Visual Recommendation Modeling for Content-Interface Click Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current news recommendation systems fail to consider the influence of the news interface on user engagement, limiting the increase in click-through rates.
Innovation Solution
A recommendation method that predicts user preference for both content and interface by using image feature data, including global and local visual impressions, through models like Bert and self-attention mechanisms, to select and recommend content and interfaces that align with user interests.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a current news recommendation system only explores news content that a user is interested in, then the system can identify user preferences, but the click-through rate of news cannot be further increased due to ignoring the influence of news interface
Solution Approach 1:
The patent merges content recommendation with interface recommendation into a unified recommendation system. The prediction model simultaneously predicts user preferences for both news content and news interface layouts by processing images that contain both elements, allowing the system to recommend optimal content-interface combinations that together maximize click-through rate rather than treating them as separate factors
Solution Approach 2:
The patent introduces a new dimension to the recommendation problem by incorporating interface visual features (global visual impression features and local visual impression features) as additional predictive factors beyond traditional content features. This transforms the recommendation from a one-dimensional content-matching problem to a two-dimensional problem that simultaneously considers content preferences and interface presentation preferences
Data Source
AI summary
Embodiments of this application disclose a recommendation method includes: obtaining a plurality of images, where each image includes one candidate interface and one type of candidate content presented by using the candidate interface; obtaining image feature data of each image; determining input for a prediction model based on user feature data of a target user and the image feature data, and then predicting a degree of preference of the target user for each image by using the prediction model; and finally, selecting, based on the degree of preference, a candidate interface and/or candidate content from candidate interfaces and candidate content that are included in the plurality of images, and then performing recommendation to the user based on the selected candidate content or candidate interface.