Potential customer identification method and device, and electronic equipment

Abstract

The invention discloses a potential customer identification method and device, and electronic equipment, and the method comprises the steps: firstly obtaining a first data set, then carrying out the preprocessing of the first data set, obtaining first preprocessing data, carrying out the processing of the first preprocessing data based on a first preset XGBoost model, obtaining a first behavior data set, and carrying out the recognition of a potential customer through the first behavior data set; the method comprises the steps that a first behavior data set is acquired, then the first behavior data set is processed through incremental feature engineering to acquire a first feature data set, finally, a customer probability value is acquired for the first feature data set through a second preset XGBoost model, and the first preset XGBoost model is different from the second preset XGBoost model, that is, potential customers are identified based on the customer probability value. Therefore, the accuracy of identifying the potential customers is improved.

Description

Technical Field

[0001] This application relates to artificial intelligence, and includes, but is not limited to, a method, apparatus, and electronic device for identifying potential customers. Background Technology

[0002] Existing technical solutions for identifying potential customers in 5G messaging marketing primarily rely on machine learning algorithms and rule engines. Specifically, by collecting user data in a 5G network environment, classification and clustering algorithms in machine learning (such as decision trees, random forests, K-means, etc.) are used to analyze user characteristics, build potential customer models, and combine these with a series of marketing rules set by the rule engine (such as purchase frequency, browsing duration, etc.) to screen and identify potential customers, thereby achieving precise marketing.

[0003] Existing solutions quantify user interests using only simple statistical features (such as click frequency), which cannot effectively distinguish between incidental behavior and genuine purchase intent, resulting in low accuracy in identifying potential customers. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus, and electronic device for identifying potential customers.

[0005] The technical solution of this application embodiment is implemented as follows: This application provides a method for identifying potential customers. The method includes: acquiring a first dataset; preprocessing the first dataset to obtain first preprocessed data; processing the first preprocessed data based on a first preset XGBoost model to obtain a first behavioral dataset; processing the first behavioral dataset through incremental feature engineering to obtain a first feature dataset; processing the first feature dataset through a second preset XGBoost model to obtain customer probability values; and identifying potential customers based on the customer probability values. The first preset XGBoost model and the second preset XGBoost model are different.

[0006] Optionally, the incremental feature engineering includes at least: weighted processing with dual decay time windows; cross-feature processing; and innovative feature processing.

[0007] Optionally, the step of processing the first behavioral dataset through incremental feature engineering to obtain the first feature dataset includes: processing the first behavioral dataset through a double decay algorithm to obtain first interest data; processing the first interest data through cross-feature processing to obtain second interest data; and processing the second interest data based on innovative features to obtain the first feature dataset.

[0008] Optionally, the step of processing the first row of the dataset using a double-decay algorithm to obtain the first interest data includes: the double-decay algorithm being: ;in, The time difference between when the action occurred and the present moment; Short-term decay coefficient; Long-term attenuation coefficient; : Mixing ratio.

[0009] Optionally, before obtaining customer probability values ​​from the first feature dataset, the second preset XGBoost model includes: generating the second preset XGBoost model based on the AFIHGD gradient strategy and the XGBoost model.

[0010] Optionally, the AFIHGD gradient strategy includes: gradient update rules: in, Fractional gradient; Traditional integer gradient; : Dynamic mixing coefficient; Adaptive fractional order.

[0011] Optionally, the AFIHGD gradient strategy further includes: .

[0012] Optionally, before processing the first preprocessed data based on the first preset XGBoost model to obtain the first row dataset, the process includes: model initialization; parameter specification selection; grid search training; and obtaining the first preset XGBoost model.

[0013] An identification device, comprising: an acquisition unit, an analysis unit, and a processing unit; the acquisition unit being configured to acquire a first dataset; the analysis unit being configured to preprocess the first dataset to obtain first preprocessed data; process the first preprocessed data based on a first preset XGBoost model to obtain a first behavioral dataset; process the first behavioral dataset through incremental feature engineering to obtain a first feature dataset; the processing unit being configured to obtain customer probability values ​​from the first feature dataset using a second preset XGBoost model, and identify potential customers based on the customer probability values; wherein the first preset XGBoost model and the second preset XGBoost model are different.

