Trending Discussion Topic Prediction on Thai Internet Forum Pantip

This study presents a method for predicting trending topics on Pantip, a popular Thai online forum. Our goal is to enhance Pantip's recommendation system by integrating Temporal Fusion Transformers (TFT), a time-series forecasting model, with BERTopic, a framework for topic modeling. The main contribution of this work is the development of a Conversation Graph and a Trending Discussion (TD) score. The Conversation Graph is constructed by analyzing both the main post and comments within a thread, capturing the discussion's structure and flow. Assuming that valuable discussions emerge naturally from the original post, we use read counts from clickstream data to compute the TD score, which serves as a key predictor. Our model achieves a precision of 0.553 in aligning with Pantip Trend topics manually selected by moderators. Experimental results reveal varying performance depending on the type of discussion: higher precision (0.773) in focused, topic-driven rooms, and lower precision (0.360) in more casual, everyday conversation rooms.