Multi-dimensional data-driven subscription service user loss risk and value combined prediction method

Abstract

The invention relates to the technical field of data processing and prediction, in particular to a multi-dimensional data driven subscription service user loss risk and value combined prediction method, and aims to solve the problems of data sparsity and high-dimensional feature conflict and multi-task target interference in existing subscription service user loss risk and value combined prediction. According to the method, multi-dimensional data of subscription service users are collected and preprocessed, and a feature extraction model fusing federated learning and a dynamic knowledge graph is constructed to solve the problem of data sparsity and high-dimensional feature conflict; a joint prediction model combining causal reasoning and a reinforcement learning dynamic weight mechanism is constructed to solve the problem of multi-task target interference; and carrying out training and prediction by utilizing the constructed model, and outputting a prediction result of the user loss risk and value. According to the method, the accuracy and reliability of subscription service user loss risk and value joint prediction are improved, and powerful support is provided for operation decision making of subscription service providers.

Description

Technical Field

[0001] The present invention relates to the field of data processing and prediction technology, and in particular to a multi-dimensional data-driven method for jointly predicting the risk and value of subscriber churn of subscription services. Background Art

[0002] In today's digital economy, subscription services have become a key business model in numerous industries, including video platforms, music software, online education, and cloud services. Accurately predicting user churn risk and assessing user value are crucial for subscription service providers. By predicting churn risk, providers can proactively implement targeted retention measures to reduce churn rates. Assessing user value helps providers optimize resource allocation, provide better services to high-value users, and identify potential high-value users.

[0003] However, existing methods for jointly predicting the risk and value of churn for subscription service users suffer from numerous problems, hindering the accuracy and reliability of predictions. Data sparsity and high-dimensional feature conflicts are particularly prominent. Behavioral data on subscription service users is often sparse. For example, some users may have only performed a small number of interactions, resulting in insufficient effective information. Furthermore, user data has numerous 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 impacting prediction results. Existing technologies often address data sparsity using simple interpolation methods, resulting in a utilization rate of less than 30% for user-content interaction data. Furthermore, fixed weight allocation in multi-task learning (e.g., a 1:1 weight ratio) results in an error rate of up to 25% in predicting the risk of churn for high-value users.

[0004] Furthermore, multi-task goal interference is a major challenge facing existing methods. User churn risk prediction and user value assessment are two related yet distinct tasks, with different objectives and evaluation criteria. During the joint prediction process, conflicting objectives may exist between the two tasks. For example, high-value users may have an increased risk of churn due to certain factors. Traditional joint prediction models often use a fixed weight distribution method, which cannot dynamically balance the importance of the two tasks. This can affect the prediction performance of one task, thereby reducing the overall joint prediction effect.

[0005] While some approaches have attempted to address these issues, most rely on conventional techniques, such as simple data aggregation and fixed-weight multi-task learning, which struggle to fundamentally overcome existing bottlenecks. Therefore, a new multidimensional data-driven approach to jointly predicting subscription service user churn risk and value is urgently needed. This approach can address issues such as data sparsity, high-dimensional feature conflicts, and multi-task interference, improve prediction accuracy and reliability, and provide more effective decision support for subscription service providers. Summary of the Invention

[0006] The present invention aims to solve the problems of data sparsity, high-dimensional feature conflicts, and multi-task objective interference in existing methods for jointly predicting the risk and value of subscription service users. It provides a multi-dimensional data-driven method for jointly predicting the risk and value of subscription service users. It increases data utilization to more than 60% through the ternary association network of the dynamic knowledge graph, and reduces the error rate to less than 10% through the dynamic weight mechanism of reinforcement learning.

