Multimodal recommendation system and method based on denoising and interest alignment

By using denoised graph convolution and user interest alignment modules, the problems of noise sensitivity and user interest ignoring in existing technologies are solved, achieving more efficient multimodal recommendation.

CN121542507BActive Publication Date: 2026-07-21SHANGHAI JIAOTONG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
SHANGHAI JIAOTONG UNIV
Filing Date
2025-11-17
Publication Date
2026-07-21

AI Technical Summary

Technical Problem

Existing recommendation technologies are sensitive to noisy samples, ignore user interests, and fail to make sufficient use of information in multimodal scenarios, resulting in a decline in recommendation performance.

Method used

We employ a denoising graph convolutional module and a user interest alignment module. Through two-stage denoising and explicit semantic alignment, we enhance the semantic distance between user and product modal features. We utilize a lightweight graph convolutional network for item embedding representation and combine it with Bayesian personalized ranking for recommendation.

Benefits of technology

It improves the robustness and recommendation performance of the model, reduces the negative impact of noise alignment, and enhances information utilization in multimodal scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121542507B_ABST
    Figure CN121542507B_ABST
Patent Text Reader

Abstract

A multi-modal recommendation system and method based on denoising and interest alignment, comprising a denoising graph convolution module, a user interest alignment module and a prediction and ranking module, wherein: the denoising graph convolution module carries out two-stage denoising on the edge and node levels to obtain enhanced representation; the user interest alignment module aligns the semantics of the user's interest and the modal of the article through contrast learning, obtains aggregated modal features after feature projection and aggregation, and realizes semantic alignment; the prediction and ranking module constructs an article semantic graph, outputs an article semantic embedding representation through a lightweight graph convolution network, obtains the final representation of the article by integrating the enhanced representation and the aggregated modal features, and calculates the Bayesian personalized ranking (BPR) to obtain the predicted score and ranking of each user for each article. The application removes noise on the structure and representation levels, obtains robust embedding, reduces the phenomenon of modal alignment but interest deviation by explicitly narrowing the semantic distance between user interest and commodity modal features.
Need to check novelty before this filing date? Find Prior Art