情緒分析與動態增強標籤於美容產品推薦系統之研究

Research on sentiment analysis and dynamic enhanced tags in beauty product recommendation system

隨著網際網路的快速發展，消費者在購買產品前普遍依賴線上評論。在美妝領域，消費者習慣透過線上平台了解產品優缺點。但龐大的文本評論數量使得閱讀與評比變得困難。本研究以線上美妝產品為例，透過網路評論與電商平台數據，建構美妝產品推薦系統。運用網路爬蟲技術蒐集多個平台的文本評論、星等評分及產品價格資訊，並對數據進行預處理，包括利用餘弦相似度公式排除虛假評論，建構美妝專屬詞典，以及運用機器學習模型與動態詞典增強提升文本探勘的準確度。
結合評論情緒分析與星等評分，重新制定美妝產品的推薦排序演算法。本研究針對銷售排名前500名的粉底類產品，共蒐集579,702筆消費者評論與2,430筆產品資訊，建立包含1,315條詞彙的美妝專屬詞典，並採用SVM、XGBoost及Random Forest進行文本預測模型訓練。初始準確度為72%，經過九次詞彙增強與模型訓練後，準確度分別提升至91%、94%與92%。推薦系統能有效降低消費者搜尋與評估產品的時間成本，提升電子口碑的準確性與可信度，並促進電商平台的點擊率與成交量，本研究整合語意情緒分析、虛假評論偵測與多輪詞典強化流程，開發出一套具備準確率、動態適應性與實證驗證成果的美妝產品推薦系統，突破了傳統系統在情緒理解、評論可信度處理與詞彙更新上的侷限。

With the rapid advancement of the internet, consumers increasingly rely on online reviews before makingpurchasing decisions. In the beauty industry, consumers tend to evaluate product pros and cons through various online platforms. However, the massive volume of textual reviews makes it difficult and time-consuming to read and assess each review effectively. This study focuses on developing an online beauty products recommendation system based on user reviews and e-commerce platform data. Using web crawling techniques, the system collects textual reviews, star ratings, and product pricing information from multiple sources. The data is then preprocessed through several steps, including removing fake reviews using cosine similarity, constructing a beauty-specific vocabulary, and applying machine learning models with vocabulary enhancement mechanisms to improvetext mining accuracy.
By integrating sentiment analysis from reviews and star ratings, the study redefines the ranking algorithm for product recommendations. Focusing on the top 500 best-selling foundation products, the dataset includes 579,702 consumer reviews and 2,430 product records. A domain-specific dictionary comprising 1,315 beauty-related terms was constructed, and classification models such as SVM, XGBoost, and Random Forest were trained. The initial prediction accuracy of 72% was improved to 91%, 94%, and 92%, respectively, after nine iterations of dynamic vocabulary enhancement and model retraining. The proposed recommendation system significantly reduces the time and effort required for consumers to search for and evaluate products. It enhances the accuracy and reliability of electronic word-of-mouth and boosts click-through and conversion rates on e-commerce platforms. By integrating semantic sentiment analysis, fake review detection, and iterative vocabulary reinforcement, this study develops a highly accurate and adaptive recommendation system with empirical validation, overcoming the limitations of traditional systems in sentiment comprehension, review credibility, and vocabulary updates.