Modeling method and device of customer conversion prediction model, customer marketing method and device, equipment and computer program product

Abstract

This application discloses a modeling method and apparatus for customer conversion prediction, a customer marketing method and apparatus, equipment, and computer program product. The method includes: constructing dynamic panel data of sample customers; performing feature engineering processing on the dynamic panel data of sample customers using a preset feature engineering strategy to obtain multi-dimensional modeling feature data of sample customers; and performing adaptive modeling using a preset classification model based on the multi-dimensional modeling feature data of sample customers to obtain a customer conversion prediction model. This application effectively captures the impact of changes in customer behavior over time on marketing conversion by constructing a dynamic panel data structure and utilizing its "intra-group correlation and inter-group independence" characteristics. Feature engineering analysis and mining based on panel data ensure the effectiveness and robustness of the mined features. Through refined adaptation of different models and data structures, the model's accuracy in identifying high-potential conversion customers is improved, providing reliable support for precision marketing.

Description

Technical Field

[0001] This application relates to the field of financial product marketing technology, and in particular to a modeling method and apparatus for a customer conversion prediction model, a customer marketing method and apparatus, equipment and computer program products. Background Technology

[0002] Precision marketing is based on the accurate positioning of potential target customer groups and relies on modern big data technology to establish personalized marketing plans, achieving a "one-size-fits-all" marketing approach. This helps enterprises achieve efficient and low-cost business development and is a core method for the digital transformation of modern enterprises.

[0003] Enterprises' digital and intelligent precision marketing requires more precise and efficient marketing communication, marketing methods that focus more on generating business results, and an increasing emphasis on the revenue generated from sales with limited marketing costs. Especially in a market environment where all industries are highly competitive, it is even more necessary to leverage big data technology to achieve precision marketing in order to improve the competitiveness of enterprises in the market.

[0004] Current marketing methods fail to accurately target consumers with genuine product needs and cannot effectively uncover customers' true requirements. Consequently, businesses cannot accurately provide consumers with the products they need most, resulting in poor marketing effectiveness. Consequently, the marketing costs of businesses do not generate maximum value, and therefore do not bring stronger competitiveness to the businesses. Summary of the Invention

[0005] This application provides a modeling method and apparatus for a customer conversion prediction model, a customer marketing method and apparatus, equipment, and computer program product to improve the modeling accuracy of the customer conversion prediction model, thereby improving the precision of customer marketing.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] In a first aspect, embodiments of this application provide a modeling method for a customer conversion prediction model, the modeling method comprising:

[0008] Construct dynamic panel data for sample customers;

[0009] The dynamic panel data of the sample customers are processed using a preset feature engineering strategy to obtain multi-dimensional modeling feature data of the sample customers.

[0010] Based on the multi-dimensional modeling feature data of the sample customers, an adaptive modeling method is used to obtain a customer conversion prediction model.

[0011] Optionally, the preset feature engineering strategy includes a feature effectiveness analysis strategy and a feature robustness analysis strategy. The step of using the preset feature engineering strategy to perform feature engineering processing on the dynamic panel data of the sample customers to obtain the multi-dimensional modeling feature data of the sample customers includes:

[0012] The feature validity analysis strategy is used to perform feature validity analysis on the dynamic panel data of the sample customers to obtain the valid feature data of the sample customers.

[0013] The feature robustness analysis strategy is used to perform feature robustness analysis on the effective feature data of the sample customers to obtain effective and robust feature data of the sample customers, which can be used as multi-dimensional modeling feature data of the sample customers.

[0014] Optionally, the feature validity analysis strategy includes a posterior probability analysis strategy based on motivational features. The step of using the feature validity analysis strategy to perform feature validity analysis on the dynamic panel data of the sample customers to obtain valid feature data of the sample customers includes:

[0015] Based on the aforementioned posterior probability analysis strategy based on motivational features, motivational features from a business perspective and multi-dimensional dynamic and static features from a data perspective are constructed.

[0016] Based on the structure of the dynamic panel data of the sample customers, the business-related driving factors and the data-related multi-dimensional dynamic and static features are used to construct the multi-dimensional driving factor feature matrix of the sample customers.

[0017] Posterior probability analysis was performed on the multidimensional motivational feature matrix of the sample customers to obtain the posterior probability analysis results of the multidimensional motivational feature matrix.

[0018] The effective feature data of the sample customers are determined based on the posterior probability analysis results of the multi-dimensional motivational feature matrix.

[0019] Optionally, the effective feature data of the sample customers includes effective feature behavior data at each time point and response variable data at the corresponding next time point. The feature robustness analysis strategy includes Rank IC analysis. The effective and robust feature data of the sample customers obtained by performing feature robustness analysis on the effective feature data of the sample customers using the feature robustness analysis strategy includes:

[0020] The Rank IC analysis method is used to calculate the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point;

[0021] The effective and robust feature data of the sample customers are determined based on the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point.

[0022] Optionally, the preset classification model includes multiple models, and the step of adaptively modeling using the preset classification model based on the multi-dimensional modeling feature data of the sample customers to obtain the customer conversion prediction model includes:

[0023] The multi-dimensional modeling feature data of the sample customers are adapted according to the type of each preset classification model to obtain the multi-dimensional modeling feature data of each adapted sample customer.

[0024] Based on the multi-dimensional modeling feature data of each adapted sample customer, adaptive modeling is performed using the corresponding preset classification model to obtain multiple customer conversion prediction models.

[0025] The customer conversion prediction models are evaluated to obtain the evaluation results of the customer conversion prediction models.

[0026] The final customer conversion prediction model is determined based on the evaluation results of multiple customer conversion prediction models.

[0027] Optionally, the preset classification model includes ordinary least squares, convolutional neural network, and long short-term memory network models. The step of adaptively modeling using the preset classification model based on the multi-dimensional modeling feature data of the sample customers to obtain the customer conversion prediction model includes:

[0028] Based on the ordinary least squares method, the multi-dimensional modeling feature data of the sample customers are tested, and modeling is performed based on the test results to obtain the first customer conversion prediction model.

[0029] Based on the convolutional neural network model, the multi-dimensional modeling feature data of the sample customers are converted into two-dimensional matrix data and modeled to obtain the second customer conversion prediction model.

