Transfer Learning Forecasting for Sparse Content Time Series
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time-series methodologies struggle to provide accurate forecasting for content titles with limited historical data, leading to uncertain and inefficient resource allocation in streaming platforms.
Innovation Solution
A glass-box transfer learning approach is employed to generate training data for forecasting models, utilizing genre archetypes and dynamically fitted curves to enhance forecasting accuracy for content titles with limited historical data, incorporating both seasonal and non-seasonal titles.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Gradient Boosting Machines (GBMs) are used for time-series forecasting, then forecasting accuracy is improved for content titles with sufficient historical data, but forecasting reliability deteriorates for content titles with limited historical data (less than 90 days)
Solution Approach 1:
The patent introduces an intermediary mechanism (data augmentation module) that generates synthetic historical data for content titles with limited data. This intermediary bridges the gap between the forecasting model's requirement for sufficient historical data and the reality of newly released titles, allowing GBMs to operate reliably even when original historical data is scarce.
Solution Approach 2:
The system performs preliminary data augmentation before forecasting, generating synthetic historical data patterns based on genre archetypes and decay curves. This preliminary action prepares adequate training data in advance, enabling the GBM model to achieve reliable forecasting performance for content titles that would otherwise have insufficient historical data.
2Reliability
If more historical data is collected for forecasting, then model training reliability is improved, but time consumption and data availability worsen for newly released titles
Solution Approach 1:
The patent creates copies of historical data patterns from similar content titles within the same genre. By copying and adapting proven historical patterns from archetype titles, the system generates synthetic training data that mimics realistic viewing behavior, eliminating the need to wait for actual historical data to accumulate for new titles.
Solution Approach 2:
The system changes key parameters of historical data patterns (such as decay rates, seasonal components, and viewing curves) to create synthetic data that matches the specific characteristics of the content title being forecasted. This parameter adaptation allows rapid generation of customized training data without requiring actual historical accumulation.
3Measurement precision
If GBM models are optimized for seasonal trends, then forecasting precision is improved for seasonal content, but adaptability deteriorates for non-seasonal content with limited data
Solution Approach 1:
The patent applies local quality by creating genre-specific forecasting configurations. Different content genres (e.g., seasonal vs. non-seasonal, movies vs. series) receive customized data augmentation parameters and archetype selections tailored to their specific characteristics. This allows the system to maintain high forecasting precision for each genre while adapting to diverse content types.
Solution Approach 2:
The system achieves universality through a unified framework that handles both seasonal and non-seasonal content. The glass-box transfer learning approach and data augmentation mechanism work across all content types, with genre-specific parameters adjusting the behavior to match each content category's characteristics, making the system versatile while maintaining precision.
Data Source
AI summary
An improved method is provided to provide efficient and accurate prediction/forecasting of inflow for content titles with limited historical data. The method may include dynamic generation of training data to be supplied to a forecasting model for predicting a performance metric of a content title of interest with limited historical data, based on the limited historical data and/or historical data of one or more other content titles with sufficient history. As such, instead of the limited historical data of the content title, the forecasting model may study from a broader range of historical data that may have similar trends as the title of interest.