A Novel Multi-Class Product Rating Prediction Model based on Enhanced Textrank Text Encoding and Human Psychology Simulation

In recent years, with the booming development of the internet and e-commerce, user comments and ratings on a particular product often determine the marketing strategies of businesses. It is believed that if businesses can collect user feedback (such as through surveys) before they rate a product, they can analyze people's product experiences at minimal cost. However, most comments are in free-form, slang, or internet language, and manually categorizing and annotating them would require significant resources, which is not feasible for small and medium -sized businesses. To address this issue, some deep learning frameworks have been used for sentiment analysis and rating prediction in text. However, these models often overlook the actual application scenario when predicting ratings. In reality, users' ratings are not solely based on their own experiences but are also influenced by the historical comments on the product. Therefore, we propose a supervised product rating prediction model. In the first stage, the improved dynamic TextRank submodule (DTsub) and a lightweight Feature Fusion submodule (FFsub) are used to construct text vectors, utilizing a sentiment word library. In the second stage, a Long Short-Term Memory (LSTM) neural network and a Feed Forward neural network (FFNN) are utilized to predict product ratings (multiclass) based on the actual application scenario using ensemble learning and voting. Additionally, we also introduced a weighted cross-entropy loss tailored to businesses product ratings in our model. Experimental results on an Amazon product review dataset show significant improvements in accuracy, F1 score, and F2 score of our proposed DTFS-LFNN model.