[0014] An electronic device includes: a memory for storing at least one set of instructions; a processor for acquiring a first dataset; preprocessing the first dataset to obtain first preprocessed data; processing the first preprocessed data based on a first preset XGBoost model to obtain a first behavioral dataset; processing the first behavioral dataset through incremental feature engineering to obtain a first feature dataset; processing the first feature dataset through a second preset XGBoost model to obtain a customer probability value, and identifying potential customers based on the customer probability value; wherein the first preset XGBoost model and the second preset XGBoost model are different.

[0015] This application provides a method, apparatus, and electronic device for identifying potential customers. First, a first dataset is acquired. Then, the first dataset is preprocessed to obtain first preprocessed data. Next, the first preprocessed data is processed based on a first preset XGBoost model to obtain a first row dataset. Then, the first row dataset is processed through incremental feature engineering to obtain a first feature dataset. Finally, a customer probability value is obtained from the first feature dataset using a second preset XGBoost model. The first preset XGBoost model and the second preset XGBoost model are different; that is, potential customers are identified based on customer probability values, thereby improving the accuracy of potential customer identification. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the potential customer identification method provided in this application embodiment; Figure 2 Another flowchart illustrating the potential customer identification method provided in this application embodiment; Figure 3 Another flowchart illustrating the potential customer identification method provided in this application embodiment; Figure 4 Another flowchart illustrating the potential customer identification method provided in this application embodiment; Figure 5 Another flowchart illustrating the potential customer identification method provided in this application embodiment; Figure 6 A schematic diagram of the identification device provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structural composition of the electronic device provided in the embodiments of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] Please refer to Figure 1 ,in, Figure 1 A flowchart illustrating an implementation of the potential customer identification method provided in this application embodiment may include: Step S101: Obtain the first dataset; Step S102: Preprocess the first dataset to obtain the first preprocessed data; Step S103: Process the first preprocessed data based on the first preset XGBoost model to obtain the first row dataset; Step S104: Process the first row dataset through incremental feature engineering to obtain the first feature dataset; Step S105: Obtain customer probability values ​​from the first feature dataset using the second preset XGBoost model, and identify potential customers based on the customer probability values; wherein, the first preset XGBoost model is different from the second preset XGBoost model.

[0019] The first dataset includes at least: user profiles, package information, and whether push notifications generated clicks; the first preset XGBoost model can be an XGBoost-trained coarse classification model; incremental feature engineering includes at least: double decay time window weighting, cross features, and innovative features; the second preset XGBoost model can be an AFIHGD-XGBoost model. Specifically, 1. Data preprocessing. Obtain 15 days of historical user data from Henan Province, including user profiles, package information, and whether push notifications resulted in clicks. Extract 200,000 data points for training a coarse-screening model, named Dataset A; extract 500,000 data points for subsequent calculations, named Dataset B (the only difference from Dataset A is that it does not include the click feature. This behavioral feature is obtained by pushing marketing information to users and monitoring whether users click on it). As shown in Table 1;

[0020] Table 1 (1) The data was processed by removing rows with missing data (leaving 189,522 records). Enumerated data types (star rating, customer type, traffic preference, etc.) were converted using labelEncode. Marketing click records were used as tag data, with clicks marked as 1 and misses marked as 0. Sample data is shown in Table 2:

[0021] Table 2 (2) The dataset is divided into a training set (151618) and a validation set (37904) according to an 8:2 ratio.

