Dimensional Isolation for Video Retention Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Video delivery services face challenges in retaining subscribers as existing prediction methods only identify users with similar characteristics to those who have canceled, making it difficult to determine specific actions to prevent cancellation.
Innovation Solution
The system generates a retention prediction score by isolating user features into individual dimensions, using prediction networks to analyze each dimension's impact and combining scores to provide a detailed analysis of cancellation probabilities, with variance and de-correlation terms to ensure meaningful and independent contributions from each dimension.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the video delivery service uses generalization of user characteristics to identify at-risk users, then the coverage of user identification is improved, but the precision of actionable insights deteriorates
Solution Approach 1:
The patent segments user characteristics into multiple independent dimensions (e.g., content consumption patterns, device usage, payment behavior, support interactions). Each dimension is evaluated separately by dedicated prediction networks, allowing the system to maintain both broad coverage and specific actionable insights by analyzing which particular dimension is driving the cancellation risk for each user.
2Reliability
If the system analyzes multiple user characteristics together, then the overall prediction accuracy is improved, but the ability to identify specific retention actions deteriorates
Solution Approach 1:
The system divides the prediction task into multiple independent dimension analyses, each handled by a separate prediction network. This segmentation maintains overall prediction reliability while making the system easier to operate for retention actions, as each dimension can be addressed with specific, targeted interventions (e.g., if content consumption dimension shows risk, recommend new content; if support interactions show risk, improve customer service).
Solution Approach 2:
The patent applies local quality by allowing different prediction networks to use different characteristics and evaluation criteria appropriate to each specific dimension. For example, content consumption dimension may focus on viewing frequency and variety, while device usage dimension may focus on device types and usage patterns. This enables tailored retention strategies for each dimension while maintaining overall prediction reliability.
3Loss of information
If the system uses individual dimension scores, then the interpretability of prediction results is improved, but the complexity of the prediction system deteriorates
Solution Approach 1:
The patent segments the prediction system into multiple independent prediction networks, each responsible for a specific user characteristic dimension. This segmentation improves interpretability by providing clear dimension scores that indicate which specific factors contribute to cancellation risk. The modular architecture manages complexity by allowing each network to focus on a single dimension, making the overall system more manageable despite having multiple components.
Data Source
AI summary
In one embodiment, a method separates subscriber features generated from subscriber interaction with a video delivery service into feature dimensions and inputs the feature dimensions into a respective prediction network. Each prediction network is trained to output a respective dimension score. The method outputs dimension scores using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a de-correlation term to control a correlation of the plurality of feature dimensions. The dimension scores are combined into a retention prediction score and an action is performed on the video delivery service for the subscriber based on the retention score.