Tools
Tools
Mini Biju, Sneha (2025) A Hybrid Sentiment and Reinforcement Learning Framework for Product Review Optimization. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (4MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (1MB) | Preview |
Abstract
Background: The beauty market is difficult to measure customer satisfaction with different products. Traditional methods of analysis are not likely to yield useful recommendations balancing contradictory factors like scent, texture, packaging, and performance. Most sentiment analysis methods only examine the general product sentiment and not the complex interdependencies between such factors.
Objectives: The main goal of this work is to build a reinforcement learning model that finds the best ways of augmenting products as per aspect-level sentiment patterns in bulk customer review datasets. Another goal is to create a dynamic framework that augments product features in a balanced way—boosting overall customer satisfaction without over-augmenting already top-rated features. This project provides an effective method for product improvement analysis and optimization through Sentimental Analysis with reinforcement learning (RL). Using a massive review dataset of over 300,000 Amazon beauty product reviews, this system identifying underperforming products and improvement tactics from reviews.
Methodology: The project employs a multi-step analytical approach beginning with data preprocessing and sentiment analysis using TextBlob. The system detects aspect-based sentiments from seven critical product features: scent, texture, packaging, efficacy, cost, durability, and usage. A confidence-driven sentiment scoring system balances review quality and alignment between ratings and sentiments. Here implements Q-learning reinforcement learning agent, which generates best-improvement plans through dynamic training using simulated product improvement scenarios.
Key Contributions: This paper introduces a novel RL dataset generation model that produces realistic transition states representing product improvement actions and consequences. The system employs offline and dynamic RL training methods, with the latter performing better by optimizing policy in real time. The Q-learning agent converges at an average reward of 1.08 through a number of training iterations and acquires effective improvement approaches successfully.
Results: Sentiment differences are considerable in review of 130 target products, with packaging being the most important area for improvement (sentiment score:-0.42). RL system can effectively discover best action sequences, where promising areas of improvement are given a high priority and over-optimization of already well-working features is avoided. By using this method-obtained high learning quality with 11.9% positive Q-values and exhaustive state exploration in 4,225 various scenarios.
Impact: The research proposes a data-driven, scalable product improvement optimization process that can be applied across many categories of consumer goods. Combining sentiment analysis and reinforcement learning allows companies to extract actionable insights for guiding targeted product upgrade, resulting in possible enhanced customer satisfaction and market performance. The ability of the system to resolve multiplicative product dimensions and obtain maximal total sentiment improvement is a fundamental break-through in data-driven product development processes.
| Item Type: | Thesis (Masters) |
|---|---|
| Supervisors: | Name Email Shahid, Abdul UNSPECIFIED |
| Uncontrolled Keywords: | Sentiment Analysis; Product Improvement; Q-Learning; Reinforcement Learning; Customer Reviews |
| Subjects: | Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing H Social Sciences > HF Commerce > Marketing > Consumer Behaviour H Social Sciences > HD Industries. Land use. Labor > Specific Industries > Cosmetics Industry |
| Divisions: | School of Computing > Master of Science in Artificial Intelligence |
| Depositing User: | Ciara O'Brien |
| Date Deposited: | 02 Jun 2026 11:23 |
| Last Modified: | 02 Jun 2026 11:23 |
| URI: | https://norma.ncirl.ie/id/eprint/9332 |
Actions (login required)
 |
View Item |