Content Recommendation Diversity via Relationship Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Content recommendation systems often fail to provide diverse content options, leading to users sticking with familiar content due to reliance on past viewing history, and applying additional constraints can result in performance issues and conflicts with personalization efforts.
Innovation Solution
A method and system that generate alternative content recommendation candidates by using user data and relationship information based on content engagement data from multiple users, expanding user profiles to include diverse metadata that differ from or overlap with existing user preferences, and applying machine learning models to create a diverse set of recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If content recommendation systems rely on past viewing history to match user profiles, then personalization is improved, but content diversity deteriorates
Solution Approach 1:
The patent segments the recommendation process into multiple independent components: a diversity generation module that creates diverse candidate content using relationship information between content metadata, and a selection module that chooses final recommendations. This segmentation allows diversity and personalization to operate in separate stages without interfering with each other.
Solution Approach 2:
The patent introduces relationship information as an intermediary element that connects content metadata without relying on user viewing history. This relationship information captures inherent connections between content items (e.g., genre relationships, thematic connections) and serves as a mediator to generate diverse recommendations independent of user past behavior.
2Object-generated harmful factors
If additional constraints are applied during content recommendation requests to promote discovery, then content diversity is improved, but system performance deteriorates
Solution Approach 1:
The patent performs preliminary generation of diverse content candidates using relationship information before the actual recommendation request is processed. By pre-computing diverse candidates and storing them, the system avoids applying complex constraints during real-time recommendation requests, thus maintaining high performance while still achieving diversity.
Solution Approach 2:
The patent extracts the diversity generation function from the main recommendation processing pipeline. The diversity generation module operates independently using relationship information, separating the diversity concern from the personalization matching process. This extraction allows each module to optimize for its specific function without burdening the other.
3Object-generated harmful factors
If additional constraints are applied during content recommendation requests, then content diversity is improved, but conflicts with personalization efforts occur
Solution Approach 1:
The patent segments the recommendation system into a diversity generation component and a personalization selection component. The diversity module generates candidates using relationship information independent of user history, while the selection module applies personalization filters. This segmentation prevents constraints from one objective from conflicting with the other objective.
Solution Approach 2:
The patent adds a new dimension to the recommendation space by introducing relationship information between content metadata items. This creates an additional dimension orthogonal to the traditional user-profile-matching dimension, allowing diversity to be achieved along a different axis that does not conflict with personalization.
Data Source
AI summary
A computer-implemented method for obtaining one or more recommendation candidates for items of content available via a content distribution system, the method comprising: obtaining user data for a selected user, wherein the user data comprise or represent user activity and/or content metadata associated with user activity; generating or otherwise obtaining relationship information for at least some content metadata associated with the available content; generating further user data from the user data using the relationship information so that the further user data comprises or represents alternative content metadata that are distinct or at most overlap with the content metadata associated with the user activity; performing a content recommendation process using at least the further user data and content information for the available content to generate one or more content recommendation candidates for the user.