匿名社群中去身份化讓使用者更真實地表達情緒，現有文獻較少探討匿名下情緒與熱門的關聯，同時匿名社群熱門機制並不透明，故本研究聚焦於匿名社群探討貼文傳播與熱門機制以彌補上述研究缺口。設計 Dcard 爬蟲機制蒐集不同週期熱門貼文；TM-LDA(Temporal-LDA)主題建模應對社群貼文詞彙稀疏問題；臺灣大型語言模型(TWLLM)與雙重提示(Prompt)以零樣本(Zero-shot)情緒標注；最後分析不同貼文特徵在 Dcard 熱門板傳播特性。為了找出 Dcard 熱門貼文關鍵特徵，用 K-prototypes 分類貼文熱門屬性，並以三種樹狀集成模型(XGBoost、LightGBM、Random Forest)結合 12 種重採樣與不採樣 Baseline 預測貼文熱門屬性，所設計方法達到約 95%整體準確率，再用 SHAP 與 Feature Importance 雙向驗證 11 個貼文特徵在熱門機制的權重與效力，結果顯示每種模型均將貼文進入熱門板所需時長與貼文進入熱門板前互動速率為最關鍵特徵。研究貢獻深入理解匿名社群貼文傳播特徵，所提出的研究架構分析不同情緒、討論話題與其他貼文特徵對三種熱門屬性的影響，為 Dcard 熱門板之運作提供了實證的機制，創作者能根據不同熱門屬性選擇適合的發文時段，精準規劃內容策略並達到預期互動量。同時在驗證集上重複上述聚類與預測流程，驗證集中三聚類分布結果、模型性能評估與特徵重要度排序皆與主資料集高度一致，驗證了流程穩健性與可遷移性。;De-identification in anonymous communities reduces users’ impression management burden and enables more authentic emotional expression. However, existing studies have paid little attention to how emotions drive post popularity in anonymous communities, and the anonymous community popularity mechanism remains underexplored. To fill this gap, this study deployed a custom Dcard web crawler to gather posts and comments across different popular cycles.Then, Temporal LDA (TM-LDA) was applied to sparse-vocabulary characteristics of Dcard posts. Next, zero-shot emotion annotation was performed using a Taiwanese large language model (TWLLM) with a dual-prompt design, and annotation consistency was confirmed via Cohen’s Kappa. Dissemination patterns of different topics and emotions on the Dcard popular board were then analyzed. To identify the key determinants of popularity, posts were clustered into three popularity-attribute types using K-prototypes. Three tree-based ensemble models (XGBoost, LightGBM, Random Forest) were trained, each combined with twelve resampling or baseline strategies to predict popularity. This approach achieved approximately 95 % Balanced Accuracy. Bidirectional interpretability analyses using SHAP values and model-based feature importance consistently highlighted “time to enter the popular board(進入熱門板所需時長)” and “initial interaction rate(進入熱門板前互動速率)” as the most critical predictors. The study advances understanding of content dissemination in anonymous communities by proposing a comprehensive framework that links emotions, discussion topics, and post features to popularity across three distinct attribute types. Practical insights are offered for content creators to optimize posting times and engagement strategies. Finally, replication of clustering and prediction on the independent validation set demonstrated high consistency in cluster structures, model performance, and feature importance rankings, confirming the robustness and transferability of the findings.
