Deep Embedding Mimicry Models for Feed Interaction Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learned models in social networking services fail to adequately capture the impact of mimicry on user interactions, leading to reduced effectiveness in determining the likelihood of user engagement with feed items.
Innovation Solution
Implementing separate mimicry machine-learned models trained using counterfactual regression algorithms, each tailored to specific item types, to estimate mimicry effects and adjust user interface elements and rankings based on individual user behaviors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing machine-learned models are used to determine user interaction likelihood, then the models can process user behavior data, but they fail to capture the mimicry effect which reduces their accuracy
Solution Approach 1:
The patent divides the modeling approach by creating separate mimicry models for different item types (e.g., images, videos, articles). Each model is specialized to capture mimicry effects specific to its item type, allowing the system to accurately measure and adapt to varying mimicry patterns across different content categories while maintaining overall prediction accuracy.
2Measurement precision
If separate mimicry models are implemented for different item types, then the accuracy of capturing mimicry effects improves, but the system complexity increases
Solution Approach 1:
The patent implements a universal framework where multiple item-type-specific mimicry models operate under a common architecture and training methodology. This allows the system to handle diverse item types with specialized models while maintaining manageable complexity through standardized processes, shared infrastructure, and consistent evaluation metrics across all models.
Solution Approach 2:
The patent applies different mimicry modeling approaches tailored to specific item types rather than using a single uniform model. Each item type (images, videos, articles) receives customized modeling attention based on its unique characteristics and mimicry patterns, allowing for higher measurement precision where needed while keeping the overall system organized and manageable.
3Productivity
If mimicry effects are incorporated into feed ranking algorithms, then user engagement likelihood improves, but the computational requirements and processing time increase
Solution Approach 1:
The patent pre-trains item-type-specific mimicry models on historical data before deployment. This preliminary action allows the models to learn and store mimicry patterns in advance, enabling faster real-time predictions during feed generation without requiring extensive computational resources at the moment of feed creation, thus reducing processing time while maintaining high user engagement accuracy.
Data Source
AI summary
In an example embodiment, a separate mimicry machine-learned model is trained for each of a plurality of different item types. Each of these models is trained to estimate an effect of mimicry for a user (i.e., a user whose user profile or other information is passed to the corresponding mimicry machine-learned model at prediction-time). The output of these models may be either used on its own to perform various actions, such as modifying a location of a user interface element of a user interface, or may be passed as input to an interaction machine-learned model that is trained to determine a likelihood of a user (i.e., a user whose user profile or other information is passed to the interaction machine-learned model at prediction-time) interacting with a particular item, such as a potential feed item.