運用內容相似度與大型語言模型提升轉推預測之表現

Abstract

近年來，人們習慣透過社群平台接收資訊或表達自身觀點，其中在 Twitter (X.com) 平台上的「轉推」行為，是指分享一則已發佈的訊息，不僅可用以傳達使用者立場，亦有助於強化個人觀點。轉推預測任務作為社群媒體中探討資訊傳播的重要研究方向，旨在提升預測準確率，並藉由分析轉推行為，深入了解使用者偏好與其背後的決策因素。為了提升預測表現，許多研究提出深度學習模型應用於轉推預測任務。與傳統機器學習方法相比，深度學習不僅免除人工特徵工程的繁瑣程序，亦能顯著提升預測效果。隨著近年來大型語言模型（LLM）的快速發展，其於文本理解、摘要生成與推理等方面展現出強大能力，並廣泛應用於各種自然語言處理領域。然而，目前針對大型語言模型在轉推預測任務中的應用仍屬少見，有待進一步探討與發展。 本文聚焦於以內容為基礎的使用者轉推行為預測，利用使用者與目標推文作者的歷史發文紀錄，透過分析目標推文與這些發文紀錄之間的相似度，以預測使用者是否會轉發該目標推文。本研究提出一個創新的預測框架，結合可進行輸入權重分析的深度學習模型，識別對預測結果影響最顯著的輸入資料，並據此調整大型語言模型的提示指令，以提升其在轉推預測任務中的表現。此方法對未來將大型語言模型應用於類似任務具有重要的參考價值與貢獻。此外，本文所提出的基於內容相似度的深度學習模型具備簡化的模型架構，能以直觀的方式進行特徵歸因分析，同時展現出優異的預測效能與執行效率，為相關領域提供一個兼具解釋性與效能的實用解決方案。

In recent years, social media platforms have become central to how people receive information and express opinions. On Twitter (X.com), retweeting—sharing an existing post—serves both to express user stance and reinforce personal views. Retweet prediction is a key research area in understanding information diffusion, aiming to improve accuracy and reveal user preferences and decision-making factors. Deep learning models have been widely adopted for this task, offering superior performance over traditional machine learning by eliminating the need for manual feature engineering. With the rapid development of large language models (LLMs), their capabilities in text understanding and reasoning have been applied across various NLP tasks. However, their use in retweet prediction remains underexplored. This study focuses on content-based retweet prediction, using the posting histories of both the user and the tweet author to analyze similarity with the target tweet. We propose a novel framework that combines a deep learning model capable of input weight analysis with prompt refinement for LLMs, improving their predictive performance. This approach offers valuable insights for applying LLMs to similar tasks. Additionally, our proposed Similarity-Based deep learning model features a simplified architecture that enables intuitive feature attribution, strong prediction performance, and efficient execution—making it a practical and interpretable solution for related research.