加權排序對比回歸學習用於資料不平衡之社群媒體熱門度預測

Abstract

社群貼文的熱門程度往往反映受眾對於內容的喜愛程度，社群網紅或廠商可透過觀察貼文的按讚數變化，制定更有效的行銷策略，進而提升社群行銷的成效。因此，如何準確預測社群貼文的熱門度是一大關鍵。然而，現實世界的社群媒體資料具備不平衡的特性，極冷門與極熱門的貼文往往只具備很少的資料量，造成預測熱門度時的失準。有鑒於近年來對比學習在特徵學習的成功，以及將對比學習概念引進回歸任務的新興趨勢，本研究提出加權排序對比迴歸(Weighted-Rank Contrastive Regression) 損失函數，以解決真實世界的回歸問題中數據不平衡的問題。我們在社群媒體熱門度預測資料集 (B. Wu et al., 2019) 上進行實驗，實驗結果顯示我們的方法優於傳統方法 (僅以 L1 損失函數進行擬合) 和當前的最先進的對比迴歸方法 Rank-N-Contrast (Zha et al., 2024)，尤其在處理高熱門貼文、缺乏負樣本數的異常值方面表現更佳。本研究所提出的加權排序對比迴歸損失函數不僅解決了社群媒體熱門度預測中的數據不平衡問題，更提供了一種可泛化的特徵學習方法，可推廣至其他任意的不平衡迴歸任務中。

Social Media Popularity Prediction (SMPP) is the task of forecasting the level of engagement a social media post will receive. In SMPP, it is crucial for understanding audience engagement and enabling targeted marketing strategies. The popularity of social media posts often reflects the audience’s preference for the content. Social media influencers or brands can design more effective marketing strategies by observing the changes in the number of likes on posts, thereby enhancing the effectiveness of social media marketing. However, the inherent imbalance in real-world social media data, where certain popularity levels are underrepresented, posed a significant challenge. In this study, we leveraged the recent success of contrastive learning and its growing integration into regression tasks, and introduced a weighted-rank contrastive regression loss to counteract the data imbalance challenges. Experiments on the Social Media Prediction Dataset (B. Wu et al., 2019) demonstrated that our method outperformed the vanilla approach (solely fit on L1 loss) and the current state of the art (SOTA) contrastive regression approach Rank-N-Contrast (Zha et al., 2024) , especially for challenging outliers with high popularity and few negative counterparts. The proposed weighted-rank contrastive regression loss not only addressed the inherent data imbalance in SMPP but also offered a robust representation learning solution that could be generalized to other real-world imbalanced regression tasks.