Multi-Branch Content Selection Model Reducing Training Bias
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Content selection models in online systems introduce bias due to training data that only includes selected content items with high prediction likelihoods, leading to underrepresentation of potentially interactive content items with low predicted likelihoods.
Innovation Solution
Training a content selection model using a multi-branched architecture that alternates between a prediction portion and an autoencoder portion, incorporating both selected and sampled content items to reduce bias, where the prediction portion is trained on selected items and the autoencoder portion is trained on unsampled items to reconstruct features, sharing common parameters to learn relationships across both sets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data includes only selected content items with high prediction likelihoods, then the model achieves high prediction accuracy for selected items, but bias is introduced causing underrepresentation of potentially interactive content items with low predicted likelihoods
Solution Approach 1:
The training data is segmented into two distinct subsets: a first subset containing selected content items with high prediction likelihoods (for supervised learning) and a second subset containing sampled content items with diverse prediction likelihoods including low likelihood items (for unsupervised learning). This segmentation allows the model to learn different patterns from different data sources, maintaining prediction accuracy while reducing bias through the separate processing of diverse content items
Solution Approach 2:
An autoencoder component is introduced as an intermediary unsupervised learning model that processes the sampled content items from the second subset. This autoencoder learns latent representations of content items without requiring interaction labels, thereby preserving information from low-likelihood items that would otherwise be excluded from training. The learned representations are then integrated with the supervised learning model, allowing bias reduction while maintaining prediction accuracy
2Adaptability or versatility
If training data is continuously updated with interaction information from selected content items, then the model adapts to user preferences over time, but bias towards already-selected content types increases
Solution Approach 1:
The continuously updated training data is segmented into two separate data streams: one stream processes interaction information from selected content items (maintaining adaptability), while the other stream processes sampled content items with diverse likelihoods (counteracting bias amplification). This segmentation ensures that adaptation to user preferences occurs without complete dominance of already-selected content types in the training distribution
Solution Approach 2:
The system implements feedback mechanisms where the model's prediction likelihoods are used to determine sampling strategies for the second subset. By sampling content items across the full range of prediction likelihoods (including low likelihood items) rather than only high likelihood items, the feedback loop incorporates diverse content types into training, preventing bias amplification while maintaining the ability to adapt to user preferences through the supervised learning component
Data Source
AI summary
An online system trains a content selection model based on a selected subset of presented content items as well as a sampled set of content items. The content selection model is configured to receive a set of features characterizing a user-content item pair and output a likelihood that the user will interact with the content item. The sampled set of content items may include content items that were not selected for display based on their likelihoods in addition to those that were selected, and may represent a wider distribution of user-content item pairs than the selected subset. By incorporating the sampled set of content items as well as the selected subset of content items in the training process, the online system can reduce bias in the content selection process such that content items similar to the unselected subset can also be adequately represented.