[0022] 2. Coarse sieve model training In marketing and promotion, user data is enormous. How to accurately filter out marketing targets from this user data has always been a pressing issue for operators. This solution utilizes a coarse-screening followed by fine-tuning approach to eliminate non-potential customers from massive amounts of data. The coarse-screening model is particularly crucial. In this solution, based on the data output from step one, an XGBoost coarse classification model is trained. This model performs a coarse-screening of potential users, narrowing down the data range required for subsequent algorithms in the automated calculation process. The process of obtaining this model is consistent with obtaining a general XGBoost classification model, as follows: (1) Initialize the XGBoost algorithm parameters, adjust the weights to address the imbalance of samples, and then input the training set data for training; (2) Use the test set data to validate the trained model using F1 scoring; (3) Use grid search to search for optimal hyperparameters, and perform hyperparameter optimization and cross-validation on the model; (4) Finally, the model that scored well on the test set was obtained, and the final model score was F1=0.84; (5) Using the trained model, input dataset B, and after calculation and filtering, potential customer data is extracted. Out of 500,000 data points, 32,174 data points are identified as potential customers; an example is shown in Table 3:

[0023] Table 3 (6) In operator marketing and promotion, the number of users after the initial screening is also very large, and there are some uncertain factors affecting users. In order to more accurately screen out high potential customers, it is also necessary to deeply explore the potential relationship between the click behavior of potential customers and the influence of uncertain factors on user appearance (mapped in the relationship of historical behavior), such as historical behavior characteristics, behavior entropy, etc. Based on the data marked as potential customers obtained in the previous step, unified marketing and promotion were carried out, and it was monitored whether customers clicked within one day after the promotion information was successfully sent. 18265 data with click behavior were screened out and named dataset C. Data samples are shown in Table 4.

[0024] Table 4 (7) Based on the dataset C output in the previous step, associate it with the user behavior data of the past 7 days. The data is updated by sliding, and the update time step is 1 day. The sample behavior data to be collected is shown in Table 5.

[0025] Table 5 The first row dataset is then processed through incremental feature engineering to obtain the first feature dataset; the first feature dataset is processed through the second preset XGBoost model to obtain customer probability values, and potential customers are identified based on the customer probability values; wherein, the first preset XGBoost model and the second preset XGBoost model are different.

[0026] Please refer to Figure 2 The method in this embodiment may include: incremental feature engineering, which at least includes: Step S201: Weighted processing with dual decay time windows; Step S202: Cross-feature processing; Step S203: Innovative feature processing.

[0027] Specifically, 3. Incremental Feature Engineering (1) Based on the user's historical 7-day behavioral data, calculate behavioral data statistics to capture the dynamic trends of user behavior and avoid the limitations of static features. The original behavioral log data is shown in Table 6:

[0028] Table 6 The processed feature sample data are shown in Table 7:

[0029] Table 7 (2) Weighted by dual decay time window Dual-decay time-window weighting is a dynamic weighting method used to quantify the importance of historical user behavior to current predictions. Its core principle is to control the decay rate of short-term and long-term behavior using two decay coefficients, thereby more flexibly capturing changing patterns of user interests. Short-term decay: rapidly reduces the impact of earlier actions, emphasizing the significance of recent actions (such as clicks within 24 hours); Long-term decay: Slowly reduce the weight of long-term behaviors while retaining signals of long-term habits (such as weekly login patterns).

[0030] a. The weighted calculation formula for the double decay time window is as follows: ; The time difference between the occurrence of the behavior and the present (in days); Short-term decay coefficient; Long-term attenuation coefficient; Mixing ratio; b. Parameter range and physical meaning: Short-term decay coefficient, which controls the rate of weight reduction of recent behavior. The larger the value, the more sensitive the model is to recent behavior. A common empirical range is (0.01, 1).

[0031] Long-term decay coefficient, which modulates the impact on long-term behavior. The smaller the value, the longer the duration of the impact of historical behavior. A common empirical range is (0.01, 1).

[0032] : A mixed weighting of short-term and long-term factors to balance the impact of near-term and long-term behavior, [0, 1], which is usually fine-tuned according to business focus.

[0033] c. Basis for parameter settings: The initial value is set to , , The optimal parameters were determined through experimental tuning, referencing [Time-Aware Recommender Systems: A Comprehensive Survey. ACM Computing Surveys, 2021.] and the distribution characteristics of user behavior activity in actual business scenarios. This ensures that the model has a high response to sudden behaviors in the past 7 days; This helps retain users' long-term stable preferences; It can take into account both short-term interest fluctuations and long-term preferences.

