User recharge prediction method and device, equipment and storage medium

Abstract

The application relates to the technical field of machine learning, and discloses a user recharge prediction method and device, equipment and a storage medium, the method comprising: associating user multi-platform behavior data through a device fingerprint algorithm, obtaining a user multi-dimensional feature set, determining a clustering number based on an elbow rule, performing clustering analysis on the user multi-dimensional feature set according to the clustering number, generating a user value sub-group label, inputting the user value sub-group label into a random forest model, and obtaining a recharge prediction result of different user sub-groups. The application comprehensively integrates multi-platform user behavior information through a device fingerprint algorithm, breaks data silos, generates a user value sub-group label through elbow rule clustering, and then inputs the label into a random forest model to predict a recharge result. The application not only considers user basic features, but also integrates value sub-group information, enriches feature dimensions, can more accurately depict different value users, and thus improves the accuracy of recharge prediction for different value user groups.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a user recharge prediction method, apparatus, device, and storage medium. Background Technology

[0002] With the advent of the digital economy era, data has become a core resource driving innovation in internet businesses and unlocking commercial value. The widespread adoption of mobile internet has led to an explosive growth in user interaction and behavioral data, providing enterprises with abundant material for accurately understanding user needs and optimizing operational strategies.

[0003] However, the current internet industry still faces challenges in the field of user recharge prediction: on the one hand, the slowdown in user growth rate and the rise in customer acquisition costs are forcing companies to rely more on data-driven refined operations; on the other hand, existing prediction methods are mostly limited to basic transaction data, resulting in a one-sided portrayal of user consumption motivation. Due to insufficient data integration and incomplete feature construction, the recharge prediction results for different value user groups are inaccurate. Summary of the Invention

[0004] The main purpose of this application is to provide a user recharge prediction method, apparatus, device and storage medium, which aims to solve the technical problem that existing user recharge prediction methods are inaccurate in predicting recharge results for different value user groups due to insufficient data integration and incomplete feature construction.

[0005] To achieve the above objectives, this application proposes a user recharge prediction method, the method comprising:

[0006] By using device fingerprinting algorithms to correlate user behavior data across multiple platforms, a multi-dimensional set of user features can be obtained.

[0007] The number of clusters is determined based on the elbow rule, and cluster analysis is performed on the user multidimensional feature set based on the number of clusters to generate user value group labels.

[0008] Input the user value group labels into the random forest model to obtain recharge prediction results for different user groups.

[0009] In one embodiment, the step of determining the number of clusters based on the elbow rule, performing cluster analysis on the user multidimensional feature set based on the number of clusters, and generating user value group labels includes:

[0010] The elbow rule is used to analyze the inflection point of the sum of squared errors of unsupervised clustering algorithms and determine the number of clusters.

[0011] Extract key behavioral features for clustering from the user's multidimensional feature set;

[0012] Based on the number of clusters, the key behavioral features are analyzed collaboratively using the unsupervised clustering algorithm to divide users into multiple groups with similar consumption characteristics.

[0013] Based on the distribution of behavioral characteristics of each group, a corresponding user value tag is added to each group to obtain a user value group tag set.

[0014] In one embodiment, the step of performing collaborative analysis on the key behavioral features based on the cluster number using the unsupervised clustering algorithm to divide users into multiple groups with similar consumption characteristics includes:

[0015] Based on the number of clusters, the weights of different features in the key behavioral features are adjusted according to the predefined user lifecycle stage labels to generate a weighted feature vector.

[0016] Determine the variance of user spending on related platforms and generate a behavioral consistency factor;

[0017] The behavior consistency factor is injected into the weighted feature vector to generate the enhanced feature space;

[0018] The enhanced feature space is processed to generate interference-resistant user group segmentation results;

[0019] By analyzing the internal density of user group segmentation results using the silhouette coefficient, group similarity authentication data is generated.

[0020] Based on the group similarity authentication data, users are divided into multiple groups with similar consumption characteristics.

[0021] In one embodiment, the step of inputting the user value group labels into a random forest model to obtain recharge prediction results for different user groups includes:

[0022] The user value grouping labels are fused with the user multidimensional feature set to construct a comprehensive feature vector;

[0023] The comprehensive feature vector is input into a pre-trained random forest model, and recharge prediction results for each user group are generated through multiple decision trees.

[0024] In one embodiment, the comprehensive feature vector includes group labels, consumption behavior features, and user attribute features. The step of inputting the comprehensive feature vector into a pre-trained random forest model and generating recharge prediction results for each user group through multiple decision trees includes:

[0025] Determine the mutual information entropy between the grouping label and the consumption behavior feature, and generate a feature interaction gain matrix based on the mutual information entropy;

[0026] The voting weights of the decision tree splitting paths are adjusted based on the user lifecycle stage tags to obtain a weighted tree path importance distribution;

[0027] Based on the weighted tree path importance distribution, the feature interaction gain matrix is ​​injected into the random forest model to generate a set of decision trees resistant to interference.