[0030] Based on the Long Short-Term Memory network model, the multi-dimensional modeling feature data of the sample customers are serialized and modeled to obtain a third customer conversion prediction model.

[0031] Secondly, embodiments of this application also provide a customer marketing method, the customer marketing method comprising:

[0032] Obtain multi-dimensional characteristic data of customers;

[0033] Based on the customer's multi-dimensional characteristic data, a customer conversion prediction model is used to make predictions and obtain the customer's conversion prediction results.

[0034] Target marketing customer information is determined based on the customer conversion prediction results.

[0035] The customer conversion prediction model is constructed based on the modeling method of the customer conversion prediction model described in any of the preceding items.

[0036] Thirdly, embodiments of this application also provide a modeling apparatus for a customer conversion prediction model, the modeling apparatus for the customer conversion prediction model comprising:

[0037] Building units are used to construct dynamic panel data for sample customers;

[0038] The feature engineering processing unit is used to perform feature engineering processing on the dynamic panel data of the sample customer using a preset feature engineering strategy to obtain multi-dimensional modeling feature data of the sample customer.

[0039] The modeling unit is used to perform adaptive modeling based on the multi-dimensional modeling feature data of the sample customers and a preset classification model to obtain a customer conversion prediction model.

[0040] Fourthly, embodiments of this application also provide a customer marketing device, the customer marketing device comprising:

[0041] The acquisition unit is used to acquire multi-dimensional feature data of customers;

[0042] The prediction unit is used to make predictions based on the customer's multi-dimensional feature data using a customer conversion prediction model, and obtain the customer's conversion prediction results.

[0043] The marketing unit is used to determine target marketing customer information based on the conversion prediction results of the customer.

[0044] The customer conversion prediction model is constructed based on the aforementioned customer conversion prediction model modeling device.

[0045] Fifthly, embodiments of this application also provide an apparatus, comprising:

[0046] A processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform a modeling method for any of the aforementioned customer conversion prediction models.

[0047] Sixthly, embodiments of this application also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the modeling method of any of the aforementioned customer conversion prediction models.

[0048] The above-mentioned technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: The modeling method of the customer conversion prediction model in the embodiments of this application first constructs dynamic panel data of sample customers; then, it uses a preset feature engineering strategy to perform feature engineering processing on the dynamic panel data of sample customers to obtain multi-dimensional modeling feature data of sample customers; finally, based on the multi-dimensional modeling feature data of sample customers, it uses a preset classification model to perform adaptive modeling to obtain the customer conversion prediction model. By constructing a dynamic panel data structure, the embodiments of this application fully utilize the "intra-group correlation and inter-group independence" characteristics of panel data to effectively capture the impact of changes in customer behavior over time on marketing conversion. Based on the panel data structure, feature engineering analysis and feature mining are performed to deduce causes from effects, ensuring the effectiveness and robustness of the mined features. In addition, through the refined adaptation of different model algorithms and data structures, the model's identification accuracy for high-potential conversion customers is significantly improved, providing reliable technical support for precision marketing. Attached Figure Description

[0049] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0050] Figure 1 This is a flowchart illustrating a modeling method for a customer conversion prediction model in an embodiment of this application.

[0051] Figure 2 This is a flowchart illustrating a customer marketing method according to an embodiment of this application;

[0052] Figure 3 This is a schematic diagram of the structure of a modeling device for a customer conversion prediction model according to an embodiment of this application;

[0053] Figure 4 This is a schematic diagram of the structure of a customer marketing device according to an embodiment of this application;

[0054] Figure 5 This is a schematic diagram of the structure of a device according to an embodiment of this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0057] Existing precision marketing methods that utilize big data modeling suffer from numerous technical shortcomings in terms of data structure, feature engineering, and model algorithms. Specifically:

[0058] (1) In terms of data structure, existing technologies usually adopt cross-sectional data structure, which does not fully consider the impact of the dynamic characteristics of the observed object on marketing conversion over time. Research, analysis and modeling cannot dynamically capture the impact of time factors on individual changes.

[0059] (2) In terms of feature engineering, existing technologies usually directly adopt existing statistical methods, such as correlation and IV value, to analyze the effectiveness of features. However, these methods have high requirements for data quality. More importantly, these methods do not fully combine the characteristics of the product to mine effective features, which leads to poor application effect of precision marketing.

[0060] (3) In terms of model algorithms, existing technologies usually directly apply machine learning models without fully considering the compatibility of each model algorithm with data structure, which leads to the trained model not being able to give full play to the advantages of the algorithm.

[0061] Based on this, embodiments of this application provide a modeling method for a customer conversion prediction model, such as... Figure 1 The diagram illustrates a flowchart of a customer conversion prediction model modeling method according to an embodiment of this application. The modeling method includes the following steps S110 to S130:

[0062] Step S110: Construct dynamic panel data for sample customers.

[0063] Panel data is data collected from different observed objects at different time periods or points in time, describing how multiple observed objects change over time. Panel data is widely used in fields such as economics, finance, and sociology. For example, in economics, panel data can be used to analyze changes in economic indicators in different regions at different points in time, helping researchers better understand the dynamic changes in economic phenomena.

[0064] Panel data in this application refers to sample data composed of multiple cross-sections taken from a time series, with sample observations selected simultaneously at these cross-sections. Alternatively, panel data is an m*n data matrix recording a specific data indicator for m objects at n time points. Panel data is characterized by "inter-group independence and intra-group correlation," allowing for better study of the dynamic behavior of each sample and its interaction with time factors.

[0065] Based on the definition of panel data, we first acquire full data for all sample customers (including converted and non-converted customers) within the observation period, covering user behavior data in mobile banking (such as browsing, clicking, and purchasing), merchant consumption behavior data, and asset and liability data. For converted customers, we align their timelines to a unified reference point, using their purchase time as the baseline (denoted as T). For example, for each converted customer, we extract dynamic data for a period prior to conversion (such as T-1, T-2, T-3, etc.) to form a time series. For non-converted customers, we use their last observation time as the baseline and extract time windows of the same length.

[0066] The dynamic characteristics (such as behavioral data and asset and liability data) of each sample customer are arranged according to time points to form an m×n two-dimensional matrix (m is the number of samples and n is the number of time points). Each feature corresponds to a time series, realizing the characteristics of "intra-group correlation" (time series of the same customer are associated) and "inter-group independence" (data of different customers are independent).