[0007] The technical solution adopted by the present invention to solve the technical problem is: a multi-dimensional data-driven subscription service user churn risk and value joint prediction method, comprising the following steps: S1. Collect multidimensional data of subscription service users and preprocess the multidimensional data; S2. Build a feature extraction model, including a horizontal federated learning framework and a dynamic knowledge graph. The horizontal federated learning framework uses homomorphic encryption technology for encrypted parameter transmission. The dynamic knowledge graph uses entity recognition to construct a user-content-behavior ternary association network and performs incremental updates based on real-time user interaction data. S3. Training a joint prediction model that incorporates causal reasoning and a reinforcement learning dynamic weighting mechanism. The causal reasoning is used to identify the causal paths between user behavior, churn risk, and user value. The reinforcement learning dynamic weighting mechanism uses the accuracy of churn risk prediction and the mean absolute error of user value prediction as reward signals, and adjusts weight distribution 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. Inputting the user features into the joint prediction model for training to obtain a trained joint prediction model; S6. After preprocessing and feature extraction, the multidimensional data of the user to be predicted is input into the trained joint prediction model, and the prediction results of the churn risk and value of the user to be predicted are output.

[0008] Specifically, in step S1, the multidimensional data includes at least one of user basic information, behavior data, content interaction data, and social relationship data.

[0009] Specifically, in step S1, the preprocessing includes at least one of data cleaning, data standardization, missing value processing and outlier processing.

[0010] Specifically, in step S2, the federated learning adopts a horizontal federated learning framework, 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 for encryption during parameter exchange.

[0011] 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 behaviors, and the relationships include the interactive relationship between users and content, and the social relationship between users.

[0012] Specifically, in step S3, the causal reasoning adopts the counterfactual reasoning method to separate the impact of endogenous and exogenous factors on churn risk and user value through intervention simulation; the counterfactual reasoning simulates the changing trend of churn risk by intervening in user behavior characteristics, such as correcting "no login for 7 consecutive days" to "daily login".

[0013] Specifically, in step S3, the reinforcement learning dynamic weight mechanism uses the churn risk prediction accuracy and the user value prediction error as reward signals, and adjusts the weights of the two tasks through the reinforcement learning algorithm.

[0014] Specifically, in step S5, the training of the joint prediction model adopts a gradient descent algorithm to optimize the model parameters by minimizing the prediction loss function.

[0015] Specifically, in step S6, the prediction result of the churn risk is the churn probability, and the prediction result of the user value is the potential value evaluation value of the user in a certain period of time.

[0016] Specifically, the subscription service includes at least one of video subscription, music subscription, reading subscription, and software service subscription.

[0017] Beneficial effects of the present invention: (1) The multi-dimensional data-driven subscription service user churn risk and value joint prediction method described in the present invention adopts federated learning and dynamic knowledge graph fusion technology to build a feature extraction model. Federated learning realizes the safe and collaborative use of cross-channel data, avoids the privacy leakage risk brought by data concentration, and makes full use of the dispersed multi-dimensional data. The dynamic knowledge graph constructs a user-content-behavior association network, and combined with the graph attention mechanism, it can effectively mine the potential features in sparse data, solves the problem of conflict between data sparsity and high-dimensional features, and improves the effectiveness of feature extraction.

[0018] (2) The present invention describes a multi-dimensional data-driven method for jointly predicting the risk and value of subscription service users’ churn. It introduces causal reasoning and reinforcement learning dynamic weight mechanism to construct a joint prediction model. Causal reasoning identifies the true causal relationship between user behavior and prediction targets, thereby 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 improving the overall performance of the joint prediction.

[0019] (3) The multi-dimensional data-driven subscription service user churn risk and value joint prediction method described in the present invention is based on multi-dimensional data drive and comprehensively considers various user information, making the prediction results more comprehensive and accurate, providing a strong decision-making basis for subscription service providers to formulate targeted operation strategies, and helping to improve user retention rate and user value.