[0028] Based on the decision tree set and user group profile coefficients, recharge prediction results are generated for each user group.

[0029] In one embodiment, the step of obtaining a user's multi-dimensional feature set by associating user behavior data across multiple platforms using a device fingerprinting algorithm includes:

[0030] By using device fingerprinting algorithms to correlate user behavior data across multiple platforms, cross-platform behavior trajectories can be obtained.

[0031] Abnormal transaction records are removed from the behavioral data to generate a clean transaction dataset;

[0032] The city information in the behavioral data is mapped to city level codes to generate structured spatial attributes;

[0033] By integrating the cross-platform behavior trajectories, the purified transaction dataset, and the structured spatial attributes, a spatiotemporal augmented dataset is obtained.

[0034] Based on the spatiotemporal augmentation dataset, multidimensional features of users are extracted to generate a multidimensional feature set.

[0035] In one embodiment, the step of obtaining cross-platform behavioral trajectories by associating user behavior data across multiple platforms using a device fingerprint algorithm includes:

[0036] Based on the device fingerprint algorithm, the device fingerprint identifier is matched with the user's behavior data on multiple platforms to generate a cross-platform user behavior mapping table.

[0037] By comparing the similarity of static attributes of users under the same device fingerprint identifier, cross-platform identity authentication data can be obtained.

[0038] Based on the cross-platform user behavior mapping table and the cross-platform identity authentication data, cross-platform behavior trajectories are obtained.

[0039] Furthermore, to achieve the above objectives, this application also proposes a user recharge prediction device, which includes:

[0040] The data preprocessing module is used to associate user behavior data across multiple platforms using a device fingerprinting algorithm to obtain a multi-dimensional set of user features.

[0041] The user segmentation module is used to determine the number of clusters based on the elbow rule, perform cluster analysis on the user multidimensional feature set according to the number of clusters, and generate user value segmentation labels.

[0042] The consumption prediction module is used to input the user value group labels into the random forest model to obtain the recharge prediction results for different user groups.

[0043] In addition, to achieve the above objectives, this application also proposes a user recharge prediction device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the user recharge prediction method as described above.

[0044] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the user recharge prediction method described above.

[0045] The technical solution proposed in this application includes: using a device fingerprint algorithm to associate user behavior data across multiple platforms to obtain a multi-dimensional user feature set; determining the number of clusters based on the elbow rule; performing cluster analysis on the multi-dimensional user feature set based on the number of clusters to generate user value group labels; and inputting the user value group labels into a random forest model to obtain recharge prediction results for different user groups. This application comprehensively integrates user behavior information from multiple platforms through a device fingerprint algorithm, breaking down data silos. It generates user value group labels through elbow rule clustering and then inputs these labels into a random forest model to predict recharge results. This approach not only considers basic user characteristics but also incorporates value group information, enriching the feature dimensions and enabling more accurate characterization of users with different values, thereby improving the accuracy of recharge prediction for different user groups. Attached Figure Description

[0046] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart illustrating the method for predicting user recharge amounts in this application, provided in Embodiment 1.

[0049] Figure 2A flowchart illustrating the second embodiment of the user recharge prediction method for this application;

[0050] Figure 3 A flowchart illustrating the third embodiment of the user recharge prediction method for this application;

[0051] Figure 4 This is a schematic diagram of the module structure of the user recharge prediction device according to an embodiment of this application;

[0052] Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the user recharge prediction method in this application embodiment.

[0053] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0055] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0056] Existing prediction methods are mostly limited to basic transaction data, resulting in a one-sided portrayal of user consumption motivation. Due to insufficient data integration and incomplete feature construction, the prediction results for recharges of different value user groups are inaccurate.

[0057] Therefore, in order to overcome the above-mentioned defects, this application provides a solution that comprehensively integrates user behavior information from multiple platforms through a device fingerprint algorithm, breaks down data silos, generates user value group labels through elbow clustering, and then inputs them into a random forest model to predict recharge results. This not only considers the basic characteristics of users, but also incorporates value group information, enriching the feature dimensions and enabling more accurate characterization of users with different values, thereby improving the accuracy of recharge prediction for different user groups with different values.

[0058] It should be noted that the executing entity of each embodiment of this application can be a computing service system with data processing, network communication, and program execution functions, such as an electronic system or a user recharge prediction system capable of implementing the above functions. The following description uses a user recharge prediction system (hereinafter referred to as "the system") as an example to illustrate the following embodiments.

[0059] Based on this, embodiments of this application provide a user recharge prediction method, referring to... Figure 1 , Figure 1 A flowchart illustrating the first embodiment of the user recharge prediction method of this application.

[0060] In this embodiment, the user recharge prediction method includes steps S10 to S30:

[0061] Step S10: Use device fingerprinting algorithm to associate user behavior data across multiple platforms to obtain a multi-dimensional feature set of the user.

