Recommender systems are crucial in e-commerce for matching users with products of interest. This paper proposes a deep learning approach that leverages representation learning to capture complex patterns in user behavior, combining long-term preferences with short-term, session-based interests. We develop a neural network model that integrates Convolutional Neural Networks (CNNs) for sequential pattern extraction with user embedding vectors representing historical preferences. Experiments are conducted on open e-commerce datasets with implicit (click/view) and explicit (purchase) feedback. The proposed model outperforms baseline recommendation techniques, achieving higher Hit Rate (HR@10) and Normalized Discounted Cumulative Gain (NDCG@10). The results demonstrate that blending users’ stable long-term tastes with their recent short-term actions leads to more accurate recommendations. This work reinforces deep learning as a powerful tool for representation learning in recommender systems, and our findings offer insights into modeling nuanced user behaviors for intelligent data-driven recommendations in online retail.