Leveraging LLM generated labels to reduce bad matches in job recommendations

What is it about?

A case study at Indeed where we use LLMs to batch generate explicit negative signals i.e. the match quality labels in our use case, and then using those LLM-generated labels in building lighter ML models that prevent bad matches in our recommendation systems at real time.

Featured Image

Photo by Mario Garcia on Unsplash

Photo by Mario Garcia on Unsplash

Why is it important?

This helps address the long-standing issue of underrepresented negative samples in our training datasets. As a result, we successfully reduced users' dislike clicks, increased jobseekers' application rates on recommended jobs, and improved employer feedback on those applications.