[0067] Record the dynamic behavior (such as mobile banking operations and transaction frequency) and asset and liability status (such as deposit balance and loan balance) of each sample i at time point t, as dynamic feature data (X). it Label whether sample i has been converted at time point t (binary label, 1 for conversion, 0 for no conversion), and use this as the response variable data (y). it ). X it With y it By combining time series data, dynamic panel data is formed, allowing the characteristics of each sample to change dynamically over time, providing time-dimensional information for subsequent analysis.

[0068] Step S120: Use a preset feature engineering strategy to perform feature engineering processing on the dynamic panel data of the sample customer to obtain multi-dimensional modeling feature data of the sample customer.

[0069] In the feature engineering stage, the main focus is on studying and analyzing all dynamic panel data of the sample customers. Predefined feature engineering strategies, such as causal feature analysis and RankIC analysis, are used to analyze the effectiveness and robustness of the features, eliminate redundant or noisy features, and remove features that are sensitive to outliers or have strong distribution dependence. Finally, features that meet the requirements of effectiveness and robustness are selected to participate in modeling, so as to improve the modeling accuracy and reliability.

[0070] Step S130: Based on the multi-dimensional modeling feature data of the sample customers, adaptive modeling is performed using a preset classification model to obtain a customer conversion prediction model.

[0071] Based on the multi-dimensional modeling feature data extracted from the dynamic panel data in the aforementioned steps, a binary classification model is established to predict the conversion probability of customers, thereby identifying the target customer group for precision marketing. Regarding the selection of model algorithms, this application fully considers the characteristics of panel data structures and various model algorithms, such as OLS (Ordinary Least Squares), CNN (Convolutional Neural Network), and LSTM (Long Short-Term Memory) models. The panel data structure is finely adjusted to adapt to the characteristics of different algorithms, ultimately constructing a customer conversion prediction model that meets the requirements of accuracy and reliability, ensuring the accuracy of identifying high-probability conversion customers.

[0072] This application's embodiments, by constructing a dynamic panel data structure, fully leverage the "intra-group correlation, inter-group independence" characteristics of panel data to effectively capture the impact of customer behavior changes over time on marketing conversion. Feature engineering analysis and feature mining are performed based on the panel data structure, deriving causes from effects to ensure the effectiveness and robustness of the mined features. Furthermore, through refined adaptation of different model algorithms to the data structure, the model's accuracy in identifying high-potential conversion customers is significantly improved, providing reliable technical support for precision marketing.

[0073] In some embodiments of this application, the preset feature engineering strategy includes a feature validity analysis strategy and a feature robustness analysis strategy. The step of using the preset feature engineering strategy to perform feature engineering processing on the dynamic panel data of the sample customers to obtain the multi-dimensional modeling feature data of the sample customers includes: using the feature validity analysis strategy to perform feature validity analysis on the dynamic panel data of the sample customers to obtain valid feature data of the sample customers; and using the feature robustness analysis strategy to perform feature robustness analysis on the valid feature data of the sample customers to obtain valid and robust feature data of the sample customers, which serves as the multi-dimensional modeling feature data of the sample customers.

[0074] Feature effectiveness analysis strategies can include, for example, posterior probability analysis based on driver-type features. The core of this approach is a reverse derivation from effect to cause, using customer conversion (response variable) as the outcome. By analyzing the behavioral differences between converted and non-converted customers in dynamic panel data, key driver features leading to conversion can be identified. For instance, in a banking marketing scenario, data such as the number of times converted customers viewed high-yield financial products and their click-through rates in the week prior to conversion can be collected to calculate the posterior probability (i.e., the proportion of converted customers exhibiting this behavior) and compare it with that of non-converted customers. This analysis identifies feature dimensions that significantly influence conversion behavior, representing the effective feature data of the sample customers.

[0075] Based on the valid feature data of the sample customers, further feature robustness analysis is carried out. For example, the strategy of feature robustness analysis can adopt the temporal stability evaluation strategy based on RankIC. Its core is to further screen out the effective features with high robustness by calculating the temporal stability of each dimension of features.

[0076] This application's embodiments innovatively employ a two-stage feature engineering method combining effectiveness analysis and robustness assessment by integrating the temporal characteristics of panel data. This overcomes the shortcomings of traditional methods in capturing dynamic quantitative relationships and temporal stability. The final selected feature set possesses both effectiveness and robustness in the time dimension, significantly improving the generalization ability of the customer conversion prediction model.

[0077] In some embodiments of this application, the feature validity analysis strategy includes a posterior probability analysis strategy based on motivational features. The step of using the feature validity analysis strategy to perform feature validity analysis on the dynamic panel data of the sample customers to obtain valid feature data of the sample customers includes: constructing motivational features from a business perspective and multi-dimensional dynamic and static features from a data perspective based on the posterior probability analysis strategy based on motivational features; constructing a multi-dimensional motivational feature matrix of the sample customers by combining the motivational features from the business perspective and the multi-dimensional dynamic and static features from the data perspective according to the structure of the dynamic panel data of the sample customers; performing posterior probability analysis on the multi-dimensional motivational feature matrix of the sample customers to obtain the posterior probability analysis result of the multi-dimensional motivational feature matrix; and determining the valid feature data of the sample customers based on the posterior probability analysis result of the multi-dimensional motivational feature matrix.

[0078] Feature validity analysis of dynamic panel data from sample customers mainly includes the following processes:

[0079] (1) Constructing motivational features:

[0080] 1) Business Understanding and Experience Summary: For target marketing products (such as insurance), analyze the driving factors of customer purchase conversion from a business perspective, focusing on identifying dynamic and changing characteristics. For example, through business research, summarize multiple motivations such as "life changes" (such as getting married or becoming a parent for the first time), "unexpected wealth" (such as bonuses), and "influence from others" (such as recommendations from relatives and friends).

[0081] 2) Integration of dynamic and static features from a data perspective: Constructing multi-dimensional customer features from a data perspective, including dynamic features (such as changes in customer assets, frequency of insurance page views, and consumption records of maternity and infant products) and static features (such as gender, age, and region). For example, dynamic features can be further refined into "number of insurance page views in the past 3 months" and "asset growth rate in the past 6 months," while static features may include "customer marital status" and "number of children."

