Project Overview

As a Data Science Intern at Lowe's, I contributed to the improvement of product information accuracy on lowes.com. By identifying inconsistencies in the data and implementing measures to optimize data quality, I successfully enhanced the overall user experience, resulting in increased sales and reduced product returns.

Project Details

Optimizing Data Quality with LLM

One of the key aspects of this project was the implementation of the LLM (Language Model) algorithm to improve data accuracy. By leveraging the capabilities of LLM, I developed strategies to identify inconsistencies in the product information available on lowes.com. Through data analysis and optimization techniques, I was able to enhance the quality of the data, ensuring that the product information displayed on the website was accurate and reliable. This, in turn, positively impacted customer satisfaction and increased sales.

Prompt Generation Approach using ChatGPT

To further enhance the performance of the LLM algorithm, I developed an iterative prompt generation approach that leveraged the capabilities of ChatGPT and human iterative feedback. This approach significantly improved the recall rate from 0.63 to 0.76. By generating prompts that effectively captured relevant information, the LLM algorithm became more efficient in identifying and rectifying inconsistencies in the product information. This iterative approach not only improved the accuracy of the algorithm but also enhanced the overall efficiency of the data optimization process.

Performance Evaluation and Optimization

As part of the project, I conducted a comprehensive evaluation of the LLM algorithm's performance in a production environment. I analyzed API responses and compared the algorithm's predictions against human labels to validate its accuracy as well as to figure out the API fail rate from OpenAI to understand business impact. This evaluation process allowed me to identify and optimize algorithm error categories leading to improved results. By fine-tuning the algorithm based on real-world data and user feedback, we were able to achieve higher levels of precision and reliability in product information accuracy.

Results

The implementation of these measures resulted in significant improvements in product information accuracy on lowes.com. By optimizing data quality with the use of LLM and leveraging the prompt generation approach, we were able to enhance the recall rate by 13%, improve algorithm performance, and reduce inconsistencies in the product information. Ultimately, this led to increased sales and a decrease in product returns, contributing to a better user experience for Lowe's customers.