Projects
I have worked on several projects in data mining in Cosmology which have been listed in publications. Here I present the other projects which I have done in industrial field.
I have built the User Feedback (UF) model, a simple model that utilizes the users’ clicked data (in www.coupang.com) to assign the advertisement keywords to the products to propose the most relevant advertised products to customer’s queries in the search result page. After training and testing carefully the model it has shown a high performance with more than 90% precision. 89% item coverage along with a query coverage from Top to Tail queries range up to rank 1M. The experiment successfully deployed to the Search production with +9% contribution in the increment of the main metric i.e. First Placement Ad GMV per customer and other important metrics like GMV per customer +3.9% and Ad GMV per Customer +15.03%.
In an online shop like www.coupang.com, Search tags are the keywords added by sellers to the products. This extra information can improve search results by customers however wrong search tags have a destructive effect on search results. Using fastText (NLP library), I have trained a model that used different features of products like title, brand, category, etc to classify the associated search tag as relevant or irrelevant to the product. The model shows 80% precision and 50% recall which is a valuable result for cleansing tags and improving search results.
Some products have images that are not suitable for some customers. These include nudity or sexual images or those related to adult products. Using some labeled product images in Coupang database, I have made a fine-tuned model based on EfficientNet to classify images into Safe and Unsafe. In addition, I have leveraged the efficiency by making a combination of fine-tuned EfficientNet and NudeNet model. The combination shows a precision more than 80% for the fashion category. This project is in progress now to be employed for other categories of products.