基於用戶影響之社交媒體流行度預測模型

Abstract

許多企業和網紅利用社交媒體來推廣其品牌形象、產品、想法和概念，社交網站上有大量的貼文，但並非每個貼文都會引起關注。因此，企業和網紅有迫切的需求想知道何種內容更有可能受到關注。然而，既有的社交媒體流行度預測模型，常常忽略用戶的影響，意即由同一用戶發布的貼文常具有類似的流行度。因此，在本研究中，我們提出一個基於用戶影響的社交媒體流行度預測框架，稱為UIA (User-Influence-Aware)，以克服過往模型的缺點。我們提出的框架包含四個模組，分別為特徵萃取、變壓器、對比學習和預測。首先，我們開發特徵萃取模組來提取視覺、標題、標籤、常用標籤分佈、貼文和用戶特徵，為了更好地整合所提取的特徵，我們應用共注意力變換器來學習視覺和標題特徵之間、標籤和常用標籤分佈特徵之間以及貼文和用戶特徵之間的交互關係，我們也利用MiniGPT-4模型生成圖像標題，以豐富貼文的文本特徵。接著，我們使用變壓器進一步強化特徵間的交互關係。然後，我們採用對比學習模組，將流行度相似的貼文表示式拉近，並將流行度差異大的貼文表示式推遠。最後，我們運用預測模組預測每個貼文的流行度。另外，我們採用協作學習框架來增加模型的泛化能力和對噪音數據的穩健性，同時，我們引入偽損失來考慮用戶的影響，以幫助模型學習每個用戶的貼文流行度分佈。實驗結果顯示，我們所提出的模型在斯皮爾曼等級相關係數和平均絕對誤差方面均優於最先進的模型。我們的模型可以幫助企業和網紅預測貼文的流行度，有效識別更有可能受關注的社交媒體內容，從而提高其貼文的能見度和觀眾的關注度。

Many businesses and influencers utilize social media platforms to promote their brand image, products, ideas, and concepts. Despite a massive number of posts being uploaded, not every post gets attention. Therefore, it is essential and desirable for businesses and influencers to identify social media contents that are more likely to become popular. Many models have been proposed for social media popularity prediction (SMP). However, they overlook the user influence that the posts made by the same user have similar popularity. Therefore, in this study, we propose a novel User-Influence-Aware framework for social media popularity prediction, called UIA, to overcome the weaknesses of previously proposed models. The proposed framework contains four modules, namely feature extraction, transformer, contrastive learning, and prediction. First, we develop the feature extraction module to extract the visual, caption, tag, frequent tag distribution, post, and user features. To better integrate the extracted features, we apply the Co-Attention Transformer to learn the interaction relationships between visual and caption features, between tag and tag distribution features, and between post and user features. Also, we exploit the MiniGPT-4 model to generate image captions to enrich the textual features of posts. Second, we use the Transformer to learn the inter-relationships among the features learned by the Co-Attention Transformer. Third, we employ the contrastive learning module to pull closer the representations of the posts with similar labels and push farther apart those with dissimilar labels. Finally, we utilize the prediction module to predict the label of each post by using the representations learned by the contrastive learning module. In addition, we adopt the collaborative learning framework to increase model generalization and robustness to noisy data. Also, we introduce the pseudo losses to consider the user influence and help our model learn the popularity distribution of each user. The experimental results demonstrate that the proposed framework outperforms the state-of-the-art models in terms of Spearman's rank correlation and mean absolute error. Our model can assist businesses and influencers in predicting the popularity of posts and effectively identifying the social media contents that are more likely to become popular, which in turn increases their post visibility and audience attention.