each child node, retaining samples with larger gradients and randomly sampling samples with smaller gradients. To reduce the influence of data distribution, a constant multiplier is introduced into the gradient data when calculating information gain. The specific algorithm steps are as follows: First, calculate the absolute value of the gradient of the data, sort the data according to the absolute value of the gradient, select the top 'a' instances with the largest gradients, and then randomly sample 'b' instances from the remaining data. Then, when calculating information gain, multiply the sampled small gradient data by a constant (1-a) / b, and repeat the iteration until the specified number of iterations is reached or convergence is achieved. The following formula is used to calculate the information gain of the split point:

[0032] in, Iterate through each split point of each feature to find d * j =argmax d V j (d) Then calculate the maximum information gain V. j (d * j ), according to feature j * Splitting point d * j Distribute the data to the left and right child nodes. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0034] Figure 1 A flowchart illustrating an optional example of a customer product prediction method based on a multi-classifier fusion model provided in this application is shown. Detailed Implementation

[0035] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0036] In this embodiment, as shown in the figure, the present invention proposes a personalized recommendation method for financial customer products based on a multi-classifier fusion model. The recommendation method includes the following steps:

[0037] S1: Source data acquisition, including data from four areas: customer profile data, behavioral time series data, product feature data, and market environment data, and storing them in the database.

[0038] S2: Data cleaning, including missing value imputation, data standardization and normalization.

[0039] S3: Feature Extraction, including customer feature construction and dimensionality reduction. Customer features are labeled into two categories. The first category is simple customer potential labels, including high-potential customers, medium-potential customers, and low-potential customers. These three categories are defined as follows: high-potential customers are those with AUM growth greater than 30% in the next X months and two new products added; medium-potential customers are those with AUM growth of 10%-30% or one new product added; low-potential customers are those with AUM growth less than 10%. The second category is composite customer potential labels, calculated using the following formula: Potential Comprehensive Score = (Absolute Growth Contribution * w1 + Relative Growth Contribution * w2 + Behavioral Gain Coefficient * w3) * Initial AUM Adjustment Factor, where the absolute growth contribution formula is the average AUM during the performance period - the average AUM during the observation period; the relative growth contribution formula is (average AUM during the performance period - average AUM during the observation period) / (average AUM during the observation period + smoothing coefficient alpha). Customer feature dimensionality reduction uses an autoencoder to convert data features into low-dimensional codes.

[0040] S4: Model building. The xgboost model and lightGBM model based on decision trees are used as the base models for the weak classifier cluster. The stacking generalization strategy is used to fuse the models. Multiple weak classifiers are built and the models are fused based on the stacking method. Combined with techniques such as cross-validation and grid search, the optimal hybrid model parameters are found.

[0041] S5: Dataset partitioning and model training, dividing the dataset into training and test sets.

[0042] S6: Model evaluation, using AUC and recall metrics to quantitatively evaluate the model's prediction results.

[0043] S7: Model application, applying the evaluated and validated data to customer product recommendations.

[0044] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for personalized customer product recommendation based on a multi-classifier fusion model, characterized in that, The method includes: Obtain feature data of a preset customer group within a set historical time range, wherein the feature data includes original feature data and feature data derived from the original feature data, and each feature data contains features of multiple dimensions. The feature data is cleaned and preprocessed to obtain initial sample feature data; Feature extraction and mapping are performed on the initial sample feature data, and a deep learning-based autoencoder algorithm is used to map multi-dimensional data to a lower dimension to obtain the model training samples of the customer group. By using multiple machine learning models and an ensemble learning strategy of stacked generalization, the model training samples are trained to obtain a product recommendation prediction model for the customer group. The trained model was evaluated for performance and its parameters were tuned, and then validated using a test dataset. The trained model is used to predict the customers and generate a product preference list, which is then sent to the terminal of the designated business personnel.

2. The step of cleaning and preprocessing the feature data to obtain the initial sample feature data includes: According to a preset feature processing method, the feature data is preprocessed to obtain the initial sample feature data of the customer group. The preset feature processing method includes one or more combinations of data standardization method, data type conversion method, non-numerical data encoding conversion method, data distribution conversion method and continuous feature binning method.

3. The method according to claim 1, characterized in that, The feature extraction of the initial sample feature data using an autoencoder algorithm includes: The initial feature data is trained using a deep learning-based autoencoder algorithm to obtain trained encoder and decoder models. The trained encoder is used to encode the multi-dimensional sample features to obtain the dimensionality-reduced feature data. The trained decoder is used to restore the output feature values ​​of the trained machine learning model to the original features, thus obtaining the original data dimensions.

4. The method according to claim 1, characterized in that, The method of training the model using multiple machine learning models and a stacked generalization ensemble learning strategy includes: Based on ensemble learning strategies and stacked generalization methods, a multi-model fusion training framework is built; Using multiple derived machine learning models based on decision trees as base learners, different parameters and dimensions are set to train the models on the training samples; The trained base learners are fused together using a stacking method to train a final model, which is then used to predict customer product preferences.

5. A customer product recommendation device, characterized in that, The device includes: The customer feature data acquisition module acquires feature data of a set customer group within a preset historical time period. The acquired feature data includes original features and features derived from the original features, and the features are features containing multiple dimensions. A data preprocessing module is used to preprocess the feature data to obtain an initial sample of the customer group; The feature extraction module is used to train the encoder model on the features of the initial sample and to perform dimension mapping on the features of the initial sample to obtain the model training samples of the customer group. The model fusion training module is used to train multiple machine learning fusion models on the training samples and output the predicted values ​​of the customer group.

6. A computer device, characterized in that, The computer device includes: Communication equipment; A memory for storing the program of the customer product forecasting method as described in any one of claims 1 to 4; A processor for loading and executing the program stored in the memory to implement the steps of the customer product prediction method as described in any one of claims 1 to 4.

Images