Cross-Type Recommendation Using a Shared Media Feature Space

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing recommendation algorithms for different types of media content, such as live streaming and short videos, operate in separate spaces and fail to utilize user behavior signals from one type to enhance recommendations in another, leading to inaccurate and unsatisfactory personalized suggestions, especially for new users.

Innovation Solution

Mapping media content types like short videos and live streaming into a shared feature space using a joint encoder, incorporating user interaction data from one type to inform recommendations for the other, and training a content understanding model to cluster and recommend content based on shared features and user behavior.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate recommendation algorithms are used for different media types, then each algorithm can be optimized for its specific media type, but user behavior signals from one media type cannot be utilized to enhance recommendations for another media type

Engineering Contradiction:
Improverecommendation accuracyVSAvoidcross-media recommendation capability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent merges separate recommendation algorithms for different media types into a unified recommendation system. This is achieved by integrating multiple media type processors that can handle different media formats (videos, images, audio) within a single system framework, allowing user behavior signals from any media type to be captured and utilized across all recommendation scenarios. The unified system enables cross-media recommendation while maintaining the ability to process media-specific features.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The recommendation system is designed with universal components that can process and analyze user interactions across multiple media types. The user behavior signal processor and feature extractor are configured to handle diverse interaction patterns (watching, liking, sharing, commenting) regardless of the media type, creating a multi-functional system that adapts to different content formats while maintaining consistent recommendation capabilities.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If separate feature spaces are used for different media types, then each media type can maintain its unique characteristics, but personalized recommendations become inaccurate especially for new users

Engineering Contradiction:
Improvemedia type differentiationVSAvoidpersonalized recommendation accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent combines multiple media type feature spaces into a unified feature space. The feature extractor is configured to extract features from different media types and map them into a common feature representation that preserves media-specific characteristics while enabling cross-media comparison. This unified feature space allows the system to leverage user behavior patterns across media types for more accurate personalized recommendations.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified feature space functions as a composite structure that integrates features from multiple media types. Different media-specific features (visual features from videos, audio features from audio content, interaction features from user behaviors) are combined into a composite feature representation that maintains the unique properties of each media type while enabling holistic user preference analysis.

Inventive Principle:
Principle #40Composite materials

Data Source

PatentEP4621657A1Method, apparatus, electronic device and program product for cross-type recommendation
Publication Date: 2025.09.24 BEIJING YOUZHUJU NETWORK TECH CO LTD
  • EP4621657A1 patent drawingFigure 1
  • EP4621657A1 patent drawingFigure 2
  • EP4621657A1 patent drawingFigure 3

AI summary

Embodiments of the present disclosure relate to a method (200), an apparatus (800), an electronic device (900), and a computer program product for cross-type recommendation. The method (200) includes determining (202) a first type of media content interacted with a user, and determining (204) a first media feature of the first type of media content. The method (200) further includes recommending (206) a second type of media content to the user based on the first media feature, wherein the first type of media content and the second type of media content belong to different types of media content and share a same feature space. According to the embodiments of the present disclosure, it can facilitate the use of the behavior signal of the user on the side of the first type of media content, and improve the accuracy of the recommendation and the satisfaction of the user.