Transformer and Deep CNN-based Product Recommendation System
Mahitha Potti, Avinash Chalumuri, Vani Golagani, DND Harini · 2023
A product recommendation system is one of the crucial components of any E-commerce platform. Content-based filtering techniques effectively provide relevant product recommendations based on user actions on the E-commerce platform. However, with the rapid growth of products and product data, content-based filtering faces several challenges, particularly in feature extraction. Conventional feature extraction methods cannot capture the wide range of product characteristics that influence consumer preferences. In addition, as the number of features extracted increases, the computation time and memory requirements also increase, making it difficult to scale the system. In this paper, we propose the use of various deep learning techniques like conventional Natural Language Processing(NLP) methods, transformer-based methods, Convolutional Neural Networks (CNNs), and Autoencoders to develop 12 models for a recommendation system. Our work aims to identify products with similar textual descriptions and visual appearances to generate accurate recommendations. The 12 models are divided into three categories based on the type of feature extraction techniques used. The proposed approach uses self-attention to extract text features and powerful image feature extraction techniques to effectively capture the product characteristics and address the limitations of traditional methods. After extensive experimentation and evaluation through A/B testing, we observed that the best-performing model is based on a transformer, CNN, and autoencoder. The model generated accurate recommendations and achieved significantly higher user ratings, with an average of 9.9 out of 10 points. Also, the latency of the model is very low.