[0020] (4) On a video subscription service dataset, the proposed method achieves 89.2% accuracy in predicting churn risk, a 12.3% improvement over existing multi-task learning methods. The MAE for user value prediction is 0.08, a 35.7% reduction over traditional regression methods. In commercial applications, this method can help subscription service providers reduce user churn by 15%-20% and increase annual revenue by more than 10%. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below with reference to the accompanying drawings and examples.

[0022] Figure 1 A flowchart of a multi-dimensional data-driven subscription service user churn risk and value joint prediction method provided by the present invention; Figure 2 This is a diagram of the feature extraction model architecture in a multi-dimensional data-driven subscription service user churn risk and value joint prediction method provided by the present invention; Figure 3 A schematic diagram of the joint prediction model mechanism in a multi-dimensional data-driven method for jointly predicting subscription service user churn risk and value provided by the present invention; Figure 4 This is a data preprocessing flowchart for a multi-dimensional data-driven subscription service user churn risk and value joint prediction method provided by the present invention. DETAILED DESCRIPTION

[0023] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0024] like Figure 1-4As shown, the multi-dimensional data-driven subscription service user churn risk and value joint prediction method of the present invention includes the following steps: S1. Collect multidimensional data of subscription service users and preprocess the multidimensional data; Specifically, the multidimensional data collected includes, but is not limited to, basic user information (such as age, gender, and region), behavioral data (such as login frequency, browsing time, and operation history), content interaction data (such as videos watched, music listened to, and articles read), and social relationship data (such as friend lists and interaction history). Preprocessing includes data cleaning to remove invalid and duplicate data; data standardization to convert data of varying magnitudes to a unified scale; missing value processing using mean filling, median filling, or context-based filling methods; and outlier processing to identify and address abnormal data through statistical analysis.

[0025] Among them, the criterion for judging invalid data is user data without any interactive behavior for three consecutive months; missing values ​​are contextually filled through the user's social relationship chain, specifically using the average behavioral characteristics of the user's top five friends to replace the missing interaction data; outliers are identified through the Z-score method, and when |Z|>3, it is determined to be an outlier and replaced with the 95th percentile value.

[0026] S2. Construct a feature extraction model that uses federated learning and a dynamic knowledge graph fusion technique. Federated learning is used to implement local training and parameter exchange for multi-dimensional data across channels. The dynamic knowledge graph is used to construct a user-content-behavior association network and to mine potential features in the association network using a graph attention mechanism. Federated Learning utilizes a horizontal federated learning framework. Each channel (e.g., different platform clients, different service regions) trains the collected data locally, builds a local feature extractor, and securely exchanges intermediate parameters (such as gradient information) through a blockchain consensus mechanism, enabling cross-channel data collaborative utilization while protecting data privacy. The parameter exchange process is encrypted using homomorphic encryption technology to ensure data privacy during parameter transmission. The blockchain consensus mechanism specifically uses the Practical Byzantine Fault Tolerance (PBFT) algorithm, which is suitable for efficient consensus verification in cross-channel scenarios.

[0027] The construction of a dynamic knowledge graph first performs entity recognition to determine entities such as users, content, and behaviors; then performs relationship extraction to identify the relationships between entities, such as the "watching" and "listening" relationships between users and content, and the "friends" and "interactions" relationships between users; and updates the knowledge graph in real time based on newly generated data. Using the graph attention mechanism, the attention weights between entities are calculated, focusing on the associations that have an important impact on predictions, thereby mining potential effective features from sparse high-dimensional data. The triggering condition for graph updates is that the user behavior data increment exceeds 10% or is automatically executed every morning; the weight of the graph attention mechanism is calculated by the formula Calculate, where and is the node feature vector, and d is the feature dimension, which is used to determine the attention weight between the user and content nodes. is the attention weight between node i and node j, quantifying the importance of the relationship between the two. Normalize the vector of dimension j so that the sum of the weights is 1. is the leaky ReLU activation function. is a trainable weight matrix with dimension [2d, k] (d is dimension, k is the dimension after linear transformation, a hyperparameter), which is used to extract the key information of splicing features.

