Decayed Popularity Content Recommendation Method
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content recommendation techniques, such as matrix-based approaches, are inefficient due to high resource requirements and fail to accurately account for popularity variations, leading to skewed similarity values and insensitivity to changes in item popularity.
Innovation Solution
A method that uses decayed popularity scores to determine similarity between content items, where popularity is adjusted over time to prioritize recent events, and pair counts are weighted by individual item popularities to correct for dominant items, generating recommendations based on these calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large matrix is formed to store user-item purchase data for recommendations, then similarity measurements between users or items can be performed, but storage and processing resources are excessively consumed
Solution Approach 1:
The patent extracts only the necessary data elements (item pairs and their co-purchase counts) from the complete user-item matrix, storing merely the essential information needed for similarity calculation rather than the entire matrix structure. This extraction approach maintains measurement capability while dramatically reducing storage requirements.
Solution Approach 2:
Instead of organizing data by users and items in a matrix format, the patent inverts the approach by organizing data around item pairs and their co-occurrence statistics. This inversion transforms the data structure from user-centric to item-centric, enabling efficient similarity computation without requiring full matrix storage.
2Ease of manufacture
If simple purchase count is used to determine item similarity, then calculations are straightforward, but highly popular and unpopular items produce skewed similarity values
Solution Approach 1:
The patent transforms the raw purchase count parameter into a normalized similarity metric by dividing the co-purchase count by the product of individual item frequencies. This parameter transformation adjusts the measurement scale to account for item popularity variations, correcting skewness while preserving computational efficiency.
Solution Approach 2:
The patent applies a correction factor that goes beyond simple counting by incorporating individual item frequency information. This partial correction approach adjusts for popularity bias without requiring complete reprocessing of all user data, achieving improved accuracy with minimal additional computational effort.
3Productivity
If traditional similarity methods are used, then recommendations can be generated, but the system is not sufficiently sensitive to changes in item popularity over time
Solution Approach 1:
The patent implements dynamic tracking of item frequencies and co-purchase counts, allowing the similarity metrics to automatically adapt as new purchase data arrives. This dynamic approach enables the system to respond to changing popularity trends in real-time while maintaining efficient recommendation generation through incremental updates rather than full recalculations.
Data Source
AI summary
Recommending content items is disclosed. In some embodiments, determining an item to recommend includes receiving an indication that a first item is of interest and selecting a second item to recommend based at least in part on a degree of similarity of the second item to the first item as determined based at least in part on a number of other instances in which both the first item and the second item have been of interest and a first decayed popularity of the first item. In some embodiments, the degree of similarity of the second item to the first item is determined based in addition on a second decayed popularity of the second item.