[0034] d. Parameter optimization experiment: Input sample data (the following example uses the historical behavior data of a single user (user ID 1001) for 7 days) is shown in Table 8:

[0035] Table 8 The weights for each time point are calculated, as shown in Table 9:

[0036] Table 9 Calculate the weighted action value, using the number of clicks as an example:

[0037]

[0038] Normalization (dividing the weighted values ​​by the sum of the total weights to obtain standardized features):

[0039] ; e. Output results, as shown in Table 10:

[0040] Table 10 f. Parameter tuning experiments: The effect of different decay coefficients was verified through grid search (based on the AUC predicted by click-through rate). The experiments showed that... , =0.1 is better. As shown in Table 11:

[0041] Table 11 (3) Cross-feature New features are generated through multiplication, division, or conditional combinations to reveal the interactions between features and improve the nonlinear expressive power of the model. The original static feature sample data is shown in Table 12.

[0042] Table 12 After cross-processing, the behavioral data samples are shown in Table 13:

[0043] Table 13 (4) Innovative characteristics a. Behavioral entropy: quantifies the unpredictability of user behavior (the higher the entropy value, the more random the behavior).

[0044] ; The percentage of this behavior within the statistical period; b. Calculation steps: Statistics on the distribution of user behavior types over the past 7 days (e.g., clicks accounted for 60%, adding to cart accounted for 20%, and browsing accounted for 20%).

[0045] Calculate the entropy value: ; c. Output examples, as shown in Table 14:

[0046] Table 14 (5) Feature Importance Verification Experiment a. Experimental Design: a) Control group: Original characteristics ("Age", "Gender", "Star rating", "Customer type", "Network age", "Traffic preference") + sum of historical seven-day behavioral data ("Number of clicks", "Number of add-to-cart", "Number of logins", "Number of page views").

[0047] b) Experimental Group 1: Original features ("Age", "Gender", "Star Rating", "Customer Type", "Network Age", "Traffic Preference") + Sum of historical seven-day behavioral data ("Number of Clicks", "Number of Add-to-Carts", "Number of Logins", "Number of Views") + Behavioral data statistics ("Mean Number of Clicks", "Variance of Number of Clicks", "Slope of Number of Add-to-Carts", "Change in Number of Logins %") + Weighted by double decay time window ("Weighted Number of Clicks", "Weighted Number of Add-to-Carts", "Weighted Number of Logins") + Cross features ("Logins × Purchases", "Add-to-Cart / Views", "High Login Flag"); c) Experimental Group 2: Original features + sum of historical seven-day behavioral data + behavioral data statistics + weighted by double decay time window + cross features + innovative features (behavioral entropy); b. Results comparison (XGBoost model) statistics, as shown in Table 15:

[0048] Table 15 Time series and cross features significantly improve model performance, with behavioral entropy becoming a key feature; (6) Examples of features are shown in Table 16:

[0049] Table 16 Input features include: "Age", "Gender", "Star Rating", "Customer Type", "Network Age", "Traffic Preference", "Number of Clicks", "Number of Adds to Cart", "Number of Logins", "Number of Views", "Mean Number of Clicks", "Variance of Number of Clicks", "Slope of Number of Adds to Cart", "% Change in Number of Logins", "Weighted Number of Clicks", "Weighted Number of Adds to Cart", "Weighted Number of Logins", "Login × Purchase", "Add to Cart / View", "High Login Flag", "Behavioral Entropy", and other 21-dimensional feature data.

[0050] Tags: Whether a click behavior is detected is used as a tag. A click behavior is marked as 1 (high potential customer), and no click behavior is marked as 0 (low potential customer).

[0051] Convert the enumerated values ​​in the data into numerical codes so that the data can be directly calculated.

[0052] The dataset was divided into a training set and a validation set in a 7:3 ratio.

