Candidate Quality Estimator

Published June 4, 2020 by Scott Hirsch

This is the fourth of five posts about machine learning at TalentMarketplace. Take a quick read of the introduction blog here.

Eric Schibli, a senior PhD candidate at Simon Fraser University, studies ion-exchange membranes for the next generation of hydrogen fuel cells. During his downtime, Eric creates and enhances machine learning programs.

Eric joined TalentMarketplace to help find the candidates that are most attractive to employers by creating a machine learning tool: The Candidate Quality Estimator. This tool uses an ensemble of neural networks to determine which candidate profiles are most likely to receive an interview request based on information in their candidate profiles. By using a semantic representation of each candidate’s personality and work experience, such as skills, project history, ideal company culture, and other parameters, numerous machine learning models were trained on historical TalentMarketplace data to recognize patterns that could determine whether or not a candidate would be likely to receive an interview requests and be hired.

The models were trained using a stratified process which consisted of iteratively using 80% of the data to train models to and make predictions about the remaining 20%. The system is designed to be self-tuning; it is capable of testing itself and automatically adjusting parameters and adapting itself to a constantly evolving pool of input data.

So how will this score be applied?

Since we started TalentMarketplace, one of the most requested features from clients was for us to provide candidates a “score” based on our pre-screening interview. The challenge here, of course, is our own internal bias - we are human, after all. Instead, we suggest candidates based on their fit with a potential employer based on the more qualitative information we capture from our interviews and interactions with them.

The Quality Estimator is a fantastic addition to our own human assessment. It is a quantitative assessment, based on data we have. Employers will be able to use a new “sort-by” function to rank the active candidate pool, in addition to their search parameters, saving time by showing the candidates they’re most likely to interview first.