• Version
• Download 14
• File Size 294.68 KB
• File Count 1
• Create Date 12 October 2025
• Last Updated 12 October 2025

Sentiment Analysis of Customer Feedback using Deep Reinforcement Learning

 

 

Prathmesh Chavan … (Department of Data Science, Dr. D. Y. Patil Arts, Commerce and Science College, Pimpri)

Rajguru Bhosale ... (Department of Data Science ,Dr. D. Y. Patil Arts, Commerce and Science College, Pimpri)

 

 

ABSTRACT
Sentiment analysis of large-scale textual data is a challenging task, particularly when dealing with imbalanced class distributions commonly found in real-world review datasets. In this work, we propose a robust pipeline that integrates SBERT-based sentence embeddings with a reinforcement learning framework for multi-class sentiment classification. Specifically, we employ a Proximal Policy Optimization (PPO) Actor-Critic model, enhanced with an imbalance-aware reward mechanism and Monte Carlo dropout, to improve prediction stability and address minority-class recognition. The pipeline first transforms customer reviews into high-dimensional contextual embeddings, which are then fed into the reinforcement learning agent to make sequential classification decisions. Experiments conducted on a sizable customer review dataset demonstrate that the proposed method achieves a weighted F1 score of 0.77 and an overall accuracy of 0.8183, indicating strong performance on majority classes while partially mitigating the effects of class imbalance. The study highlights the effectiveness of combining pre-trained  embeddings  with  policy- gradient reinforcement learning, offering a practical approach for real-world sentiment analysis tasks. Further, we discuss the implications of reward shaping and dropout for stability, providing insights into designing scalable and reliable sentiment classification systems.

Keywords:
Sentiment Analysis, SBERT Embeddings, Deep Reinforcement Learning, Proximal Policy Optimization (PPO), Imbalanced Data

 

 

Download