[0053] Please refer to Figure 3 The method in this embodiment may include: processing the first row dataset through incremental feature engineering to obtain a first feature dataset, including: Step S301: Process the first row of the dataset using a double decay algorithm to obtain the first interest data; Step S302: Process the first interest data through cross-feature processing to obtain the second interest data; Step S303: Process the second interest data based on the innovative features to obtain the first feature dataset.

[0054] The method in this embodiment may include: processing the first row of the dataset using a double decay algorithm to obtain first interest data, including: the double decay algorithm is as follows: ;in, The time difference between when the action occurred and the present moment; Short-term decay coefficient; Long-term attenuation coefficient; : Mixing ratio.

[0055] The method in this embodiment may include: before obtaining customer probability values ​​from the first feature dataset using a second preset XGBoost model, the method includes: generating a second preset XGBoost model based on the AFIHGD gradient strategy and the XGBoost model. The method in this embodiment may include: the AFIHGD gradient strategy includes: gradient update rules: in, Fractional gradient; Traditional integer gradient; : Dynamic mixing coefficient; Adaptive fractional order.

[0056] The method in this embodiment may include: the AFIHGD gradient strategy, and also includes: .

[0057] Specifically, 4. Potential customer identification (improved XGBoost algorithm) (1) Obtain the enhanced feature data from step 3, as shown in Table 17:

[0058] Table 17 These are: "Age", "Gender", "Star Rating", "Customer Type", "Online Age", "Traffic Preference", "Number of Clicks", "Number of Adds to Cart", "Number of Logins", "Number of Views", "Mean Number of Clicks", "Variance of Number of Clicks", "Slope of Number of Adds to Cart", "Change in Number of Logins %", "Weighted Number of Clicks", "Weighted Number of Adds to Cart", "Weighted Number of Logins", "Login × Purchase", "Add to Cart / Views", "High Login Flag", "Behavioral Entropy", and Potential Customer Tags (1 for high potential, 0 for low potential). (2) Dynamic weights - balancing sample differences An adaptive weight adjustment mechanism is introduced to assign dynamic weights to a minority group of samples (high-potential users) (data samples obtained through steps 1 to 3. First, a coarse-grained screening model is obtained through historical marketing data. The coarse-grained model identifies new customer data tags (high-potential customers). Marketing promotions are carried out on users identified as high-potential customers. Customers are monitored for click behavior. Those with click behavior are marked as high-potential customers, and those without click behavior are marked as low-potential customers. Then, the sample data required here is obtained through feature processing. For details, please refer to steps 1 to 3).

[0059]

[0060] Values ​​are 0 or 1; : , where N is the number of samples; Initialize to 0.01; Function: High-potential users are given significantly higher weight than low-potential users, alleviating the problem of category imbalance.

[0061] Using the sample mean to illustrate: Adaptive balancing: The weights are dynamically adjusted based on the proportion of positive class samples.

[0062] If positive classes are extremely rare High-potential users Approaching (Significantly increase weight).

[0063] If there are more positive classes ( The weight differences narrowed.

[0064] Compared to traditional methods: Fixed weights cannot adapt to changes in data distribution.

[0065] This method is achieved through... It reflects the degree of data imbalance in real time.

[0066] (3) Improved (Adaptive Fractional-Integer Mixed-Order Gradient Descent) AFIHGD-XGBoost a. The core formula and principle of AFIHGD: Gradient update rule: ; Fractional gradient (calculated by weighting historical gradients to capture long-term dependencies); Traditional integer gradient (fast convergence); Dynamic mixing coefficients (controlling the weights of fractional and integer orders); Adaptive fractional order (adjusted according to gradient changes); Fractional gradient calculation: ; ; The weighted sum of historical gradients, with more recent gradients having larger weights, but distant gradients still retain some influence. Adaptive mechanism: Dynamic order : ; ; The gradient changes drastically ( ), (Enhancing long-term memory of fractional orders) When the gradient is gentle ( ), (Degenerates into a traditional gradient) Mixing coefficient :

[0067] Early training :

[0068] Early training :

[0069] b. Specific implementation in XGBoost The original objective function of XGBoost, and the loss function of traditional XGBoost (second-order Taylor expansion).

