Dynamic Strategy Optimizer (DSO): Application In Enhancing New User Engagement in Hybrid Recommender System

This paper introduces the Dynamic Strategy Optimizer (DSO), a novel hybrid approach that combines sampling with dynamic optimization to improve decision-making performance. The primary contribution is the DSO’s ability to optimize multi-armed bandits (MAB) for higher rewards and significantly reduced computational time, outperforming traditional methods like Thompson Sampling, Epsilon Greedy, and A/B Testing. Additionally, we propose a novel hybrid recommender system (RS) architecture, scalable for large datasets, integrating collaborative filtering, neural network classification, and MAB techniques to address the cold-start problem in real-time recommendation systems. Evaluations on the MovieLens25m and JesterJoke1.7 datasets show that DSO achieves superior reward accumulation and computational efficiency, with times as low as $1.59 \mathrm{it} / \mathrm{s}$ on JesterJoke1.7 and $8.41 \mathrm{~s} / \mathrm{it}$ on MovieLens25m, making it an efficient solution for large-scale adaptive systems.