When companies want to recruit for a new job, they get 1000's of candidates.. The human has to sort through by hand for each resume! This is horribly inefficient, and there has to be a better way.
What it does
There is a better way!
LSTM - Long Short Term Memory. Basically means that our neural network will remember past inputs, and take them into account when deciding what to do with a new input.
ATS - Application Tracking System. When a job posting gets 1000 applicants, these applicants enter into an ATS that keeps track of all associated information, such as resumes, cover letter, answers to unique questions, where they are at during the interview process, who has spoken to them, decision recommendations, keyword filtering, etc...
Culture fit vs technical fit - When interviewing with a new company, they judge you on cultural and technical fit. Technical fit is "do you have the skills" (can you code in C++, do you know python, etc). Culture fit is much more subjective, and it is "can I work with this candidate for at least 2 years? ". Currently, there is no technology to help recruiters judge culture fit which results in an EXTREMELY inconsistent system that has copious amounts of human selective bias.
DeepHire uses Siamese Recurrent Neural Networks with LSTM, and detailed emotion analysis in order to effectively perform as a first interviewer and to judge culture fit. It integrates with a common ATS, Jazz, to provide an extremely smart filter so that only qualified candidates with a good culture fit will ever be shown to the recruiter.
How we built it
We use a Siamese Network with and LSTM modification. A siamese network is basically 2 identical machine learning models, and helps judge the similarity between items (pass in a known input to the first model, pass in an unknown input to the second, and compare outputs to see how close the unknown is to the known). After judging similarity, we run the FRQ responses through sentiment analysis to give more information to the recruiter.
Then, the top candidates are chosen and given priority during the filtering process.
Challenges we ran into
The siamese networks were being extremely finnicky, and we had to provide a lot of training data to them. This was a challenge at first because the type of data we needed is not easy to get!
Parallelizing the tasks to improve performance was also a challenge. At first, the neural nets would only utilize one CPU core of a Macbook (slow!).
Accomplishments that we're proud of
Optimizing for speed, integrating with a real ATS that is widely used, and giving meaningful results. It is far too easy to give junk results using machine learning, but we are proud to say our results were pretty accurate (~87%)!
What we learned
a LOT about how neural networks work, and all the variations there are.
What's next for DeepHire
Allowing the recruiter to talk to a chat bot AI, where this AI will gently ask questions in order to figure out what the company values. Later, these values will be used to generate novel questions for the candidates to answer, and the responses will be compared in order to find the most like-minded candidates.