A novel Transformer based E-Commerce Recommendation System

---

This was part of our final Major Project for the Course AI-705(Recommendation Systems)

We present to you our context-aware product recommendation system solution which redefines the capabilities of traditional systems by enhancing contextual relevance. Unlike conventional systems that require users to enter specific product queries to retrieve search results confined to a particular type, our approach allows users to express their needs in natural language. For instance, users can input requests like

"Suggest me some good handmade products" or "I have a birthday party tomorrow, suggest me some products."

Our advanced model, trained on a dataset of approximately 54,000 products spanning various categories, comprehends the context of these inputs and recommends products accordingly.

• Our context-aware recommendation system offers significant advantages over traditional methods. By utilizing advanced natural language processing (NLP) techniques using BERT transformer model, it accurately interprets user intent and context, providing personalized and relevant suggestions that enhance user satisfaction.
• Unlike conventional systems confined to single categories, our approach presents a diverse range of products across categories, encouraging broader exploration and a richer shopping experience. This system improves user engagement, leading to higher satisfaction, loyalty, and conversion rates.
• Even with a relatively small dataset, our model delivers impressive results, suggesting exponential performance gains with larger datasets. This scalability benefits new or smaller e-commerce platforms. Additionally, our system's ability to drive cross-category sales and uncover hidden demand patterns boosts sales, taps into new market segments, and optimizes inventory management, enhancing revenue streams and profitability for e-commerce platforms.
• Our solution uses the product description, title, features, category data which is preprocessed and converted to a meaningful sentence on which we then apply the power of TF-IDF algorithm to give a kickstart with contexts to a pre-trained BERT transformer which is trained and fine-tuned on the information about products which we meaningfully generated. Post training, we give our input in Natural Language to the trained model and it recommends us products.
• This is a very robust solution and it has the potential to be very powerful once trained on huge amount of data.
• Please Refer to the project report to understand how we have implemented and how much promising our model is. It will work not only for E-Commerce but for any large scale Recommendation Systems.