A multi-dimensional data-driven method for joint prediction of subscription service user churn risk and value

Abstract

This invention relates to the field of data processing and prediction technology, specifically a multi-dimensional data-driven method for jointly predicting the risk and value of churn in subscription services. It aims to address the problems of data sparsity and high-dimensional feature conflicts, as well as interference from multiple task objectives, present in existing methods for jointly predicting the risk and value of churn in subscription services. This method collects and preprocesses multi-dimensional data from subscription service users, constructs a feature extraction model integrating federated learning and dynamic knowledge graphs to handle the data sparsity and high-dimensional feature conflicts, and constructs a joint prediction model combining causal reasoning and a dynamic weighting mechanism of reinforcement learning to address the interference from multiple task objectives. The constructed model is then used for training and prediction, outputting the predicted results of user churn risk and value. This invention improves the accuracy and reliability of jointly predicting the risk and value of churn in subscription services, providing strong support for the operational decisions of subscription service providers.

Description

Technical Field

[0001] This invention relates to the field of data processing and prediction technology, specifically a multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services. Background Technology

[0002] In today's digital economy, subscription services have become a crucial business model for many industries, such as video platforms, music software, online education, and cloud services. For subscription service providers, accurately predicting churn risk and assessing user value are of paramount importance. By predicting churn risk, providers can take targeted retention measures in advance to reduce churn rates; while assessing user value helps providers optimize resource allocation, providing better services to high-value users and identifying potential high-value users.

[0003] However, existing methods for jointly predicting the risk and value of churn in subscription services suffer from numerous problems, limiting the accuracy and reliability of the predictions. Among these, the conflict between data sparsity and high-dimensional features is a particularly prominent issue. Subscription service user behavior data is often sparse; for example, some users may have only performed a limited number of interactions, resulting in insufficient effective information. Simultaneously, user data has many dimensions, including basic information, behavioral data, and content preferences. Complex correlations and conflicts may exist between high-dimensional features, making it difficult for traditional data processing and feature extraction methods to effectively utilize this data, thus affecting prediction performance. Current technologies often employ simple interpolation methods to address data sparsity, utilizing less than 30% of user-content interaction data. Furthermore, fixed weight allocation (e.g., a 1:1 weight ratio) in multi-task learning leads to a prediction error rate as high as 25% for high-value user churn risk.

[0004] Furthermore, interference from multiple task objectives is a major challenge faced by existing methods. User churn risk prediction and user value assessment are two related but distinct tasks with different objectives and evaluation criteria. During joint prediction, there may be conflicting objectives between the two tasks. For example, high-value users may experience increased churn risk due to certain factors. Traditional joint prediction models often employ fixed weight allocation methods, failing to dynamically balance the importance of the two tasks. This can negatively impact the predictive performance of one task, thereby reducing the overall effectiveness of the joint prediction.

[0005] Currently, while some methods attempt to address the aforementioned issues, most employ conventional techniques such as simple data aggregation and fixed-weight multi-task learning, which fail to fundamentally overcome existing bottlenecks. Therefore, a novel, multi-dimensional, data-driven method for jointly predicting subscription service user churn risk and value is urgently needed to resolve issues such as data sparsity and high-dimensional feature conflicts, as well as interference from multi-task objectives, thereby improving the accuracy and reliability of predictions and providing more effective decision support for subscription service providers. Summary of the Invention

[0006] This invention aims to address the problems of data sparsity and high-dimensional feature conflict, as well as interference from multiple task objectives, in existing methods for jointly predicting the risk and value of user churn in subscription services. Instead, it provides a multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services. This method improves data utilization to over 60% through a ternary association network of dynamic knowledge graphs and reduces the error rate to below 10% through a dynamic weighting mechanism of reinforcement learning.

[0007] The technical solution adopted by this invention to solve its technical problem is: a multi-dimensional data-driven method for jointly predicting the risk and value of subscription service user churn, comprising the following steps:

[0008] S1. Collect multidimensional data from users who subscribe to the service, and preprocess the multidimensional data;

[0009] S2. Construct a feature extraction model, including a horizontal federated learning framework and a dynamic knowledge graph; the horizontal federated learning framework uses homomorphic encryption technology to encrypt and transmit parameters, and the dynamic knowledge graph constructs a user-content-behavior ternary association network through entity recognition and performs incremental updates based on real-time user interaction data;

[0010] S3. Train a joint prediction model, which introduces causal inference and reinforcement learning dynamic weighting mechanism. The causal inference is used to identify the causal path between user behavior and churn risk and user value. The reinforcement learning dynamic weighting mechanism uses the churn risk prediction accuracy and the mean absolute error of user value prediction as reward signals and adjusts the weight allocation in real time through the PPO algorithm.