[0082] 3) Feature Matrix Construction: Integrate business driver features and data features according to a dynamic panel data structure to form a multi-dimensional driver-type feature matrix. For example, with customer ID as the row and time slice as the column, fill in the values ​​of each driver feature (such as "gender" as a binary feature and "insurance browsing count" as a continuous feature) to construct a time-series-based feature matrix.

[0083] (2) Verification of motivational features (posterior probability analysis)

[0084] Based on the panel structure feature matrix data of the motivation class constructed in the previous step, the validity of each motivation class feature in the feature library is verified in turn. For example, this can be achieved by calculating the Bayesian posterior conditional probability, specifically including the following steps:

[0085] 1) Select a specific feature and arrange the driving and response variables for each customer in chronological order;

[0086] 2) Set the time of the first occurrence of the driving behavior as T. After all customers' times T are aligned, trace back to the time period of T+12 months to see if all customer response variables have made a purchase conversion, and record the time point of customers who have converted.

[0087] 3) Calculate the posterior probability corresponding to the driver feature, that is, the probability of the response variable under the selected driver condition, and calculate the time interval from the first appearance of the driver to the conversion for each customer.

[0088] For example, if customer A first viewed the insurance page on January 1, 2023, then T = 2023-01-01. Using T as the baseline, we trace back 12 months (T+12 months) to record whether the customer converted (e.g., purchased insurance) during this period and to calculate the conversion time. We then calculate the percentage of customers who purchased insurance within 12 months of their first visit to the insurance page (e.g., 720,000 out of 3.9 million customers purchase, a conditional probability of 18.5%). Simultaneously, we calculate the average and median time from the first appearance of the motivation to conversion. For example, the average interval for first-time insurance purchases is 66 days, with a median of 1 day; the average interval for repeat purchases is 82 days, with a median of 11 days.

[0089] Finally, features with conditional probabilities significantly higher than the random level (e.g., >10%) and concentrated time interval distributions were retained as effective driving features. For example, through effectiveness analysis, several important features were ultimately extracted, such as unexpected wealth, life changes (new marriage, becoming parents for the first time), and the influence of others.

[0090] (3) Application of motivational features

[0091] 1) Quantitative Business Explanation: The results of the causal feature analysis can provide good business explanation for marketing products. Feeding back the posterior probability and time interval results to the business department can support the quantitative verification of business intuition.

[0092] 2) Modeling feature input: Effective driving features can be used as highly explanatory variables to input into machine learning models (such as LSTM), replacing traditional inefficient features.

[0093] To facilitate understanding of the above embodiments, examples of the results of the motivation feature analysis are further provided:

[0094] (1) Causal relationship analysis between insurance page views and purchases:

[0095] The statistical data covers the period from May 31, 2021 to April 30, 2023 (two years).

[0096] A total of 3.9 million customers viewed the insurance page, of which 720,000 purchased insurance products, representing a conditional probability of 18.5%.

[0097] In terms of interval duration, the average interval for first-time purchases was 66 days, with a median of 1 day; the average interval for repeat purchases was 82 days, with a median of 11 days.

[0098] (2) Causal relationship analysis of the consumption of maternity and infant products in the past three months and the purchase:

[0099] The statistical data covers the period from May 31, 2021 to April 30, 2023 (two years).

[0100] A total of 1.99 million customers purchased maternity and infant products, of which 360,000 purchased insurance products, representing a conditional probability of 19%.

[0101] In terms of interval length, the average interval for first-time purchases was 221 days, with a median of 180 days; the average interval for repeat purchases was 206 days, with a median of 165 days.

[0102] This application's embodiments construct a driver-based feature system driven by both business and data, and combine it with posterior probability analysis methods to achieve accurate mining and quantitative verification of customer conversion drivers. Compared to traditional feature engineering methods, this application overcomes the limitations of static features, can capture the temporal correlation between dynamic behavior and conversion results, and quantifies feature effectiveness and statistically analyzes conversion time through Bayesian conditional probability calculation, thus endowing features with strong business interpretability.

[0103] In some embodiments of this application, the effective feature data of the sample customers includes effective feature behavior data at each time point and response variable data at the corresponding next time point. The feature robustness analysis strategy includes Rank IC analysis. The step of using the feature robustness analysis strategy to perform feature robustness analysis on the effective feature data of the sample customers to obtain the effective and robust feature data of the sample customers includes: calculating the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point using the Rank IC analysis method; and determining the effective and robust feature data of the sample customers based on the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point.

[0104] For dynamic panel data, extract effective customer characteristic behavior data (such as "number of times insurance pages were viewed in the past month" and "asset growth rate") for each time point (e.g., monthly, quarterly) and response variable data (e.g., "whether insurance was purchased") for the next time point. For example, customer A's characteristic behavior data in January 2023 was "viewed insurance pages 5 times", and their response variable in February 2023 was "purchased" (marked as 1) or "not purchased" (marked as 0).

[0105] At each time point t, extract all customers' valid characteristic behavior data (X_t) for a certain dimension and the response variable data (Y_{t+1}) for the next time point t+1 to form a cross-sectional dataset. For example, the "number of times viewed" and "purchase status" of all customers in January 2023 and February 2023, respectively. Perform rank transformation on X_t and Y_{t+1} (sort by numerical value and assign a rank), and calculate the Spearman rank correlation coefficient (Rank IC).

[0106] Iterate through all time points and calculate the Rank IC time series for each feature. For example, the Rank IC of the feature "Number of Views" fluctuated between 0.38 and 0.45 between January 2023 and December 2024, while the Rank IC of the feature "Asset Growth Rate" fluctuated between -0.12 and 0.08.

[0107] Calculate the mean, standard deviation, and significance (e.g., p-value) of the Rank IC. Set thresholds for the Rank IC mean and standard deviation, and filter features that simultaneously meet the threshold requirements. For example, retain features with a Rank IC mean > 0.3 and a standard deviation < 0.1, and remove features with a mean close to 0 or those that fluctuate drastically.

[0108] This application, by introducing Rank IC analysis, addresses the problem of traditional feature selection methods neglecting time-varying stability, achieving accurate evaluation of feature robustness in dynamic panel data. Compared to existing technologies, this application can capture the time-varying patterns of the relationship between features and response variables, selecting features that maintain strong correlation at different time points, significantly improving the model's out-of-sample generalization ability.