[0028] S3. Construct a joint prediction model that incorporates causal reasoning and a reinforcement learning dynamic weighting mechanism. The causal reasoning is used to identify the causal paths between user behavior and churn risk and user value, and the reinforcement learning dynamic weighting mechanism is used to adjust the weight distribution of churn risk prediction tasks and user value prediction tasks in real time. Causal reasoning employs counterfactual reasoning, simulating various intervention scenarios to analyze the true causal relationship between user behavior, churn risk, and user value, eliminating false correlations. For example, analyzing the relationship between a user's browsing of a competitor's product page and churn risk can determine whether this behavior directly leads to increased churn risk or whether other intervening factors are involved. Specifically, counterfactual reasoning is performed by intervening in user behavior characteristics. For example, modifying "not logging in for seven consecutive days" to "logging in daily" can simulate the impact of exogenous factors (such as weekends and holidays) on churn risk.

[0029] The reinforcement learning dynamic weight mechanism regards churn risk prediction and user value prediction as two tasks, and uses the prediction results of the two tasks (such as churn risk prediction accuracy and user value prediction error) as reward signals. It dynamically adjusts the weights of the two tasks in model training through reinforcement learning algorithms (such as Q-learning, PPO, etc.). When the prediction effect of one of the tasks decreases, its weight is increased to improve performance.

[0030] The quantization formula of the reward signal is , where Acc is the churn prediction accuracy, MAE is the mean absolute error of user value prediction, is the initial weight coefficient; the weight adjustment frequency is performed once after each batch of training, and real-time adjustment is triggered when the reward signal fluctuation exceeds 5%.

[0031] S4. Input the preprocessed multidimensional data into the feature extraction model to extract features and obtain user features; The preprocessed multidimensional data is input into the feature extraction model. The federated learning part aggregates the local features of each channel, and the dynamic knowledge graph part extracts entity association features through the graph attention mechanism, and finally integrates them to obtain comprehensive user features.

[0032] S5. Inputting the user features into the joint prediction model for training to obtain a trained joint prediction model; The user features are divided into training set and test set, and the joint prediction model is trained using the training set. The gradient descent algorithm is used, and the prediction loss function (such as the cross entropy loss function for churn risk prediction and the mean square error loss function for user value prediction) is used as the optimization target. The model parameters are continuously adjusted, and the model performance is evaluated through the test set until the model reaches the preset performance indicators. The joint loss function is specifically: ,in Using cross entropy loss (formula: ,in, is the true label of the churn risk, is the user churn probability predicted by the model), Using mean square error loss (formula: ,in, is the true value of the user value, is the user value predicted by the model, and Continuous values ​​of the same dimension. To calculate The number of samples at the time of . ), where Dynamically adjusted weight parameters for reinforcement learning. The cross entropy loss for churn risk prediction measures the difference between the predicted probability and the true label. It is the mean square error loss of user value prediction, which measures the deviation between the predicted value and the true value.

[0033] S6. After preprocessing and feature extraction, the multidimensional data of the user to be predicted is input into the trained joint prediction model, and the prediction results of the churn risk and value of the user to be predicted are output.

[0034] The multidimensional data of the user to be predicted is preprocessed according to the method of step S1, and then input into the trained joint prediction model after the features are extracted by the feature extraction model in step S2. The model outputs the user's churn probability (used to represent the churn risk) and the potential value assessment value within a certain period of time (used to represent the user's value).

[0035] 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.

[0036] Step S1: Collect multi-dimensional data on video subscription service users, including age, gender, login time, video viewing time, video genre, comment and interaction history, and friend list. Preprocess this data by removing duplicate login records, normalizing viewing time to a value between 0 and 1, and filling in missing age information with the average age of users of the same gender. Identify and address abnormal viewing time data (e.g., extreme values ​​significantly exceeding the average viewing time).