[0011] S4. Input the preprocessed multidimensional data into the feature extraction model to extract features and obtain user features;

[0012] S5. Input the user features into the joint prediction model for training to obtain the trained joint prediction model;

[0013] S6. After preprocessing and feature extraction, the multidimensional data of the users to be predicted is input into the trained joint prediction model, and the prediction results of the churn risk and value of the users to be predicted are output.

[0014] Specifically, in step S1, the multidimensional data includes at least one of user basic information, behavioral data, content interaction data, and social relationship data.

[0015] Specifically, in step S1, the preprocessing includes at least one of data cleaning, data standardization, missing value handling, and outlier handling.

[0016] Specifically, in step S2, the federated learning adopts a horizontal federated learning framework, where each channel trains a feature extractor locally and exchanges intermediate parameters through a blockchain consensus mechanism; the blockchain consensus mechanism adopts a practical Byzantine fault-tolerant algorithm, and homomorphic encryption technology is used to encrypt the parameter exchange.

[0017] Specifically, in step S2, the construction of the dynamic knowledge graph includes entity recognition, relationship extraction, and graph update. The entities include users, content, and behavior, and the relationships include the interaction relationship between users and content, and the social relationship between users.

[0018] Specifically, in step S3, the causal reasoning adopts the counterfactual reasoning method, which simulates the impact of endogenous and exogenous factors on churn risk and user value through intervention; the counterfactual reasoning simulates the changing trend of churn risk by intervening in user behavior characteristics, such as changing 'not logging in for 7 consecutive days' to 'log in daily'.

[0019] Specifically, in step S3, the reinforcement learning dynamic weighting mechanism uses the churn risk prediction accuracy and user value prediction error as reward signals, and adjusts the weights of the two tasks through a reinforcement learning algorithm.

[0020] Specifically, in step S5, the training of the joint prediction model adopts the gradient descent algorithm to optimize the model parameters by minimizing the prediction loss function.

[0021] Specifically, in step S6, the predicted result of the churn risk is the churn probability, and the predicted result of the user value is the potential value assessment value of the user within a certain period.

[0022] Specifically, the subscription service includes at least one of video subscription, music subscription, reading subscription, and software service subscription.

[0023] The beneficial effects of this invention are:

[0024] (1) The multidimensional data-driven subscription service user churn risk and value prediction method described in this invention uses federated learning and dynamic knowledge graph fusion technology to construct a feature extraction model. Federated learning realizes the secure collaborative use of cross-channel data, avoids the privacy leakage risk brought about by data centralization, and makes full use of scattered multidimensional data. Dynamic knowledge graph constructs a user-content-behavior association network, which, combined with graph attention mechanism, can effectively mine potential features in sparse data, solve the problem of data sparsity and high-dimensional feature conflict, and improve the effectiveness of feature extraction.

[0025] (2) The multi-dimensional data-driven subscription service user churn risk and value prediction method described in this invention introduces causal reasoning and reinforcement learning dynamic weight mechanism to construct a joint prediction model. Causal reasoning identifies the real causal relationship between user behavior and prediction target, improving the reliability of prediction. The reinforcement learning dynamic weight mechanism can adjust the weight in real time according to the prediction effect of the task, solving the problem of interference between multiple task targets, enabling churn risk prediction and user value prediction to promote each other and improve the overall performance of joint prediction.

[0026] (3) The multidimensional data-driven method for predicting the risk and value of user churn in subscription services described in this invention is based on multidimensional data-driven methods and comprehensively considers various types of user information, making the prediction results more comprehensive and accurate. It provides a strong basis for subscription service providers to formulate targeted operational strategies and helps to improve user retention rate and user value.

[0027] (4) On the video subscription service dataset, the churn risk prediction accuracy of the present invention reaches 89.2%, which is 12.3% higher than the existing multi-task learning methods; the MAE of user value prediction is 0.08, which is 35.7% lower than the traditional regression method. In commercial applications, it can help subscription service providers reduce user churn rate by 15%-20% and increase annual revenue by more than 10%. Attached Figure Description

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] Figure 1 A flowchart illustrating a multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services, as provided by this invention.