[0062] It should be noted that device fingerprinting algorithms are a technology that generates unique identifiers by analyzing the hardware characteristics of a user's device (such as device model and operating system version), software environment (such as browser type and plugin information), and network behavior (such as IP address and access time). This algorithm can track user behavior across platforms, for example, by linking a user's browsing history and payment data on e-commerce platforms with their social media interactions to form a feature set covering multiple dimensions such as transaction amount, order frequency, occupation, and address.

[0063] In this step, raw data from different platforms is first collected, and user identities are unified through device fingerprinting technology. Then, static attributes and dynamic behaviors are extracted to generate a structured feature dataset. For example, in terms of static attributes, occupations are extracted and classified into 5 categories and quantified. In terms of dynamic behaviors, the average recharge amount in the past 7 days is calculated through a sliding window, thereby generating the corresponding structured feature dataset.

[0064] It should be noted that, in the above process, data collection and preprocessing includes the following:

[0065] (1) Data Sources: Collect user transaction data (recharge amount, order time, etc.) and static attribute data (occupation, address, etc.) from internet platforms, covering a time span of at least 12 months, with a sample size of no less than 100,000 records. This can be: a) Product outbound sales data from an e-commerce company's ERP outbound system, including 32 fields such as outbound order number, internal order number, unique code (ICCID), store name, payment date, and product name, obtained using Python; b) Recharge order record data from e-commerce platform websites, including operator order details data, with main fields including recharge amount, order time, and recharge method; c) User personal information details data provided by an e-commerce company, mainly including user occupation and address data.

[0066] (2) Data Cleaning: After obtaining the raw data, the time field format is standardized to ensure that the timestamp field can be directly used for subsequent interval calculations. During the cleaning process, records with zero or negative recharge amounts are deleted, abnormal transactions are removed, and user data with a high proportion of missing key information are removed to ensure the representativeness and stability of the analysis. Duplicate circuit card identification code records are merged to ensure the completeness of each user's transaction trajectory. Data cleaning may include: a) handling missing values: deleting records with missing key fields (such as user ID, recharge amount); b) removing outliers: filtering abnormal transactions with recharge amounts of 0 or negative; c) standardization: standardizing the time format to timestamp and mapping the address data to city levels.

[0067] Step S20: Determine the number of clusters based on the elbow rule, and perform cluster analysis on the user multidimensional feature set according to the number of clusters to generate user value group labels.

[0068] It should be noted that the elbow rule determines the optimal value of K by calculating the trend of the sum of squares due to error (SSE) as a function of the number of clusters K.

[0069] Step S30: Input the user value group labels into the random forest model to obtain the recharge prediction results for different user groups.

[0070] It should be noted that the random forest model is an ensemble learning method that constructs multiple decision trees and uses voting to arrive at the final prediction result.

[0071] The system combines user segmentation labels (such as high-value users, users at risk of churn) with standardized features (such as occupational one-hot encoding, sliding window features) as the input feature set. After dividing the system into training and testing sets, a random forest model is trained. In the prediction phase, the model outputs the predicted recharge amount for each group over a future period based on the user segmentation features. For example, the average recharge amount for the high-value user group over the next 30 days.

[0072] This embodiment comprehensively integrates user behavior information from multiple platforms through a device fingerprint algorithm, breaking down data silos. It generates user value group labels through elbow clustering and then inputs them into a random forest model to predict recharge results. This not only considers basic user characteristics but also incorporates value group information, enriching the feature dimensions and enabling more accurate characterization of users with different values, thereby improving the accuracy of recharge prediction for different user groups with different values.

[0073] Furthermore, after obtaining the predicted recharge amounts for each user segment, a cross-platform resource demand mapping model is constructed to convert the predicted values ​​into physical resource scheduling instructions in real time. Specifically: based on the predicted peak recharge amount for high-value user groups over a preset number of days, combined with the concurrent operation density in their historical cross-platform behavior trajectory (such as transaction requests on payment platforms + clicks on social media sharing), the required server resources are dynamically calculated. When a sudden drop in predicted recharge amount for a user group at risk of churn is identified, but activity on social media platforms increases, CDN (Content Delivery Network) edge node expansion is triggered to preload disaster recovery bandwidth resources in the geographically concentrated areas of users.

[0074] In one implementation, step S20 in this embodiment may include: analyzing the inflection point of the sum of squared errors of the unsupervised clustering algorithm using the elbow rule to determine the number of clusters; extracting key behavioral features for clustering from the user multidimensional feature set; performing collaborative analysis on the key behavioral features using the unsupervised clustering algorithm based on the number of clusters to divide users into multiple groups with similar consumption characteristics; and adding corresponding user value tags to each group according to the distribution of behavioral characteristics of each group to obtain a user value grouping tag set.

[0075] In its implementation, the elbow rule calculates the SSE (Solution-Oriented Sequence) for different numbers of clusters K in an unsupervised clustering algorithm (such as K-means) and observes the trend of SSE as K increases. When K is small, SSE decreases rapidly as K increases; when K exceeds a certain critical value, the rate of decrease in SSE slows down significantly, forming an inflection point similar to an "elbow". The K value corresponding to this inflection point is the optimal number of clusters.

