E-commerce platforms face a persistent challenge in effectively categorizing and recommending products from low-frequency categories, often lacking sufficient labeled data for traditional machine learning models. This paper introduces a robust ensemble-based few-shot learning recommendation framework designed to enhance personalization by addressing the challenge of low-frequency product interactions. The method improves recall and ranking of rarely interacted products by integrating session-product data fusion, threshold-based frequency labeling, multimodal feature engineering (text embeddings, numerical attributes), and stacked ensemble learning. Experimental validation demonstrates superior performance over existing models, particularly for low-frequency items, highlighting the potential of FSL and advanced ranking systems for addressing sparsity in e-commerce data.