Eye-Tracking Recommendation Using Visual Line Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems struggle to accurately reflect user preferences due to potential erroneous manipulations in user logs and challenges in determining user interests based on visual line motion, particularly in online shopping scenarios.
Innovation Solution
A recommendation device that extracts feature vectors from visual line trajectories, generates user contexts, computes content scores using contextual bandit algorithms, and outputs recommendations based on these scores, without relying on explicit content information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If user logs (clicks, page transitions) are used for recommendation, then recommendation can be implemented, but the data may be affected by erroneous manipulation and cannot sufficiently reflect user preference
Solution Approach 1:
The patent introduces visual line trajectory as an intermediary indicator to indirectly measure user interest. Instead of directly relying on potentially erroneous explicit actions (clicks), the system uses the visual line trajectory - a more reliable proxy that reflects genuine user attention and interest without being susceptible to manipulation or errors in explicit interaction data.
2Measurement precision
If visual line motion is used to determine user interest, then direct measurement of attention is possible, but it is difficult to determine what the user has been watching in online shopping scenarios
Solution Approach 1:
The patent transforms the visual line trajectory data into a different dimensional representation - extracting features such as trajectory length, number of direction changes, and time spent in specific regions. This dimensional transformation converts complex motion data into quantifiable metrics that preserve the essential information about user attention while enabling computational analysis and recommendation.
3Measurement precision
If explicit content information is used for recommendation, then specific user interests can be identified, but the system becomes complex and requires determining what user has been watching
Solution Approach 1:
The patent extracts only the essential features from visual line trajectory data - specifically trajectory length, number of direction changes, and regional time spent - rather than processing the complete detailed motion path. This extraction approach captures the core information needed for recommendation while significantly reducing system complexity and computational requirements.
Data Source
AI summary
A recommendation device includes an acquisition unit acquiring a visual line image obtained by imaging the trajectory of the visual line of a user browsing a content displayed on a user terminal, an extraction unit extracting a feature vector indicating a feature amount of the visual line image, a generation unit generating a correspondence relationship between the content and the trajectory of the visual line of the user as a user context on the basis of the feature vector, a computation unit computing the score of each of a plurality of contents, which are recommendation candidates, by using the user context, and an output unit outputting a recommendation result of a content selected on the basis of the score.