Revolutionizing ecommerce product categorization with LLM (GPT)

See how we addressed an e-commerce platform’s challenge of managing diverse product descriptions by implementing a solution leveraging the GPT 3.5 language model. The transformative approach streamlined model upgrades, reducing maintenance time and significantly cutting down model retraining time from two weeks to two days. Despite a slight dip in accuracy, this resulted in enhanced operational efficiency and substantial cost savings for the client, showcasing the solution’s effectiveness in revolutionizing ecommerce product categorization.

Challenge

The ecommerce marketplace faced was challenged with managing the variability in product categorization and sub-categorization of descriptions provided by sellers, leading to difficulties in accurate product classification within the marketplace’s taxonomy. Initially reliant on manual tagging, the marketplace later automated the categorization process using a Natural Language Processing (NLP) text classification-based Machine Learning (ML) model. Trained on previously tagged products, the model aimed to accurately classify product categories.

Approach and solution

Netscribes conducted an audit of the previous solution, identifying key challenges:

• The ever-changing nature of products and categories impacted model accuracy.
• Human Quality Control (QC) and frequent model retraining were required.
• Taxonomy alterations incurred significant costs for re-tagging and retraining.

We proposed a transformative approach integrating the GPT 3.5 language model and Few-Shot prompt loading techniques. This solution utilized dynamic taxonomy, product titles, and descriptions to enhance the accuracy of category predictions.

Results delivered

The implemented solution led to:

• Streamlined model upgrades and dynamic taxonomy changes
• Reduced time for maintenance and support, enhancing overall operational efficiency
• Despite a slight drop in model accuracy from ~85% to ~78%, the turnaround time for model retraining decreased from ~2 weeks to ~2 days, resulting in substantial cost savings.

Related reading: Catalog scoring and quality seller support for an e-commerce marketplace