[0076] The user's multidimensional feature set includes static attributes (such as occupation and address) and dynamic behaviors (such as consumption frequency and recent consumption time). The system filters features strongly correlated with user consumption behavior based on business needs. For example, it extracts recent consumption time (R), consumption frequency (F), and consumption amount (M) from the RFM model (a tool for measuring customer value and customer benefit generation), or extracts the average recharge amount over the past 7 days from sliding window features, forming a subset containing only key behavioral features to reduce the computational complexity of clustering.

[0077] After determining the number of clusters K, an unsupervised clustering algorithm is used to perform collaborative analysis on key behavioral features. That is, using the distance in the feature space (such as Euclidean distance) as a metric, users with similar features are grouped into the same category, ultimately generating K groups with clear differences in consumption characteristics, such as highly active users and users with low consumption frequency.

[0078] After segmenting users, analyze the distribution of behavioral characteristics of each group (such as average spending amount and spending frequency range). Based on the statistical regularity of the characteristic distribution, define value tags for each group. For example, mark groups with high spending amount and stable frequency as "high-value users" and groups with recent spending time that has been longer and the frequency has decreased as "users at risk of churn". This will ultimately form a set of user value segmentation tags covering all groups.

[0079] This embodiment uses the elbow rule to determine the number of clusters, avoiding inaccurate analysis due to too many or too few clusters. Extracting key behavioral features can reduce data redundancy and improve analysis efficiency. Collaborative analysis can more accurately classify users into similar groups, and adding value tags can intuitively present user value.

[0080] As one implementation method, the above-mentioned group segmentation steps in this embodiment may include: based on the number of clusters, adjusting the weights of different features in key behavioral features according to predefined user lifecycle stage labels to generate a weighted feature vector; determining the user's consumption variance on associated platforms to generate a behavioral consistency factor; injecting the behavioral consistency factor into the weighted feature vector to generate an enhanced feature space; processing the enhanced feature space to generate an interference-resistant user group segmentation result; analyzing the internal density of the user group segmentation result using silhouette coefficients to generate group similarity authentication data; and dividing users into multiple groups with similar consumption characteristics based on the group similarity authentication data.

[0081] In practice, the weights of key behavioral features are dynamically adjusted based on predefined labels for user lifecycle stages (such as new users, mature users, and declining users). Differentiated weights are assigned to features (such as recent purchase time R and purchase frequency F) for different lifecycle stages. For example, users in the declining stage pay more attention to the weight of R, while users in the mature stage focus on the weight of F, ultimately generating a weighted feature vector that reflects the characteristics of the user stage.

[0082] Furthermore, by calculating the variance of user spending amounts or frequencies across multiple platforms (such as e-commerce and payment), the cross-platform consistency of user behavior is quantified. User spending data from each platform is extracted, and variance values ​​are calculated. Smaller variance indicates more stable behavior, generating a behavior consistency factor in the range of 0-1. This consistency factor is injected into the feature vector through linear weighting or non-linear mapping, for example, by multiplying the weighted F feature by the consistency factor, generating a new feature space that considers both user stage and cross-platform behavior. The enhanced feature space is then denoised and standardized to reduce the interference of outliers or missing values ​​on the clustering results. Z-score standardization or quantile-based truncation methods can be used to eliminate differences in feature dimensions and filter outout points, ensuring that the clustering algorithm generates robust group partitioning results in a stable data environment.

[0083] By calculating the silhouette coefficient of samples in each user group, the intra-group density and inter-group separation are assessed. For each sample, the average distance (a) to other samples in the same group and the average distance (b) to the nearest sample in another group are calculated. The silhouette coefficient = (ba) / max(a,b). The closer the value is to 1, the more reasonable the group segmentation is, generating group similarity authentication data. Based on the group similarity authentication data, groups with silhouette coefficients higher than a threshold are filtered out, and subgroups with too low similarity are merged to ensure that each group has clear differences in consumption characteristics (such as users with high consumption frequency but low amount, and users with low frequency but high amount), thus completing the final user segmentation.

[0084] This embodiment improves clustering accuracy by adjusting feature weights to better match the characteristics of users in different value groups. The behavioral consistency factor reflects the stability of user consumption and enhances feature distinguishability. Anti-interference processing avoids the influence of abnormal data. Silhouette coefficient analysis ensures that the internal group division is tight and the external group division is separate, making the group division more reasonable.

[0085] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. On this basis, refer to Figure 2 Step S30 may include steps S301 to S302:

[0086] Step S301: The user value grouping label and the user multidimensional feature set are fused to construct a comprehensive feature vector.

[0087] It should be understood that horizontally concatenating user value segmentation labels with user multidimensional feature sets can transform the segmentation labels into numerical features through label encoding (such as one-hot encoding), and then merging them with the original multidimensional features column by column to form a comprehensive feature vector containing user value categories and behavioral attributes. For example, the label "high-value user" can be encoded as [1,0,0], which, together with consumption frequency and occupational characteristics, constitutes a new feature input.

