Recommendation Device Using Story Generation for Unexpected Content
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems lack unexpectedness in presenting information to users, as they primarily focus on products related to the content of images users upload, failing to provide novel or stimulating recommendations.
Innovation Solution
A recommendation information presentation device that analyzes user images using machine learning models to generate analysis information, which is then used to create stories that describe fictitious events, allowing for the selection of recommendation information from a database that matches the story's themes, thereby presenting unexpected yet relevant recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If recommendation information is generated based on simple image content analysis, then the system is easy to operate and quick to process, but the recommendation information lacks unexpectedness and novelty
Solution Approach 1:
The recommendation system is segmented into multiple independent modules: image analysis module, story generation module (with motif selection and sentence generation), and recommendation information generation module. This segmentation allows each module to specialize in one function, improving overall system adaptability while maintaining ease of operation through modular design.
Solution Approach 2:
A story generation module is introduced as an intermediary between image analysis and recommendation generation. This intermediary transforms simple image content into rich contextual narratives, adding unexpectedness and adaptability to recommendations without requiring users to directly interact with complex algorithms.
2Adaptability or versatility
If a complex machine learning model for story creation is used, then unexpectedness and novelty in recommendations are improved, but device complexity increases
Solution Approach 1:
The complex story generation process is segmented into discrete steps: motif selection, sentence generation, and story assembly. Each step uses a specialized sub-model, distributing computational complexity across multiple simpler components rather than requiring one monolithic complex model.
Solution Approach 2:
Motifs are pre-selected and prepared before story generation begins. Image analysis results are pre-processed into structured formats suitable for story generation. These preliminary actions reduce the complexity of the main generation process by preparing data in advance.
3Measurement precision
If personalized story generation is implemented, then recommendation information becomes more relevant to user personality, but processing time increases
Solution Approach 1:
The system generates a limited number of candidate stories (e.g., 3-5) rather than exhaustively generating all possible stories. This partial action approach provides sufficient personalization precision while significantly reducing processing time compared to complete generation.
Solution Approach 2:
User personality traits are extracted and stored from image analysis results in advance. When generating recommendations, the system retrieves pre-extracted personality information rather than re-analyzing images, reducing processing time while maintaining precision.
4Measurement precision
If multiple types of analysis information are generated from images, then the quality and personalization of recommendations improve, but the quantity of data to be processed increases
Solution Approach 1:
The system extracts only the most relevant analysis information for story generation, such as key personality traits and dominant image themes. Less critical data is excluded or summarized, reducing data quantity while maintaining recommendation precision.
Solution Approach 2:
Image analysis results are pre-processed and filtered before being passed to story generation. Redundant or less important features are removed in advance, reducing the data burden on subsequent processing stages while preserving essential information for personalized recommendations.
Data Source
AI summary
Provided are a recommendation information presentation device, an operation method of a recommendation information presentation device, and an operation program of a recommendation information presentation device capable of presenting recommendation information filled with unexpectedness to a user. A CPU of an image management server includes a second analysis unit, a creation unit, an information acquisition unit, and a distribution control unit. The second analysis unit analyzes an image to generate analysis information. The creation unit inputs the analysis information into a model for story creation and causes a story configured of a set of sentences describing a fictitious event based on the analysis information to be output from the model for story creation. The information acquisition unit selects recommendation information according to the story. The distribution control unit presents the recommendation information to the user.