運用多模態深度學習方法針對品牌進行微網紅排名

Abstract

在網際網路時代，人們容易受到社群媒體上追隨的網紅影響，而追隨某些潮流，尤其是擅長使用社交網絡的年輕用戶。因此，「網紅行銷」逐漸取代傳統行銷，成為品牌在社群媒體上行銷的重要策略與手段，「網紅行銷」泛指品牌與網紅合作，以在社群平台上張貼贊助的貼文等方式來促銷產品或服務。 在網紅之中，「微網紅」泛指擁有較少追蹤者，並在特定領域擁有專業知識的網紅。由於「微網紅」在特定領域的專精、高互動性的追蹤者、較高的顧客信任度及高性價比，越來越多品牌選擇進行「微網紅行銷」，以觸及理想的目標客群。品牌為了達到成功的「微網紅行銷」，如何選擇適當的網紅便成為重要的問題。過去的研究大多關於「網紅偵測」，試圖由許多社群媒體使用者中找出具有影響力的網紅；較少針對各品牌，分別推薦其相關的網紅。 本研究提出一個多模態深度學習方法，根據品牌在社交媒體上發表的內容，預測網紅排名。我們進一步加入先前未利用的特徵（如帳號互動、主題標籤資訊等），以及特定的負樣本抽樣策略來提升排名模型的表現。根據實驗結果，我們提出的方法在整體評估指標上，展示了其優於現有表現最佳的方法。

In the age of the Internet, individuals are significantly influenced by the content creators (i.e., influencers) they follow on social media platforms, particularly young users who are adept at navigating social networks. Therefore, influencer marketing has emerged as a powerful replacement for traditional marketing and has become a crucial strategy for brands to execute their marketing campaigns. “Influencer marketing” denotes the collaboration between brands and influencers to promote products or services primarily through sponsored posts or other ways on social platforms. Among influencers, micro-influencers are defined as those influencers with fewer followers but with high-level expertise on distinct topics. Due to their specialized knowledge, high engagement followers, high customer trust, and cost-effectiveness, an increasing number of brands are opting for micro-influencer marketing to effectively reach their target audiences. For brands to achieve successful micro-influencer marketing, the selection of suitable influencers becomes a critical prerequisite. Previous studies predominantly concentrate on influencer detection for identifying influencers within a pool of social media users. Fewer studies consider individual brands and recommend micro-influencer rankings for each brand. In this research, we propose a multimodal deep learning approach to predict influencer ranking for given brands based on their content on social media. We further incorporate previously unexploited features such as account interactions and hashtag information, along with employing a specific negative sampling strategy to enhance ranking performance. Experimental results provide empirical evidences of the effectiveness of our proposed method, outperforming the state-of-the-art methods across various evaluation metrics.