[0088] Step S302: Input the comprehensive feature vector into the pre-trained random forest model, and generate recharge prediction results for each user group through multiple decision trees.

[0089] Furthermore, the comprehensive feature vector is input into the trained random forest model, which consists of multiple decision trees through ensemble learning. Each decision tree makes independent predictions based on the group labels and behavioral features in the feature vector. Finally, the prediction results of all trees are aggregated by voting or mean to generate a deterministic recharge prediction value for each user group.

[0090] This embodiment constructs a comprehensive vector by fusing labels and features, which can make full use of user grouping information and multi-dimensional behavioral data to improve feature richness.

[0091] The comprehensive feature vector includes group labels, consumption behavior features, and user attribute features. As one implementation, step S302 in this embodiment may include: determining the mutual information entropy between the group labels and the consumption behavior features, and generating a feature interaction gain matrix based on the mutual information entropy; adjusting the voting weights of the decision tree splitting paths according to the user lifecycle stage labels to obtain a weighted tree path importance distribution; injecting the feature interaction gain matrix into a random forest model based on the weighted tree path importance distribution to generate an interference-resistant decision tree set; and generating recharge prediction results for each user group based on the decision tree set and the user group profile coefficients.

[0092] In practice, the strength of the association between the cluster labels and consumer behavior features can be quantified by calculating the mutual information entropy between them. Information theory methods are used to calculate the mutual information value between the cluster labels and each consumer behavior feature; a larger value indicates a stronger explanatory power of the feature for the cluster. A feature interaction gain matrix is ​​then constructed, with the matrix elements representing the combined mutual information values ​​between features.

[0093] Furthermore, the system dynamically adjusts the voting weights of decision tree splitting paths in the random forest by incorporating user lifecycle stage tags. Differentiated weights are assigned to splitting paths for different lifecycle stages, and the weights of all decision tree splitting paths are statistically analyzed to generate a weighted tree path importance distribution, which is used to optimize the model's sensitivity to lifecycle features. During decision tree growth, high-correlation feature combinations in the feature interaction gain matrix are prioritized for splitting. Simultaneously, the splitting threshold is adjusted based on the weighted path importance to suppress interference from low-value or noisy features. Ultimately, a set of decision trees containing high-value feature interactions is generated, improving the model's adaptability to complex data scenarios.

[0094] Then, combining the prediction results of the decision tree ensemble with the user group profile coefficients, the recharge prediction value for each user group is output. That is, firstly, each user group is independently predicted using the decision tree ensemble to generate an initial prediction result, and then the user group profile coefficient (reflecting the density and separation within the group) is introduced to weight and correct the prediction result. This generates a recharge prediction result that takes into account both feature interactions and group consistency.

[0095] This embodiment mines potential correlations between features by using mutual information entropy and feature interaction gain matrix to improve feature utilization efficiency, adjusts decision tree voting weights to consider differences in user lifecycles to make the model more in line with changes in user behavior, improves model stability by using anti-interference decision tree set, and generates more accurate prediction results by combining silhouette coefficient.

[0096] Based on the first embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 Step S10 may include steps S101 to S105:

[0097] Step S101: Use device fingerprinting algorithm to associate user behavior data across multiple platforms to obtain cross-platform behavior trajectories.

[0098] It should be understood that the system uses device fingerprinting algorithms to correlate and match user behavior data across different platforms. That is, by using the device ID as a unique identifier, it integrates behavioral data such as login time, operation type, and page access path from various platforms to form a sequence of user behavior trajectories across platforms. For example, it records the entire process of user A browsing products on an e-commerce platform, sharing links on a social platform, and finally completing the transaction on a payment platform.

[0099] Step S102: Remove abnormal transaction records from the behavioral data to generate a clean transaction dataset.

[0100] Furthermore, in the process of eliminating abnormal transaction records, a deep verification based on the spatiotemporal consistency features of users' cross-platform behavior trajectories is introduced. Specifically, a spatiotemporal graph convolutional neural network can be used to place the transaction to be judged within the user's complete cross-platform behavior trajectory sequence for contextual analysis. The system calculates and verifies whether the spatiotemporal features of the transaction at the time of occurrence (such as the speed of geographical location switching and the coherence of cross-platform operation logic) are consistent with the user's historical normal behavior patterns in the spatiotemporal dimension. For example, if the geographical location of a transaction jumps a physically impossible long distance in a very short period of time, or if the platform operation sequence triggered by the transaction seriously violates the user's inherent cross-platform behavior habits, such as a sudden high frequency of small-amount transactions on a payment platform, but no related browsing or interaction records on e-commerce or social platforms during the same period, even if its single feature does not trigger the traditional abnormal threshold, the system will judge it as an abnormal transaction and eliminate it based on this significant break in spatiotemporal behavior logic, in order to overcome the blind spot of traditional single-point detection methods in identifying new, complex, or deliberately imitative fraudulent behaviors.

