Contrastive Learning Embeddings for Search Recommendation Diversity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content recommendation systems struggle to provide diverse and engaging related content recommendations, often resulting in monotone user experiences due to reliance on semantic context alone, which may lead to user disengagement.
Innovation Solution
The approach involves supervising the training of pair embeddings using contrastive learning and pairwise co-click signals, combining a two-tower model architecture with a cascaded multilayer perceptron model to generate more representative learned embeddings, thereby improving recommendation quality and user engagement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If semantic context alone is used for recommendation, then recommendation relevance is improved, but content diversity deteriorates
Solution Approach 1:
The recommendation system is segmented into multiple independent towers (query tower and content tower) that process different features separately. The query tower handles semantic context while the content tower handles diversity through contrastive learning, allowing each component to optimize its specific function without compromising the other.
Solution Approach 2:
The system changes the parameter space by introducing contrastive loss training that operates on embedding vectors. By transforming the recommendation problem into a contrastive learning problem in embedding space, the system can simultaneously optimize for relevance (through co-click signals) and diversity (through contrastive loss that pushes apart similar items).
2Measurement precision
If semantic context alone is used for recommendation, then recommendation accuracy is improved, but user engagement deteriorates
Solution Approach 1:
The system incorporates feedback through contrastive loss training that uses co-click signals as ground truth. User interaction data (co-clicks) is fed back into the training process to continuously refine the embedding representations, ensuring that recommended content not only appears accurate but also maintains user interest over time.
Solution Approach 2:
By changing the training objective to contrastive loss that incorporates user engagement signals, the system transforms the optimization target from mere accuracy to accuracy-plus-engagement. The contrastive loss function penalizes recommendations that are accurate but boring, while rewarding those that are both accurate and engaging.
3Measurement precision
If contrastive learning with pairwise co-click signals is applied, then recommendation quality is improved, but model complexity increases
Solution Approach 1:
The complex recommendation task is segmented into two independent towers with specific functions. The query tower focuses on semantic understanding while the content tower focuses on contrastive learning for diversity. This segmentation allows each tower to be optimized independently, reducing the overall complexity compared to a monolithic approach.
Solution Approach 2:
The contrastive learning framework serves multiple functions simultaneously: it trains embeddings for relevance, generates diversity through contrastive loss, and incorporates user feedback through co-click signals. This multi-functionality reduces the need for separate specialized components, thereby managing complexity while improving quality.
Data Source
AI summary
Aspects of the present disclosure relate to a providing related content recommendations in response to a user search query by supervising the training of pair embeddings using contrastive learning and pairwise co-click signals. The approach combines a two tower model architecture with a cascaded multilayer perceptron model to enable the adoption of variable combinations of input features and more representative learned pair embeddings. The learned embeddings undergo supervised contrastive loss training to generate a related content recommendation model, which is subsequently evaluated using both online and offline metrics. The related content recommendation model can provide results to search queries that improve recommendation quality and increase user engagement, thereby ultimately enhancing long term user experience.