Media Title Search Using Vector Distance Metrics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing media streaming services face challenges in providing relevant search results and content recommendations when records of past user behavior are lacking, as current methods rely heavily on user history, which can be unhelpful in such scenarios.
Innovation Solution
A computer-implemented method using a distance metric in a vector space to quantify similarity between media titles based on metadata tags, correcting for under-tagging and redundant tagging by non-linear scaling and considering statistically significant tags, to generate search results and recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If search results and recommendations rely on past user behavior records, then personalization and relevance are improved, but the system becomes unhelpful when such records are lacking
Solution Approach 1:
The patent introduces metadata tags as an intermediary between media titles and user preferences. Instead of directly relying on user behavior records, the system uses tags (genres, actors, directors, etc.) to represent media titles and computes distances between tagged objects. This intermediary layer enables the system to provide relevant results even when user history is unavailable, as the tagged space allows for similarity computation based on object properties rather than user behavior patterns.
2Loss of information
If the system uses metadata tags to represent media titles, then content understanding is improved, but tagging errors (under-tagging and redundant tagging) reduce accuracy
Solution Approach 1:
The patent transforms the tagging problem by changing the parameter space from binary tag presence/absence to continuous distance metrics in a tagged space. Instead of treating tags as discrete categories, the system computes distances between tagged objects based on their tag compositions, allowing for nuanced similarity measurements that account for partial overlaps and hierarchical relationships between tags. This parameter transformation makes the system robust to tagging imperfections.
Solution Approach 2:
The patent replaces traditional information retrieval mechanics (keyword matching, exact tag matching) with a geometric approach using distance metrics in a tagged space. Instead of mechanically comparing tags for exact matches, the system uses distance computations that naturally handle partial matches, missing tags, and redundant tags by measuring the 'distance' between tag compositions. This substitution of mechanics provides robustness to tagging errors.
3Productivity
If the system computes distances between all tagged objects, then comprehensive search results are generated, but computational complexity increases
Solution Approach 1:
The patent segments the computational problem by organizing objects in a tagged space where distances can be computed efficiently using the structure of tags. Instead of computing pairwise distances between all objects in a brute-force manner, the system leverages the hierarchical and categorical structure of tags to group and organize computations, reducing the overall computational complexity while maintaining comprehensiveness.
Data Source
AI summary
Techniques are described for generating search results and content recommendations using a distance metric on a space of media titles. In one embodiment, each media title may be associated with metadata tags and represented by a vector which indicates those associated tags. The distance metric may measure distance as an angle between vectors representing media titles in a bent vector space that accounts for correlations between tags. Further, a non-linear scaling may be applied to the vectors representing media titles to correct for under-tagging and redundant tagging. Based on the distance metric, a search or recommendation application may generate search results and/or content recommendations and cause the same to be presented to a user.