Scholarlink

<p>Social-commerce influencer content increasingly combines portraits, product displays, brand symbols, promotional copy, OCR text, and platform interaction signals, making value assessment and risk warning a visual computing problem rather than only a marketing analytics problem. Existing studies usually rely on questionnaire variables, global image-text features, or interaction statistics and therefore provide limited auditable evidence for explaining why a post is valuable, over-commercialized, or risky. This paper proposes IP-VisionNet, a visual cue-guided multimodal network for interpretable value-risk assessment of social-commerce influencer content. The model reorganizes people, products, logos, on-screen text, scenes, and styles into a structured visual cue matrix, learns separate value and risk paths on the same cue set, aligns them with post-text, OCR text, interaction statistics, and contextual metadata, and predicts brand identification, source trust, perceived over-commercialization, authenticity skepticism, purchase intention, the content risk-adjusted value index, and warning level. Experiments are conducted on a self-built multimodal dataset containing 460 valid social-commerce posts, using 368 samples for training and 92 samples for testing under a unified 8:2 setting. IP-VisionNet achieves a Macro-F1 of 78.46 for four-level warning classification, which is 2.38 percentage points higher than the strongest compared baseline, and obtains lower RMSE on value, risk, purchase intention, and content risk-adjusted value estimation. The explanation module provides path heatmaps, OCR-region evidence, occlusion sensitivity, and high-risk retrieval results, allowing risk judgments to be traced back to observable visual and textual cues. The repository at <a href="https://github.com/Yanhe-ai/SCPFAN">https://github.com/Yanhe-ai/SCPFAN</a> provides the source code, preprocessing pipeline, model configuration, train/test split protocol, numerical result files, and reproducibility instructions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

IP-VisionNet: visual cue-guided multimodal value-risk assessment for social-commerce influencer content

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • He Yan

摘要

Social-commerce influencer content increasingly combines portraits, product displays, brand symbols, promotional copy, OCR text, and platform interaction signals, making value assessment and risk warning a visual computing problem rather than only a marketing analytics problem. Existing studies usually rely on questionnaire variables, global image-text features, or interaction statistics and therefore provide limited auditable evidence for explaining why a post is valuable, over-commercialized, or risky. This paper proposes IP-VisionNet, a visual cue-guided multimodal network for interpretable value-risk assessment of social-commerce influencer content. The model reorganizes people, products, logos, on-screen text, scenes, and styles into a structured visual cue matrix, learns separate value and risk paths on the same cue set, aligns them with post-text, OCR text, interaction statistics, and contextual metadata, and predicts brand identification, source trust, perceived over-commercialization, authenticity skepticism, purchase intention, the content risk-adjusted value index, and warning level. Experiments are conducted on a self-built multimodal dataset containing 460 valid social-commerce posts, using 368 samples for training and 92 samples for testing under a unified 8:2 setting. IP-VisionNet achieves a Macro-F1 of 78.46 for four-level warning classification, which is 2.38 percentage points higher than the strongest compared baseline, and obtains lower RMSE on value, risk, purchase intention, and content risk-adjusted value estimation. The explanation module provides path heatmaps, OCR-region evidence, occlusion sensitivity, and high-risk retrieval results, allowing risk judgments to be traced back to observable visual and textual cues. The repository at https://github.com/Yanhe-ai/SCPFAN provides the source code, preprocessing pipeline, model configuration, train/test split protocol, numerical result files, and reproducibility instructions.