Multi-source data modeling and intelligent recommendation system for visual communication based on attention mechanism

View at publisher Summarise with VIRAT
Authors: Jiaqi Liang
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

To address issues such as semantic fragmentation of multi-source heterogeneous data and difficulties in capturing user intent within visual communication design scenarios, this paper proposes a multi-source data modeling framework for visual communication (DHA-VDM) that integrates a dynamic hierarchical attention mechanism. Firstly, a differential gated dynamic weight allocation strategy is constructed to adaptively balance the attention contribution of users' historical behavior and real-time query, and alleviate the static weight rigidity; Secondly, a cross-modal semantic alignment module based on contrastive learning is designed, and the consistent mapping of text description and visual features in public semantic space is realized through joint projection and symmetrical contrast loss. Thirdly, the ontology library of design domain is introduced as external knowledge, and the concept similarity bias is embedded in attention calculation to suppress the interference of irrelevant features. The system is tested on Behance 50 K and Adobe 30 K data sets, and Recall@10 reaches 0.258, which is 41.0% higher than the best baseline VBPR. The ablation experiment and visual analysis verify the synergistic effectiveness of dynamic attention, cross-modal alignment and domain knowledge enhancement.

Fields of Study & Subjects

Recommender Systems and Techniques Multimodal Machine Learning Applications Information Retrieval and Search Behavior

Publication details

Journal / source
IET conference proceedings.
Volume
2025
Issue
47
Pages
1045–1050
Publication date
1 March 2026
Language
EN