Contextual Product Recommendation Using Transformer-Based Models: Uncovering Product Dependencies in Transactional Data

Understanding the relationships between products within an invoice can significantly enhance the accuracy and effectiveness of recommendation systems in e-commerce. In this paper, we propose a novel approach to model transactional data using transformer architectures, treating invoices as analogous to sentences and products as words. Using the Online Retail dataset, we construct sequences of products for each invoice and train a Generative Pre-trained Transformer (GPT)-based model with multi-head attention to uncover latent relationships based on product co-occurrence within transactions. This approach captures contextual dependencies, enabling the model to predict complementary and alternative products more effectively than traditional methods. For example, products frequently purchased together can reveal nuanced contextual patterns that are often overlooked by standard recommendation techniques. Our results demonstrate the potential of natural language processing techniques for transactional data and show that this method leads to more accurate, diverse, and context-aware recommendations. These improvements can directly impact e-commerce platforms by increasing cross-selling opportunities, enhancing customer satisfaction, and driving sales through more personalized shopping experiences.