隨著電子商務平台快速發展，線上評論已成為消費者降低購買不確定性的重要依據，同時也逐漸成為觀察商品市場反應與成長潛力的重要訊號來源。過去相關研究多著重於評論數量、評分或一般銷售量預測，較少直接針對商品是否會在後續期間出現爆發性成長進行早期辨識。基於此，本研究以台灣電子商務平台MOMO購物網之商品評論資料為基礎，建構一套以線上評論多維訊號為核心的商品爆發性成長預測框架，並將研究任務設定為二元分類問題，以判斷商品於特徵截止日後是否具有成長為爆品之可能。 本研究以商品級別作為分析單位，最終納入3，729筆有效商品樣本，其中爆發性成長商品為338筆，非爆發性成長商品為3，391筆。為符合早期預測情境，本研究設定特徵截止日為2025年6月25日，所有特徵皆僅根據截止日前可取得之商品資訊與評論資料建構，標籤則依據截止日後之銷售快照變化判定，以避免資料洩漏。特徵工程方面，本研究建構五類商品評論特徵，包含基礎量化特徵、TF-IDF詞項文本特徵、時間動態特徵、語意特徵，以及行為與資訊結構特徵，以比較不同特徵面向對爆品預測任務之貢獻。 實驗結果顯示，在模型比較階段，XGBoost相較於LightGBM、Logistic Regression與Linear SVM，在F1-score與Precision／Recall平衡上表現較佳，因此作為後續主要分析模型。在TF-IDF詞項文本特徵增益驗證中，基礎量化特徵加入TF-IDF 500後，Test F1-score由0.6906提升至0.7121，顯示評論文本中的詞項資訊能補充基礎量化特徵所未涵蓋的訊號。進一步比較不同擴充特徵面向後，加入時間動態特徵與TF-IDF詞項文本特徵的模型取得最高Test F1-score（0.7626），顯示評論熱度在時間軸上的變化，有助於提升模型對爆品樣本的實際辨識能力；加入行為與資訊結構特徵的模型則取得最高Test AUC（0.9746）與最高Recall（0.7941），表示評論群體的結構性訊號有助於提升整體區辨能力與爆品樣本覆蓋率。相較之下，語意特徵單獨加入後，於固定測試集上的F1-score未明顯提升，其增益效果相對有限。 研究結果顯示，商品上市初期的線上評論具有作為爆品早期預測訊號之可行性。其中，評論熱度的時間演化對實際分類平衡具有較直接的改善趨勢，評論群體的資訊結構則有助於提高爆品樣本覆蓋率。不過，主要模型組合之McNemar’s test結果皆未達統計顯著水準，因此本文將不同特徵面向之改善解讀為固定測試集上的實務分類改善趨勢，而非統計上已確認之模型優勢。整體而言，本研究補充既有電子商務評論研究多偏重一般銷售預測之不足，並提出一套整合基礎量化、詞項文本、時間動態、語意內容與行為結構訊號的多維評論預測框架，可作為平台與商家進行商品篩選、行銷資源配置與早期市場判斷之參考。;With the rapid development of e-commerce platforms， online reviews have become an important source of information for consumers to reduce purchase uncertainty. At the same time， they have gradually become useful signals for observing product market response and growth potential. Previous studies have mostly focused on review volume， ratings， or general sales prediction， while relatively fewer studies have directly addressed the early identification of whether a product will experience explosive growth in a subsequent period. Based on this research gap， this study uses product review data from MOMO， a Taiwanese e-commerce platform， to construct an early prediction framework for product explosive growth based on multidimensional online review signals. The research task is formulated as a binary classification problem to determine whether a product is likely to become a high-growth product after the feature cutoff date. This study adopts the product level as the unit of analysis. The final dataset contains 3，729 valid product samples， including 338 products with explosive growth and 3，391 products without explosive growth. To align with the early prediction setting， the feature cutoff date is set as June 25， 2025. All features are constructed only from product information and review data available before the cutoff date， while the labels are determined based on changes in post-cutoff sales snapshots to avoid data leakage. In terms of feature engineering， this study constructs five types of product review features， including quantitative features， TF-IDF term-based textual features， temporal dynamics features， semantic features， and behavioral and structural features. These feature dimensions are used to compare the contributions of different review signals to the high-growth product prediction task. The experimental results show that， in the model comparison stage， XGBoost outperforms LightGBM， Logistic Regression， and Linear SVM in terms of F1-score and the balance between Precision and Recall. Therefore， XGBoost is selected as the main model for subsequent analysis. In the TF-IDF term-based textual feature gain validation stage， after adding TF-IDF 500 to the quantitative features， the Test F1-score increases from 0.6906 to 0.7121. This result indicates that term-level information in review texts can supplement signals not captured by basic quantitative features. In the comparison of expanded feature dimensions， the model incorporating temporal dynamics features and TF-IDF term-based textual features achieves the highest Test F1-score of 0.7626， suggesting that changes in review activity over time help improve the actual identification of high-growth products. The model incorporating behavioral and structural features achieves the highest Test AUC of 0.9746 and the highest Recall of 0.7941， indicating that structural signals in review groups help improve overall discriminative ability and the coverage of high-growth product samples. In contrast， when semantic features are added alone， the Test F1-score does not clearly improve， suggesting that their incremental effect is relatively limited in the fixed test set. The findings show that online reviews in the early stage of a product’s lifecycle are feasible signals for predicting high-growth products. Among the examined feature dimensions， the temporal evolution of review activity shows a more direct improvement trend in classification balance， while the information structure of review groups helps increase the coverage of high-growth product samples. However， the McNemar’s test results for the main model comparisons are not statistically significant. Therefore， the improvements observed in different feature dimensions are interpreted as practical performance improvement trends in the fixed test set， rather than statistically confirmed model advantages. Overall， this study addresses the limitation of prior e-commerce review research that has mainly focused on general sales prediction and proposes a multidimensional review-based prediction framework that integrates quantitative， term-based textual， temporal dynamics， semantic， and behavioral-structural signals. The proposed framework can serve as a reference for platforms and merchants in product screening， marketing resource allocation， and early market assessment.
