Unified Generalized Linear Mixed Model for Content Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current content evaluation models for online content recommendation suffer from a lack of personalization, leading to irrelevant content being prioritized over relevant content, resulting in excessive resource consumption and maintenance overhead due to the use of separate layers for first pass rankers and main content evaluation models.
Innovation Solution
Implementing a partial generalized linear mixed model as a first pass ranker and a full generalized linear mixed model as the main content evaluation model, which selects candidate content items based on filter scores and recommendation scores, respectively, to improve relevance and reduce maintenance overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a separate first pass ranker layer is used before the main content evaluation model, then candidate content items can be pre-filtered to improve efficiency, but maintenance overhead increases due to needing to maintain and update both layers separately
Solution Approach 1:
The patent merges the first pass ranker and main content evaluation model into a single unified content evaluation model. The model uses a two-stage filtering process internally (first pass filter and second pass filter) but maintains a single model structure, eliminating the need to maintain separate layers. This resolves the contradiction by combining the functionality of both layers while preserving the efficiency benefits of staged filtering.
2Power
If a separate first pass ranker is used, then processing load can be reduced by filtering candidates early, but experiment bandwidth is consumed by diagnostic simulations on the ranker, slowing down experiments on the main model
Solution Approach 1:
By merging the first pass ranker into the main content evaluation model as an integrated unified model with staged filtering, the patent eliminates the need for separate diagnostic simulations on a standalone ranker. Experiment bandwidth is preserved because all experiments occur within a single model framework, resolving the contradiction between processing load reduction and experiment speed.
3Device complexity
If traditional content evaluation models are used without personalization, then model simplicity is maintained, but relevant content is downgraded in favor of irrelevant content in recommendation displays
Solution Approach 1:
The patent applies local quality by implementing user-specific random effects parameters that personalize content evaluation for each user. The unified content evaluation model incorporates both population-level fixed effects and user-specific random effects, allowing the model to adapt its predictions to individual user preferences while maintaining the overall model structure. This resolves the contradiction by adding personalization locally without fundamentally changing the global model architecture.
Data Source
AI summary
In some embodiments, a computer system selects a first subset of candidate content items based on their filter scores that are generated based on a partial generalized linear mixed model comprising a baseline model and a user-based model, with the baseline model being a generalized linear model, and the user-based model being a random effects model based on user actions by the target user directed towards reference content items related to the candidate content items. In some embodiments, the computer system then selects a second subset from the first subset based on recommendation scores that are generated based on a full generalized linear mixed model comprising the baseline model, the user-based model, and an item-based model, with the item-based model being a random effects model based on user actions directed towards the candidate online content item by reference users related to the target user.