[0101] It should be understood that when applying spatiotemporal graph convolutional neural networks for abnormal transaction filtering, the system first constructs user cross-platform behavioral trajectories (such as e-commerce browsing, social sharing, and payment operations) as spatiotemporal graph nodes, and calculates the connection weights between nodes based on the time interval and geographical distance of the behavior. The system aggregates behavioral features of adjacent platform nodes through spatial graph convolutional layers (e.g., the conflict between the high-speed rail time from Beijing to Shanghai and the 13-minute operation interval), and analyzes the dynamic patterns of the operation sequences through temporal convolutional layers. Finally, through multi-layer spatiotemporal feature fusion, the system outputs an abnormal probability score for the transaction to be verified (e.g., high-frequency transfers on payment platforms are judged to have an abnormal probability of 92% due to violations of physical movement logic). Based on this score, the system directly removes the transaction record, generating a cleaned dataset that retains only operations with consistent spatiotemporal behavioral logic.

[0102] Furthermore, hard features extracted based on preset rules (such as transaction amount exceeding limits or time intervals being too short), soft anomaly probability scores output by the anomaly detection model, and the aforementioned cross-platform spatiotemporal consistency verification results are all mapped to a unified multi-dimensional feature space. In this space, based on the specific scenario of the current transaction (such as platform type, time period, and user's historical risk level) and the confidence levels between different types of features, the weights of different feature dimensions in the final decision are dynamically adjusted in real time, thereby filtering out abnormal transaction records multiple times.

[0103] Step S103: Map the city information in the behavioral data to the city level code to generate structured spatial attributes.

[0104] The city names or IP location information in user behavior data are mapped to a predefined city-level coding system (e.g., first-tier cities are coded as 1, second-tier cities as 2, and so on). City-level labels are obtained through geographic information databases or third-party APIs, transforming unstructured city information into structured numerical codes. For example, "Beijing" is mapped to "1" and "Wuhan" is mapped to "2", facilitating subsequent correlation analysis between spatial features and consumption behavior.

[0105] Step S104: Integrate the cross-platform behavior trajectory, the purified transaction dataset, and the structured spatial attributes to obtain a spatiotemporal augmented dataset.

[0106] Based on the timeline, user operation records and transaction data on different platforms are arranged in chronological order and associated with spatial attributes such as city level codes to generate a spatiotemporal augmented dataset containing four-dimensional information of "time-platform-behavior-space". For example, user A completed a transaction on an e-commerce platform in a first-tier city (code 1) at 14:00 on October 1, 2023.

[0107] Step S105: Extract multidimensional features of the user based on the spatiotemporal augmentation dataset to generate a multidimensional feature set.

[0108] Multidimensional features of users are extracted from the spatiotemporal augmented dataset, covering statistical, temporal, and spatial features. Statistical features may include transaction frequency and average amount; temporal features may include the most recent transaction time and transaction interval volatility; spatial features may include the resident city level and cross-city transaction frequency. All features are concatenated into a multidimensional feature vector. For example, user A's feature vector may contain [transaction frequency = 5 times / month, average amount = 200 yuan, most recent transaction time = 2023-10-01, resident city level = 1], forming a complete set of user multidimensional features.

[0109] This embodiment comprehensively presents user behavior paths through cross-platform behavior trajectories, eliminates abnormal transaction records to ensure data quality, makes the data more analyzable through structured spatial attributes, integrates information from multiple aspects through spatiotemporal augmented datasets, and extracts multidimensional features to more accurately characterize users.

[0110] As one implementation method, step S101 in this embodiment may include: matching the device fingerprint identifier with the user's behavior data on multiple platforms based on the device fingerprint algorithm to generate a cross-platform user behavior mapping table; comparing the similarity of static attributes of users under the same device fingerprint identifier to obtain cross-platform identity authentication data; and obtaining a cross-platform behavior trajectory based on the cross-platform user behavior mapping table and the cross-platform identity authentication data.

[0111] In the specific implementation, a mapping relationship table of "device ID-platform-behavioral data" is constructed using the device fingerprint identifier as an index. For example, if a user with device ID "DEV123" has browsing records on an e-commerce platform, likes on a social platform, and transaction records on a payment platform, all of these are associated with the same user through this ID, forming a cross-platform user behavior mapping table.

[0112] For multi-platform behavioral data associated with the same device fingerprint identifier, user static attributes (such as registered mobile phone number, email address, ID card number, and shipping address) are extracted, and the consistency of these attributes across different platforms is compared using similarity algorithms (such as string matching, edit distance, and Jaccard similarity coefficient). If the user's static attributes (such as mobile phone number and address) under the same device ID are highly similar or completely consistent across multiple platforms, they are determined to be the same user, and a cross-platform identity authentication label (such as "authentication passed" or "authentication score = 0.9") is generated. If the attribute differences are large, it is marked as "pending confirmation" or the authentication score is reduced, ultimately forming a cross-platform identity authentication dataset.

