Personalized Dynamic Marketing Optimization: Integrating Hybrid Recommender Systems, Machine Learning and LLM concepts for E-commerce

Riji Sudheer, Amritha (2025) Personalized Dynamic Marketing Optimization: Integrating Hybrid Recommender Systems, Machine Learning and LLM concepts for E-commerce. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (417kB) | Preview

Abstract

The accelerated expansion of e-commerce has fueled the demand for customized marketing strategies that are capable of responding to the unique customer tastes while optimizing pricing in real time. The Personalized Dynamic Marketing Optimizer (PDMO) is the system that this study introduces, which can be used to maximize customer engagement and revenue through personalized messaging, product suggestions, and dynamic price optimization. Two alternative methods are tried -a Reinforcement Learning (RL)– based PDMO integrated with the Llama-3.1 large language model, and a hybrid machine learning-driven PDMO combining a Neural Recommender, XGBoost pricing, and the mistral language model. RL-based solution characterizes the marketing optimization as a sequential decision problem, with an agent trained for the optimal strategy through customer interaction simulation with engagement and conversion reward signals. The hybrid solution combines deep learning–based recommendation, tree-boosting regression for price optimization, and generative marketing message generation using Mistral. Both processes employ advanced natural language processing (NLP) for personalized message generation and maintain stringent compliance with price ceilings to maintain customer trust and market competitiveness. Ranked on recommendation relevance, pricing accuracy (MAE, R²), message quality (BLEU, perplexity), and campaign metrics (open rate, CTR, conversion rate, revenue), the hybrid pipeline performs well in short-term stability with excellent pricing compliance (97%) and recommendation accuracy (0.92), fueled by the strong feature handling of XGBoost, albeit constrained by conservative revenue ($42.10). On the other hand, the RL-LLaMA-3.1 pipeline generates higher revenue ($324.98) with dynamic exploration at the cost of low pricing compliance (9.49%) and volatility (MAE 233.42), pointing to SAC’s compromises during unstable training. The research contributes a comparative analysis of two state-of-the-art personalization pipelines, shedding light on their respective strengths and compromises for both researchers and practitioners. Future work will study the integration of reinforcement learning with neural recommendation to balance adaptability and predictive accuracy in a single framework.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Chikkankod, Arjun UNSPECIFIED
Subjects:	H Social Sciences > HF Commerce > Electronic Commerce Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning H Social Sciences > HF Commerce > Marketing
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	26 Aug 2026 10:57
Last Modified:	26 Aug 2026 10:57
URI:	https://norma.ncirl.ie/id/eprint/9660

Actions (login required)

View Item