UCM: Personalized Document-Level Sentiment Analysis Based on User Correlation Mining

Jiayue Qiu, Ziyue Yu, Wuman Luo

Abstract: Personalized document-level sentiment analysis (PDSA) is important in various fields. Although various deep learning models for PDSA have been proposed, they failed to consider the correlations of rating behaviors between different users. It can be observed that in the real-world users may give different rating scores for the same product, but their rating behaviors tend to be correlated over a range of products. However, mining user correlation is very challenging due to real-world data sparsity, and a model is lacking to utilize user correlation for PDSA so far. To address these issues, we propose an architecture named User Correlation Mining (UCM). Specifically, UCM contains two components, namely Similar User Cluster Module (SUCM) and Triple Attributes BERT Model (TABM). SUCM is responsible for user clustering. It consists of two modules, namely Latent Factor Model based on Neural Network (LFM-NN) and Spectral Clustering based on Pearson Correlation Coefficient (SC-PCC). LFM-NN predicts the missing values of the sparse user-product rating matrix. SC-PCC clusters users with high correlations to get the user cluster IDs. TABM is designed to classify the users’ sentiment based on user cluster IDs, user IDs, product IDs, and user reviews. To evaluate the performance of UCM, extensive experiments are conducted on the three real-world datasets, i.e., IMDB, Yelp13, and Yelp14. The experiment results show that our proposed architecture UCM outperforms other baselines.