[0109] In some embodiments of this application, the preset classification model includes multiple models. The step of adaptively modeling using the preset classification models based on the multi-dimensional modeling feature data of the sample customers to obtain a customer conversion prediction model includes: adapting the multi-dimensional modeling feature data of the sample customers to the type of each preset classification model to obtain adapted multi-dimensional modeling feature data of each sample customer; performing adaptive modeling using the corresponding preset classification model based on the adapted multi-dimensional modeling feature data of each sample customer to obtain multiple customer conversion prediction models; evaluating the multiple customer conversion prediction models to obtain evaluation results for the multiple customer conversion prediction models; and determining the final customer conversion prediction model based on the evaluation results of the multiple customer conversion prediction models.

[0110] In terms of model algorithm selection, this application fully considers the characteristics of panel data structures and model algorithms, selecting several representative binary classification models as preset classification models. These include OLS (Linear Regression) models, suitable for scenarios where there is a linear relationship between features and target variables, fitting coefficients using the least squares method; CNN (Convolutional Neural Network) models, which utilize convolutional layers to extract local temporal or spatial features, suitable for capturing local patterns in data (such as short-term fluctuations in customer behavior sequences); and LSTM (Long Short-Term Memory) models, which handle long-sequence dependencies through gating mechanisms, suitable for analyzing long-term trends in customer dynamic characteristics over time (such as changes in purchasing behavior over six consecutive months). The panel data structure is finely adjusted to adapt to the characteristics of different algorithms, meeting the modeling requirements of different models.

[0111] The various classification models constructed above can be evaluated, for example, by calculating evaluation metrics for each model on an independent test set (such as data from the next 3 months):

[0112] (1) Precision: The proportion of samples that are predicted to be positive but are actually positive (e.g., if the model predicts 1,000 potential customers and 600 of them actually make a purchase, the precision is 60%).

[0113] (2) Recall: The proportion of samples that are actually positive that are correctly predicted (e.g., if 800 people actually made a purchase and the model predicts 600 people, the recall rate is 75%).

[0114] (3) AUC (Area under the ROC curve): A comprehensive measure of the model’s classification ability at different thresholds (AUC=0.85 means that the model has an 85% probability of ranking positive samples before negative samples).

[0115] Based on the evaluation results of the above three dimensions, the model with higher precision, recall, and AUC was selected as the final customer conversion prediction model.

[0116] This application integrates three heterogeneous models—OLS, CNN, and LSTM—and combines data adaptation processing and a dynamic evaluation and optimization mechanism to achieve accurate modeling of customer conversion prediction problems. Compared to single-model solutions, this application fully considers the characteristics of different models and evaluates their advantages, significantly improving the practicality and business value of the model's implementation.

[0117] In some embodiments of this application, the preset classification model includes ordinary least squares, a convolutional neural network model, and a long short-term memory network model. The step of adaptively modeling using the preset classification model based on the multi-dimensional modeling feature data of the sample customers to obtain a customer conversion prediction model includes: performing data verification on the multi-dimensional modeling feature data of the sample customers based on the ordinary least squares method and modeling based on the data verification results to obtain a first customer conversion prediction model; converting the multi-dimensional modeling feature data of the sample customers into two-dimensional matrix data and modeling based on the convolutional neural network model to obtain a second customer conversion prediction model; and performing serialization processing on the multi-dimensional modeling feature data of the sample customers and modeling based on the long short-term memory network model to obtain a third customer conversion prediction model.

[0118] (1) Ordinary Least Squares (OLS) Modeling Process

[0119] 1) Data validation and preprocessing:

[0120] Unit root test: Perform an ADF test on the multi-dimensional modeling feature data of sample customers (such as average monthly spending and number of page views over six consecutive months) to determine whether the time series is stationary. If a unit root exists (such as an increase in average monthly spending year by year), perform first-order differencing (such as calculating the difference in spending between adjacent months) to ensure that the data meets the stationarity assumption of OLS.

[0121] Granger causality test: Analyzes the causal relationship between a feature variable (such as "number of ad impressions") and the target variable ("whether to purchase"). For example, if the lagged term of "number of ad impressions" has a significant effect on "purchase behavior" (p<0.05), then that feature is retained as a modeling input.

[0122] Fixed effects model selection: In a panel data structure, the fixed effects model and the random effects model are compared using the Hausman test. If the test result is significant (p<0.05), the fixed effects model is selected to control for individual heterogeneity (such as the inherent consumption preferences of different customers).

[0123] 2) Modeling and Output:

[0124] Using the tested characteristic variables (such as the average monthly spending amount after difference and the lagged term of the number of ad exposures) as independent variables and "whether to purchase" (0 / 1) as the dependent variable, an OLS binary logistic regression model (which maps the linear output to probability through the Sigmoid function) is constructed as the first customer conversion prediction model.

[0125] (2) Convolutional Neural Network (CNN) Modeling Process

[0126] In CNN models, to adapt to the characteristics of the CNN algorithm, panel data is adjusted for suitability, that is, all features observed in all samples over time are considered. and response variables The obtained panel data is transformed into two-dimensional matrices resembling black and white images (i.e., at one observation point, all features and response variables of all observed samples constitute a two-dimensional input matrix), which is then input into the CNN model for training. This process is repeated, inputting the sample matrix from each time point into the model sequentially to complete training, ultimately resulting in a trained binary classification CNN model, which is the second customer conversion prediction model mentioned above.

[0127] (3) Modeling process of Long Short-Term Memory Network (LSTM)

[0128] In the special RNN model of Long Short-Term Memory (LSTM), in order to adapt to the characteristics of the LSTM algorithm, the panel data also needs to be adjusted for suitability. This involves serializing the panel data, that is, processing all features of each observation sample. and response variables The data is expanded chronologically and transformed into sequential data (i.e., a single observation sample, all features, and response variables arranged chronologically to form a sequence of data), which is then input into an LSTM model for training. The model uses data from before time t (i.e., the output of the hidden layer) representing customer purchase conversions. ) and the current independent variable Response variables The models were trained together, and the third customer conversion prediction model was finally obtained.