[0113] Furthermore, the system combines cross-platform user behavior mapping tables with cross-platform identity authentication data to filter and integrate reliable user behavior records, generating complete cross-platform behavior trajectories. Only behavior records from the identity authentication data that are "authenticated successfully" or have an authentication score higher than a threshold are retained. The same user's operations on different platforms are arranged in chronological order (e.g., browsing on an e-commerce platform first, sharing on a social media platform, and finally making a transaction on a payment platform), forming a time-series behavior trajectory that includes "time-platform-behavior type-device ID". For example, user A browses products on an e-commerce platform at 10:00 on October 1, 2023, shares a link on a social media platform at 10:30, and completes a transaction on a payment platform at 11:00. These three events are linked into a continuous trajectory through the device ID "DEV123".

[0114] This embodiment uses device fingerprinting algorithm matching to accurately associate user behavior across multiple platforms, breaking down data silos. It compares the similarity of static attributes for identity authentication, ensuring data accuracy and avoiding data confusion. The obtained cross-platform behavior trajectory can comprehensively reflect the user's behavioral habits and consumption preferences on different platforms.

[0115] The user recharge prediction method in this application has the following advantages:

[0116] 1. Multi-dimensional User Profile: By integrating dynamic RFM behavioral characteristics with static attributes such as occupation and city tier, the constructed user profile can more comprehensively depict user value. Experimental data shows that high-value user groups account for more than 50% of consumption. Accurately identifying such users can provide a clear basis for enterprises to focus their resources, improving the user identification accuracy by more than 40% compared to traditional single-dimensional analysis.

[0117] 2. Refined User Segmentation: Combining the RFM model and K-means clustering algorithm, users are segmented into four groups, each with significantly different characteristics. Among them, users at risk of churn account for 9.2%, exhibiting a clear characteristic of having a longer recent purchase time, facilitating targeted recall strategies. Calculations show that this can improve the retention rate of this user group by 25%-30%.

[0118] 3. High-precision prediction: By comparing three models, Random Forest, GBDT, and LightGBM, Random Forest was found to perform best in the Internet recharge prediction scenario. Its RMSE value is 43.80 and MAE value is 12.89, which reduces the prediction error by 10%-15% compared with other models. It can more accurately predict the amount of recharge by users in the future, providing reliable data support for enterprises' inventory management, marketing budget allocation, etc.

[0119] 4. Standardized Process: A complete standardized process is established from data cleaning and feature engineering to model training and result application. It can handle large-scale datasets containing 1 million records and has flexible handling of missing and outlier values, enabling the system to adapt to the sales forecasting needs of different Internet platforms (such as e-commerce and digital content platforms), and has strong versatility and scalability.

[0120] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the user recharge prediction method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0121] This application also provides a user recharge prediction device, please refer to... Figure 4 The user recharge prediction device includes:

[0122] Data preprocessing module 10 is used to associate user behavior data across multiple platforms using a device fingerprint algorithm to obtain a set of multi-dimensional user features;

[0123] User segmentation module 20 is used to determine the number of clusters based on the elbow rule, perform cluster analysis on the user multidimensional feature set according to the number of clusters, and generate user value segmentation labels.

[0124] The consumption prediction module 30 is used to input the user value group labels into the random forest model to obtain the recharge prediction results for different user groups.

[0125] This application provides a user recharge prediction device, for reference... Figure 5 This diagram illustrates a structural schematic suitable for implementing a user recharge prediction device according to embodiments of this application. The user recharge prediction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the user recharge prediction device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. The communication device 1009 allows the user recharge prediction device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows user recharge prediction devices with various systems, it should be understood that it is not required to implement or possess all of the systems shown. More or fewer systems may be implemented alternatively.

[0126] 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.

[0127] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the user recharge prediction method in the above embodiments.

[0128] The aforementioned computer-readable storage medium may be included in the user recharge prediction device; or it may exist independently and not assembled into the user recharge prediction device.

[0129] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the content of this application specification and drawings, or direct / indirect applications in other related technical fields, are included in the scope of protection of this application.

Claims

1. A method for predicting user recharge, characterized in that, The user recharge prediction method includes: By using device fingerprinting algorithms to correlate user behavior data across multiple platforms, a multi-dimensional set of user features can be obtained. The number of clusters is determined based on the elbow rule, and cluster analysis is performed on the user multidimensional feature set based on the number of clusters to generate user value group labels. Input the user value group labels into the random forest model to obtain the recharge prediction results for different user groups; The steps of determining the number of clusters based on the elbow rule, performing cluster analysis on the user multidimensional feature set based on the number of clusters, and generating user value group labels include: The elbow rule is used to analyze the inflection point of the sum of squared errors of unsupervised clustering algorithms and determine the number of clusters. Extract key behavioral features for clustering from the user's multidimensional feature set; Based on the number of clusters, the key behavioral features are analyzed collaboratively using the unsupervised clustering algorithm to divide users into multiple groups with similar consumption characteristics. Based on the distribution of behavioral characteristics of each group, add corresponding user value tags to each group to obtain a set of user value group tags; The step of performing collaborative analysis on the key behavioral features based on the cluster number using the unsupervised clustering algorithm to divide users into multiple groups with similar consumption characteristics includes: Based on the number of clusters, the weights of different features in the key behavioral features are adjusted according to the predefined user lifecycle stage labels to generate a weighted feature vector. Determine the variance of user spending on related platforms and generate a behavioral consistency factor; The behavior consistency factor is injected into the weighted feature vector to generate the enhanced feature space; The enhanced feature space is processed to generate interference-resistant user group segmentation results; By analyzing the internal density of user group segmentation results using the silhouette coefficient, group similarity authentication data is generated. Based on the group similarity authentication data, users are divided into multiple groups with similar consumption characteristics.