[0070] in: (First-order gradient) = (Second-order gradient) AFIHGD gradient modification:

[0071] in: Fractional gradient (historical gradient weighted):

[0072]

[0073] Adaptive parameters:

[0074]

[0075] Node split gain calculation:

[0076] L, R, and I are the core symbols in XGBoost split evaluation, respectively Represents the left and right child nodes and the parent node.

[0077] Split gain measures the improvement of the objective function before and after splitting. AFIHGD modifies gradient calculation to integrate long-term historical information into the splitting strategy.

[0078] c. Comparison with traditional XGBoost To comprehensively evaluate the advantages of AFIHGD-XGBoost in high-potential customer identification and real-time marketing, an in-depth comparison and verification were conducted from three aspects: model accuracy, business conversion, and system response.

[0079] a) Comparison experiment of model accuracy and generalization ability The input data is the input data of Experiment 2 in the feature importance verification experiment (namely: "Age", "Gender", "Star Rating", "Customer Type", "Network Age", "Traffic Preference", "Number of Clicks", "Number of Add-to-Carts", "Number of Logins", "Number of Views", "Mean Number of Clicks", "Variance of Number of Clicks", "Slope of Number of Add-to-Carts", "Change in Number of Logins %", "Weighted Number of Clicks", "Weighted Number of Add-to-Carts", "Weighted Number of Logins", "Login × Purchase", "Add-to-Cart / Views", "High Login Flag", "Behavioral Entropy".

[0080] AFIHGD (Adaptive Fractional-Integer Mixed-Order Gradient Descent) significantly improves the convergence efficiency and generalization ability of models on high-dimensional sparse behavioral data by dynamically adjusting the order and step size of gradient descent. Unlike traditional XGBoost, which uses a fixed order, AFIHGD can adaptively capture the long-term memory effect of user behavior (such as the resurgence of low-frequency but high-value dormant users) while efficiently fitting short-term intensive behavior (such as click bursts during promotions), effectively balancing local optima and global convergence.

[0081] b) Business Implementation Results Actual conversion rate of the top 5% high-potential customer group: Compared with traditional XGBoost, AFIHGD-XGBoost improves the conversion rate of the high-potential customer list identified by 21%, significantly improving marketing ROI.

[0082] High-value user recall performance: Among users with "sudden behavioral changes" (such as suddenly adding high-value items to their cart), the target data recall rate of AFIHGD-XGBoost+ incremental features reached 93%+, which is significantly better than the traditional model (66%).

[0083] Model convergence speed: Under the same number of training rounds, AFIHGD-XGBoost improves convergence speed by 10%, supporting higher frequency model iteration and real-time update requirements.

[0084] c) Real-time incremental characteristics and system response analysis The system's end-to-end response capability is shown in Table 18:

[0085] Table 18 After each key user action (such as click, add to cart, or browse), the system can complete incremental feature updates and refresh the list of high-potential customers within 3 seconds, supporting marketing outreach at the minute or even second level, meeting the core needs of real-time intelligent marketing.

[0086] d. Model output examples and interpretability For binary classification problems, we can output the probability value of a target being a potential customer, thus distinguishing the degree of probability that the target is a high-potential customer. In marketing promotion, marketers can use this high-potential customer probability, combined with behavioral data, to conduct more detailed marketing efforts. An example is shown in Table 19:

[0087] Table 19 5. Customer Service Marketing (1) Hierarchical rules High-value potential customers: Output probability in step 4

[0088] Mid-value potential customers: Output probability in step 4

[0089] Low-value potential customers: Output probability in step 4

[0090] (2) Tiered Marketing Strategy Strategy design example: High-value potential customer group: one-on-one seat marketing.

[0091] Mid-value potential customer group: Combine with package deals and push luxury goods promotional coupons.

[0092] Low-value potential customer group: Plain text push, push frequency = 1 time / 3 days, activation-type benefit push.