[0030] Figure 2 The feature extraction model architecture diagram in the multi-dimensional data-driven subscription service user churn risk and value joint prediction method provided by the present invention is shown below.

[0031] Figure 3A schematic diagram of the joint prediction model mechanism in the multi-dimensional data-driven subscription service user churn risk and value joint prediction method provided by the present invention;

[0032] Figure 4 The flowchart illustrates the data preprocessing process in a multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services, as provided in this invention. Detailed Implementation

[0033] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0034] like Figure 1-4 As shown, the multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services, as described in this invention, includes the following steps:

[0035] S1. Collect multidimensional data from users who subscribe to the service, and preprocess the multidimensional data;

[0036] Specifically, the collected multidimensional data includes, but is not limited to, basic user information (such as age, gender, and region), behavioral data (such as login frequency, browsing duration, and operation records), content interaction data (such as videos watched, music listened to, and articles read), and social relationship data (such as friend lists and interaction records). Preprocessing includes data cleaning to remove invalid and duplicate data; data standardization to convert data of different magnitudes to a uniform scale; missing value handling using mean imputation, median imputation, or context-based imputation methods; and outlier handling, identifying and processing abnormal data through statistical analysis.

[0037] The criteria for invalid data are user data that has no interaction behavior for three consecutive months; missing values ​​are filled in context by the user's social relationship chain, specifically by replacing the missing interaction data with the average behavioral characteristics of the user's top 5 friends; outliers are identified by the Z-score method, and when |Z|>3, they are judged as outliers and replaced with the 95th percentile value.

[0038] S2. Construct a feature extraction model. The feature extraction model adopts a fusion technology of federated learning and dynamic knowledge graph. The federated learning is used to realize local training and parameter exchange of multi-dimensional data across channels. The dynamic knowledge graph is used to construct a user-content-behavior association network and mine potential features in the association network through graph attention mechanism.

[0039] Federated learning employs a horizontal federated learning framework, where each channel (such as different platform clients or different service regions) trains its local feature extractor on the collected data. Intermediate parameters (such as gradient information) are securely exchanged through a blockchain consensus mechanism, enabling collaborative utilization of cross-channel data while protecting data privacy. The parameter exchange process uses homomorphic encryption to ensure data privacy and security during parameter transmission. The blockchain consensus mechanism specifically employs the Practical Byzantine Fault Tolerance (PBFT) algorithm, suitable for efficient consensus verification in cross-channel scenarios.

[0040] The construction of a dynamic knowledge graph first involves entity recognition, identifying entities such as users, content, and behaviors. Then, relationship extraction is performed to identify relationships between entities, such as the "viewing" and "listening" relationships between users and content, and the "friends" and "interactions" relationships between users. The knowledge graph is updated in real time based on newly generated data. A graph attention mechanism is used to calculate attention weights between entities, focusing on associations that significantly impact prediction, thereby mining potential effective features from sparse, high-dimensional data. Graph updates are triggered when user behavior data increments exceed 10% or are automatically executed daily at midnight. The weights of the graph attention mechanism are calculated using a formula. Calculation, where and Let be the node feature vector, and d be the feature dimension, used to determine the attention weights between user and content nodes. The attention weight between node i and node j quantifies the importance of their relationship. It normalizes the vector of dimension j so that the sum of the weights is 1. It is a leaky ReLU activation function. It is a trainable weight matrix with dimensions [2d, k] (d is... The dimension of is k, which is the dimension after linear transformation (a hyperparameter), used to extract key information of the spliced ​​features.

[0041] S3. Construct a joint prediction model, which introduces causal reasoning and a reinforcement learning dynamic weighting mechanism. The causal reasoning is used to identify the causal path between user behavior and churn risk and user value. The reinforcement learning dynamic weighting mechanism is used to adjust the weight allocation of the churn risk prediction task and the user value prediction task in real time.

[0042] Causal reasoning employs counterfactual reasoning methods. By simulating different intervention scenarios, it analyzes the true causal relationship between user behavior and churn risk and user value, eliminating spurious correlations. For example, it analyzes the relationship between a user's browsing of competitor pages and churn risk, determining whether this behavior directly leads to an increase in churn risk or whether other intermediate factors exist. Specifically, it uses counterfactual reasoning by intervening in user behavior characteristics, such as changing the behavior of 'not logging in for 7 consecutive days' to 'logging in daily,' simulating the impact trend of exogenous factors (such as weekends and holidays) on churn risk.

