An E-Commerce Based Personalized Health Product Recommendation System Using CNN-Bi-LSTM Model

Abstract

With the increasing complexity of contemporary E-commerce recommender systems, personalized health product recommendations have become a challenging task. Existing methods predominantly rely on latent natural language features extracted from product descriptions, lacking transparency and user engagement. While recent itemrepresentation-learning algorithms integrate pre-defined attribute data, they still struggle to provide accurate recommendations and detailed explanations due to sporadic relationships between stated qualities and the recommendation process. To address these challenges, we propose an E-commerce based personalized health product recommendation system using a CNN-Bi-LSTM model. Our system leverages pre-trained CNN-based transfer learning models, such as AlexNet, Google Net, and ResNet-50, for predicting health products, while the Bi-LSTM acts as a numerical function to provide ratings. The CNN-Bi-LSTM recommendation system utilizes attribute-specific representation extraction methods, resulting in a more robust compatibility model and improved recommendation effectiveness for health products. The proposed approach is evaluated on two datasets: The health-product dataset and the Flipkart health product dataset. The health-product dataset comprises 20,726 products, including health supplements (14,871 items) and fitness-related products (13,663 items) recommended by experts. The Flipkart health product dataset contains 21,889 healthcare products with essential information like name, price, MRP, discount, and ratings. Comparative analysis against existing methods, such as W-RNN and Improved SVM, demonstrates the superiority of our CNN-Bi-LSTM model. The evaluation metrics, including accuracy (95.52%), recall (88.31%), and user coverage (89.51%), highlight the effectiveness of our recommendation system in providing more accurate and relevant personalized health product recommendations. For the aforementioned dataset, the AUC and hit rate of the proposed method are contrasted with those of existing methods. The findings were gathered to demonstrate that the suggested approach performs better when compared to the current recommendation systems.