Content Similarity Modeling for Metadata-Rich Streaming Recommendations

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing content streaming systems struggle to effectively recommend similar content based on detailed text metadata, such as titles, synopses, genres, and hashtags, limiting the accuracy of content recommendations.

Innovation Solution

A method and device that utilize a language model trained on masked language models (MLM) to predict synopsis and hashtag information, followed by secondary training to enhance the model's ability to determine similarity between content items, using cosine similarity algorithms to calculate vector similarities based on average pooling of hidden layer vectors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a language model is trained using masked language modeling on synopsis and hashtag information, then the accuracy of content similarity determination is improved, but the device complexity and training time increase

Engineering Contradiction:

Improvecontent similarity determination accuracyVSAvoidlanguage model training complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The training process is divided into two distinct stages: first training on synopsis information to capture detailed content semantics, then training on hashtag information to capture categorical characteristics. This segmentation allows the model to learn different aspects of content similarity separately, improving overall accuracy while managing computational complexity through staged learning.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system performs preliminary training on synopsis information before training on hashtag information. This preliminary action establishes a foundation of detailed content understanding that enhances subsequent hashtag-based training, allowing the model to leverage both detailed semantics and categorical labels for improved similarity determination.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If detailed text metadata (synopsis, hashtags, genre) is used for content recommendation, then recommendation accuracy is improved, but the processing time and computational resources increase

Engineering Contradiction:

Improverecommendation accuracyVSAvoidprocessing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The metadata processing is segmented into multiple training phases where the model learns to represent different types of metadata (synopsis, hashtags, genre) separately. This allows the system to efficiently process and combine these different metadata types during inference without treating them as a single complex processing task.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system extracts key information from detailed text metadata and represents it as condensed vector embeddings through the language model. By extracting and compressing the essential semantic information into vectors, the system reduces the amount of raw data that needs to be processed during recommendation while maintaining high accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

3Adaptability or versatility

If a two-step learning process is implemented for language model training, then the model's ability to capture both synopsis and hashtag information is improved, but the training duration increases

Engineering Contradiction:

Improvemodel's ability to handle multiple metadata typesVSAvoidtraining duration

Core Design Contradiction:

Adaptability or versatilityVSDuration of action of moving object

Solution Approach 1:

The training process is segmented into two sequential steps: first training on synopsis information to capture detailed content semantics, then training on hashtag information to capture categorical characteristics. This segmentation enables the model to learn different aspects of content representation separately, improving versatility while managing training duration through structured learning phases.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system performs preliminary training on synopsis information before training on hashtag information. This preliminary action establishes a foundation of detailed content understanding that enhances subsequent hashtag-based training, allowing the model to efficiently incorporate multiple metadata types through a structured learning sequence.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20250371070A1Method and device for providing similar content in content streaming system

Publication Date: 2025.12.04 TVING CO LTD

AI summary

Provided are a method and device for providing similar content in a content streaming system. A method of operating a server in a content streaming system may comprise obtaining first sequence-type text data including information included in first metadata of a first content item, obtaining second sequence-type text data including information included in second metadata of a second content item, determining a first vector corresponding to the first sequence-type text data and a second vector corresponding to the second sequence-type text data using a language model learned based on synopsis information included in metadata of content items, determining similarity between the first content item and the second content item using the first vector and the second vector, and providing a content list including at least one content item including the second content item selected based on the similarity.