Understanding Brand's Association: Fine-Grained Multi-Modal Content Discovery for Brands

Shuhan Qi, Chengkai Huang, Yifan Li

Abstract: With the increasing popularity of short video plat-forms, the massive UGC provides high-quality materials for brand advertising. Recently, personalized content discovery methods have been proposed to recommend appropriate posts for brands to attract potential consumers. However, there are some drawbacks of them. Firstly, these methods only use images but ignore the multimodal information, which is not suitable for video-based brand advertising. Secondly, the generated features of brands are too coarse to distinguish the subtle differences from similar brands. In this paper, we propose a fine-grained multimodal content discovery (FGMCD) framework, which uses a large number of multimodal data to recommend informative videos for brands. Further, we designed a new hierarchical model, which adopts fine-grained features of different levels to make distinctions of similar brands exactly. We also collect a high-quality dataset with 50 car brands from Instagram and the experiments show the effectiveness of our method.