[0093] (3) Comparison of marketing results data, as shown in Table 20:

[0094] Table 20 Please refer to Figure 4 The method in this embodiment may include: processing the first preprocessed data based on a first preset XGBoost model to obtain the first row dataset, including: Step S401: Model initialization; Step S402: Parameter specifications are selected; Step S403: Grid search training; Step S404: Obtain the first preset XGBoost model.

[0095] Please refer to Figure 5 The method in this embodiment may include: First, an XGBoost coarse-screening model is trained based on historical marketing data (user profiles, package information, marketing results) to complete the initial screening of potential customers. Second, the behavior of reached users is tracked for 7 days, and incremental features are constructed, including statistical features (mean / variance of clicks, slope of add-to-cart frequency, etc.), double-decay time-weighted features, behavioral cross features, and behavioral entropy features. Then, the XGBoost model is innovatively optimized using the AFIHGD (Adaptive Fractional-Integer Mixed-Order Gradient Descent) algorithm. Its dynamic adjustment of the gradient descent order and step size can adapt to the sparse distribution of behavioral features and suppress data noise through fractional order constraints, enabling the model to respond to behavioral changes in real time and distinguish users with fine granularity. Finally, the system achieves a significant effect of improving CTR by 52.4% (due to the accurate characterization of interest concentration by behavioral entropy features) and reducing conversion cost by 22% (thanks to the dynamic optimization of resource allocation by AFIHGD), forming a set of intelligent marketing potential customer identification methods that combine "coarse-screening" and "fine-screening" and link features and algorithms. The core innovations of this scheme include dynamic incremental feature engineering and AFIHGD (Adaptive Fractional-Integer Mixed-Order Gradient Descent): The "dual decay time window weighted" feature construction method accurately captures the decay characteristics of user behavior through differentiated processing of recent and long-term data. This method innovatively introduces a differentiated weighting mechanism for recent and long-term behavior. By applying different decay weights to the time series data of user behavior, it effectively captures the dynamic process of changes in user interest over time and improves the model's ability to perceive the timeliness of user behavior.

[0096] Develop behavioral entropy features to quantify the uncertainty of user behavior in marketing scenarios: This method is based on entropy theory and quantifies the uncertainty of user behavior and the concentration of interest in different marketing stages, thereby providing the model with a new descriptive dimension about the distribution of user behavior and the focus of interest, which significantly enhances the ability to refine the identification of potential customer interests.

[0097] Explanation: Point 1 focuses on capturing the "dynamic changes in user behavior over time," emphasizing time sensitivity; point 2 focuses on the "dispersion and concentration of user behavior," emphasizing the uncertainty and information content of behavior. Both points, from the complementary perspectives of "time" and "information," comprehensively enhance the ability to express features.

[0098] Constructing multi-dimensional cross-features to achieve deep correlation between browsing, clicking, adding to cart and other behaviors: By performing high-order cross-combinations of various user behavior data such as browsing, clicking, adding to cart, etc., we can explore deep-level correlation features between different behaviors, further enrich the feature space of the model, and improve the ability to identify complex user behavior patterns.

[0099] An adaptive fractional-integer hybrid gradient descent mechanism is proposed to address the local optima and global convergence problems in high-dimensional sparse behavioral data by dynamically adjusting the order and step size of gradient descent. An optimization mechanism for dynamically adjusting the order and step size of XGBoost gradient descent is proposed, enabling the model to adaptively find the optimal convergence path based on high-dimensional sparse behavioral data. This overcomes the local optimum trap and improves the real-time response capability to behavioral changes.

[0100] Note: Point 3 focuses on how to improve the efficiency of information utilization of raw data through "feature engineering," highlighting the deep interaction of features; point 4 focuses on "model optimization algorithms" to improve the model's training and generalization capabilities in high-dimensional data scenarios. Both work synergistically from the feature and algorithm levels respectively to maximize model performance.