[0129] This application's embodiments integrate three heterogeneous models—OLS, CNN, and LSTM—and combine them with data adaptation adjustments and dynamic evaluation and optimization mechanisms to achieve accurate modeling of customer conversion prediction problems. Compared to single-model solutions, the multi-model evaluation framework provides business users with flexible choices, significantly improving the practicality and business value of model implementation.

[0130] This application also provides a customer marketing method, such as... Figure 2 The diagram illustrates a customer marketing method according to an embodiment of this application, which includes the following steps S210 to S230:

[0131] Step S210: Obtain multi-dimensional feature data of the customer;

[0132] Step S220: Based on the multi-dimensional characteristic data of the customer, a customer conversion prediction model is used to make a prediction and obtain the customer conversion prediction result.

[0133] Step S230: Determine the target marketing customer information based on the customer's conversion prediction results;

[0134] The customer conversion prediction model is constructed based on the modeling method of the customer conversion prediction model described in any of the preceding items.

[0135] Based on the feature engineering analysis results of the modeling phase of the aforementioned embodiments, it is possible to determine the driver-type feature dimensions that are strongly correlated with and robust to customer conversion. For example, these can cover customer behavior characteristics in mobile banking, such as browsing and clicking on mobile banking operations; consumption behavior characteristics, such as users' consumption behavior in different types of merchants; and customer asset characteristics, such as recent asset deposits and whether there are large fluctuations.

[0136] Based on the aforementioned multi-dimensional driving characteristics determined during the modeling phase, customer feature data across these dimensions is collected through methods such as system tracking. According to the data input requirements of the customer conversion prediction model trained in the previous embodiment, the multi-dimensional customer feature data is adapted and adjusted, then input into the customer conversion prediction model for conversion prediction, yielding the customer's conversion probability. Finally, based on the set conversion probability threshold, customers with high conversion probabilities are selected as target customers for marketing.

[0137] The customer marketing method in this application uses a pre-trained customer conversion prediction model to predict conversion probability, which improves the accuracy of customer conversion probability prediction and thus improves the precision of customer marketing.

[0138] In summary, the key points of this application are mainly as follows:

[0139] (1) The precision marketing modeling method proposed in this application is a complete systematic approach based on panel data analysis and modeling. It fully utilizes the characteristics of the time factor in panel data, and combines dynamic features, static features, and response variables to carry out feature engineering and model construction. This allows the past actions and behaviors of the observed samples to be linked to whether they make a purchase or conversion now. Both the research analysis and modeling can dynamically capture the impact of the time factor on individual changes.

[0140] (2) In the feature engineering stage, this application abandons traditional techniques such as correlation analysis and IV value analysis, and instead uses panel data structures entirely, employing both posterior probability and RankIC methods for feature engineering analysis. A causal analysis method closely integrated with marketing products is used to find effective features for modeling by deriving causes from effects. Furthermore, taking into full account the time factor characteristics of panel data structures and the stability of features, this application also uses the RankIC method to analyze the robustness of features. Both analysis methods are fully coupled with the panel data structure.

[0141] (3) The method proposed in this application is the first to apply multiple neural network modeling algorithms to panel data, and the panel data is adapted to the characteristics of different neural network model algorithms. This adaptation is also the key innovation of this application.

[0142] The main technical effects of this application include:

[0143] (1) In terms of data structure, the precision marketing method proposed in this application is based on dynamic panel data for analysis and modeling. The data structure is a fusion of cross-sectional data and time series data. This application also fully considers the impact of the dynamic characteristics of the observed objects on marketing conversion over time. The research, analysis and modeling process can dynamically capture the impact of time factors on individual changes, making up for the shortcomings of existing precision marketing modeling techniques.

[0144] (2) Regarding feature engineering, this application uses a panel data structure and employs two methods, posterior probability and RankIC, for feature engineering analysis. The posterior probability method fully integrates the characteristics of the product, using cause-and-effect reasoning to mine effective features, thus avoiding the impact of data quality on feature selection in traditional methods. In addition, considering the time factor of panel data, the features analyzed by the RankIC method further ensure the discovery of robust and effective features. In terms of feature selection, this application fully compensates for the deficiencies of existing precision marketing feature analysis.

[0145] (3) Regarding the model algorithm, this application is based on a panel data structure. According to the characteristics of different algorithms, the panel data is adapted and adjusted to fully leverage the advantages of the algorithm and the panel data structure, resulting in higher model accuracy and better marketing application effects. In terms of the model algorithm, this application fully fills the gap in existing neural network model algorithms for modeling on panel data, allowing the advantages of both the model algorithm and the data structure to be fully utilized.

[0146] This application embodiment also provides a modeling apparatus 300 for a customer conversion prediction model, such as... Figure 3 The diagram shows a structural schematic of a modeling device for a customer conversion prediction model according to an embodiment of this application. The modeling device 300 for the customer conversion prediction model includes:

[0147] Building unit 310 is used to build dynamic panel data for sample customers;

[0148] The feature engineering processing unit 320 is used to perform feature engineering processing on the dynamic panel data of the sample customer using a preset feature engineering strategy to obtain multi-dimensional modeling feature data of the sample customer.

[0149] Modeling unit 330 is used to perform adaptive modeling using a preset classification model based on the multi-dimensional modeling feature data of the sample customers to obtain a customer conversion prediction model.

[0150] In some embodiments of this application, the preset feature engineering strategy includes a feature validity analysis strategy and a feature robustness analysis strategy. The feature engineering processing unit 320 is specifically used to: perform feature validity analysis on the dynamic panel data of the sample customer using the feature validity analysis strategy to obtain the valid feature data of the sample customer; and perform feature robustness analysis on the valid feature data of the sample customer using the feature robustness analysis strategy to obtain the valid and robust feature data of the sample customer, which serves as the multi-dimensional modeling feature data of the sample customer.

[0151] In some embodiments of this application, the feature validity analysis strategy includes a posterior probability analysis strategy based on motivational features. The feature engineering processing unit 320 is specifically used to: construct motivational features from a business perspective and multi-dimensional dynamic and static features from a data perspective based on the posterior probability analysis strategy based on motivational features; construct a multi-dimensional motivational feature matrix for the sample customer by combining the motivational features from the business perspective and the multi-dimensional dynamic and static features from the data perspective according to the structure of the sample customer's dynamic panel data; perform posterior probability analysis on the multi-dimensional motivational feature matrix of the sample customer to obtain the posterior probability analysis result of the multi-dimensional motivational feature matrix; and determine the valid feature data of the sample customer based on the posterior probability analysis result of the multi-dimensional motivational feature matrix.

