基於異構網路的產品中心化網紅推薦框架

Abstract

隨著社群媒體的興起，網紅行銷已成為品牌推廣產品的常見策略。近年來的研究也越來越聚焦於利用機器學習來幫品牌找到適合的網紅。然而，品牌在推出新產品時，往往面臨到不知道該找誰業配的困境，這也可能進一步影響到行銷活動的成效。此外，目前大多數研究都集中於分析網紅的個人檔案，或是網紅與品牌之間的適配度，卻忽略了被推廣產品在網紅推薦中的角度。因此本研究構建了一個異構網路，整合了品牌、產品和網紅的數據，每個產品會與推出它的品牌連 結，也會跟曾經業配過它的網紅連結。我們的模型會先透過鄰居的特徵聚合生成該節點的向量。接著利用節點向量來計算產品與網紅之間的連結機率，最後再透過排序來取得 top-k 的推薦網紅。此方法在網紅推薦上展現出更好的表現。此外，大多圖模型應用都是 transductive learning，沒辦法預測沒看過的點，而本研究也有納入 inductive learning 的實驗，確保我們的模型在實際應用上，不僅可以有好的表現，也可以針對新的產品提供選項，有效的解決網紅推薦中的冷啟動問題。我們用公開的資料集驗證了模型的有效性，為品牌提供有價值的商業見解。

With the rise of social media, influencer marketing has become a common strategy for brands to promote their products. Recent studies have increasingly focused on using machine learning to identify influencers for brands. However, when brands launch new products, they often face challenges in finding suitable influencers. This difficulty can hinder the effectiveness of their marketing campaigns. Moreover, most studies concentrate on analyzing influencer profiles or the compatibility between influencers and brands, neglecting the perspective of the promoted product in influencer recommendations. This study addresses this gap by constructing a heterogeneous graph that integrates brand, product, and influencer data. Each product is connected to the brand that launched it and also has an edge to the influencer who has promoted it before. Our model first generate embeddings for each node by aggregating neighbors’ features. It then utilizes the node embeddings to calculate the probability of connections between product and influencer nodes. Finally, the top-k influencer recommendations are obtained by ranking these probability scores. This approach has demonstrated outperformed performance in influencer recommendations. Additionally, while most graph-based applications rely on transductive learning, which cannot predict unseen nodes, our study also includes inductive learning experiments. This ensures that our method not only performs well in real-world applications but also provides options for new products, effectively solving the cold-start problem in influencer recommendation. The effectiveness of our model is validated using public datasets, offering valuable business insights for brands.