Metadata and Review-Based Hybrid Apparel Recommendation System Using Cascaded Large Language Models

The commercial success of online retailing platforms relies on sophisticated technological advancements of recommender systems. As online marketplaces gain popularity it becomes crucial to provide accurate product suggestions to make personalized shopping experience more exciting for the online customer. The evolution of recommendation systems from using collaborative and content-based approaches to more advanced deep learning-based approaches provide motivation as well the necessity to leverage more advanced tools for creating recommendation systems more powerful. Among variety of e-commerce products, this work concentrates on the apparels section, specifically coming under men's shirt category. In this work, we introduce a Hybrid Recommender System that combines content based and item-based collaborative filtering strategies. To implement the content-based filtering approach we employ text-representation techniques. These methods carefully analyze how users interact with product specifications to understand their nuanced preferences. We also propose a cascading transformer mechanism for generating contextually aware representations using sentence transformers derived from Large Language Models (LLMs) T5 and BERT. To gauge the emotional tone and sentiment associated with product reviews, we make use of rule-based and embedding based sentiment analysis models. A weighted score mechanism has been used to determine the final recommendations by the proposed system. Finally, we discuss the efficiency of the proposed filtering technique and compare various text representation methods along with popular baseline transformer models. The introduced hybrid recommendation system with the proposed cascaded model was evaluated and found to give recommendations with 94.76% similarity score while FastText achieved 66.90%. Doc2Vec and Word2Vec resulted in recommendations with 46.62% and 47.70% similarity respectively.