A Deep Text Mining-Based Cosmetics Consumer Sentiment Analysis Model

With the rapid development of e-commerce, consumer reviews have become crucial channels for businesses to obtain market insights and user feedback. In this study, we propose a sentiment analysis method based on deep text mining, targeting cosmetic product reviews from the JD.com platform. Firstly, comments are preprocessed using Jieba segmentation and TF-IDF feature extraction combined with N -gram models, along with class-weight adjustment to address class imbalance. Experiments involve traditional machine learning methods, with Logistic Regression achieving an initial class- 0 F1-score of 0.8135. We further compare the influence of different stopword strategies, including generic stopwords, Chinese sentiment-specific stopwords, and a customized stopword list constructed based on word frequency and N-gram analysis. Results show the updated stopword list performs similarly to not using any stopwords in Random Forest. Additionally, deep learning models (LSTM and CNN) achieve initial class-0 F1-scores of 0.8331 and 0.8271, respectively, improving to 0.8598 and 0.8521 after Bayesian optimization, while Logistic Regression slightly improves to $\mathbf{0. 8 1 3 9}$. Overall, results indicate that TF-IDF text representation combined with deep learning and Bayesian optimization significantly improves negative sentiment prediction, offering valuable insights for future large-scale, cross-domain sentiment analysis research.