[0037] Step S2: Build a feature extraction model. Using a horizontal federated learning framework, video platform clients in different regions serve as participants. Each participant trains a 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 viewing and commenting. Relationships include "view" relationships between users and videos, and "friend" relationships 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 viewed video types, extracting potential interest features.

[0038] Step S3: Construct a joint prediction model. Counterfactual reasoning is used to analyze the causal relationship between a user's multiple days of absence 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 these two tasks, increasing the weight when the user value prediction error is large.

[0039] Step S4: Input the pre-processed multidimensional data into a feature extraction model to obtain user features including basic user features, behavioral features, interest features, etc.

[0040] Step S5: Divide user features into a training set and a 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 parameters are verified by the test set. When the accuracy of churn risk prediction and user value prediction reaches the preset requirements, the training is stopped to obtain the trained model.

[0041] Step S6: For newly registered users to be predicted, their registration information and initial viewing behavior data are collected. After preprocessing and feature extraction, this data is input into the trained model to output the user's churn probability and potential value assessment over a period of time. Based on the prediction results, video subscription service providers can promote promotions to users with high churn risk and provide exclusive services to high-value users, thereby improving user retention and overall revenue.

[0042] Experimental verification Experimental data: 3 months of user behavior data from a video platform, 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.). The experimental results are shown in Table 1 below: Table 1 Comparison table of the method proposed in this invention, the single-task churn prediction model and the fixed-weight multi-task model:

[0043] Experimental results show that the present invention is significantly superior to existing methods in both prediction accuracy and error control.

[0044] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above-described embodiments. The above-described embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional data-driven method for jointly predicting subscription service user churn risk and value, characterized by: The following steps are involved: S1. Collect multidimensional data of subscription service users and preprocess the multidimensional data; S2. Build a feature extraction model, including a horizontal federated learning framework and a dynamic knowledge graph. The horizontal federated learning framework uses homomorphic encryption technology for encrypted parameter transmission. The dynamic knowledge graph uses entity recognition to construct a user-content-behavior ternary association network and performs incremental updates based on real-time user interaction data. S3. Training a joint prediction model that incorporates causal reasoning and a reinforcement learning dynamic weighting mechanism. The causal reasoning is used to identify the causal paths between user behavior, churn risk, and user value. The reinforcement learning dynamic weighting mechanism uses the accuracy of churn risk prediction and the mean absolute error of user value prediction as reward signals, and adjusts weight distribution 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. Inputting the user features into the joint prediction model for training to obtain a trained joint prediction model; S6. After preprocessing and feature extraction, the multidimensional data of the user to be predicted is input into the trained joint prediction model, and the prediction results of the churn risk and value of the user to be predicted are output.

2. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method according to claim 1, characterized in that: In step S1, the multidimensional data includes at least one of user basic information, behavior data, content interaction data, and social relationship data.

3. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method according to claim 1, characterized in that: In step S1, the preprocessing includes at least one of data cleaning, data standardization, missing value processing and outlier processing.

4. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method according to claim 1, characterized in that: In step S2, the federated learning adopts a horizontal federated learning framework, 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 for encryption during parameter exchange.

5. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method 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 interactive relationship between users and content, and the social relationship between users.

6. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method according to claim 1, characterized in that: In step S3, the causal reasoning adopts the counterfactual reasoning method to separate the impact of endogenous factors and exogenous factors on churn risk and user value through intervention simulation; the counterfactual reasoning simulates the changing trend of churn risk by intervening in user behavior characteristics.

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

8. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method according to claim 1, characterized in that: In step S5, the training of the joint prediction model adopts a gradient descent algorithm to optimize the model parameters by minimizing the prediction loss function.

9. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method according to claim 1, characterized in that: In step S6, the prediction result of the churn risk is the churn probability, and the prediction result of the user value is the potential value evaluation value of the user in a certain period of time.

10. The multi-dimensional data-driven subscription service user churn risk and value joint prediction method 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