[0152] In some embodiments of this application, the effective feature data of the sample customers includes effective feature behavior data at each time point and response variable data at the corresponding next time point. The feature robustness analysis strategy includes Rank IC analysis. The feature engineering processing unit 320 is specifically used to: calculate the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point using the Rank IC analysis; and determine the effective and robust feature data of the sample customers based on the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point.

[0153] In some embodiments of this application, the preset classification model includes multiple models, and the modeling unit 330 is specifically used for: adapting the multi-dimensional modeling feature data of the sample customers according to the type of each preset classification model to obtain multi-dimensional modeling feature data of each adapted sample customer; performing adaptive modeling using the corresponding preset classification model based on the multi-dimensional modeling feature data of each adapted sample customer to obtain multiple customer conversion prediction models; evaluating the multiple customer conversion prediction models to obtain evaluation results of the multiple customer conversion prediction models; and determining the final customer conversion prediction model based on the evaluation results of the multiple customer conversion prediction models.

[0154] In some embodiments of this application, the preset classification model includes ordinary least squares, a convolutional neural network model, and a long short-term memory network model. The modeling unit 330 is specifically used for: performing data verification on the multi-dimensional modeling feature data of the sample customers based on the ordinary least squares method and modeling according to the data verification results to obtain a first customer conversion prediction model; converting the multi-dimensional modeling feature data of the sample customers into two-dimensional matrix data and modeling based on the convolutional neural network model to obtain a second customer conversion prediction model; and performing serialization processing on the multi-dimensional modeling feature data of the sample customers and modeling based on the long short-term memory network model to obtain a third customer conversion prediction model.

[0155] It is understood that the modeling device for the customer conversion prediction model described above can implement each step of the modeling method for the customer conversion prediction model provided in the foregoing embodiments. The relevant explanations regarding the modeling method for the customer conversion prediction model are applicable to the modeling device for the customer conversion prediction model, and will not be repeated here.

[0156] This application embodiment also provides a customer marketing device 400, such as Figure 4 As shown, a structural schematic diagram of a customer marketing device according to an embodiment of this application is provided. The customer marketing device 400 includes:

[0157] Acquisition unit 410 is used to acquire multi-dimensional feature data of customers;

[0158] Prediction unit 420 is used to make predictions based on the customer's multi-dimensional feature data using a customer conversion prediction model, and obtain the customer conversion prediction result.

[0159] Marketing unit 430 is used to determine target marketing customer information based on the conversion prediction results of the customer;

[0160] The customer conversion prediction model is constructed based on the aforementioned customer conversion prediction model modeling device.

[0161] It is understood that the above-mentioned customer marketing device can realize each step of the customer marketing method provided in the foregoing embodiments. The relevant explanations of the customer marketing method are applicable to the customer marketing device and will not be repeated here.

[0162] Figure 5 This is a schematic diagram of the structure of a device according to an embodiment of this application. For example... Figure 5 As shown, the device includes one or more processors (or processing units), and may also include one or more memories coupled to the processors, and may also include a communication module coupled to the processors.

[0163] A communication module can be used to communicate with other devices or apparatuses, such as sending or receiving data and / or signals. A communication module may have at least one communication module for communication. A communication module may include any interface necessary for communicating with other devices. Exemplarily, a communication module may be a transceiver, circuit, bus, module, or other type of communication module.

[0164] The processor may include, but is not limited to, one or more of the following: a general-purpose computer, a special-purpose computer, a microcontroller, a digital signal processor (DSP), or a controller-based multi-core controller architecture. The device may have multiple processors, such as application-specific integrated circuit (ASIC) chips, which are time-dependent on a clock synchronized with the main processor.

[0165] The memory may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, at least one of the following: read-only memory (ROM), electrically programmable read-only memory (EPROM), flash memory, hard disk, compact disc (CD), digital video disc (DVD), or other magnetic and / or optical storage. Examples of volatile memories include, but are not limited to, at least one of the following: random access memory (RAM), or other volatile memories that do not persist during the duration of a power outage.

[0166] A computer program consists of computer-executable instructions that are executed by an associated processor. Programs can be stored in ROM. A processor can perform any appropriate action and processing by loading the program into RAM.

[0167] Possible implementations of this application can be achieved through a program, enabling the communication device to execute any of the processes discussed in the foregoing embodiments. Possible implementations of this application can also be achieved through hardware or a combination of software and hardware.

[0168] In some implementations, the program may be tangibly contained in a computer-readable storage medium, which may include in a device (such as in memory) or other storage device accessible by the device. The program may be loaded from the computer-readable storage medium into RAM for execution. The computer-readable storage medium may include any type of tangible non-volatile memory, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc.

[0169] This application also provides a computer-readable storage medium storing computer instructions or program code thereon, which, when executed by a processor, causes the processor to perform the methods and functions involved in any of the above embodiments. A computer-readable medium can be any tangible medium that contains or stores a program for or relating to an instruction execution system, apparatus, or device. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. More detailed examples of computer-readable storage media include electrical connections with one or more wires, magnetic media (e.g., disks, floppy disks, hard disks, magnetic tapes, magnetic storage devices), optical media (e.g., optical storage devices, DVDs), semiconductor media (e.g., solid-state drives), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), or any suitable combination thereof.

[0170] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. Embodiments of this application also provide at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. This computer program product includes one or more computer-executable instructions, such as instructions included in a program module, which execute in a device on a target real or virtual processor to perform the processes, methods, and functions involved in any of the above embodiments. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0171] This application also proposes a computer program product, including a computer program or instructions that, when run on a computer, cause the computer to perform the processes, methods, and functions described in the above embodiments. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided as needed. The machine-executable instructions for the program modules can be executed locally or in a distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0172] Generally, the various embodiments of this application can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software, which can be executed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of this disclosure are shown and described as block diagrams, flowcharts, or represented using some other illustration, it should be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as, as non-limiting examples, in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0173] It should be noted that although embodiments of this application have been described above with reference to the accompanying drawings, these embodiments are not independent of each other, and they can be combined to obtain other embodiments. The methods, situations, categories, and classifications of embodiments in this application are only for the convenience of description and should not constitute a special limitation. Various methods, categories, situations, and features in embodiments can be combined with each other if logically consistent. The various embodiments of this application can be arbitrarily combined to achieve different technical effects. The embodiments of this application will not list various combinations.

