TỐI ƯU HÓA HIỆU SUẤT HỆ THỐNG GỢI Ý ĐỘNG TRONG MÔI TRƯỜNG DỮ LIỆU LỚN BẰNG HỌC TĂNG CƯỜNG VÀ KỸ THUẬT GIẢM CHIỀU SÂU

Abstract

This paper proposes a hybrid framework to improve stability and computational efficiency. We
incorporate a Deep Dimensionality Reduction technique using Autoencoder to compress the sparse matrix
into a low-density feature space, significantly reducing the computational load. We then model the realtime recommendation process as a Markov Decision Process (MDP) and apply a Reinforcement Learning
(RL) algorithm, namely Deep Q-Network (DQN), to learn policies that optimize user interaction rewards.
Experimental results on large-scale datasets (e.g., MovieLens 20M) demonstrate that this method not
only reduces the processing latency by 40% (from 150ms to 90ms) but also simultaneously improves the
recommendation accuracy (Precision and Recall) by 12% compared to baseline models