2. The user recharge prediction method as described in claim 1, characterized in that, The step of inputting the user value group labels into a random forest model to obtain recharge prediction results for different user groups includes: The user value grouping labels are fused with the user multidimensional feature set to construct a comprehensive feature vector; The comprehensive feature vector is input into a pre-trained random forest model, and recharge prediction results for each user group are generated through multiple decision trees.

3. The user recharge prediction method as described in claim 2, characterized in that, The comprehensive feature vector includes group labels, consumption behavior features, and user attribute features. The step of inputting the comprehensive feature vector into a pre-trained random forest model and generating recharge prediction results for each user group through multiple decision trees includes: Determine the mutual information entropy between the grouping label and the consumption behavior feature, and generate a feature interaction gain matrix based on the mutual information entropy; The voting weights of the decision tree splitting paths are adjusted based on the user lifecycle stage tags to obtain a weighted tree path importance distribution; Based on the weighted tree path importance distribution, the feature interaction gain matrix is ​​injected into the random forest model to generate a set of decision trees resistant to interference. Based on the decision tree set and user group profile coefficients, recharge prediction results are generated for each user group.

4. The user recharge prediction method as described in claim 1, characterized in that, The step of obtaining a user's multi-dimensional feature set by associating user behavior data across multiple platforms using a device fingerprinting algorithm includes: By using device fingerprinting algorithms to correlate user behavior data across multiple platforms, cross-platform behavior trajectories can be obtained. Abnormal transaction records are removed from the behavioral data to generate a clean transaction dataset; The city information in the behavioral data is mapped to city level codes to generate structured spatial attributes; By integrating the cross-platform behavior trajectories, the purified transaction dataset, and the structured spatial attributes, a spatiotemporal augmented dataset is obtained. Based on the spatiotemporal augmentation dataset, multidimensional features of users are extracted to generate a multidimensional feature set.

5. The user recharge prediction method as described in claim 4, characterized in that, The step of obtaining cross-platform behavior trajectories by associating user behavior data across multiple platforms using a device fingerprint algorithm includes: Based on the device fingerprint algorithm, the device fingerprint identifier is matched with the user's behavior data on multiple platforms to generate a cross-platform user behavior mapping table. By comparing the similarity of static attributes of users under the same device fingerprint identifier, cross-platform identity authentication data can be obtained. Based on the cross-platform user behavior mapping table and the cross-platform identity authentication data, cross-platform behavior trajectories are obtained.

6. A user recharge prediction device, characterized in that, The user recharge prediction device includes: The data preprocessing module is used to associate user behavior data across multiple platforms using a device fingerprinting algorithm to obtain a multi-dimensional set of user features. The user segmentation module is used to determine the number of clusters based on the elbow rule, perform cluster analysis on the user multidimensional feature set according to the number of clusters, and generate user value segmentation labels. The consumption prediction module is used to input the user value group labels into the random forest model to obtain the recharge prediction results for different user groups. The user segmentation module is further used to analyze the inflection point of the sum of squared errors of the unsupervised clustering algorithm using the elbow rule to determine the number of clusters; extract key behavioral features for clustering from the user multidimensional feature set; based on the number of clusters, perform collaborative analysis on the key behavioral features using the unsupervised clustering algorithm to divide users into multiple groups with similar consumption characteristics; and add corresponding user value tags to each group according to the distribution of behavioral characteristics of each group to obtain a user value segmentation tag set. The user segmentation module is further configured to: adjust the weights of different features in key behavioral features based on the number of clusters and predefined user lifecycle stage labels to generate a weighted feature vector; determine the user's consumption variance on associated platforms and generate a behavioral consistency factor; inject the behavioral consistency factor into the weighted feature vector to generate an enhanced feature space; process the enhanced feature space to generate interference-resistant user group segmentation results; analyze the internal density of the user group segmentation results using silhouette coefficients to generate group similarity authentication data; and classify users into multiple groups with similar consumption characteristics based on the group similarity authentication data.

7. A user recharge prediction device, characterized in that, The user recharge prediction device includes: a memory, a processor, and a user recharge prediction program stored in the memory and executable on the processor. When the user recharge prediction program is executed by the processor, it implements the user recharge prediction method as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium stores a user recharge prediction program, which, when executed by a processor, implements the user recharge prediction method as described in any one of claims 1 to 5.

Images