Latent Vector Content Selection with Limited Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing online systems face challenges in accurately predicting user interest in new content items due to limited feedback, leading to inaccurate content recommendation models.
Innovation Solution
The system generates latent vector spaces to represent user interactions and content items, allowing for the prediction of user engagement by combining vectors from different object types and content categories, thereby reducing reliance on existing user interactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional content recommendation models are used that rely on feedback signals from user interactions, then the model can be trained to predict user interest, but the model accuracy deteriorates when limited feedback is available for new content items or specific users
Solution Approach 1:
The patent segments the content item representation into multiple latent vectors, each corresponding to different objects associated with the content item (e.g., external page, internal page, video). This segmentation allows the model to leverage feedback data from related objects even when direct feedback for the specific content item is limited, thereby improving prediction accuracy with limited feedback data.
Solution Approach 2:
The patent creates a universal latent vector space that can represent multiple types of objects (external pages, internal pages, videos, etc.) using a common embedding dimensionality. This universal representation allows feedback signals from diverse object types to be integrated and transferred, enabling the model to generalize better when feedback for a specific content item is scarce.
2Reliability
If the model is trained only on direct user interactions with specific content items, then the training data is specific and relevant, but the model cannot effectively predict interactions for new content items without existing feedback
Solution Approach 1:
The patent performs preliminary action by pre-training embedding models on large sets of related objects (external pages, internal pages, videos) before the content item needs prediction. These pre-trained latent vectors capture general patterns of user interactions with similar objects, which can then be quickly applied to new content items without requiring extensive feedback data for each specific item.
Solution Approach 2:
The patent introduces latent vectors as intermediaries between user interactions and content item predictions. Instead of directly modeling user-content interactions, the system uses latent vectors of associated objects as mediators to transfer and aggregate feedback signals, enabling reliable predictions for new content items by leveraging patterns from related objects.
Data Source
AI summary
To select the content to be presented to the user, a first latent vector is determined for a content item based on a first object associated with the content item. A second latent vector is determined for the content item based on a second object associated with the content item. A content item vector is then determined based on the first and second latent vectors. Furthermore, a user vector is determined based on interactions of the user with the first set of content objects and the second set of content objects. A score indicative of the likelihood of the user interacting with the content item is determined based on the content item vector and the user vector.