Enhancing Recommendation Systems with Autoencoder-SVD and Transformer-Based Summarization: A Sentiment-Aware Approach Using GPT-2 and VADER

Abstract

Recommender systems play a crucial role in enhancing user experience on social networks. However, traditional methods face challenges in processing user information, such as the inability to effectively analyze textual reviews and the weakness in learning nonlinear and complex relationships between users and items. Additionally, many of these approaches fail to optimally integrate explicit information (such as, ratings) and implicit information (such as, sentiment analysis), leading to reduced recommendation accuracy. This paper proposes a novel approach to improve recommendation accuracy by integrating Autoencoder-SVD with language model-based summarization. In this method, the GPT-2 language model is employed to extract rich summaries from user textual reviews, while VADER is utilized for rule-based sentiment analysis. Furthermore, Autoencoder-SVD enables dimensionality reduction of the user-item rating matrix, and a transformer network enhances the recommendation process. The evaluation results on the Amazon Fine Foods and Amazon Clothes datasets indicate that the proposed model achieves significant improvements over benchmark methods such as SVD and NMF, reducing MAE to 0.3 on both datasets and lowering RMSE to 0.66 and 0.63, respectively. These findings demonstrate that the effective integration of explicit and implicit information within the proposed framework can enhance recommendation accuracy compared to previous methods.