[0043] The reinforcement learning dynamic weight mechanism treats churn risk prediction and user value prediction as two tasks, and uses the prediction performance of the two tasks (such as the accuracy of churn risk prediction and the error of user value prediction) as reward signals. The weights of the two tasks in model training are dynamically adjusted through reinforcement learning algorithms (such as Q-learning, PPO, etc.). When the prediction performance of one task decreases, its weight is increased to improve performance.

[0044] The quantification formula for the reward signal is: Where Acc is the churn prediction accuracy and MAE is the mean absolute error of user value prediction. The initial weight coefficients are used; the weight adjustment frequency is once after each batch of training, and real-time adjustment is triggered when the reward signal fluctuation exceeds 5%.

[0045] S4. Input the preprocessed multidimensional data into the feature extraction model to extract features and obtain user features;

[0046] The preprocessed multidimensional data is input into the feature extraction model. The federated learning part aggregates the local features from each channel, and the dynamic knowledge graph part extracts entity association features through the graph attention mechanism. Finally, the data is fused to obtain comprehensive user features.

[0047] S5. Input the user features into the joint prediction model for training to obtain the trained joint prediction model;

[0048] User features are divided into training and test sets. The joint prediction model is trained using the training set. A gradient descent algorithm is employed, with the prediction loss function (e.g., cross-entropy loss function for churn risk prediction, mean squared error loss function for user value prediction) as the optimization objective. Model parameters are continuously adjusted, and model performance is evaluated using the test set until the model reaches the preset performance metrics. The joint loss function is specifically... ,in Cross-entropy loss is used (formula: ,in, A true label for the risk of attrition. (The probability of user churn predicted by the model). Mean squared error loss (formula: ,in, The true value of user value The user value predicted by the model, and Continuous numerical values ​​with the same dimensions. For calculation The number of samples at that time. The weight parameters are dynamically adjusted to enhance learning. Cross-entropy loss is used to predict churn risk, measuring the difference between the predicted probability and the true label. The mean squared error loss for user value prediction measures the deviation between the predicted value and the actual value.

[0049] S6. After preprocessing and feature extraction, the multidimensional data of the users to be predicted is input into the trained joint prediction model, and the prediction results of the churn risk and value of the users to be predicted are output.

[0050] The multidimensional data of the user to be predicted is preprocessed according to the method in step S1, and then the features are extracted by the feature extraction model in step S2. The data is then input into the trained joint prediction model. The model outputs the user's churn probability (to represent churn risk) and the potential value assessment value within a certain period (to represent user value).

[0051] This embodiment takes a video subscription service as an example and applies the method of the present invention to jointly predict user churn risk and value.

[0052] Step S1: Collect multidimensional data from video subscription service users, including user age, gender, login time, video viewing duration, video type, comment and interaction history, friend list, etc. Preprocess this data by removing duplicate login records, standardizing viewing duration to a value between 0 and 1, filling missing age information with the average age of users of the same gender, and identifying and handling abnormal viewing duration data (such as extreme values ​​significantly higher than the average viewing duration).

[0053] Step S2: Construct the feature extraction model. A horizontal federated learning framework is adopted, with video platform clients from different regions as participants. Each participant trains its feature extractor locally and exchanges gradient parameters through a blockchain consensus mechanism. Simultaneously, a dynamic knowledge graph is constructed. Entities include users, video content, and behaviors such as watching and commenting. Relationships include the "watching" relationship between users and videos, and the "friend" relationship between users. The knowledge graph is updated based on real-time user behavior data. A graph attention mechanism is used to calculate the attention weights between users and frequently watched video types to extract potential interest features.

[0054] Step S3: Construct a joint prediction model. Counterfactual reasoning is used to analyze the causal relationship between users' consecutive days of inactivity and churn risk, eliminating the influence of random factors. A reinforcement learning dynamic weighting mechanism uses the accuracy of churn risk prediction and the mean squared error of user value prediction as rewards. The PPO algorithm adjusts the weights of the two tasks, increasing the weight when the user value prediction error is large.

