Personalized Feed Ranking via Segmented GAME Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current social networking algorithms rely on single machine learned models that are biased due to shared global weights, leading to inaccurate personalized predictions as user interactions are influenced by data from other members, failing to capture individual user preferences and item attractions effectively.
Innovation Solution
Implementing a Generalized Additive Mixed Effect (GAME) model that uses three separate models - a global model, a per-viewer model, and a per-actor model - to account for individual user and item attributes, interactions, and context, allowing for personalized predictions by incorporating member-level and actor-level regression coefficients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single machine learned model with global weights is used for feed ranking, then the model structure is simple and computationally efficient, but the predictions are biased because all members share the same weights and data from other members impacts score calculations
Solution Approach 1:
The patent segments the single global model into multiple independent models: a global model that learns general patterns from all user data, and per-user models that learn individual user preferences. Each user has their own model instance with personalized weights, eliminating the bias caused by shared global weights while maintaining computational efficiency through model sharing and regularization.
2Use of energy by moving object
If personalized scores are calculated using a single machine learned model, then computational resources are conserved, but individual user preferences cannot be effectively captured due to bias from other members' data
Solution Approach 1:
The patent applies local quality by giving each user a personalized model with unique weights tailored to their specific preferences and behavior patterns. While the global model structure and shared features provide computational efficiency, the per-user customization ensures that individual preferences are captured accurately without being influenced by other users' data.
3Ease of manufacture
If global weights are shared across all members in a single model, then the model is easier to train and maintain, but the scores are biased due to data from other members impacting calculations
Solution Approach 1:
The patent segments the monolithic global model into a hierarchical structure with a global model for general patterns and per-user models for individual preferences. This segmentation isolates user data so that each user's score calculation is based on their own personalized weights, eliminating bias from other members' data while maintaining ease of training through shared model components and regularization techniques.
Data Source
AI summary
In an example embodiment, a GLMix model is utilized that models viewers and actors of feed items. This allows for random effects of individual viewers and actors to be taken into account without introducing biases. Additionally, in an example embodiment, predictions/recommendations are made more accurate by using three models, which are then combined, instead of a single GLMix model. Each of these models has different granularities and dimensions. A global model may model the similarity between user attributes (e.g., from the member profile or activity history) and item attributes. A per-viewer model may model user attributes and activity history of actors on feed items. A per-actor model may model user attributes and activity history of the viewers of feed items. The per-actor model may therefore, rely on information regarding how and what type of viewers interacted with items acted on by the particular actor.