Exploring Emotion Trends in Product Reviews: A Multi-modal Analysis with Malicious Comment Filtering and User Privacy Protection

In this paper, we introduce an innovative approach that combines multi-modal sentiment analysis with mechanisms for safeguarding user information security. Our objective is to effectively combat harmful comments and precisely forecast trends in product sentiment. Initially, we detect and secure sensitive user information by implementing regular expressions to ensure user confidentiality and security. First, we apply statistical analysis and K-means++ algorithms to screen users who post malicious reviews. Next, we develop a novel multi-modal sentiment analysis and prediction model that incorporates the pre-trained BERT model and Swin Transformer model for feature extraction from comment text and image data. Furthermore, the expressive capability of image features is enhanced with the aid of the SENet model. We input the image features, improved by the SENet model, as well as the text features, extracted by the BERT model, into the Transformer model for fusion, and the classification probability is determined using the SoftMax function. We employ the Prophet method to combine sentiment indicators and time series features, which allows us to predict the upcoming sentiment trends of product reviews. The algorithm is implemented in the evaluation of the Amazon book review datasets and is compared to other algorithms such as Bert, yielding an accuracy of 94.15%. This study is of significant value for enabling real-time monitoring of product sentiment trends, filtering malicious reviews, and enhancing both product management and user experience.