[0055] Step S4: Input the preprocessed multidimensional data into the feature extraction model to obtain user features including basic user features, behavioral features, interest features, etc.

[0056] Step S5: Divide user features into training set and test set. The training set is used to train the joint prediction model. The gradient descent algorithm is used to optimize the model parameters. The model is verified through the test set. When the churn risk prediction accuracy and user value prediction accuracy reach the preset requirements, training is stopped and the trained model is obtained.

[0057] Step S6: For newly registered users to be predicted, collect their registration information and initial viewing behavior data. After preprocessing and feature extraction, input the data into the trained model to output the user's churn probability and potential value assessment value in the future. Video subscription service providers can use the prediction results to push promotional activities to users with high churn risk and provide exclusive services to high-value users, thereby improving user retention rate and overall revenue.

[0058] Experimental verification

[0059] Experimental data: User behavior data from a video platform over three months was used, including basic information of 100,000 users, 50,000 pieces of content data, and 2 million user-content interaction records (such as clicks, favorites, payments, etc.).

[0060] The experimental results are shown in Table 1 below:

[0061] Table 1. Comparison of the proposed method with single-task churn prediction models and fixed-weight multi-task models:

[0062]

[0063] Experimental results show that the present invention is significantly superior to existing methods in both prediction accuracy and error control.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services, characterized in that, Includes the following steps: S1. Collect multidimensional data from users who subscribe to the service, including user basic information, behavioral data, content interaction data, and social relationship data; The multidimensional data is preprocessed, including data cleaning, data standardization, missing value handling, and outlier handling. S2. Construct a feature extraction model, including a horizontal federated learning framework and a dynamic knowledge graph; the horizontal federated learning framework uses homomorphic encryption technology to encrypt and transmit parameters, and exchanges intermediate parameters through a blockchain consensus mechanism; the dynamic knowledge graph constructs a user-content-behavior ternary association network through entity recognition, and performs incremental updates based on real-time user interaction data; S3. Train a joint prediction model, which introduces causal reasoning and a reinforcement learning dynamic weighting mechanism. The causal reasoning is used to identify the causal path between user behavior and churn risk and user value. The causal reasoning adopts a counterfactual reasoning method to simulate the changing trend of churn risk by intervening in user behavior characteristics. The reinforcement learning dynamic weighting mechanism uses the churn risk prediction accuracy and the mean absolute error of user value prediction as reward signals, and adjusts the weight allocation in real time through the PPO algorithm. S4. Input the preprocessed multidimensional data into the feature extraction model to extract features and obtain user features; S5. The user features are input into the joint prediction model for training. The training of the joint prediction model adopts the gradient descent algorithm to optimize the model parameters by minimizing the prediction loss function. S6. After preprocessing and feature extraction, the multidimensional data of the users to be predicted is input into the trained joint prediction model, and the prediction results of the churn risk and value of the users to be predicted are output.

2. The multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services according to claim 1, characterized in that: In step S2, the federated learning adopts a horizontal federated learning framework, where each channel trains its feature extractor locally and exchanges intermediate parameters through a blockchain consensus mechanism. The blockchain consensus mechanism uses a practical Byzantine fault-tolerant algorithm, and homomorphic encryption is used to encrypt the parameter exchange.

3. The multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services according to claim 1, characterized in that: In step S2, the construction of the dynamic knowledge graph includes entity recognition, relationship extraction, and graph update. The entities include users, content, and behavior, and the relationships include the interaction relationship between users and content, and the social relationship between users.

4. The multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services according to claim 1, characterized in that: In step S3, the causal reasoning adopts the counterfactual reasoning method, which uses intervention simulation to separate the impact of endogenous factors and exogenous factors on churn risk and user value.

5. The multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services according to claim 1, characterized in that: In step S3, the reinforcement learning dynamic weight mechanism uses the churn risk prediction accuracy and user value prediction error as reward signals, and adjusts the weights of the two tasks through a reinforcement learning algorithm.

6. The multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services according to claim 1, characterized in that: In step S6, the predicted result of the churn risk is the churn probability, and the predicted result of the user value is the potential value assessment value of the user within a certain period.

7. The multi-dimensional data-driven method for jointly predicting the risk and value of user churn in subscription services according to claim 1, characterized in that: The subscription service includes at least one of video subscription, music subscription, reading subscription, and software service subscription.

Images