Session-Level Randomization for Social Network Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
On-line social networks face challenges in reducing biases in training data for their news feed ranking models due to the reliance on historical interactions, which can lead to skewed representations of user interests.
Innovation Solution
Implementing session-level randomization by designating a portion of member sessions as random sessions, where updates are presented using random ordering or perturbed ranking scores, to collect diverse interaction data that countersact biases in training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If historical interaction data is used for training news feed ranking models, then the models can learn from real user behavior patterns, but the training data becomes biased and skewed towards certain user interests
Solution Approach 1:
The patent applies inversion by deliberately presenting updates in random order or with perturbed ranking scores during designated random sessions, reversing the normal ranked presentation approach. This allows the system to collect unbiased interaction data that reflects genuine user interest rather than preferences shaped by existing ranking algorithms, thereby improving fairness and diversity in training data while maintaining model accuracy through separate training and serving pathways.
2Adaptability or versatility
If updates are presented in random order during sessions, then diverse user interactions are collected, but the user experience deteriorates compared to personalized ranked feeds
Solution Approach 1:
The patent segments user sessions into two distinct types: random sessions for data collection and normal sessions for user consumption. During random sessions, updates are presented in random order or with perturbed scores to collect diverse interaction data. During normal sessions, the personalized ranking model presents optimally ordered updates for best user experience. This segmentation allows the system to simultaneously achieve data diversity and user experience quality without compromising either.
Solution Approach 2:
The patent introduces an intermediary mechanism that selectively applies randomization only to specific sessions designated for training data collection, while leaving the majority of user sessions unaffected. This intermediary approach allows the system to gather diverse interaction data without exposing all users to degraded experience, thereby balancing data diversity requirements with user experience quality.
3Ease of operation
If a portion of sessions is designated as random sessions for data collection, then training data bias is reduced, but system complexity increases
Solution Approach 1:
The patent applies partial action by designating only a specific portion of user sessions (e.g., a small percentage) as random sessions for training data collection, rather than applying randomization to all sessions. This partial approach reduces system complexity compared to universal randomization while still gathering sufficient diverse interaction data to improve training fairness and model robustness.
Data Source
AI summary
A news feed system of an on-line social network system news utilizes a relevance model to determine which updates from an inventory of updates are to be presented to a member on their news feed page. The relevance model is trained using historical data that reflects interactions of members of the on-line social network system with items in their respective news feed pages. In order to reduce potential biases in the historical data that is used to train the relevance model, the news feed system designates a certain portion of all member sessions to be random sessions. The news feed generated for a member during a random session includes updates that are selected and/or ordered for presentation using one or more randomization techniques.