[0101] Please refer to Figure 6 The apparatus of this embodiment may include the following structure: Unit 601 is used to acquire the first dataset. Analysis unit 602 is used to preprocess the first dataset to obtain first preprocessed data; process the first preprocessed data based on a first preset XGBoost model to obtain a first row dataset; and process the first row dataset through incremental feature engineering to obtain a first feature dataset. The processing unit 603 is used to obtain customer probability values ​​from the first feature dataset through the second preset XGBoost model, and to identify potential customers based on the customer probability values; wherein the first preset XGBoost model is different from the second preset XGBoost model.

[0102] Please refer to Figure 7 This embodiment of the present application also discloses an electronic device, which includes at least one processor 701, and at least one memory 702 and a bus 703 connected to the processor 701; wherein the processor 701 and the memory 702 communicate with each other through the bus 703; the processor 701 is used to call program instructions in the memory 702 to execute the above-mentioned potential customer identification method.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for identifying potential customers, characterized in that, The method includes: Obtain the first dataset; The first dataset is preprocessed to obtain the first preprocessed data; The first preprocessed data is processed based on the first preset XGBoost model to obtain the first row dataset; The first behavior dataset is processed by incremental feature engineering to obtain the first feature dataset; The first feature dataset is processed by a second preset XGBoost model to obtain customer probability values, and potential customers are identified based on the customer probability values. The first preset XGBoost model is different from the second preset XGBoost model.

2. The method according to claim 1, characterized in that, The incremental feature engineering includes at least: Weighted processing with dual decay time windows; Cross-feature processing; Innovative feature processing.

3. The method according to claim 2, characterized in that, The step of processing the first behavior dataset through incremental feature engineering to obtain the first feature dataset includes: The first row of the dataset is processed using a double decay algorithm to obtain the first interest data; The first interest data is processed by cross-feature processing to obtain the second interest data; The second interest data is processed based on the innovative features to obtain the first feature dataset.

4. The method according to claim 3, characterized in that, The first row of the dataset is processed using a double-decay algorithm to obtain the first interest data, including: The double attenuation algorithm is as follows: ; in, The time difference between when the action occurred and the present moment; Short-term decay coefficient; Long-term attenuation coefficient; : Mixing ratio.

5. The method according to claim 4, characterized in that, Before obtaining customer probability values ​​from the first feature dataset, the second preset XGBoost model includes: The second preset XGBoost model is generated based on the AFIHGD gradient strategy and the XGBoost model.

6. The method according to claim 5, characterized in that, The AFIHGD gradient strategy includes: Gradient update rule: in, Fractional gradient; Traditional integer gradient; : Dynamic mixing coefficient; Adaptive fractional order.

7. The method according to claim 6, characterized in that, The AFIHGD gradient strategy also includes: 。 8. The method according to claim 1, characterized in that, Before processing the first preprocessed data based on the first preset XGBoost model to obtain the first row of dataset, the process includes: Model initialization; Parameter specifications selected; Grid search training; Obtain the first preset XGBoost model.

9. An identification device, characterized in that, The device includes: an acquisition unit, an analysis unit, and a processing unit. The acquisition unit is used to acquire the first dataset; The analysis unit is configured to preprocess the first dataset to obtain first preprocessed data; process the first preprocessed data based on a first preset XGBoost model to obtain a first row dataset; and process the first row dataset through incremental feature engineering to obtain a first feature dataset. The processing unit is used to obtain customer probability values ​​from the first feature dataset using a second preset XGBoost model, and to identify potential customers based on the customer probability values; wherein the first preset XGBoost model is different from the second preset XGBoost model.

10. An electronic device, characterized in that, include: Memory, used to store at least one set of instructions; The processor is used to acquire the first dataset; The first dataset is preprocessed to obtain the first preprocessed data; The first preprocessed data is processed based on the first preset XGBoost model to obtain the first row dataset; The first behavior dataset is processed by incremental feature engineering to obtain the first feature dataset; The first feature dataset is processed by a second preset XGBoost model to obtain customer probability values, and potential customers are identified based on the customer probability values. The first preset XGBoost model is different from the second preset XGBoost model.

Images