Contrastive Learning Media Discovery Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recommender systems face challenges in user exploration and discovery of new media content due to data sparsity, particularly for 'cold-start' items with limited user interactions, leading to echo chambers and reduced user satisfaction.
Innovation Solution
A two-tower model with contrastive learning approach using feature-level and instance-level augmentation to enhance user and episode embeddings, increasing the training set size and improving recommendations for new content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If collaborative filtering systems model user interests based on historical interactions, then recommendation accuracy for known items is improved, but the ability to discover new and cold-start items deteriorates
Solution Approach 1:
The system performs preliminary actions by pre-processing episode features (transcripts, metadata, audio signals) into embeddings before user interactions occur. This allows cold-start episodes to have pre-computed representations that can be immediately used for recommendation, eliminating the waiting period for accumulated interaction data.
Solution Approach 2:
The patent introduces an intermediary contrastive learning module that bridges the gap between content-based episode features and user preferences. This intermediary learns to align episode embeddings with user embedding spaces through contrastive optimization, enabling the system to recommend episodes based on content similarity even without historical interaction data.
2Adaptability or versatility
If recommender systems rely on historical interaction data, then personalization is improved, but user exploration and diversity of recommendations deteriorate
Solution Approach 1:
The recommendation score is segmented into multiple independent components: content-based similarity score, user preference score, and exploration bonus. This segmentation allows the system to balance personalization and exploration by weighting different components, preventing the echo chamber effect while maintaining personalization through the user preference component.
Solution Approach 2:
The system dynamically adjusts the exploration bonus based on episode popularity and user interaction patterns. For less popular episodes, the exploration bonus is higher, encouraging discovery. As episodes gain popularity, the bonus decreases, allowing the system to adaptively balance between recommending known popular content and exploring new content.
3Productivity
If the training set size is small due to limited user interactions with new shows, then model training efficiency is improved, but recommendation performance for cold-start items deteriorates
Solution Approach 1:
The system creates synthetic training examples by copying and augmenting existing episode features. Through contrastive learning, it generates positive pairs (user-episode interactions) and negative pairs (user-non-interactions) from limited actual interaction data, effectively multiplying the training signal without requiring additional user interactions.
Solution Approach 2:
The patent changes the parameter space by transforming raw episode features (transcripts, metadata, audio) into embedding vectors through neural network encoders. This parameter transformation allows the model to learn from fewer examples by operating in a compressed, meaningful representation space where patterns are more readily detectable.
Data Source
AI summary
The various implementations described herein include methods and devices for media discovery. In one aspect, a method includes obtaining a pre-trained recommender model that has been trained using contrastive learning with feature-level augmentation and instance-level augmentation. The method further includes generating, via the model, a user embedding based on features of the user and generating, via the model, a respective episode embedding for each episode of a plurality of episodes, each respective episode embedding based on features of the corresponding episode. The method also includes generating, via the model, a respective similarity score (corresponding to a latent similarity between the user embedding and each respective episode embedding) for each episode, the respective similarity score, and ranking the episodes in accordance with the respective similarity scores. The method further includes recommending the highest ranked episode to the user.