[0174] Furthermore, although the operation of the methods of this disclosure is described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowcharts may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps. It should also be noted that the features and functions of two or more devices according to this disclosure may be embodied in one device. Conversely, the features and functions of one device described above may be further divided and embodied by multiple devices.

[0175] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0176] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A modeling method for a customer conversion prediction model, characterized in that, The modeling method for the customer conversion prediction model includes: Construct dynamic panel data for sample customers; The dynamic panel data of the sample customers are processed using a preset feature engineering strategy to obtain multi-dimensional modeling feature data of the sample customers. Based on the multi-dimensional modeling feature data of the sample customers, an adaptive modeling method is used to obtain a customer conversion prediction model.

2. The modeling method for the customer conversion prediction model according to claim 1, characterized in that, The preset feature engineering strategy includes a feature effectiveness analysis strategy and a feature robustness analysis strategy. The process of using the preset feature engineering strategy to perform feature engineering on the dynamic panel data of the sample customers to obtain multi-dimensional modeling feature data of the sample customers includes: The feature validity analysis strategy is used to perform feature validity analysis on the dynamic panel data of the sample customers to obtain the valid feature data of the sample customers. The feature robustness analysis strategy is used to perform feature robustness analysis on the effective feature data of the sample customers to obtain effective and robust feature data of the sample customers, which can be used as multi-dimensional modeling feature data of the sample customers.

3. The modeling method for the customer conversion prediction model according to claim 2, characterized in that, The feature validity analysis strategy includes a posterior probability analysis strategy based on motivational features. The effective feature data of the sample customers obtained by performing feature validity analysis on the dynamic panel data of the sample customers using the feature validity analysis strategy includes: Based on the aforementioned posterior probability analysis strategy based on motivational features, motivational features from a business perspective and multi-dimensional dynamic and static features from a data perspective are constructed. Based on the structure of the dynamic panel data of the sample customers, the business-related driving factors and the data-related multi-dimensional dynamic and static features are used to construct the multi-dimensional driving factor feature matrix of the sample customers. Posterior probability analysis was performed on the multidimensional motivational feature matrix of the sample customers to obtain the posterior probability analysis results of the multidimensional motivational feature matrix. The effective feature data of the sample customers are determined based on the posterior probability analysis results of the multi-dimensional motivational feature matrix.

4. The modeling method for the customer conversion prediction model according to claim 2, characterized in that, The effective feature data of the sample customers includes effective characteristic behavior data at each time point and response variable data at the corresponding next time point. The feature robustness analysis strategy includes Rank IC analysis. The effective and robust feature data of the sample customers obtained by using the feature robustness analysis strategy to perform feature robustness analysis on the effective feature data of the sample customers includes: The Rank IC analysis method is used to calculate the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point; The effective and robust feature data of the sample customers are determined based on the correlation coefficient between the effective feature behavior data at each time point and the response variable data at the corresponding next time point.

5. The modeling method for the customer conversion prediction model according to claim 1, characterized in that, The preset classification model includes multiple models. The step of adaptively modeling using the preset classification model based on the multi-dimensional modeling feature data of the sample customers to obtain the customer conversion prediction model includes: The multi-dimensional modeling feature data of the sample customers are adapted according to the type of each preset classification model to obtain the multi-dimensional modeling feature data of each adapted sample customer. Based on the multi-dimensional modeling feature data of each adapted sample customer, adaptive modeling is performed using the corresponding preset classification model to obtain multiple customer conversion prediction models. The customer conversion prediction models are evaluated to obtain the evaluation results of the customer conversion prediction models. The final customer conversion prediction model is determined based on the evaluation results of multiple customer conversion prediction models.

6. The modeling method for the customer conversion prediction model according to claim 5, characterized in that, The preset classification model includes ordinary least squares, convolutional neural network, and long short-term memory network. The step of adaptively modeling using the preset classification model based on the multi-dimensional modeling feature data of the sample customers to obtain the customer conversion prediction model includes: Based on the ordinary least squares method, the multi-dimensional modeling feature data of the sample customers are tested, and modeling is performed based on the test results to obtain the first customer conversion prediction model. Based on the convolutional neural network model, the multi-dimensional modeling feature data of the sample customers are converted into two-dimensional matrix data and modeled to obtain the second customer conversion prediction model. Based on the Long Short-Term Memory network model, the multi-dimensional modeling feature data of the sample customers are serialized and modeled to obtain a third customer conversion prediction model.

7. A customer marketing method, characterized in that, The customer marketing methods include: Obtain multi-dimensional characteristic data of customers; Based on the customer's multi-dimensional characteristic data, a customer conversion prediction model is used to make predictions and obtain the customer's conversion prediction results. Target marketing customer information is determined based on the customer conversion prediction results. The customer conversion prediction model is constructed based on the modeling method of the customer conversion prediction model according to any one of claims 1 to 6.

8. A modeling apparatus for a customer conversion prediction model, characterized in that, The modeling apparatus for the customer conversion prediction model includes: Building units are used to construct dynamic panel data for sample customers; The feature engineering processing unit is used to perform feature engineering processing on the dynamic panel data of the sample customer using a preset feature engineering strategy to obtain multi-dimensional modeling feature data of the sample customer. The modeling unit is used to perform adaptive modeling based on the multi-dimensional modeling feature data of the sample customers and a preset classification model to obtain a customer conversion prediction model.

9. A customer marketing device, characterized in that, The customer marketing device includes: The acquisition unit is used to acquire multi-dimensional feature data of customers; The prediction unit is used to make predictions based on the customer's multi-dimensional feature data using a customer conversion prediction model, and obtain the customer's conversion prediction results. The marketing unit is used to determine target marketing customer information based on the conversion prediction results of the customer. The customer conversion prediction model is constructed based on the modeling device of the customer conversion prediction model described in claim 8.

10. An apparatus comprising: processor; And a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the modeling method of the customer conversion prediction model of any one of claims 1 to 6, and the customer marketing method of claim 7.

11. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the modeling method of the customer conversion prediction model according to any one of claims 1 to 6, and the customer marketing method according to claim 7.

Images